Jasa Konsultasi Disertasi Keuangan (1) dan Perbankan S3, hubungi WA 0821 2230 7021: CONTOH PAPER MENGENAI ALGORITHM, INTELLIGENT SYSTEM, COMPUTATIONAL

Seminar Nasional Teknologi Informasi 2011 A8

IMPLEMENTASI METODE TABU SEARCH UNTUK

PENJADWALAN KELAS

Ade Trisnawati 1), Iriansyah BM Sangadji 2), Sely Karmila 3)

1), 2), 3)Teknik Informatika Sekolah Tinggi Teknik PLN

Jl. Lingkar Luar Duri Kosambi Jak Bar 11750

Email: ade.trisnawati@gmail.com, iriansyach@gmail.com

ABSTRACT

Scheduling is an activity to allocate a number of resources available. This activity is done to ensure that planning can go well with the time and energy used efficiently.

Taboo search algorithm will keep the best solution is not lost by storing the best solution and continue to search by last solution. This algorithm is forbidden to use solutions that have been used so as to avoid useless repetition.

The solution of this paper is the schedule of implementation of the model generated using taboo search

Key words

Taboo search, Schedulling , Implementation Methode

1. Pendahuluan

Penjadwalan merupakan kegiatan untuk mengalokasikan sejumlah sumber daya yang tersedia. Kegiatan ini dilakukan untuk memastikan bahwa perencanaan dapat berjalan dengan baik dengan waktu dan tenaga yang digunakan secara efisien. Banyak model-model penjadwalan yang telah digunakan. Berbagai algoritma telah banyak juga dimanfaatkan.

Penjadwalan kelas merupakan kegiatan yang mengawali pergantian semester di setiap perguruan tinggi. Proses ini harus memperhitungkan banyaknya mata kuliah, ketersediaan ruang, dan rentang waktu yang digunakan. Inti dari penjadwalan kuliah adalah menjadwalkan beberapa komponen yang terdiri dari mata kuliah, ruang, dan waktu dengan memperhatikan sejumlah batasan dan syarat tertentu. Permasalahan yang dihadapi penjadwal terletak pada lebih banyaknya mata kuliah yang harus dijadwalkan daripada ruang yang tersedia, kesesuaian kebutuhan perkuliahan dengan fasilitas ruangnya, kapasitas ruang yang harus sesuai dengan jumlah mahasiswa, serta keinginan pengajar untuk mengajar pada suatu hari atau jam tertentu. Distribusi jadwal perkuliahan juga diharapkan dapat merata tiap harinya

untuk setiap kelas. Pekerjaan penjadwalan mata kuliah ini akan semakin berat jika melibatkan semakin banyak kelas perangkatannya.

Pada saat ini sebagian besar pembuatan jadwal masih belum memiliki proses yang lebih efisien. Hal ini dibuktikan dengan masih digunakan cara; taf tata usaha akan mendata seluruh mata kuliah yang akan dibuka pada suatu semester. Kemudian mencatat prioritas-prioritas yang akan diberikan terhadap suatu mata kuliah. Baru kemudian mengatur penjadwalan, dan membuat laopran jadwal kuliah dengan menggunakan Microsoft Excel. Proses seperti ini membutuhkan banyak variable seperti dosen, mata kuliah, hari, ruang dan waktu (ketersediaan dosen), masing-masing variable terdiri dari sejumlah dosen, enampuluhan mata kuliah dengan 156 sks, 5 hari, dan terdapat 16 ruangan. Selain itu proses ini membutuhkan ketelitian dan waktu pengerjaan yang tidak singkat, sehingga seringkali terjadi jadwal yang bentrok yang menyebabkan suasana belajar mengajar terhambat.

Diharapkan dengan digunakannya algoritma Tabu Search akan diperoleh optimasi penjadwalan yaitu kondisi dimana terjadi kombinasi terbaik untuk pasangan mata kuliah dan dosen pengajar secara keseluruhan, tidak ada permasalahan bentrokan jadwal, serta ketersediaan ruang yang cukup dan sesuai secara fasilitas untuk seluruh mata kuliah yang ada. Berdasarkan uraian tersebut maka dalam penelitian ini akan dijelaskan bahwa dengan bantuan algoritma Tabu Search penyusunan penjadwalan mata kuliah dapat dioptimalkan. Program dapat mencari solusi penjadwalan pada waktu yang dapat digunakan baik oleh dosen, kelas maupun ruangan yang terlibat dalam suatu mata kuliah.

2. Tujuan

1. Mengimplementasikan Algoritma Tabu Search dalam penjadwalan perkuliahan.

2. Menghasilkan solusi penjadwalan yang optimal.

A8 Seminar Nasional Teknologi Informasi 2011

3. Teori

3.1. Penjadwalan

Penjadwalan menurut Rosani Ginting[1] adalah pengalokasian sumber daya pada objek-objek yang ada pada ruang waktu dan bergantung pada kendala-kendala yang sedemikian sehingga sedapat mungkin memenuhi sekumpulan sasaran yang diinginkan. Secara sederhana, penjadwalan dapat diartikan sebagai pengalokasian sumber-sumber daya yang tersedia pada ruang waktu yang ada sehingga memenuhi kondisi-kondisi tertentu. Tujuannya adalah untuk memaksimalkan suatu proses dengan tetap menjaga agar tidak melanggar constraint yang berlaku pada proses yang bersangkutan. Penjadwalan kuliah merupakan salah satu masalah optimasi, dimana definisi optimasi adalah sarana untuk mengekspresikan model matematika yang bertujuan memecahkan masalah dengan cara yang tepat

Constraint

Di dalam penjadwalan dikenal 2 macam constraint ( persyaratan) yaitu Hard Constraint dan Soft Constraint. Constraint penjadwalan perkuliahan yang berlaku tentunya berbeda tergantung dari aturan dan tata cara perkuliahan dari institusi yang bersangkutan. Constraint pada umumnya terbagi menjadi dua jenis yaitu hard constraint dan soft constraint. Hard constraint adalah batasan yang wajib untuk dipenuhi atau tidak boleh dilanggar, sementara soft constraint adalah batasan yang masih memberi toleransi terhadap pelanggaran, namun sebisa mungkin diminimalisir pelanggarannya. Contoh hard constraint dan soft constraint pada kasus penjadwalan perkuliahan adalah :

1. Mata kuliah yang berbeda tidak boleh dijadwalkan pada waktu dan tempat yang sama, serta mahasiswa tidak boleh menghadiri 2 mata kuliah pada waktu yang sama (hard constraint).

2. Dalam satu hari mahasiswa tidak boleh menghadiri lebih dari 3 x timeslot kuliah berturut-turut,harus ada jeda minimal 1 timeslot (soft constraint).

Perhitungan Constraint

(kls)i

r = S1+ S2 Sn

Dimana :

Cji = Nilai cost constraint ke-j pada kelas mta kuliah ke-i

(kls)i = Jumlah kelas matakuliah ke-i

S = Total nilai cost

r = Nilai cost total yang dihasilkan

3.2. Tabu Search

Menurut F Glover[2] Tabu search adalah teknik local search yang memilih langkah berikutnya (neighbor-solution) berdasarkan constraint dan pinalty. Setiap constraint yang ada akan didefenisikan dengan sebuah nilai pinalty yang akan dikenakan apabila constraint itu terlanggar. Metode ini menggunakan Tabu List untuk menyimpan sekumpulan solusi yang baru saja dievaluasi. Selama proses optimasi, pada setiap iterasi, solusi yang akan dievaluasi akan dicocokkan terlebih dahulu dengan isi Tabu List untuk melihat apakah solusi tersebut sudah ada pada Tabu List. Apabila solusi tersebut sudah ada pada Tabu List, maka solusi tersebut tidak akan dievaluasi lagi pada iterasi berikutnya. Sejumlah pengembangan yang baru dari Tabu Search memungkinkan sebuah solution diterima kembali menjadi current solution apabila nilai pinalty keseluruhan jauh lebih kecil daripada solusi sebelumnya menurut batasan-batasan tertentu.

Glover mengatakan bahwa prosedur TS ini dapat ditemukan dalam tiga pola (scheme) utama. Pola pertama adalah adanya penggunaan struktur memori berbasiskan atribut-atribut fleksibel yang dirancang untuk membolehkan sebuah kriteria evaluasi dan hasil pencarian di masa lalu dieksploitasi lebih mendalam. Pola ini menjadikan TS berbeda dengan aplikasi lain yang menggunakan struktur memori yang rigid (kaku) atau tanpa menggunakan struktur memori (seperti simulated annealing). Pola kedua adalah penggunaan mekanisme atau kondisi yang dapat membatasi atau membebaskan suatu proses pencarian yang sedang berlangsung. Pola kedua ini dikenal sebagai mekanisme tabu restriction dan aspiration criteria. Pola ketiga adalah pelibatan suatu fungsi memori dengan rentang waktu yang berbeda yakni berupa memori jangka pendek (short term memory) dan memori jangka panjang (long term memory) untuk menjalankan strategi intensifikasi dan diversifikasi dalam proses pencarian solusi. Strategi intensifikasi adalah strategi pencarian yang mengarahkan/ mengfokuskan pencarian pada suatu area tertentu, sedangkan strategi diversifikasi adalah strategi pencarian yang mengarahkan pencarian pada area baru. Setiap constraint yang ada akan didefinisikan sebagai sebuah nilai pinalti yang akan dikenakan apabila constraint ini dilanggar.

Prosedure dari metode TS adalah :

1. Pilih inisialisasi solusi i dalam S. Set i*=i dan k=0.

2. Set k=k+ 1 dan buat subset V* dari solusi dalam N(i,k) sedemikian sehingga kondisi TS tr(i,m) ) Tr

Seminar Nasional Teknologi Informasi 2011 A8

dilanggar (r=i,...,t) atau sedikitnya satu dari kondisi dipenuhi a (i,m) A (i,m) terpenuhi (r=i,...,a).

3. Pilih yang terbaik j = i ⨁ m dalam V* (yang didasarkan pada atau fungsi yang sudah dirumuskan) dan set i = j.

4. Jika f(i) < f(i*) maka set i*=i.

5. Perbaharui tabu dan kondisi aspirasi

6. Jika tercapai kondisi pemberhentian, hentikan.Jika

tidak, kembali ke langkah 2.

Algoritma Tabu search

Pilih solusi awal i dalam S

Gambar 1 Flowchart Algoritma Tabu search

Keterangan :

S = himpunan semesta dari fungsi tujuan

i = solusi awal

k = iterasi

V* = nilai subset optimal dari N(i,k)

j = solusi tetangga terbaik dari V*

f(i) = nilai fungsi dengan variabel i

neighbour solution tersebut harus ditolak dan proses dilanjutkan dengan memeriksa neighbour solution yang baru. Dengan demikian dengan adanya tabu list dapat menghindari cycles (iterasi berjalan terus tanpa memperbaiki kualitas solusi).

Tabu list bisa disimpan dalam bentuk array atau tree,

tergantung dari masalah dan implementasi yang

diinginkan. Ukuran tabu list dapat mengikuti aturan berikut :

1. Menggunakan ukuran tabu list statis. Ukuran tabu list

tidak berubah sesuai dengan yang telah ditentukan sebelumnya.

2. Ukuran tabu list disesuaikan dengan masalah. Semakin besar masalah, semakin besar juga ukuran tabu list yang dibutuhkan.

3. Mengubah ukuran tabu list sesuai dengan kualitas

solusi. Move adalah suatu cara untuk menghasilkan

solusi yang baru di dalam sebuah neighborhood. Dalam penjadwalan matakuliah move bisa berupa :

a) Pertukaran nilai dalam 2 baris atau kolom yang berbeda

b) Pergeseran nilai yang ada dalam suatu baris dan kolom

Perpindahan nilai yang ada pada suatu baris atau kolom ke baris atau kolom lainnya.

4. Metodologi

Menentukan data input sistem Pemberian nilai cost pada tiap parameter Mempresentasikan dengan tabu search

Analisis Kebutuhan Sistem

Tabu list digunakan untuk menyimpan n solusi yang sudah dikunjungi sebelumnya. Jadi setiap current solution yang ditinggalkan akan dicatat dan dimasukkan ke dalam sebuah daftar yang disebut tabu list. Tabu list juga digunakan untuk menghindari diterimanya kembali solusi sebelumnya menjadi current solution pada looping selanjutnya. Sebelum sebuah neighbour solution diperiksa, terlebih dahulu harus diperiksa apakah solusi tersebut sudah pernah masuk ke dalam tabu list. Jika sudah maka

A8 Seminar Nasional Teknologi Informasi 2011

Alat dan Bahan

Alat Penelitian

Alat – alat yang digunakan dalam penelitian meliputi :

1. Spesifikasi Hardware

a. Prosesor Intel(R) Atom(TM) CPU N475 @1,83GHz 1,83 GHz

b. RAM 1 GB

c. HDD 320 GB

d. Keyboard dan Mouse

e. Monitor

f. Printer

2. Spesifikasi Software

a. Sistem operasi yang digunakan, yaitu Windows 7 Home Premium

b. Microsoft office Word 2007, Microsoft Visio 2007, SQL Server Management Studio Ekspress 2005, Microsoft Visual Studio 2010.

Bahan Penelitian

Adapun bahan penelitian yang digunakan dalam pembuatan makalah ini, dan dijadikan acuan oleh penulis adalah sebagai berikut :

1. Daftar data dosen berupa kode dosen, nama dosen dan gelar.

2. Daftar data mata kuliah berupa kode mata kuliah, nama mata kuliah, semester dan SKS.

3. Daftar data ruang, berupa kode ruang dan kapasitas.

5. Analisis

Constrain yang berlaku :

1. Proses pembuatan jadwal kuliah mengenai penetapan waktu, mata kuliah dan dosen yang mengajar di lakukan oleh sekretaris jurusan sedangkan penetapan dilakukan oleh oleh staf tata usaha Jurusan Teknik Informatika.

2. Pengaturan jadwal kuliah dilakuakan setiap tahun sebanyak dua periode, yaitu :

a. Semester ganjil : awal September – akhir februari tahun berikutnya.

b. Semester genap : awal maret – akhir agustus tahun yang sama.

3. Data yang digunakan adalah nama dosen, nama mata kuliah, waktu kesediaan dosen mengajar, jumlah kelas dan ruangan yang tersedia.

a. Jumlah dosen berjumlah 31 orang yang terdiri dari 22 orang dosen tetap dan 9 orang dosen tidak tetap.

b. Jumlah mata kuliah adalah 68 mata kuliah dengan 146 sks.

Semester gasal: 34 mata kuliah dan 79 sks Semester genap: 34 mata kuliah dan 77 sks

c. Waktu perkuliahan yang diperbolehkan adalah Senin - Jumat pukul 08.00 – 16.10 WIB.

d. Jumlah ruang kelas ada 17 ruang yang terdiri dari 12 ruang belajar kelas dan 5 laboratorium.

4. Program kuliah satu SKS setara dengan 50 menit kegiatan tatap muka terjadwal, merupakan perkuliahan dimana dosen memberikan bahan kuliah di kelas untuk dipahami dan dimengerti.

Menentukan Nilai Cost parameter

Perhitungan cost pada algoritma Tabu Search telah dirancang untuk menghasilkan nilai 0, apabila kelas tersebut sudah memiliki komposisi dosen dan waktu kuliah yang paling baik, yang tidak melanggar semua constraint yang ada. Sebaliknya apabila nilainya mendekati 1, dapat disimpulkan bahwa jadwal yang dimiliki kelas tersebut mengandung banyak pelanggaran soft constraint. Rumus untuk menghitung nilai cost :

Nd Nh



Cdh

d 1 h 1

. Nh

Nk Nh



k1 h1

. Nh

r SdSk

Keterangan dimana :

Cdh : Jumlah cost untuk dosen ke d pada hari ke h

Ckh: Jumlah cost untuk kelas ke k pada hari ke h dimana cost a < x < b ,0

b < x < c , (x-b/c-b)

c < x < d,1

Sk : Total nilai cost pada soft constraint kelas

Sd : Total nilai cost pada soft constraint dosen

Nk : Jumlah kelas

Nh : Jumlah hari

Nd : Jumlah seluruh dosen

r : Nilai cost total yang dihasilkan

Agar solusi yang diinginkan tercapai, tiap pelanggaran soft constraint didefinisikan sebagai nilai cost yang berbeda. Hal ini dilakukan karena tingkat keharusan untuk

Seminar Nasional Teknologi Informasi 2011 A8

memenuhi sebuah soft constraint berbeda antara satu dengan yang lain. Soft constraint yang lebih perlu untuk dipenuhi diberi bobot yang lebih tinggi dari pada soft constraint yang dipandang lebih bisa ditoleransi. Berikut adalah nilai cost untuk masing-masing pelanggaran soft constraint:

Jadwal mengajar seorang dosen dalam sehari maksimal 6 slot waktu.

Merepresentasi dengan Tabu Search

1. Cari kelas yang memiliki nilai cost yang terbesar atau paling banyak melanggar soft constraint.

2. Cari kombinasi komponen utama yang memenuhi hard constraint.

3. Periksa apakah neighbour solution sudah tercatat dalam Tabu List, bila sudah kembali kelangkah ke dua.

4. Alokasikan komponen–komponen utama

penjadwalan.

5. Catat neighbour solution ke dalam Tabu List.

6. Periksa nilai cost secara keseluruhan apakah telah memenuhi syarat untuk berhenti, jika belum maka ulangi langkah 1.

7. Pembahasan

Hasil Skenario Pengujian

Skenario pengujian yaitu mengobservasi parameter-parameter yang digunakan pada Algoritma Tabu Search (TS) pada Soft Constraint, yaitu Random search, TSMaxIter dan Target Cost. Hal ini bertujuan untuk melihat bagaimana pengaruh masing-masing parameter pada algoritma, sehingga untuk menghasilkan solusi yang optimal bisa menggunakan parameter yang sesuai. Ketentuan-ketentuan yang digunakan dalam skenario ini yaitu:

a. Data yang digunakan yaitu data semester genap tahun ajaran 2010/2011.

b. Nilai-nilai parameter yang digunakan untuk pengujian yaitu Random search, TSMaxIter dan Target Cost.

C. Masing-masing skenario dilakukan 10 kali uji untuk mendapatkan hasil pengujian yang valid.

Tabel 1Hasil Pengujian

no Random TSMaxiter Rata-rata Cost

1 100 5 0,345296

2 500 5 0,219553

3 300 7 0,618906

4 500 7 0,219546

5 300 8 0,319268

6 500 8 0,119103

7 300 9 0,218819

8 500 9 0,119268

9 50 10 0,119478

10 300 10 0,019466

11 500 10 0,219368

12 1000 10 0,319608

Tabel 2 Detil Pengujian

No HD HR HM SD SK Total Hard Total Soft Total

Cost

1 0 0 0 0,00185 0,01843 0 0,02028 0,02028

2 0 0 0 0 0,02003 0 0,02003 0,02003

3 0 0 0 0,00062 0,01923 0 0,01985 0,01985

4 0 0 0 0,00062 0,02003 0 0,02065 0,02065

5 0 0 0 0,00062 0,01603 0 0,01665 0,01665

6 0 0 0 0,00123 0,01923 0 0,02046 0,02046

7 0 0 0 0 0,01843 0 0,01843 0,01843

8 0 0 0 0 0,02003 0 0,02003 0,02003

9 0 0 0 0 0,01843 0 0,01843 0,01843

10 0 0 0 0,00062 0,01923 0 0,01985 0,01985

jumlah 0 0 0 0,00556 0,1891 0 0,19466 0,19466

rata-rata 0 0 0 0,000556 0,01891 0 0,0194

66 0,019466

Keterangan nilai pinalti pada tabel diatas adalah sebagai berikut :

o Hd adalah besarnya nilai pinalti (pelanggaran) hard constraint dosen.

o Hr adalah besarnya nilai pinalti (pelanggaran) hard constraint ruang

o Hm adalah besarnya nilai pinalti (pelanggaran) hard constraint kelas mahasiswa

o Sd adalah besarnya nilai pinalti (pelanggaran) soft constraint dosen (dalam himpunan fuzzy).

o Skk adalah besarnya nilai pinalti (pelanggaran) soft constraint kelas kuliah (dalam himpunan fuzzy) .

A8 Seminar Nasional Teknologi Informasi 2011

- Hard Constraint dan Total pelanggaran Hard (hard constraint Dosen, Mahasiswa (Kelas Kuliah) dan Ruang)

Dari 10 kali percobaan pada tabel diatas, tidak ada satupun pelanggaran pada hard constraint dosen,mahasiswa (kelas kuliah) dan ruang semuanya bernilai 0. Nilai cost tergantung dari solusi awal yang dihasilkan yang dilakukan secara random dan metoda pencarian tetangga yaitu move dan swap.

- Soft Constraint Dosen

Dari 10 kali percobaan pada tabel diatas ,nilai rata-rata

soft constraint dosen adalah 0,000556 nilai ini terendah jika dibandingkan dengan nilai rata-rata soft constraint mahasiswa. Hal ini disebabkan karena jadwal jam mengajar setiap dosen lebih sedikit dibandingkan jadwal kuliah mahasiswa dalam satu hari. Nilai terendah dari soft constraint dosen adalah 0.

- Soft Constraint Mahasiswa (Kelas Kuliah Mahasiswa). Dari 10 kali percobaan pada tabel diatas,nilai rata-rata soft constraint Mahasiswa adalah 0,01891 nilai ini tinggi jika dibandingkan dengan nilai rata-rata soft constraint dosen. Hal ini disebabkan karena jumlah jam atau jadwal kuliah mahasiswa dalam sehari lebih banyak dibandingkan jumlah jam atau jadwal mengajar dosen.Nilai terendah dari soft constraint mahasiswa (kelas kuliah) adalah 0,01603.

- Total pelanggaran Soft Constraint

Rata-rata total pelanggaran soft constraint adalah 0,019466. Total pelanggaran soft constraint tertinggi pada percobaan ke 4 adalah 0,02065, sedangkan pelanggaran soft constraint terendah pada percobaan ke 5 adalah 0,01665.

Analisis Parameter

Berdasarkan hasil perbandingan 10 kali pengujian grafik terdapat rata-rata cost terendah yaitu kombinasi parameter yang menghasilkan pelanggaran paling sedikit ada pada pengujian nomor 10, dimana Random search = 300 dan TSMaxiter = 10 menghasilkan rata-rata cost = 0,019466. Berikut ini adalah grafik dari pengujian tersebut(diambil salah satu dari 10 kali ujicoba).

Gambar 3 Cost Soft Constrain

Gambar 4 Cost Hard Constrain

Berdasarkan gambar 3 dan 4 bahwa parameter yang mempengaruhi Tabu Search adalah jumlah iterasi. Tetapi hal ini juga dipengaruhi oleh probabilitas nilai random yang dibangkitkan pada setiap yang diinputkan untuk mendapatkan hasil penjadwalan kuliah menuju arah yang optimal. Karena inisialisasi awal selalu diawali dengan

jadwal yang di generate secara random.

8. Kesimpulan

1. Jadwal yang dihasilkan dalam setiap percobaan adalah berbeda, maka hasil yang didapatkan akan bervariasi setiap run time, karena inisialisasi awal selalu diawali dengan jadwal yang di generate secara random.Kualitas solusi yang didapatkan juga bervariasi, maka soft constraint pada kasus ini sulit mendapatkan jadwal yang valid dengan menghasilkan nilai pelanggaran 0.

2. Dalam pengujian yang dilakukan pada kasus Implementasi Penjadwalan Kegiatan Belajar Mengajar dengan Algoritma Tabu Search (TS) menggunakan Soft Constraint,menemukan parameter yang baik yaitu Random = 300 dan TSMaxiter = 10, dimana hasil rata-rata hard constraint adalah 0 dan hasil rata-rata soft constraint adalah 0,019466, hasil soft constraint tersebut telah memenuhi penjadwalan ke arah yang optimal.

REFERENSI

[1] Glover,F, 1990, “Tabu Search, Part II”,ORSA Journal on Computing, Vol 2 No 1 pp 4 – 32

[2] Glover,F, 1989,” Candidate List strategies and Tabu search”, CAAI Research report, University Of Colorado, Boulder.

[3] Glover,F, and greenberg HJ, 1989, “New Approaches for Heuristic Search. A bilateral link-age with Artificial Intelligence, European Journal Of Operational Research, Vol 39 No 2 p 119 – 130.

[4] Hertz A and de Werra,D., 1987, “Using Tabu search Techniques for graph coloring”, Computing, Vol 29 p 345 – 351.

[5] Hertz A and de Werra,D.D. Forthcoming, 1987,” The Tabu search Meta Heuristic : How We Use it”, Annals of Mathematics and Artificial Intelligence.

Arduino and Processing Workshop

http://www.cla.purdue.edu/vpa/etb/

Fabian Winkler

rev. 3

Required software/hardware/materials for this workshop:

Description Get it from Jameco.com

Arduino board, preferably Duemilanove (w/ USB cable) www.arduino.cc

Arduino software

www.arduino.cc

Processing software

www.processing.org

Breadboard PN 20601

Pushbutton switch, SPST PN 174414

Resistor (1 – 10kΩ) PN 690865

Some hookup wire PN 36792

Processing

What is Processing?

From http://www.processing.org: Processing is an open source programming language and environment for people who want to program images, animation, and interactions. It is used by students, artists, designers, researchers, and hobbyists for learning, prototyping, and production. It is created to teach fundamentals of computer programming within a visual context and to serve as a software sketchbook and professional production tool. Processing is developed by artists and designers as an alternative to proprietary software tools in the same domain.

Toolbar

Tabs

Text editor

Message area Console

Winkler, Processing/Arduino workshop, p.2

Basic Tips&Tricks

• Mark a word in the Processing code and ctrl click (Macintosh) on it to display information about it in the Processing reference.

• Where can I find my files?

Each sketch resides in its own folder where the main program file is located with the ending “.pde”. You can browse to this folder by choosing Sketch > Show Sketch Folder from the Processing menu.

Example:

Sketch name: “Sketch_01", the directory for the sketch will be called "Sketch_01",

the main file will be called "Sketch_01.pde".

• Please read “Processing – A Programming Handbook for Visual Designers and Artists” by Casey Reas and Ben Fry if you are interested in a thorough introduction to the software and its functions. You can also follow the online tutorials at: http://www.processing.org/learning/

Extending Processing’s functionality – libraries

Processing has become an extremely powerful scripting environment for the inclusion of almost any type of media – mainly through the concept of libraries. These libraries consist of subroutines and code that extend the functionality of Processing, often to include a particular kind of media (such as video, sound or 3D graphics) or functionality (communications, interface design, fullscreen playback, etc.).

To install a library, unzip the file into the “libraries” folder in your sketchbook (on the Macintosh you find the sketchbook in your home directory > Documents > Processing). If you do not already have a libraries folder in your sketchbook, create this folder manually and unzip the files in there. In general, most libraries have instructions on the website where you can download them from, following these directions is the easiest way to install a library.

Processing and the Arduino board

For a thorough introduction to the Arduino board see: Banzi, Massimo. Getting Started with Arduino (Make: Projects). Sebastopol, CA: Make Books, 2008.

Additionally, you can also read pp. 633-657 in “Processing – A Programming Handbook for Visual Designers and Artists” by Casey Reas and Ben Fry for an excellent introduction on how to connect the Arduino board with Processing and for code and circuit examples.

Winkler, Processing/Arduino workshop, p.3

The following examples are very basic, for more advanced Processing/Arduino interactions read: http://www.arduino.cc/playground/Interfacing/Processing

In detail, the next 9 examples illustrate the following ideas:

1) triggering content in Processing reading a pushbutton switch (p. 5)

2) animating a series of still images in Processing (p. 6)

3) animating a sequence of images based on pushing a button (p. 7)

4) randomly triggering images based on pushing a button (p. 8)

5) playing back sounds by pressing a button (p. 9)

6) playing back a Quicktime movie by pressing a button (p. 10)

7) connecting more than one digital (ON/OFF) sensor (p. 11)

8) controlling content in Processing with one analog sensor (p. 19)

9) multiple analog sensors, virtual etch-a-sketch (p. 20)

(1) Reading a Pushbutton:

This is the most basic setup that allows Processing to respond to events in the external

world. Depending on the state of the button a square changes its color.

Circuit – using Fritzing (http://fritzing.org) to visualize the connections between the Arduino and the components on the breadboard:

Winkler, Processing/Arduino workshop, p.4

Code for the Arduino board:

int switchPin = 7;

int LEDPin = 13;

void setup() {

pinMode (switchPin, INPUT);

pinMode (LEDPin, OUTPUT);

Serial.begin(9600);

}

void loop() {

if (digitalRead(switchPin) == HIGH) {

Serial.print(0, BYTE);

digitalWrite(LEDPin, LOW);

} else {

Serial.print(1, BYTE);

digitalWrite(LEDPin, HIGH);

}

delay(100);

}

Processing Code:

import processing.serial.*;

Serial port;

int val;

void setup() {

size(400, 400);

noStroke();

frameRate(10);

println(Serial.list());

// print a list of all available ports

port = new Serial(this, Serial.list()[0], 9600);

// choose the port to which the Arduino is connected

// on the PC this is usually COM1, on the Macintosh

// this is usually tty.usbserial-XXX

}

void draw() {

if (0 < port.available()) {

val = port.read();

}

background(204);

if (val == 0) {

fill(255, 127, 0); // fill orange

} else {

fill(0, 170, 0); // fill green

}

rect(50, 50, 300, 300);

}

Winkler, Processing/Arduino workshop, p.5

(2) Animating a sequence of images in Processing (without the Arduino board) – in order for this example to work you need to “add” all of the frames of your animation to the sketch first. Go to: Sketch > Add file... and then choose the frames you want to display. These will be copied into a folder called “data” which resides in your sketch folder.

int numFrames = 15; // The number of frames in the animation

int frame = 0;

PImage[] images = new PImage[numFrames];

void setup()

{

size(400, 300);

frameRate(10);

images[0] = loadImage("money_0000.jpg"); images[1] = loadImage("money_0001.jpg"); images[2] = loadImage("money_0002.jpg"); images[3] = loadImage("money_0003.jpg"); images[4] = loadImage("money_0004.jpg"); images[5] = loadImage("money_0005.jpg"); images[6] = loadImage("money_0006.jpg"); images[7] = loadImage("money_0007.jpg"); images[8] = loadImage("money_0008.jpg"); images[9] = loadImage("money_0009.jpg"); images[10] = loadImage("money_0010.jpg"); images[11] = loadImage("money_0011.jpg"); images[12] = loadImage("money_0012.jpg"); images[13] = loadImage("money_0013.jpg"); images[14] = loadImage("money_0014.jpg");

// If you don't want to load each image separately

// and you know how many frames you have, you

// can create the filenames as the program runs.

// The nf() command does number formatting, which will

// ensure that the number is (in this case) 4 digits.

// for(int i=0; i<numFrames; i++) {

// String imageName = "PT_anim" + nf(i, 4) + ".gif";

// images[i] = loadImage(imageName);

//}

}

void draw()

{

if (frame<numFrames-1) {

frame = (frame+1);

} else {

frame = 0;

}

image(images[frame], 0, 0);

}

Winkler, Processing/Arduino workshop, p.6

(3) Animating a sequence of images based on pushing a button The Arduino code and circuit setup is the same as in example 1 Processing code:

import processing.serial.*;

Serial port;

int val = 0;

int oldval = 0;

int numFrames = 15; // The number of frames in the animation

int frame = 0;

PImage[] images = new PImage[numFrames];

void setup()

{

size(400, 300);

frameRate(10);

// If you don't want to load each image separately

// and you know how many frames you have, you

// can create the filenames as the program runs.

// The nf() command does number formatting, which will

// ensure that the number is (in this case) 4 digits.

// for(int i=0; i<numFrames; i++) {

// String imageName = "PT_anim" + nf(i, 4) + ".gif";

// images[i] = loadImage(imageName);

//}

println(Serial.list());

// print a list of all available ports

port = new Serial(this, Serial.list()[0], 9600);

// choose the port to which the Arduino is connected

// on the PC this is usually COM1, on the Macintosh

// this is usually tty.usbserial-XXX

}

void draw()

{

if (0 < port.available()) {

Winkler, Processing/Arduino workshop, p.7

val = port.read();

}

if (val != oldval && val == 1) {

// the line above makes sure we advance only one frame with

// each pressing of the button

if (frame<numFrames-1) {

frame = (frame+1);

} else {

frame = 0;

}

image(images[frame], 0, 0);

oldval = val;

}

(4) Randomly triggering images based on pushing a button.

This example code is very similar to the previous one, only now the images are not

played back sequentially but based on the value of a random number.

import processing.serial.*;

Serial port;

int val = 0;

int oldval = 0;

int numFrames = 15; // The number of frames in the animation

int rand_frame = 0;

PImage[] images = new PImage[numFrames];

void setup()

{

size(400, 300);

frameRate(10);

// If you don't want to load each image separately

// and you know how many frames you have, you

// can create the filenames as the program runs.

// The nf() command does number formatting, which will

// ensure that the number is (in this case) 4 digits.

// for(int i=0; i<numFrames; i++) {

Winkler, Processing/Arduino workshop, p.8

// String imageName = "PT_anim" + nf(i, 4) + ".gif";

// images[i] = loadImage(imageName);

//}

println(Serial.list());

// print a list of all available ports

port = new Serial(this, Serial.list()[0], 9600);

// choose the port to which the Arduino is connected

// on the PC this is usually COM1, on the Macintosh

// this is usually tty.usbserial-XXX

}

void draw()

{

if (0 < port.available()) {

val = port.read();

}

if (val != oldval && val == 1) {

// the line above makes sure we advance only one frame with

// each pressing of the button

rand_frame = int(random(numFrames-1));

}

image(images[rand_frame], 0, 0);

oldval = val;

}

(5) Playing back sounds by pressing a button.

This example plays back a sound (and loops it) when the button is pressed. When the button is released the sound stops and it continues to play where it was paused the next time the button is pressed. This example uses the “minim” library (see: http://code.compartmental.net/tools/minim/). You might have to check the Javadocs at: http://code.compartmental.net/minim/javadoc/ to find out all about the methods you can use with certain classes, such as the AudioPlayer class.

import processing.serial.*;

import ddf.minim.*;

Minim minim;

AudioPlayer mySound;

Serial port;

int val = 0;

int oldval = 0;

int playback_pos = 0;

void setup() {

size(400, 400);

noStroke();

frameRate(10);

println(Serial.list());

// print a list of all available ports

port = new Serial(this, Serial.list()[0], 9600);

// choose the port to which the Arduino is connected

// on the PC this is usually COM1, on the Macintosh

Winkler, Processing/Arduino workshop, p.9

// this is usually tty.usbserial-XXX

minim = new Minim(this);

mySound = minim.loadFile("expressway.aiff");

// make sure you the file “expressway.aiff” in your “data” folder

fill(255, 127, 0);

}

void draw() {

if (0 < port.available()) {

val = port.read();

}

background(204);

if (val != oldval && val == 1) {

// the line above makes sure we only trigger the sound once

// when the pushbutton is pressed & held down

fill(0, 170, 0); // fill green

mySound.play(playback_pos);

// print("played back at: ");

// println(playback_pos);

// continues to playback where the sound was paused by

// remembering the playback head's position

mySound.loop();

}

if (val != oldval && val == 0) {

// the line above makes sure we only pause the sound once

// when the pushbutton is released

fill(255, 127, 0); // fill orange

playback_pos = mySound.position();

// print("paused at: ");

// println(playback_pos);

// saves the position of the playback head when paused

mySound.pause();

}

rect(50, 50, 300, 300);

oldval = val;

}

void stop()

{

// always close Minim audio classes when you are done with them

mySound.close();

minim.stop();

super.stop();

}

(6) Playback a Quicktime movie.

This example shows how you can play back a Quicktime video by pressing a button connected to the Arduino board. Whenever the button is pressed the video starts playing back, when the button is released the video pauses.

Winkler, Processing/Arduino workshop, p.10

import processing.video.*;

import processing.serial.*;

Serial port;

Movie myMovie;

int val = 0;

void setup() {

size(320, 240);

background(0);

myMovie = new Movie(this, "spooky2.mov");

myMovie.loop();

println(Serial.list());

// print a list of all available ports

port = new Serial(this, Serial.list()[0], 9600);

// choose the port to which the Arduino is connected

// on the PC this is usually COM1, on the Macintosh

// this is usually tty.usbserial-XXX

}

void draw() {

background(255);

if (0 < port.available()) {

val = port.read();

}

image(myMovie, 0, 0);

if (val == 0) {

myMovie.speed(0);

} else {

myMovie.speed(1);

}

// Called every time a new frame is available to read

void movieEvent(Movie m) {

m.read();

}

(7) Connecting more than one digital (ON/OFF) sensor

I developed the following code for a student in an earlier AD101 course who wanted to read more than one digital sensor with the Arduino board and trigger separate events in Processing – one for each sensor. The following code is designed for 7 digital sensors connected to pins 5-11 on the Arduino board:

Arduino code:

int inA = 5;

int inB = 6;

int inC = 7;

int inD = 8;

int inE = 9;

int inF = 10;

Winkler, Processing/Arduino workshop, p.11

int inG = 11;

int current_inA = 0;

int previous_inA = 0;

int current_inB = 0;

int previous_inB = 0;

int current_inC = 0;

int previous_inC = 0;

int current_inD = 0;

int previous_inD = 0;

int current_inE = 0;

int previous_inE = 0;

int current_inF = 0;

int previous_inF = 0;

int current_inG = 0;

int previous_inG = 0;

int LEDPin = 13;

void setup() {

pinMode (inA, INPUT); pinMode (inB, INPUT); pinMode (inC, INPUT); pinMode (inD, INPUT); pinMode (inE, INPUT); pinMode (inF, INPUT); pinMode (inG, INPUT);

pinMode (LEDPin, OUTPUT);

Serial.begin(9600);

}

void loop() {

// read all sensors:

current_inA = digitalRead(inA); current_inB = digitalRead(inB); current_inC = digitalRead(inC); current_inD = digitalRead(inD); current_inE = digitalRead(inE); current_inF = digitalRead(inF); current_inG = digitalRead(inG);

// SENSOR A

if (current_inA != previous_inA) {

// send out serial data only if input changes

// this is to avoid serial port overflow

switch (current_inA){

case HIGH:

Serial.print("A"); // identifier for sensor A

Serial.print(0);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

case LOW:

Serial.print("A"); // identifier for sensor A

Winkler, Processing/Arduino workshop, p.12

Serial.print(1);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

}

// SENSOR B

if (current_inB != previous_inB) {

// send out serial data only if input changes

// this is to avoid serial port overflow

switch (current_inB){

case HIGH:

Serial.print("B"); // identifier for sensor B

Serial.print(0);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

case LOW:

Serial.print("B"); // identifier for sensor B

Serial.print(1);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

}

// SENSOR C

if (current_inC != previous_inC) {

// send out serial data only if input changes

// this is to avoid serial port overflow

switch (current_inC){

case HIGH:

Serial.print("C"); // identifier for sensor C

Serial.print(0);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

case LOW:

Serial.print("C"); // identifier for sensor C

Serial.print(1);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

}

// SENSOR D

if (current_inD != previous_inD) {

// send out serial data only if input changes

// this is to avoid serial port overflow

switch (current_inD){

case HIGH:

Winkler, Processing/Arduino workshop, p.13

Serial.print("D"); // identifier for sensor D

Serial.print(0);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

case LOW:

Serial.print("D"); // identifier for sensor D

Serial.print(1);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

}

// SENSOR E

if (current_inE != previous_inE) {

// send out serial data only if input changes

// this is to avoid serial port overflow

switch (current_inE){

case HIGH:

Serial.print("E"); // identifier for sensor E

Serial.print(0);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

case LOW:

Serial.print("E"); // identifier for sensor E

Serial.print(1);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

}

// SENSOR F

if (current_inF != previous_inF) {

// send out serial data only if input changes

// this is to avoid serial port overflow

switch (current_inF){

case HIGH:

Serial.print("F"); // identifier for sensor F

Serial.print(0);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

case LOW:

Serial.print("F"); // identifier for sensor F

Serial.print(1);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

}

Winkler, Processing/Arduino workshop, p.14

// SENSOR G

if (current_inG != previous_inG) {

// send out serial data only if input changes

// this is to avoid serial port overflow

switch (current_inG){

case HIGH:

Serial.print("G"); // identifier for sensor G

Serial.print(0);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

case LOW:

Serial.print("G"); // identifier for sensor G

Serial.print(1);

Serial.print(10, BYTE); // ASCII 10 = newline character, used to

// separate the data strings

break;

}

previous_inA = current_inA; previous_inB = current_inB; previous_inC = current_inC; previous_inD = current_inD; previous_inE = current_inE; previous_inF = current_inF; previous_inG = current_inG;

}

Processing code:

import processing.serial.*;

String buff = "";

String temp = "";

int NEWLINE = 10; // this is the ASCII code for a newline character

int input_A = 0;

int input_B = 0;

int input_C = 0;

int input_D = 0;

int input_E = 0;

int input_F = 0;

int input_G = 0;

Serial port;

void setup() {

size(660, 480);

noStroke();

Winkler, Processing/Arduino workshop, p.15

background(204);

println(Serial.list());

// print a list of all available ports

port = new Serial(this, Serial.list()[0], 9600); // choose the port to which the Arduino is connected // on the PC this is usually COM1, on the Macintosh // this is usually tty.usbserial-XXX

}

void draw() {

while (port.available() > 0) { serialEvent(port.read());

// call to function "serialEvent" --> see below }

// do something with SENSOR A input:

switch (input_A){

case 0:

fill(255, 255, 255);

break;

case 1:

fill(255, 127, 255);

break;

}

rect(30, 120, 60, 240);

// do something with SENSOR B input:

switch (input_B){

case 0:

fill(255, 255, 255);

break;

case 1:

fill(255, 127, 255);

break;

}

rect(120, 120, 60, 240);

// do something with SENSOR C input:

switch (input_C){

case 0:

fill(255, 255, 255);

break;

case 1:

fill(255, 127, 255);

break;

}

rect(210, 120, 60, 240);

// do something with SENSOR D input:

switch (input_D){

case 0:

fill(255, 255, 255);

break;

case 1:

Winkler, Processing/Arduino workshop, p.16

fill(255, 127, 255);

break;

}

rect(300, 120, 60, 240);

// do something with SENSOR E input:

switch (input_E){

case 0:

fill(255, 255, 255);

break;

case 1:

fill(255, 127, 255);

break;

}

rect(390, 120, 60, 240);

// do something with SENSOR F input:

switch (input_F){

case 0:

fill(255, 255, 255);

break;

case 1:

fill(255, 127, 255);

break;

}

rect(480, 120, 60, 240);

// do something with SENSOR G input:

switch (input_A){

case 0:

fill(255, 255, 255);

break;

case 1:

fill(255, 127, 255);

break;

}

rect(570, 120, 60, 240);

}

void serialEvent(int serial) {

if(serial != NEWLINE) {

buff += char(serial);

} else {

if (buff.length() > 1) {

temp = buff.substring(0, buff.length()-(buff.length()-1));

// this isolates just the first letter to identify the sensor

}

if (temp.equals("A") == true) { // identifies sensor A value temp = buff.substring(1, buff.length());

input_A = int(temp); // get sensor A value; }

if (temp.equals("B") == true) { // identifies sensor B value temp = buff.substring(1, buff.length());

input_B = int(temp); // get sensor B value;

Winkler, Processing/Arduino workshop, p.17

}

if (temp.equals("C") == true) { // identifies sensor C value temp = buff.substring(1, buff.length());

input_C = int(temp); // get sensor C value; }

if (temp.equals("D") == true) { // identifies sensor D value temp = buff.substring(1, buff.length());

input_D = int(temp); // get sensor D value; }

if (temp.equals("E") == true) { // identifies sensor E value temp = buff.substring(1, buff.length());

input_E = int(temp); // get sensor E value; }

if (temp.equals("F") == true) { // identifies sensor F value temp = buff.substring(1, buff.length());

input_F = int(temp); // get sensor F value; }

if (temp.equals("G") == true) { // identifies sensor G value temp = buff.substring(1, buff.length());

input_G = int(temp); // get sensor G value; }

}

Winkler, Processing/Arduino workshop, p.18

Analog Sensors

(8) controlling content in Processing with one analog sensor

The following example illustrates how you can control 256 lines in a Processing window with a potentiometer connected to the Arduino board. If the potentiometer is at its minimum limit, the Processing sketch will draw just one line on the left hand side of the screen. The more the potentiometer’s knob is turned the more lines will be drawn on the screen until the whole screen is filled with 256 lines (potentiometer at its maximum limit). This example uses a for loop to iterate through a set of numbers ranging from 0 to the current number that the Arduino board sends from the potentiometer. This number determines the number of lines drawn on the screen and their distribution. We also use this number to gradually change the color of the lines drawn to the screen. The code for the Arduino board is the following:

int val;

int inputPin = 2; // this is the pin to which we

// connect the potentiometer

void setup() {

Serial. begin(9600);

}

void loop() {

val = analogRead(inputPin)/4; // the potentiometer’s range is from

// 0-1023 but we only need numbers from

Winkler, Processing/Arduino workshop, p.19

// 0-255.

Serial.print(val, BYTE);

delay (100);

}

Processing code:

import processing.serial.*;

Serial port;

int val;

void setup() {

size(1000, 500);

//noStroke();

frameRate(10);

println(Serial.list());

port = new Serial(this, Serial.list()[0], 9600);

// choose the port that the Arduino is

// connected to, on a Macintosh choose

// the tty.usbserial port

}

void draw() {

if (0 < port.available()) {

val = port.read();

}

background(204);

for (int i = 0; i < val+1; i++) {

// this is the “for” loop, see pp.61

// in the book “Processing” by Reas and Fry

stroke(255-i, 0, i); // choose the color of the line based on the // potentiometer’s value

line (i*5, 100, (i*5)+10, 400); // draw the lines based on the

// potentiometer’s value

}

println(val); // prints out values from Arduino board in // the console, just to see what Processing // actually receives from the Arduino board

}

(9) multiple analog sensors

This example uses two potentiometers to recreate the popular etch-a-sketch toy in Processing. The example is adapted from Christian Nold’s code and was used for a project in my Physical computing class in 2008.

Arduino code:

int potPin1 = 4; // select the input pin for potentiometer 1

int int potPin2 = val1 = 0; 5; // select the input pin for potentiometer 2

// variable to store the value coming from potentiometer 1

Winkler, Processing/Arduino workshop, p.20

int val2 = 0;

// variable to store the value coming from potentiometer 2

void setup() {

Serial.begin(9600);

}

void loop() {

val1 = analogRead(potPin1); // read the value from sensor 1

val2 = analogRead(potPin2); // read the value from sensor 2

Serial.print("A");

// Example identifier for sensor 1, send as character

Serial.print(val1); // send sensor 1 value as decimal number

Serial.print(10, BYTE); // send ASCII string "10"

// ASCII 10 = newline character,

// used to separate the data strings

Serial.print("B");

// Example identifier for sensor 2, send as character

Serial.print(val2); // send sensor 2 value as decimal number

Serial.print(10, BYTE); // send ASCII string "10"

// ASCII 10 = newline character,

// used to separate the data strings

}

Processing code:

/* Virtual Etch A Sketch using 2 Potentiometers

This program reads two analog sensors via serial and draws their

values as X and Y

Processing Code

Christian Nold, 22 Feb 06

slightly modified by Fabian Winkler

March 2008

*/import processing.serial.*;

String buff = "";

String temp = "";

float temporary = 0.0;

float screenwidth = 0.0;

float xCoordinate = 0;

float yCoordinate = 0;

int val = 0;

int NEWLINE = 10;

Serial port;

Winkler, Processing/Arduino workshop, p.21

void setup()

{

size(200, 200);

strokeWeight(10); // fat

stroke(255);

smooth();

println(Serial.list());

port = new Serial(this, Serial.list()[0], 9600);

// change [0] to the correct number of your serial port

// e.g. tty.usbserial on the Mac and COM1 or COM 2 on the PC

}

void draw()

{

fill(0,2); // use black with alpha 2

rectMode(CORNER);

rect(0,0,width,height);

while (port.available() > 0) {

serialEvent(port.read());

}

point(xCoordinate, yCoordinate);

}

void serialEvent(int serial)

{

if(serial != NEWLINE) {

buff += char(serial);

} else {

if (buff.length() > 1) {

temp = buff.substring(0, buff.length()-(buff.length()-1));

// this isolates just the beginning character of the sensor

// identified

if(temp.equals("A") == true) { //sensor A value

temp = buff.substring(1, buff.length());

temporary = float(temp);

xCoordinate = width/(1024/temporary);

println(xCoordinate);

}

if(temp.equals("B") == true) { //sensor B value

temp = buff.substring(1, buff.length());

temporary = float(temp);

yCoordinate = height/(1024/temporary);

println(yCoordinate);

Winkler, Processing/Arduino workshop, p.22

A Real Estate Transaction Process Based on Blockchain

An Enhanced Real Estate Transaction

Process Based on Blockchain Technology

Emergent Research Forum Paper

Dr. Atefeh Mashatan

Ted Rogers School of Management

Ryerson University

amashatan@ryerson.ca

Zachary Roberts

Ted Rogers School of Management

Ryerson University

zroberts@ryerson.ca

Abstract

We discuss the current state of the Canadian real estate market and the impact blockchain technology could have on it. We start with a review of the current popular and scholarly literary landscape relating to blockchain technology and its real estate applications. Special focus is given to the impact the technology could have on transaction costs, transaction times and fraud deterrence in this market. Preliminary recommendations are provided in the form of blockchain based bidding and transaction systems.

Keywords: Blockchain, real estate, shared ledger, transaction process, bidding process, fraud.

Introduction

Over the past few years, the Canadian Real Estate market has been growing significantly as a proportion of total Canadian GDP (Argitis 2016). In major Canadian housing markets like Toronto and Vancouver increasing prices are beginning to make housing affordability a major concern (Kelly, et al. 2017). Transaction costs in Canada can range from between 11-22% of the total purchase price (Global Property Guide 2016). This data implies that improvements in transaction costs could have a significant effect on housing affordability for Canadian citizens.

Numerous reports indicate that fraud is a serious problem in the Canadian real estate market. The closed-bidding system currently used makes it hard to prove wrongdoing by brokers acting fraudulently (Malik and Foxcroft 2016). Other types of real estate fraud also occur in Canada, with title fraud being the most devastating to the victims financially (Myrick 2013). At present, little information exists on the total savings that could come from revamping the existing system. This research project intends to examine if changing the business process which currently records property transactions to one based on blockchain technology would significantly reduce the transactions costs, transaction times, and reduce fraud in the Canadian market. The project will attempt to identify the benefits, the likely costs of implementation, and give recommendations with potential alternative options.

This paper will address the challenges within the existing system and recommend an innovative business process using emerging blockchain technology. The literature review that informs the recommendations was conducted by systematically reviewing the available popular and scholarly literature on blockchain and the Canadian real estate market. The recommended process will be designed to increase efficiency, transparency, and fraud resistance. At present this paper is focused primarily on the Canadian real estate market. However, the techniques and solutions used here can be expanded or adopted to other alike markets. This paper is currently a work in-progress but will later be expanded to cover all the questions introduced here and examine the impact on international markets.

Blockchain Technology

Blockchain is a distributed ledger technology, the design of which was first released publicly in 2008 (Iansiti and Lakhani 2017). A blockchain ledger is a record of transactions that is shared across a network of users and individually verified by each participant (Swan 2015). Transactions are grouped together and tied to all previous transactions using a cryptographic technique called hashing (Swan 2015). Blocks of transactions linked together in this way make altering the record nearly impossible (Swan 2015). The difficulty is further

Twenty-third Americas Conference on Information Systems, Boston, 2017 1

Organizational Transformation & Information Systems

increased by the distributed nature of the system, which would require an alteration to be made simultaneously across all the network’s participants to be successful – something that becomes more difficult as the user base grows (Swan 2015).

Blockchain first became famous for its use as the primary technology enabling the Bitcoin cryptocurrency, launched in 2009 (Swan 2015). Blockchain technology has continued to be developed for use in numerous different applications and industries. The technology has the potential to disrupt existing markets because it has solved a fundamental problem with creating a digital economy: How do you prove ownership of an asset in an environment where everything can be copied or modified? With this problem solved, the internet can be used as the foundation to build out a large-scale transaction processing platform.

Many global financial institutions have been investing in blockchain technology in recent years. The blockchain start-up R3 leads a technology partnership of over 75 financial institutions with the mission to research applications of blockchain technology on banking and finance (R3 2017). The partnership includes some of the world’s largest banks such as JP Morgan, Bank of America, HSBC, and RBC (Williams-Grut 2015). The investment into the technology is likely due to predicted improvements available from implementations of blockchain technology in financial transactions like reduced transaction times, lower transaction costs, and decreased risk (Deloitte 2017).

The functionality of blockchain has been expanded to enable secured, decentralized contracts and the ability to track real-world asset ownership (Bheemaiah 2015). Smart contracts are digital agreements that can be enforced automatically without the use of 3rd parties or manual intervention (Morrison 2016). Blockchain enabled smart contracts can help simplify complex transactions by removing requirements for escrow, physical signatures, lawyers and manual remittance (Morrison 2016). Ownership of physical and digital assets can be tracked using the blockchain framework. By using a method called “colored coins”, an asset is assigned a unit on a blockchain, and whoever owns that specific digital unit is the owner of the underlying asset (Mizrahi 2015). This can be applied to different types of physical ownership like cars, art, or real estate property (Mizrahi 2015). Later in this paper we present a blockchain model for real estate transactions to allow for this type of digital ownership tracking of Canadian property.

Blockchain for Real Estate

Real estate is a uniquely good target for blockchain technology because it has a complex transaction process that is built to prevent fraudulent behavior and enable strict ownership protection. These traits required by the real estate market are where blockchain is uniquely strong. Blockchain can be enabled to improve the transparency a system enabling regulators to catch and prevent fraudulent behavior. The architecture of blockchain also enables the creation of an immutable record which can be trusted to not be modified or lost, something which is not possible with traditional electronic or paper records.

In Canada, transaction costs for real estate can include title search fees, land transfer taxes, appraisal costs, legal fees, agent fees, listing fees, estoppel certificate fees, and notary fees (Ratehub 2017). While a blockchain system may not be able to eliminate all transaction costs, it could likely simplify the process in most jurisdictions. For example, a blockchain system with all previous transaction information could simplify the title search process to reduce the need for title insurance (Speilman 2015). Already, some jurisdictions are beginning to implement blockchain based title tracking systems including Georgia, Honduras and Ghana (Dale 2016). A blockchain system is also attractive because it could reduce the total transaction time to complete a sale. The real estate industry has historically been slow to adopt new technologies and so in many areas, real estate transactions are competed primarily through paper-based processes (Kelly, et al. 2017). Digitizing and automating the processes with smart contracts could significantly reduce waiting times (Swan 2015).

In Canada, real estate fraud can come in many forms. Per the Globe and Mail, title fraud usually occurs when a criminal pretends to be the legitimate owner of a property, takes out credit against it, then flees with the funds (Myrick 2013). A criminal may also sell a property they do not own, though this is a rarer form of title fraud. These forms of real estate fraud are likely the most damaging for property owners (Myrick 2013). A recent investigation by CBC Marketplace revealed that another form of fraud perpetrated by real estate brokers may be more common. This type of fraud can occur when a broker represents both buyers and sellers of a property. Brokers can withhold bid information from sellers to earn a second commission from buyers they represent, which fraudulently lowers the price the seller receives. (Malik and Foxcroft 2016).

2 Twenty-third Americas Conference on Information Systems, Boston, 2017

A Real Estate Transaction Process Based on Blockchain

Blockchain systems could likely be designed to help reduce both these types of frauds by providing greater transparency to title ownership records and the bidding process.

Preliminary Recommendations

The expected contribution to the existing body of knowledge is a more complete understanding of the impact the technology will have on the real estate market with clear recommendations for decision makers in each jurisdiction studied. The primary research question we want to answer is, “Does it make sense for governments to invest in a blockchain based real estate processes?” While the research described in this paper is still ongoing, it is evident now that the Canadian real estate market could benefit from blockchain technology. The final research paper will attempt to quantify the financial benefits and estimate the return on investment for each jurisdiction. Here, we introduce our preliminary recommendations which include an outline for two blockchain systems designed to improve the bidding and transaction processes. For each of the proposed systems, a diagram is included to help visualize the process flows.

A Transparent Bidding Process Based on Blockchain Technology

The current bidding process is primarily paper-based and controlled by the agent representing the property seller. In the current system, bids are submitted by the buyers’ agents to the seller’s agent who compiles them and communicates the offers to the seller. Counter-offers are similarly communicated through the agents to the buyers. In Canada, agents are legally allowed to represent both buyers and sellers in a single real estate transaction (Malik and Foxcroft 2016). This practice enables the “double-ending” fraud discussed earlier. Real estate agents can give confidential information about the existing bids to buyers they represent to allow that buyer to win by a small margin and effectively double the agent’s commission at the expense of the seller and potential buyers who are represented by their agents and are willing to pay more. While this practice is illegal, in the current system this type of fraud is very difficult to police and likely quite common (Malik and Foxcroft 2016).

Transparent Bidding Blockchain (TBB) Process

A blockchain based system can help to improve the control of information in the bidding process. A system like one outlined in Figure 1, named the Transparent Bidding Blockchain (TBB) process, would allow for the bidding information to be collected independent from the Seller’s Agent by means of securely capturing the bids on a blockchain. Select bidding information could then be withheld from the Seller’s Agent until bidding is closed, but still communicated to the Seller in real time, decreasing the potential for fraud.

A TBB system could improve the bidding process from the perspective of Buyers as well. The Seller could opt to allow Buyers to view selected information following their bid. A Buyer could perhaps view the winning bid amount and be given the option to submit a second offer. Buyers could check on the status of their bid through the TBB without needing to speak with the Seller Agent as they do now. This increases the systems communication efficiency.

A TBB system could also enable conditional bids that execute automatically within the bidding and transaction processes. This could take the form of time-sensitive offers which automatically expire, or a bid conditional on a successful inspection of the property. The accepted bid’s conditions could be loaded into the transaction blockchain, like the one discussed below, and be used as the basis for a smart contract.

Twenty-third Americas Conference on Information Systems, Boston, 2017 3

Organizational Transformation & Information Systems

An Enhanced Real Estate Transaction Process Based on Blockchain Technology

The current real estate transaction process contains many steps and participants. This is due to the high value of the transactions which necessitates intermediaries to be involved to establish trust and to confirm things like financing and title ownership are in good standing. Some of the steps involved in the typical real estate transaction include: securing financing, property appraisal, purchasing title insurance, hiring a lawyer, transferring funds, and creating the purchase contract. The manual transmission of data between the participants for all these steps makes the process take a long time. In addition, while the participants are waiting for each step to be completed, the only way they can know about the status and progress is by asking for an update by those executing the individual tasks.

Smart Property Ledger (SPL)

Deploying the concept of “colored coins”, we propose the Smart Property Ledger (SPL) system outlined in Figure 2. This potential replacement for the current system is built upon a blockchain ledger with smart contract functionality. Throughout the transaction process, conditions of the smart contracts are met, like financing approval, approved by the users and passed to the next user. A blockchain system like this can help to improve efficiency in the transaction process. The SPL allows for all data to be transmitted in real time to all approved users with minimal latency. This will likely help to reduce the overall transaction time by reducing much of the waiting done between steps.

Figure 2 – Smart Property Ledger (SPL)

The SPL system would also improve transparency for the users. Users would be notified when approvals are given in the process, and when they are required to make an action. Participants would no longer need to call their agents or lawyers to get an update on the status of their transaction – they could simply log into the SPL and see for themselves. The existing system requires that a lengthy title search be conducted to ensure the property title being sold is free of defects. In our proposed system, the requirement for this search should be greatly reduced. The SPL will contain an easily accessible history of all previous

4 Twenty-third Americas Conference on Information Systems, Boston, 2017

A Real Estate Transaction Process Based on Blockchain

transactions on any asset, i.e., the property being sold, and enable a title search to be done much more quickly and with more confidence in its accuracy and legality.

Conclusion

This paper has discussed the current state of the Canadian real estate market and the impact blockchain technology could have on it. Special focus was given to the impact the technology could have on transaction costs, transaction times and fraud deterrence in this market. Blockchain based bidding and transaction systems were introduced which could help to improve the transaction experience for each participant by increasing transparency and the accuracy of the data. This research is a work in-progress. The scope will later be expanded to quantify the benefits and costs of the systems introduced here. The techniques and solutions will also later be applied to other markets outside of Canada with similar real estate systems in place. When the research described here is complete, it may have a significant impact on practical implementations of blockchain technology within the studied jurisdictions and could point to further research in industries that have similar characteristics to the real estate industry.

REFERENCES

Argitis, T. 2016. "Canada’s economy is growing at the slowest pace in 60 years and the only thing holding us up is housing." Financial Post, August 2. (http://business.financialpost.com/news/economy/canadas-economy-is-growing-at-the-slowest-pace-in-60-years-and-the-only-thing-holding-us-up-is-housing).

Bheemaiah, K. 2015. “Block Chain 2.0: The Renaissance of Money”. Wired. January. (https://www.wired.com/insights/2015/01/block-chain-2-0).

Dale, B. 2016. “Three Small Economies Where Land Title Could Use Blockchain to Leapfrog the US”. Observer. May 5. (http://www.observer.com/2016/10/benben-factom-bitfury-ghana-georgia-honduras/).

Deloitte. 2017. “Blockchain technology: 9 benefits & 7 challenges. Accessed February 12, 2017. (https://www2.deloitte.com/nl/nl/pages/innovatie/artikelen/blockchain-technology-9-benefits-and-7-challenges.html)

Global Property Guide. 2016. “Buying Property In Canada”. November 24. (www.globalpropertyguide.com/North-America/Canada/Buying-Guide).

Iansiti, M., and Lakhani K.R. 2017. "The Truth About Blockchain." Harvard Business Review, January-February Edition: 118-127. (https://hbr.org/2017/01/the-truth-about-blockchain).

Kelly, H.F., Billingsley, A.C., Warren, A. and Kramer, A. 2017. “PwC and the Urban Land Institute: Emerging Trends in Real Estate 2017”. Washington, D.C.: PwC and the Urban Land Institute.

Malik, S., and Foxcroft, T. 2016. “Real estate agents caught breaking the rules on Marketplace's hidden camera”. CBC News. November 3. (http://www.cbc.ca/news/business/real-estate-agents-caught-breaking-the-rules-on-marketplace-s-hidden-camera-1.3825841).

Mizrahi, A. 2015. "A blockchain-based property ownership registry." ChromaWay Corporate Website.

Accessed February 9, 2017. (http://chromaway.com/papers/A-blockchain-based-property-registry.pdf).

Morrison, A. 2016. “Blockchain and smart contract automation”. PwC Technology Forecast. (https://www.pwc.com/us/en/technology-forecast/blockchain.html).

Myrick, C. 2013. “Top 6 real estate scams – and how to avoid them”. The Globe and Mail. August 9. (http://www.theglobeandmail.com/real-estate/mortgages-and-rates/top-6-real-estate-scams-and-how-to-avoid-them/article13108985).

R3. 2017. “About - R3”. (http://www.r3cev.com/about).

Ratehub. 2017. “Closing Costs Overview”. February 11. (https://www.ratehub.ca/closing-costs-overview). Speilman, A. 2015. "Blockchain: digitally rebuilding the real estate industry." Masters Dissertation, Massachusetts Institute of Technology. (http://hdl.handle.net/1721.1/1 06753).

Swan, M. 2015. Blockchain. Sebastopol, CA. O’Reilly Media, Inc. (http://w2.blockchain-tec.net/blockchain/blockchain-by-melanie-swan.pdf).

Williams-Grut, O. 2015. "Blockchain: R3 membership hits 42, as it looks to non-banks." Business Insider. December 17. (http://www.businessinsider.com/blockchain-r3-membership-hits-42-as-it-looks-to-non-banks-2015-12).

Twenty-third Americas Conference on Information Systems, Boston, 2017 5

The Qualitative Report

Volume 18 | Number 13 Article 2

4-1-2013

Tasers and Community Controversy: Investigating

Training Officer Perceptions of Public Concern

Over Conducted Energy Weapons

Joseph De Angelis

Regis University

Brian Wolf

University of Idaho, bwolf@uidaho.edu

Follow this and additional works at: http://nsuworks.nova.edu/tqr

Part of the Quantitative, Qualitative, Comparative, and Historical Methodologies Commons, and the Social Statistics Commons

Recommended APA Citation

De Angelis, J., & Wolf, B. (2013). Tasers and Community Controversy: Investigating Training Officer Perceptions of Public Concern Over Conducted Energy Weapons. The Qualitative Report, 18(13), 1-20. Retrieved from http://nsuworks.nova.edu/tqr/vol18/iss13/ 2

This Article is brought to you for free and open access by the The Qualitative Report at NSUWorks. It has been accepted for inclusion in The Qualitative Report by an authorized administrator of NSUWorks. For more information, please contact nsuworks@nova.edu.

Tasers and Community Controversy: Investigating Training Officer Perceptions of Public Concern Over Conducted Energy Weapons

Abstract

Over the last several decades, “Tasers,” “stun guns” and other conducted energy devices (CEDs) have become a widely adopted, though publicly controversial, form of police restraint technology. While there is a growing body of research on the physiological effects of these types of weapons, less attention has been devoted to the social effects of this technology. This paper draws on in - depth interviews with a stratified random sample of police training officers from two states (n=27) to explore the effect that community controversy over the use of CEDs has had on police organizational practices. In particular, we explore how police training officers: (a ) Represent the sources of recent community controversies relating to CEDs; (b ) Characterize the effects that community controversy has on officer practices and policy development.

Keywords

Use-of-Force, Tasers, Controversy, Community, Police Training, In - Depth Interviews

Creative Commons License

This work is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 4.0 License.

Acknowledgements

We would like thank Brandon Long for his outstanding research assistance on this project. This research was

partially funded by the University of Idaho’s, 2009-2010 Faculty Seed Grant Program.

This article is available in The Qualitative Report: http://nsuworks.nova.edu/tqr/vol18/iss13/2

The Qualitative Report 2013 Volume 18, Article 26, 1-20

http://www.nova.edu/ssss/QR/QR18/wolf26.pdf

Tasers and Community Controversy:

Investigating Training Officer Perceptions of Public Concern

Over Conducted Energy Weapons

Joseph De Angelis

Regis University, Denver, Colorado USA

Brian Wolf

University of Idaho, Moscow, Idaho USA

Over the last several decades, “Tasers,” “stun guns” and other conducted energy devices (CEDs) have become a widely adopted, though publicly controversial, form of police restraint technology. While there is a growing body of research on the physiological effects of these types of weapons, less attention has been devoted to the social effects of this technology. This paper draws on in-depth interviews with a stratified random sample of police training officers from two states (n=27) to explore the effect that community controversy over the use of CEDs has had on police organizational practices. In particular, we explore how police training officers: (a) Represent the sources of recent community controversies relating to CEDs; (b) Characterize the effects that community controversy has on officer practices and policy development. Keywords: Use-of-Force, Tasers, Controversy, Community, Police Training, In-Depth Interviews

Perhaps one of the greatest changes over the last decade to police use of force policies and tactics has occurred with the widespread adoption of TasersTM, “stun guns” and other conducted energy devices (CEDs). While adoption estimates vary, the Government Accounting Office (GAO) reported that more than half of law enforcement agencies in the U.S. have deployed some form of CED (GAO, 2005). And even though CEDs have become a widely utilized form of restraint technology, these types of police technology have continued to generate public controversy. For example, international human rights organizations (e.g., Amnesty International), police watchdog groups (e.g., local “Copwatch” groups), and civil rights organizations (e.g., American Civil Liberties Union, 2005) have strongly opposed the adoption of Tasers, and have routinely tied Taser usage to police in-custody deaths (see Amnesty International, 2004, 2008). Moreover, the controversy over these types of devices has been sustained by popular news media accounts of questionable CEDs deployments involving children, college students, pregnant women, protesters, and the mentally ill. As a result, a polarizing and emotional public debate has emerged in relation to the use of CEDs by police agencies. On one side, human rights groups and activists organizations have called into question the legitimacy and safety of weapons (Amnesty International, 2008). One the other side, CEDs manufacturers have strongly argued for the safety and effectiveness of these devices (see Taser International, Inc., 2008).

Strangely, while there has been a growing body of research on deployment patterns and medical affects of conducted energy restraint devices, there has been comparatively little research on the public controversy surrounding CEDs or its effect on police policy or training (Kaminski, 2009; McEwen, 1997; Thomas, Collins, & Lovrich, 2010). As a result, this paper situates police training officers within this debate and examines their perceptions of the controversy. In particular, we examine how police training officers perceive CEDs and how they make sense of the controversy surrounding Tasers. In addition to examining their

2 The Qualitative Report 2013

attitudes on the more controversial aspects of the Taser, we are also interested in how comfortable officers are with the weapon and how it has impacted their jobs as police officers.

Police Less Lethal Force and CEDs

Communities call on police officers to perform a wide array of social functions, ranging from providing front line social service roles to intervening in violent situations. Of all duties associated with policing, however, the capacity to use force is identified as one of the core functions of the police (Bittner, 1980). Indeed, police officers are the most visible instrument of state power a citizen may encounter. For police to perform their role, it is necessary for them to have the power to use force to restrain violent subjects. While departments typically have polices and guidelines on use of force procedures, officers have tremendous discretion as to when, and to what degree, force may be used (Alpert & Dunham, 2004). Yet, despite the need to use force, the problem of excessive force has been a recurring source of public controversy for police departments. For example, high profile uses of force by the police was often a catalyst for urban unrest during the 1960’s (Walker, 2005). The 1991 beating of Rodney King was another high profile incident of police violence that provoked civil unrest and outrage. The fatal shooting of Timothy Thomas, the fifteenth African-American shot by police in five years, sparked riots in Cincinnati in 2001.

Due to controversies over police use of force and the political fallout that results from police shootings, police departments have long been interested in adopting “less lethal” means of subduing resisting suspects (Adams & Jennison, 2007). To solve the dilemma of maintaining the legitimacy of police use of force while ensuring officer safety, departments have long sought a more “humane” and less harmful means of restraining suspects. As part of this trend, conducted energy devices (CEDs) have emerged as a nearly standard piece of police equipment over the last decade. In particular, a subcategory of CEDs called electromuscular disruptors has come to play a crucial role in modern police practice. These types of CEDs work by transmitting a rapidly pulsed high voltage/low amperage that overrides a subjects skeletal muscles, inducing temporary paralysis and significant, but fleeting, physical discomfort. While the technology behind electro-muscular disruption has been around for decades, it is only in the last decade or so that these devices have seen widespread use in policing. The most commonly known CED (the Taser X26) is manufactured by Taser International, Inc. In fact, the Taser International’s products have so thoroughly dominated the less lethal market that the public, media, and police officers generally refer to CEDs as “Tasers” (Wolf, Pressler, & Winton, 2009).

Even though CEDs were intended to function as a more humane means of restraining combative subjects, there has been considerable controversy related to the safety of these kinds of devices. In particular, media accounts of officers using CEDs on disabled, elderly, mentally ill, and other individuals posing no immediate threat have become common stories reported in the media. Moreover, a number of human rights organizations have highlighted more than one hundred cases where an in-custody death has been temporally associated with a Taser deployment (see Amnesty International, 2004). In addition, police watchdog groups have suggested that the introduction of Tasers and other CEDs into police practice has widened the net of force used by police departments or that officers may be using CEDs to administer unconstitutional pre-judicial corporal punishments (see Amnesty International, 2004; Wolf & DeAngelis, 2011). Some have also suggested that officers may be developing an over reliance on CEDs when other options, such as verbal control or hands on techniques, might be more appropriate in certain circumstances (Alpert & Dunham, 2010, p. 253).

A large proportion of the scholarship on CEDs has focused on either its effectiveness at incapacitating aggressive subjects (Government Accountability Office, 2005; Meyer &

Joseph De Angelis and Brian Wolf 3

Greg, 1992; White & Ready, 2007) or their health effects, especially cardiac and the secondary physical injuries to subjects resulting from a sudden loss in motor function (Fish & Geddes, 2001; Ho, Miner, Lakireddy, Bultman, & Heegaard, 2006; Kornblum & Reddy, 1991; Kosgrove, 1985; Levine, Sloan, Chan, Vilke, & Dunford, 2005; McDaniel, Stratbucker, Nerheim, & Brewer, 2000; McDaniel, Stratbucker, & Smith, 2000; Ordog, Wasserberger, Schlater, & Balasubramanium, 1987; Vilke & Chan, 2007;). Overall, while Tasers are not necessaryly medically benign, most of the preliminary medical research on Tasers has generally supported the idea that Tasers are less likely to physically injure healthy subjects than the use of hands-on physical force (hands, feet, or fists), impact weapon, or canines (Alpert & Dunham, 2010; Smith, Kaminski, Rojek, Alpert, & Mathis, 2007; Taylor & Woods, 2010). Moreover, a growing body of research has demonstrated that injuries to officers tend to drop when police departments introduce CEDs (Charlotte-Mecklenburg Police Department, 2006; Jenkinson, Neeson, & Bleetman, 2006).

Research Questions

While the health consequences of Tasers are increasingly understood, we know far less about the organizational or social effects that the spread of this technology has had on police departments and police-community relations. More importantly, while the viewpoints of anti-Taser activists and Taser manufacturers have received widespread coverage in the mainstream media, less attention has been devoted to officer attitudes towards CEDs. As researchers, we became interested in the topic of police use of CED’s because it represented a site of controversy where questions of police use of force and more “humane” forms of use-of-force tactics were debated. From our perspective, this public controversy is a useful cultural site within which we can explore the deeper social and cultural tensions that exist in relation to the use of force by institutions of formal social control. While there is an increasing body of criminal justice and policy-oriented research on the technical and medical aspects of CEDs, we know very little about how the controversy is constructed and internalized by those who use the device -- police officers. We felt a qualitative approach using in-depth interviews was the best way to examine how officers may interpret the meanings and controversy associated with CED’s, and how that may shape grounded policing practices.

As a result of the lack of research on officer perspectives on CEDs and community controversy, we developed the following research questions for our study:

RQ1: How do officers view the effectiveness, safety and accountability mechanisms associated with these types of restraint devices?

RQ2: How do training officers understand and make sense of the public controversy over CEDs?

RQ3: What can be done to mitigate the problems and controversies related to the deployment of CEDs?

To answer these questions, we conducted in-depth telephone interviews with police use-of-force training officers in two states.

Data and Methods

The research methods for this study were reviewed and approved by Ohio University’s Institutional Review Board (IRB) on August 21, 2008 (Ohio University was Joseph De

4 The Qualitative Report 2013

Angelis’ home institution at the time the interviews were completed). After receiving IRB approval, we conducted in-depth phone interviews with a stratified random sample of police use-of-force training officers in two states. We chose to interview training officers because of their familiarity with police technology, and especially less lethal restraint technology. Moreover, use-of-force training officers are intimately familiar with the training practices of their department and are usually well informed about their departments’ policies and procedures.

We interviewed police training officers in two U.S. states, Ohio and Idaho. We chose these states for two reasons. First, since the investigators on this project were working at universities in Ohio and Idaho, we believed that the training officers’ familiarity with the universities would have a positive influence on their willingness to consent to an interview. Second, both of these states are not commonly the site of policing research. Due to the unique characteristics of each state we devised the following sampling strategy: In Ohio, 25 departments were identified using a random sample of municipal police departments from two strata based on municipal population. Using the US Census’ 2004 Population Estimates, we randomly selected 15 departments from cities with 10,000 to 50,000 residents and 10 from cities with more than 50,000 residents. Since there is far less diversity in the state of Idaho in terms of municipal population, the Idaho survey included a random sample of fifteen municipalities with more than 10,000 residents. We did not conduct interviews with officers from departments located in municipalities with less then 10,000 residents because we found during the pilot stage of this project that very small departments tended not to have dedicated use–of-force training officers. As a result, we chose to focus on municipal departments in medium-to-large cities.

Table 1. Interview Response Rate

Response Rate

Departments in Sample 40

Responded 27

Refused 1

Non-response 12

Response Rate 68%

After the departments were selected, we took several steps to solicit interviews. First, we sought to identify the departments training officer(s) from the departments’ websites (this information is available on ~ 25% of the departments websites). If the department did not post this information on their website, we contacted the department through the use of their general line and asked for the name and contact information (phone and email) of their department’s use-of-force training officer. Once we secured the training officer’s contact information, we contacted them by phone and sought to gain informed consent to conduct the interview. At the beginning of the interview, we explained the purpose of the study, the length of time the interview would likely take, and that the interview would be voluntary and confidential. Once the officer gave their informed consent, we conducted the interview. Once the interview was completed, it was transcribed for analysis. Of the 40 police departments contacted, 27 (68%) participated in the interviews (interview times ranged from 35 minutes to two-and-a-half hours). All of the officers that completed interviews worked for departments that had adopted or were in the process of adopting Tasers. Table 1 summarizes the overall interview response rate.

Joseph De Angelis and Brian Wolf 5

As this paper is primarily concerned with how officers describe their experiences with CEDs, how they view the controversy over CEDs, and what officers believe can be done to reduce this controversy, we opted for a research approach that allows for the exploration of in-depth meaning and contextual ideas associated with the deployment of CEDs. This type of research focus requires a qualitative method of analysis that is somewhat unconventional in police research. Recently, some quantitative research has begun to emerge that examines patterns of CED adoption, policy, field use, and the efficacy of these types of devices (see Thomas, Collins, & Lovrich, 2010). While such research is extremely important and has helped enhance our understanding of police less lethal force options, this work aims to fill in some of the rich detail in officer attitudes that can be missed by quantitative surveys. From our perspective, qualitative data analysis enables us to delve much further into how CEDs are understood by officers that use the technology (and in this case, officers that train other officers in how to deploy this type of technology). This type of qualitative research permits us to explore officer attitudes inductively, allowing for themes to emerge based on the rich narrative descriptions of officers (Patton, 2005). From this perspective, the descriptive accounts of police work given by police training officers may provide insight into officer attitudes that may not be anticipated in advance when developing large sample, quantitative surveys. Based on our research questions and the subsequent collection of interview data, we were able to transcribe, code, and analyze the data to investigate recurrent themes and subthemes in officer perspectives on Tasers and community controversies.

After the interview process was completed and each interview transcribed, we conducted an interpretive thematic analysis of the interviews with the assistance of NVivo 9TM software (QSR International, 2010). After receiving the completed transcriptions, we began the analysis by first reading through (“eyeballing”) each of the transcribed interviews multiple times, making relatively short, general notes about recurrent themes (Bernard, 2000). The goal of the initial readings was to develop a sense of the overall context of officers’ comments, as well as identify the broadest possible themes before we began to formally code each individual expression of themes and subthemes into distinct units (Thompson & Barrett, 1997). After we developed a sense of the broad thematic contours of the data, we began to more systematically and selectively code each text segment that matched each individual theme. As we read and re-read the data, we initially marked individual themes. However, as we were able to identify clusters of related thematic ideas, we began to re-code text segments hierarchically into themes and subordinate subthemes. For example, a theme initially coded as the “Public does not understand Tasers” was later coded as either misunderstandings due to “sensational” or “inadequate” media reporting.

While we did not explicitly adopt Grounded Theory as our theoretical approach, we sought to ensure reliability in our coding by utilizing the process of constant comparison (Dye, Schatz, Rosenberg, & Coleman, 2000; Strauss, 1987). Each individually coded text segment was compared to previously coded segments with the same classification across each of the cases. Over time, this led to substantial revision in the organization of the coding scheme. Once we were comfortable that we had identified the major themes and subthemes, we quantified the number of times each theme appeared across our cases. We opted to do this not in an effort to draw inferences to some larger population of training officers, but only to contextualize the relative frequency with which officers in this sample invoked particular themes during their interviews. These frequencies are represented in the charts found in the rest of this paper.

6 The Qualitative Report 2013

Findings

In order to develop some background and build rapport with each officer, we initiated interviews with a series of general questions about the makeup and organization of the officer’s department. With these questions, we collected basic demographic information, whether the department had adopted some form of CED, patterns of use, training and existing policies regarding CEDs (and particularly Tasers). Before asking officers specific questions about CEDs, we queried about the department’s use-of-force policies and procedures. We also asked officers about the presence of a community-policing program and how they would characterize their relationship with the community. The final parts of the interview probed training officers’ understandings and viewpoints of the more controversial aspects of the CEDs. While specific questions were asked verbatim of each training officer, the interview was designed to be free flowing and as conversational as possible. Based on these interviews we were able to identify several common themes among the officers we contacted. While the general departmental information can be useful for understanding organizational context, the rich descriptive data contained in individual officer responses provides useful insight into training officer opinions and perceptions of the controversy, or lack thereof, in their communities.

After obtaining some basic information about the composition of the officer’s department, we asked about the policies and training procedures the departments had implemented. Based on the officers’ response, we ascertained that a little more than half of all uniformed officers in the departments we surveyed were certified to carry the Taser (with wide variations). Table 2 presents the department size, police carrying, and the training used required by departmental policy. Initial training typically lasted between six and eight hours, with an average of 6.8 hours. All but two departments required annual or ongoing recertification averaging 3.1 hours of instruction.

Table 2. Department Characteristics and Taser Training Hours

Mean SD Range

Number of officers (n=27) 170 349 23 - 1780

Percent officers carrying Tasers (n=25) 51.9% 28.5 8%-100%

Training hours (n=25) 6.8 2.6 4 to 8

Recertification hours (n=25) 3.1 1.2 0 to 4

Table 3 presents a summary of Taser adoption patterns and training procedures reported by the training officers we interviewed. The Taser X26 is, by far, the most popular model of CED used with 19 of 25 departments using it exclusively. Several officers (4) mentioned that their department was considering adding other types of CEDs, such as the longer range Taser X12 shotgun round. We also had discussions with officers about some more of the specific aspects of the training procedures. For instance, we asked if officers were shocked as part of the initial training. All but five officers affirmed that officers were shocked by a Taser to allow them to experience the effects of CEDs firsthand, but most said it was a voluntary part of the training. A number of training officers who said it was voluntary also stated that nearly all trainees elect to have a CED used on them. Internal training was the preferred method of certification and most training officers were trained and certified by Taser, Inc., who subsequently trained the department’s officers.

Joseph De Angelis and Brian Wolf 7

Table 3. Patterns of Taser Adoption, Training, Injury Reduction and Safety

N Percentage

Department has Taser?

Yes 25 92%

No 1 4%

Adopting 1 4%

Which model?

X-26 only 19 73%

M-26 only 1 4%

Both 6 23%

Are officers shocked?

Voluntary 13 56%

Yes 5 22%

No 5 22%

How are officers Trained?

Internal 8 42%

Internal (Taser Master Trainer) 8 42%

External 3 16%

Reduction in Officer Injury

Yes 12 52%

Yes, anecdotal 2 9%

No 3 13%

Don’t Know/Unclear 6 26%

Safety Concerns

No 19 76%

Yes, cardiac 2 8%

Yes, accidental discharge 2 8%

Yes, suspect falling 1 4%

Yes, excited delirium 1 4%

Officer Views of the Benefits of CEDs

After obtaining departmental information and probing about the patterns of deployment and training procedures, we asked questions about each training officer’s feelings about the use of CEDs. One central area we inquired about was the officer’s perception of the overall safety of the device. We found that officer safety was a primary reason for favoring the adoption of CEDs (See Table 3). Suspect and citizen safety was also routinely mentioned by training officers. In addition to open-ended responses, we also asked each training officer directly “do you think Tasers are safe?” Overwhelmingly, officers voiced a belief that CEDs were safe devices. Fully three-quarters (76%) of officers had no safety concerns about CEDs and a majority indicated that that the adoption of these devices resulted in an overall pattern of injury reduction. During the interviews, officers would often elaborate on their reasons for believing the device is safe.

I have no concerns at all over the safety of the technology. It’s been proven safe... safe enough out here and all its deployments, and everything that I’ve read and studies they’ve done, the Taser’s never actually caused a death yet. (Officer M)

8 The Qualitative Report 2013

In fact, many officers were quick to refer to the same studies and literatures certifying CEDs as safe and effective device. Other officers have mentioned that they had the CED used on them during training and that was the proof that the CED is safe: “I’ve taken 12 full exposures to the Taser including a 20 second continuous exposure. And, I find that, you know, as soon as it’s over, I have recovered within a second of the effects of it.” (Officer K) Indeed, many officers used their own experience in being shocked by a CED (as part of the standardized training procedure for certification to use the device) as evidence for the safety of this type of technology.

In only a handful of instances (six total), did officers mention or imply that they were not sure if CEDs were safe or not. Even in these few occasions, it was a vague concern of unknown effects or uncertain risks,:

I think there’s still the unknown out there. As far as, you know, I think the human body is such a fascinating thing, that you know, you will never know everything about it, and will know, never know, and everybody’s different (Officer H).

Several officers did mention the subject of in-custody deaths, but they also made the point that “other factors” were responsible for the deaths, including factors such as excited delirium, positional asphyxia, drug use, and other medical problems. Again, officers referred to studies that seem to indicate this. For example, one officer argued that: “People have died after they’ve been Tased, but autopsies have proven...that it wasn’t the Taser – it was because of drugs or something else in their system” (Officer M). Another officer (J) mentioned being concerned about potential adverse effects. Yet, the officer was also quick to point out that deaths were being caused by factors other than CEDs:

You can’t help when you hear those situations where you hear people that are dying from that, but they, they’ve been very consistent and clear that it’s always, there’s other factors involved. You’re dealing with someone with some cocaine, psychosis, positional asphyxia situations...medical problems that are only, that happened to be a factor in that case. And, they’re maintaining and have successfully supported the fact that the Taser was not the result of that. But, you know, common sense, you still can’t get a little, you can’t help but get a little concerned.

We found that despite some occasional, but ambiguous concerns about the safety of CEDs, officers generally felt that theses devices were a safe piece of police restraint technology, especially when adequate training and accountability procedures are in place.

Besides injury reduction, one of the ways that officers represented the safety of CEDs was by emphasizing the device’s deterrent effect, which can be used to reduce the risk of harm to both officers and suspects brought about by hands-on control techniques. In particular, officers talked about the deterrent effects of simply unholstering and aiming a CED. For example, Taser International’s X26 has a laser guide that is used to aim the weapon. The recognition, on both sides of the Taser, of the meaning of the “red dot,” is said to alter suspect behavior, often preventing an escalation of tense situations into something dangerous. The deterrent effect of the “red dot” was mentioned, in detail, by four of the officers that we interviewed, as in the case of this officer’s description:

I can tell you circumstances where people have seen the red dot from the Taser and just automatically given up because of the tool itself is intimidating.

Joseph De Angelis and Brian Wolf 9

Whether they’ve had personal experience being tased before, it’s a deterrent tool. But you know, the biggest benefit has just really been that, it doesn’t, you know, both thesuspect and the officer don’t get into that physical combat, and so it just reduces injuries to both parties uh, involved. (Officer N)

Officers also mentioned the deescalating potential of the weapon. As the prior quote demonstrates, some officers argued that the simple act of unholstering and pointing CEDs can have a deterrent effect, which enables officers to avoid using much more serious forms of force. Another officer further describes the deterrent effect of drawing the CEDst:

I know that the controversy has definitely affected public opinion as far as most, you know, we’ve actually had, you know, like I said, about a dozen deployments. But, we’ve drawn the taser and pointed it at a lot of people. And, as soon as they see it, see the laser dot, they, we’ve had several comments where, ‘Oh, that’s one of those tasers. I give up.’ So, which has actually worked out well for us. (Officer K)

Table 4. Open-ended categorical themes mentioned by police training officers of Taser strengths and limitations (Poisson frequency)

Category Frequency

Main benefit of Taser

Officer and Citizen Injury Reduction 6

Officer Injury Reduction (only) 3

Deterrent Effect 4

Avoid contact or “hands off” 3

“Equalizer” for female officers 2

Effective situations

Resistant/Noncompliant Suspect 13

Suicidal Person 5

Emotionally Disturbed 4

Domestic Situations 4

Alcohol Related 4

Prevent fleeing 2

Animals 1

Situations Not Effective

Thick clothing 8

Large/strong people 3

In crowds/bystander close contact 5

Flammable situations 2

Distance 2

When suspect too inebriated 3

Drawbacks

Overdependence/reliance 5

Extra burden 5

Officer injury (AD) 1

Officer misuse 2

Deskilling 1

No Drawbacks 7

10 The Qualitative Report 2013

In all, officers viewed CEDs as effective use-of-force devices that aids police in a variety of situations. Table 4 tabulates several off the common themes regarding the benefits, efficacy and limitations of CEDs based on our interviews with officers. From the point of view of the police officer, CEDs become a vehicle by which officers can de-escalate potentially volatile encounters without resorting to dangerous types of physical force. Overall, the officers argued that this aspect of CEDS made policing safer for both officers and suspects. Also of note is that two officers (male) felt CEDs were an “equalizer” for female officers against larger or stronger individuals. Besides subduing noncompliant suspects, officers found CEDs to be effective in a variety of situations including when used on emotionally disturbed individuals, suicidal persons, and in domestic situations.

Clearly, safety and injury reduction are represented as a central benefit of CEDs, in addition to the utility of these devices for controlling noncompliant suspects. Yet, while training officers indicated near universal approval of CEDs in terms of their the tactical merits and usefulness, a number of the same officers were also quick to add that CEDs were just “another tool” within their available use-of-force options (mentioned six times), and that this tool is not without limitations. When asked, most officers mentioned situations where CEDs would not be effective and several officers noted drawbacks including the problem of officers becoming over-dependent on these devices, carrying extra items in their belts, and the possibility of misuse. Training officers were also cognizant of the limitations of CEDS and spoke of the reduced efficacy of the device on people with heavy clothing or suspects who are too large to be incapacitated by the device. Some also discussed the safety of deploying the device in flammable situations, near bystanders, or if the suspect is at risk of falling. Seven other officers said that they saw no drawbacks at all about the device. In sum, officers voice widespread support in the use of CEDs as safe and effective weapon.

The Perception of the Controversy

Throughout our interviews with police training officers, we found that police were keenly aware of the contentious public debate over CEDs. During the later half of our interviews with officers, we sought to identify why police training officers thought CEDs were a controversial police technology. To accomplish this, we asked about citizen complaints regarding CEDs and the interaction between the police and the community, as well as how officers would characterize the public’s knowledge of the devices. While some officers were hesitant to talk about the contentious aspects of CEDs, most willingly volunteered their points of view, and others talked about it with only a little prompting.

Table 5 summarizes some our findings regarding police perceptions of community relations and public knowledge of CEDs. Overwhelmingly officers reported having a positive relationship with the community, and nearly two-thirds reported that their department had regular dialogues with the community. Almost half of training officers reported receiving citizen complaints about the Taser (11), and 4 out of 24 departments that addressed this issue said that a formal complaint had been filed by a community member complaining about the inappropriate use of a CED. Despite reporting warm relations with the community, half of all officers who responded said that the public was not well informed about CEDs. From here, we sought to examine how training officers perceived the relationship between public perceptions and controversy over CEDs.

Joseph De Angelis and Brian Wolf 11

Table 5. Prevalence of Citizen Complaints, Police-Community Relations and Officer Perceptions of Citizen Knowledge About Tasers

Response Frequency Percentage

Citizen Complaints About Taser?

Yes 11 46%

No 13 54%

Relationship to community

Very Favorable 10 38%

Generally Favorable 14 54%

Neutral 1 4%

“Mixed” 1 4%

Regular Dialogues With Community

Yes 17 63%

No 10 37%

Public well informed

Yes 3 13%

Somewhat 6 25%

No, but improving 3 13%

No/Not at all 12 50%

Received Media Attention?

Yes 13 59%

No 9 41%

Media Attention Positive or Negative?

Positive 8 57%

Mixed 4 29%

Neutral 1 7%

Negative 1 7%

To elicit officers understanding of this controversy, we asked, “Why do you think the Taser is controversial?” Based on officer responses, we found several central themes relating to the source and reasons for the controversy. These items are tabulated in Table 6 and elaborated on in the next three sub-sections. We also interviewed officers about the perceived misconceptions the public may have in relation to CEDs.

According to the officers we interviewed, the most common misconception held by the public relates to a lack of understanding about CED technology. Twelve officers said that the public thinks that CEDs electrocute suspects. An additional five officers specifically mentioned that the newness of a technology and the perception of change could instigate controversy. As one officer describes the source of the controversy:

Any, any time you use new technology, and this happens when they introduced OC spray, years and years ago, it happened when police officers switched from revolvers to semi-automatic pistols, because the increased number of bullets, any time you change, and then when you throw in the mix of using what is perceived as electricity, then it becomes highly controversial. (Officer J)

From the point of view of this officer and three other officers, the introduction of CEDs is no different than the introduction of prior force technologies in policing. The skepticism among public is thought to be a result of the novel nature of the use-of-force technology. While this was mentioned regularly, many of the officers who elaborated on the

12 The Qualitative Report 2013

problem of controversy felt that the public misunderstood the nature of the technology, and often misunderstand how CEDs have helped to improve the force options available to officers.

You need to explain to ‘em that, you know, people, you know, the rules, you know...the rules are that young men die, and number two – you can’t do anything about it... People wantin’ to fight the policeman and unfortunately, sometimes they die... And then, prior to the, the tasers, we’re doin’ the same thing that Wyatt Earp did – we either beat ‘em or we shot ‘em. We’ve not had that much change within a 150 years ‘til the tasers came about. (Officer E)

After eliciting feelings such as this, we then probed to officers for their viewpoints on CEDs and the roots of the controversy. We found that the officers we interviewed tended to see the controversy over CEDs as stemming from a variety sources. How officers view each of these sources is worth examining in some detail. What follows is a description of the three main themes that we found in how the training officers explained the source of the controversy, which we describe as the misinformed public, the sensational media, and the agitation of activist groups with a political agenda.

Table 6. Open-ended categorical themes mentioned by police training officers for sources and solutions to the Taser controversy. (Poisson frequency)

Category Frequency

Public Misconceptions (open ended)

Thinks Electrocution 12

Thinks Lethal 4

Doesn’t understand technology 5

Media 1

Barbs 1

Reasons For Controversy (open ended)

Media 8

Misinformation (within general public) 9

Interest groups 4

Because it is a new technology 5

Use of force 2

Lethality 2

Corporal punishment 1

How to reduce controversy?

Education 9

Public Dialogue 6

Media 3

Training 3

The Misinformed Public

The first and most commonly mentioned aspect of the CED controversy was the role of the public’s misconceptions about police work and use-of-force. As we probed officer responses about the controversy, we found that many of them were quick to mention that they

Joseph De Angelis and Brian Wolf 13

felt that the public does not understand police work, in general, and are specifically misinformed about CEDs.

[It is] public misinterpretation of, of events associated with [the Taser]... It’s like, you know, officers respond to somethin’, they don’t want to go hands-on, and the risk of injury to the subject or themselves. They use a taser. Boom! You know, that guy dies and, you know, it’s a whole big investigation...I’d say there’s a misconception that a Taser has a high probability of injury or death. I’d say that’s probably about the biggest misconception... there’s a misconception that it’s a cure-all for anything. (Officer A)

The above reflects a common misgiving about how the public perceives police work; the public does not understand the reality of what police do. This officer believes that the public is naive about the way force is used in the field, and therefore misinformed about the effects of the weapon. In addition to mentioning a lack of understanding in the public about police work, another officer also mentioned a lack of awareness of the technical specifications and medical effects of the high voltage/low amperage electrical current used in CEDs:

You know, and they just, they don’t understand that it’s not really, they think that the person is being electrocuted. Um, well, people think that a wall socket is a 110 volts and that kills people. So, so they think 50,000 volts is just electrifin’ ‘em, and torturin’ ‘em, and killin’ ‘em. (Officer B)

Not only did officers mention that they felt that the public was misinformed about the technical aspects of CEDs, they often mistake these devices as the cause of force-related suspect fatalities, when really the death is caused by other factors that are unrelated to CEDs. As Officer C explains: “They think that the Taser kills people. Uh, and they don’t take into account... they have lethal amounts of uh, controlled substances in their, their person, and then they end up dying.”

Even though almost half of officers interviewed mentioned the “problem” of a misinformed public, this was not always the perception among all officers. One Ohio officer said that they thought the public had “gotten better over the years” adding that:

Before we went to them, uh, we approached the media, we approached public safety committees, we approached private groups, we approached the groups that historically are at odds with police, uh, and offered them the opportunity to come in, to ask us questions, to see demonstrations, to look at the technology, to ask, you know, unrestricted, you know, access to our people on it, to ask them about that. And, that’s why I think we’ve had little or no uh, controversy here in [Ohio city] about the use of tasers. (Officer K)

The above officer demonstrates an understanding of community policing principles where police departments can work and interact with the public and community groups in order to potentially ease public concern over police practices. This shows a belief that controversy might be mitigated though greater openness and dialogue with the community. Besides the public, officers discussed several other social actors in greater depth. These actors include the media and activist groups.

14 The Qualitative Report 2013

The Sensational Media

As the previous passage indicates, the training officers routinely mentioned that the media plays a major role in shaping the public’s understanding about CEDs and other aspects of police use of force. While not always the case, police can often have an antagonistic relationship with the media. Conflicts of interest, differing institutional goals, and critical reporting have all been sources of this antagonism. For example, Officer C stated directly: “We don’t have a good relationship with [local newspaper].” Not only may there be strain between the media and police, officers often pointed out that it was the media’s fault that the public was ill-informed about CEDs.

I think a lot of it has to do with terminology. I think there’s a lot of hype by the media. I think in some areas, I think you have pandering to the minority community, especially if the officer who uses the, the Taser, or any kind of force, shooting – whatever, you know, if it’s a, if it’s a white officer and you have a minority criminal, a suspect, that gets played up a lot. And, I think that the newspaper uses the wrong terminology when they say the Taser, you know, delivers, you know, like they’re electrocuting people with this handheld device...Plus, the newspapers want to sell papers, so that old expression – if it “bleeds, it leads.” Like, on the news, or on TV, or in the paper. The big splash when somebody goes belly up after they get tased. (Officer T)

Of course, the very imperatives of police work calls for media attention. The spread of CEDs has only accentuated this coverage. Officers also mentioned the “spectacle” (Officer B) of CEDs, combined with it being a “new technology” (Officers M, K, C) that “always get’s a lot of coverage.” (Officer D). To some officers, the media coverage of CEDs is similar to stories about the introduction of OC spray a generation before. Still, officers mention the subject of deaths, use on vulnerable populations (e.g., pregnant women), and cases of “excited delirium” as something that the media seizes on.

Anytime dealin’ with mental illness, cases of excited delirium, anything, involving in-custody deaths, all of those things have always been uh, hot button topics just because, you know, they lend itself towards scandal or to some... abuse that is always been characterized in the media, historically. (Officer R)

While officers had much to say about the media, one final area worth mentioning relates to the aforementioned perception of a lack of knowledge in the general public about the technical aspects of CEDs, which they see as being reinforced by inadequate news media reporting. Officers often mentioned how the phrase “50,000 volts” (specific number mentioned by eight officers) is cited in media accounts of Taser International’s CEDs. This number, many officers felt, was technically misleading and reflects ignorance in the media about the engineering of CEDs.

Because, the media always puts it in their story that 50,000 volts of electricity. And, and they use words, they use words like shocked, and stun gun, and stuff like that. And, I think the media gives the Taser a bad rap because as soon as the Taser makes the initial contact it automatically decreases from 50,000 volts to 12,000 volts and the pulses per second goes from 19 pulses per second to 17

Joseph De Angelis and Brian Wolf 15

pulses per second. But, the media kind of forgets to put that stuff in. They’re, all you ever hear and see in the media is 50,000 volts. (Officer P)

In all, these depictions of media accounts by police training officers reflect a view that the media sensationalizes stories related to CEDs. These narratives show how many officers think the media “gets it wrong” and contributes to a misinformed general public.

While some officers spoke in negative terms about the capacity of the media to perpetuate misinformation about the Taser, others reported that the media could serve a productive role in educating the public about the positive features of Tasers and other CEDs. Specifically, a number of officers (5) mentioned how the media could be used to “get the word out” and educate the public about CEDs. Opening up to reporters, as well as other communities, could mitigate conflicts with the media and, by extension, the general public at large.

I think they need to keep the media informed. I think, you know, like we did, we had the media come... We had a newspaper guy take the ride and we had a female reporter take the ride. And, the basis of their whole story was, you know, “Hey. Just, just comply. You don’t, you don’t want to do this.” (Officer P)

For these officers, while the media could promote controversy, they could also serve as a mechanism by which those controversies could be defused. By demonstrating how CEDs are used and giving them access to the police perspective on the role they can play in police work, these officers seemed to believe that they could foster a more cooperative, less contentious relationship with the public.

Activist Groups with an Agenda

The final social actor we have found in training officers’ accounts of the controversies over CEDs was the role of activist and social justice groups. In our interviews, activists were depicted by many officers as outsider groups pushing a political agenda. Commonly mentioned groups include Amnesty International and the ACLU. Both of these organizations have issued publications critical of CEDs and have publicly questioned the safety and legitimacy of such weapons in police work. Some officers have mentioned such “liberal” groups have created conflicts of interest within the community and contribute to a degree of misperception and antagonism between the police and community. For example:

And, uh, the president of [Social Justice Group] is [a Local College graduate], alumni. (laughter) And, when Police Department A was trying to adopt Tasers they ran into big problems because there, I mean, there’s a large uh, you know, a large political base in, at Local College students and, obviously, in faculty, and alumni that are faculty, and everything. (Officer E)

Also mentioned were groups that had a political agenda, expressed open hostility toward police, and did not care about the difficulties that officers face in the field.

So, it’s the liberals out there that, that basically want to tie officer’s hands and don’t care about officer injuries, they just care about perception that we’re usin’ it to punish people. (Officer O)

16 The Qualitative Report 2013

Such liberal groups were thought to have other motives and were not depicted positively by police officers who mentioned them.

You know, we’ve got people and organizations out here, like Amnesty International, that it’s not gonna matter what we use, they’re not gonna like it. For whatever reason. Uh, you know, uh, PETA’s involved – only because Taser, in the past, has done testing on pigs. But Amnesty International is probably the biggest opponent to the taser. But, they also oppose the use of pepper spray... (Officer M)

Like the media, training officers said that these groups also play a role in spreading misinformation to the general public. Misunderstandings of the technology along with baiting by activist groups was repeatedly mentioned by one officer:

Misinformation, lack of educating the public, and then just groups like the ACLU that really don’t want any use of force done. That coupled with the fact that a lot of um, coroners and doctors don’t fully understand the technology um, you know, just puts out a lot of misinformation out there. (Officer B)

Amnesty International is mentioned by several officers as being a source of wrong information being seeded in the general public. Another officer’s account represents this sentiment in stating:

Well, the misconception is that electricity is causing death and there have been a number of in-custody death uh, situations that they related after the use of Tasers. Uh, Amnesty International is by far the biggest one that says, “Tasers are causing in-custody deaths.” But, you know, the technology doesn’t support that. And the legal findings don’t support that. Even the medical findings don’t indicate that. (Officer K)

Our data shows that officers regularly mentioned the importance of activist groups in misleading the public and stoking community controversies. Officers reported that these groups are a source of false information who are intent on “tying the hands” of police and limiting the force options available to police.

Discussion and Conclusion

In this study, we were concerned with exploring how police use-of-force training officers viewed the role that CEDs play in policing. Overall, we found that the training officers held a favorable view of CEDs and the role they play in bringing about the compliance of resisting suspects. The training officers noted not just the value of the incapacitating effects of CED’s but also their deterrent effects as well. Echoing the language of CED manufacturers, the officers were quick to discuss the relative safety of the device, along with the advantages of CED’s relative to other force options. In all, the officers in our study expressed a generally cautious, but welcoming, view of CEDs. While these results indicate a positive viewpoint on CED’s, the officers readily acknowledged the dissenting opinions and negative portrayals of the devices. Accordingly, officer perceptions of the community controversy over CEDs was the second area we examined in our interview responses.

Joseph De Angelis and Brian Wolf 17

This research probed police training officers’ understanding and description of the controversial aspects of Tasers and other similar CEDs. While the officers reported good relations with the community and widespread support from citizens, the officers did refer to some tensions with specific groups within the community. In particular, officers reported problems related to public understanding of the Taser, media sources having interests in portraying Tasers in a somewhat sensationalistic way, and activist groups that are hostile toward any use of force by the police. Based on our interviews, the officers argued that the problems and controversies associated with CEDs were not the result, necessarily, of the technical features of the devices themselves, but rather were a part of larger sociological forces that shaped the interaction between the police and the community actors. Overall, the officers argued that the controversy stemmed from several fundamental misunderstandings that the public holds about police work and the medical effects of CEDs. They also believed that these misunderstandings had been fostered and reinforced through the agitation by several community actors, particularly the media and activist groups. In general, they argued that the media and various activist groups stoked fears of CEDs by exploiting misinformation and sensational stories about Taser abuses and injury. Crucially, though, a number of the officers argued that the dynamics of the public controversy over CEDs is similar to the public controversy over other types of less lethal force technologies, such as pepper sprays. Implicit in their comments is the idea that the current controversy results less from technical features/medical effects of CEDs, and more from fundamental and broader tensions and anxieties that exist in the community over how and when the police use force.

Of course, even though the officers spoke in somewhat negative and frustrated terms about the role that the media and activists play in generating controversy and promoting misinformation, they did point to a few areas where lessons can be learned about how departments can respond to the controversy over CEDs. For example, several officers mentioned that their departments had successfully avoided most of the public controversy over CEDs by conducting vigorous public outreach campaigns targeting community groups. From this perspective, local departments may be able to counteract some of the negative features of national-level debate over CEDs by reaching out to local residents and community groups to explain how and when their officers use these devices. By holding community meetings and scheduling demonstrations to showcase the limited medical effects of these devices, these officers implied that controversy can be managed and is not necessarily inevitable at the local level.

Similarly, in addition to outreach to community groups, a number of officers noted that their departments had been successful in shifting local reporting in relation CEDs. While many of the officers we spoke with did mention the somewhat unhelpful role that the media can play in fostering misinformation, a number of the officers argued that the media can play a positive role in promoting a more adequate understanding of how CEDs function in police practice. By conducting outreach with reporters, bringing them on ride-alongs, performing demonstrations, allowing them to sit through trainings (and also allowing them to experience a cycle of the Taser themselves), these officers indicated that local media reporting can become a vehicle by which the public is informed about the benefits of CEDs for officers (and resisting suspects). Moreover, these officers implied that these types of media outreach activities can improve the deterrent effects of CEDs by publicly highlighting the effectiveness (and painful nature) of such of restraint devices. So while officers did seem to be somewhat frustrated with the media, overall, they did seem to indicate that the media, with the right kind of outreach, can play a positive role in informing the public about the benefits of CEDs and the importance of less lethal force options in local police policies and practices.

This research has aimed to explore previously unexamined questions regarding how police officers perceive conflicts in the community over CEDs and other less lethal force

18 The Qualitative Report 2013

options. The use of in-depth interviews enabled us to explore how training officers depict changing use of force technologies and their relationship to the larger community. This project represents an attempt to move discussions of force alternatives beyond the national level debates between anti-taser activists and CED manufacturers and to take seriously the experience of officers who use these devices in the field. Our research however, raises a number of important questions that should be explored in future research. For example, how has the controversy influenced how officers use the CEDs? Clearly the training officers we interviewed were aware of the public controversy over CEDs. Yet, it is still unclear how these types of public controversies filter down to ground-level force policies and practice. Have departments changed how they integrate CEDs into their force policies or training practices based on public concern over CED use? Have line officers begun to use CEDs and other similar less lethal force options differently as a result of the public attention to these types of technologies? Even though there is a small yet growing body of research on how police departments are classifying CEDs within their formal use-of-force policies (see Alpert & Dunham, 2010), we still do not know whether these larger cultural conflicts over less lethal force technologies are influencing how officers understand and actually use force in grounded social encounters with resisting suspects.

This research reports on the qualitative results of in-depth interviews with police training officers. While this is a useful approach for inductively understanding how a small number of officers make sense of the controversy over CEDs, there are important limitations to this type of research. In particular, while we utilized sampling procedures to reduce the impact of selection bias, the small sample sizes limit the generalizeablility of this research. It cannot be assumed that the attitudes of the officers we interviewed reflect the attitudes of police training officers nationally. However, even though there are limitations to this type of work, we believe that qualitative research can fill in gaps left by large sample surveys and help us to understand in more detail how local actors such as police trainings officers make sense of broader controversies over police less lethal force technology.

References

Adams, K., & Jennison, V. (2007). What we do not know about police use of Tasers.

Policing: An International Journal of Police Strategies & Management, 30, 447-465. Alpert, G., & Dunham, R. (2004). Understanding police use of force: Officers, suspects, and

reciprocity. New York, NY: Cambridge University Press.

American Civil Liberties Union. (2005). Stun gun fallacy: How the lack of TASER regulation endangers lives. Retrieved from http://www.aclunc.org/issues/criminal_justice/police_practices/asset_upload_file389_ 5242.pdf

Amnesty International. (2004). Excessive and lethal force? Amnesty International's concerns about deaths and ill-treatment involving police use of Tasers. Retrieved from http://www.amnesty.org/en/library/info/AMR51/139/2004.

Amnesty International. (2008). Amnesty International's concerns about Taser® use: Statement to the U.S. Justice Department inquiry into deaths in custody. Posted January 21, 2008 (http://www.amnestyusa.org/document.php?id=engamr511512007).

Bernard, H. R. (2000). Social research methods: Qualitative and quantitative approaches. Thousand Oaks, CA: Sage Publications.

Bittner, E. (1980). The functions of the police in modern society. Cambridge, MA: Olegeschlager, Gunn & Hain.

Charlotte-Mecklenburg Police Department. (2006). Taser project: First year—full deployment study. Charlotte, N.C.: Charlotte-Mecklenburg Police Department.

Joseph De Angelis and Brian Wolf 19

Dye, J. F., Schatz, I. M., Rosenberg, B. A., & Coleman, S. T. (2000). Constant comparison method: A kaleidoscope of data. The Qualitative Report, 4, Retrieved from http://www.nova.edu/ssss/QR/QR3-4/dye.html.

Fish R., & Geddes, L. (2001). Effects of stun guns and tasers. Lancet, 358, 687-688. Government Accountability Office. (2005). Use of Tasers by select law enforcement agencies GAO-05-464. Retrieved from http://www.gao.gov/new.items/d05464.pdf.

Ho, J., Miner, J., Lakireddy, D., Bultman, L., & Heegaard, W. (2006). Cardiovascular and physiologic effects of conducted electrical weapon discharge in resting adults. Academy of Emergency Medicine, 13, 589-595.

Jenkinson, E., Neeson, C., & Bleetman, A. (2006). The relative risk of police use-of-force options: Evaluating the potential for deployment of electronic weaponry. Journal of Clinical Forensic Medicine, 13, 229-241.

Kaminski, R. (2009). Research on conducted energy devices: Findings, methods, and a possible alternative. Criminology & Public Policy, 8, 903-913.

Komblum, R. N., & Reddy, S. K. (1991). Effects ofthe TASER in fatalities involving poHce confrontation. Journal of Forensic Science, 36, 434-438.

Kosgrove E. (1985). The Taser weapon: A new emergency medicine problem. Annals of Emergency Medicine, 14, 1205-1208.

Levine, S., Sloane, C., Chan, T., Vilke, G., & Dunford, J. (2005). Cardiac monitoring of subjects exposed to the TASER. Academic Emergency Medicine, 12, 113-117.

McDaniel, W. C., Stratbucker, R. A., Nerheim, M., & Brewer, J. E. (2005). Cardiac safety of neuromuscular incapacitating defense devices. Pacing and Clinical Electrophysiology, 28, 284–287.

McDaniel, W. C., Stratbucker, R. A., & Smith, R. (2000). Surface application of Taser stun guns does not cause ventricular fibrillation in canines. Proceedings of the Annual International Conference of the IEEE EMBS, Chicago, IL: Engineering in Medicine & Biology Society.

McEwen, T. (1997). Policies on less-than-lethal force in law enforcement agencies. Policing: An International Journal of Police Strategies & Management, 20, 39-59.

Meyer, A., & Greg, W. (1992). Nonlethal weapons: where do they fit? Part II, Journal of California Law Enforcement, 26, 53-58.

Ordog, G., Wasserberger, J., Schlater, T., & Balasubramanium, S. (1987). Electronic gun (Taser) injuries. Annals of Emergency Medicine, 16, 73-78.

Patton, M. Q. (2005). Qualitative research & evaluation methods (3rd ed.). Thousand Oaks, CA: Sage

QSR International. (2010). NVivo 9: Qualitative Data Analysis Software. Doncaster, Australia: QSR International Pty Ltd.

Smith, M. R., Kaminski, R. J., Rojek, J., Alpert, G. P., & Mathis, J. (2007). The impact of conducted energy devices and other types of force and resistance on officer and suspect injuries. Policing: An International Journal of Police Strategies & Management, 30, 426-446.

Strauss, A. (1987). Qualitative analysis for social scientists. Cambridge: Cambridge University Press.

Thompson, S. M., & Barrett, P. A. (1997). Summary oral reflective analysis: Method for interview data analysis in feminist qualitative research. Advances in Nursing Science, 20, 55-65.

Taser International, Inc. (March 26, 2008). Law Enforcement Overview. Retrieved from http://www.taser.com/Pages/le_overview.aspx.

Taylor, B., & Woods, D. J. (2010). Injuries to officers and suspects in police use-of force cases: A quasi-experimental evaluation. Police Quarterly, 13, 260–289.

20 The Qualitative Report 2013

Thomas, K., Collins, P., & Lovrich, N. (2010). Conducted energy device use in municipal policing: Results of a national survey on policy and effectiveness Assessments. Police Quarterly, 13, 290-315.

Vilke, G. M., & Chan, T. C. (2007). Less lethal technology: Medical issues. Policing: An International Journal of Police Strategies & Management, 30, 341-57.

Walker, S. (2005). The new world of police accountability. Thousand Oaks, CA: Sage.

Wolf, B., & De Angelis, J. (2011). Tasers, accountability, and less lethal force: Keying in on the contentious construction of police electroshock weapons. International Journal of Criminology and Sociological Theory, 4, 657-673.

Wolf, R., Pressler, T., & Winton, M. (2009). Campus Law Enforcement Use-of-Force and Conducted Energy Devices. Criminal Justice Review, 34(1), 29-43.

White, M. D., & Ready, J. (2007). The TASER as a Less Lethal Force Alternative: Findings on Use and Effectiveness in a Large Metropolitan Police Agency. Police Quarterly, 9, 170-191.

Author Note

Joseph De Angelis is Affiliate Faculty in the Department of Criminology at Regis University in Denver, Colorado. His research explores issues relating to legitimacy and criminal justice, police misconduct, and citizen oversight of police.

Brian Wolf is an Assistant Professor of Sociology at the University of Idaho. His research has centered on complex organizations and systems of justice. In addition to his work on police, he has conducted research in environmental criminology and issues of global conflict.

We would like thank Brandon Long for his outstanding research assistance on this project. This research was partially funded by the University of Idaho’s, 2009-2010 Faculty Seed Grant Program.

Correspondence should be forwarded to Brian Wolf, University of Idaho, PO Box 441110 Moscow, Idaho 83844-1110, bwolf@uidaho.edu

Article Citation

De Angelis, J., & Wolf, B. (2013). Tasers and community controversy: Investigating training officer perceptions of public concern over conducted energy weapons. The Qualitative Report, 18(Art. 26), 1-20. Retrieved from http://www.nova.edu/ssss/QR/QR18/wolf26.pdf

MBA Construction and

Real Estate

M Construction and Real Estate Programme S ecification

Summary Programme Details

inal ard

ard: M

itle of final Programme Construction and Real Estate

Credit oints: 180

evel of a ard Q HEQ : 7

Intermediate a ard s

Intermediate a ard 1: Postgraduate Diploma in Construction and Real Estate

Credit oints: 120

evel of a ard Q HEQ : 7

Intermediate a ard 2: Postgraduate Certificate in Built Environment Studies

Credit oints: 60

evel of a ard Q HEQ : 7

Validation

Validating institution: niversity College of Estate Management CEM

aculty Management and Vocational

Date of last validation: ovember 2013

Date of ne t eriodic revie : ovember 2018

Professional accreditation

ccrediting body: Royal Institution of C artered Surveyors RICS

Date of last accreditation: ovember 2015

Date of ne t eriodic revie : Marc 2017

ccrediting body: e C artered Institute of uilding CIO

Date of last accreditation: ovember 2014

Date of ne t eriodic revie : ovember 2019

ccrediting body: C artered ssociation of uilding Engineers C E

Date of last accreditation: ugust 2015

Date of ne t eriodic revie : ugust 2020

ccrediting body: C artered Management Institute CMI

Date of last accreditation: uly 2015

Date of ne t eriodic revie : uly 2018

Miscellaneous

Q benc mar statement Master’s Degrees in usiness and Management Q 2015

Date of commencement of first delivery Se tember 2014

Duration 2 years standard route or 18 mont s accelerated route

M Construction and Real Estate Programme S ecification

Ma imum eriod of registration 9 years

C S Code n/a

Programme Code PM CCRS Standard Route PM CCR

ccelerated Route

Ot er coding as re uired n/a

Programme Overview

Rationale

Our M , accredited by Royal Institution of C artered Surveyors RICS , C artered Institute of uilding CIO , C artered Management Institute CMI and C artered ssociate of

uilding Engineers C E , is a leading s ecialist M delivered by su orted online learning t at focuses on business and management rinci les it in t e international construction and real estate sectors. e rogramme attracts real estate and construction

rofessionals from all continents, offering virtual net or ing and e c anges of e eriences in different international conte ts. Students engage it sub ect e erts in modules covering

ey sub ects t at blend tec nical no ledge in ro erty and construction sub ects it core management disci lines suc as finance, organisational leaders i, strategic management of c ange and mar eting.

e rogramme is suitable for ambitious and e erienced construction and real estate

rofessionals in t e ublic and rivate sectors. is stimulating, c allenging and s ecialist

M rogramme is suitable for current and as iring senior managers o ant to im rove

t eir leaders i s ills t roug strategic management and reac more senior ositions. e

fle ible and life-c anging rogramme increases t e career o tions of its graduates, articularly in international mar ets.

Entry re uirements

Entrants to t is rogramme normally are re uired to ave attained one of t e follo ing:

• a ac elor Degree it onours at lo er second standard 2.2 and t ree years’ experience are required; or

• a evel 5 ualification as defined by rame or for Hig er Education Qualifications for England, ales and ort ern Ireland HEQ lus 5

years relevant e erience, t o of ic s ould be at a ro riate senior management level or

• a rofessional ualification lus 5 years relevant e erience - t o of ic

s ould be at a ro riate senior management level.

• International Students Englis language re uirements:

ll CEM rogrammes are taug t and assessed in Englis . ou ill

t erefore be re uired to demonstrate ade uate roficiency in t e language before being admitted to a course.

o rade or above in Englis anguage or iterature at CSE or

its e uivalent

o rade 6.0 or above, it at least 6.5 in t e reading and riting modules, in t e International Englis anguage esting System

M Construction and Real Estate Programme S ecification

IE S test administered by t e ritis Council in t e Social

Sciences academic module

o 88 or above in t e Internet o tion, 230 or above in t e com uter

based o tion or 570 or above in t e a er-based o tion, of t e

eac ing of Englis as a oreign anguage OE test

o rade C or above in se of Englis at /S evel.

Note: applicants with a Bachelor’s degree that has been taught and

e amined in t e Englis medium can be considered for entry in t e

absence of t e ualifications detailed above.

Students may a ly to enter t e rogramme in eit er semester.

Recognition of rior certificated learning RPC or recognition of rior e eriential learning

RPE routes into t e Programme

CEM olicy and rocedures for Recognition of Prior E eriential earning RPE and

Recognition of Prior Certificated earning RPC are set out in t e CEM Code of Practice:

dmissions and Recognition of Prior earning. is olicy statement ta es recedence in

any suc decision.

RPE may be used for admission onto a level 7 rogrammes in accordance it t e entry

re uirements stated in t e section above. Ho ever, RPE and RPC do not normally enable

transfer of credit into a level 7 rogramme nor enable e em tion from any com onent on

t ese rogrammes.

Programme rogression

Successful com letion of your M may enable you to ta e a P D/ MP il or to conduct researc .

ard Regulations

• e M Construction and Real Estate is conferred u on successful com letion of

180 credits of study.

• Postgraduate Di loma in Construction and Real Estate is conferred u on successful com letion of 120 credits of study.

• Postgraduate Certificate in uilt Environment Studies is conferred u on successful

com letion of 60 credits of study.

ards are conferred in accordance it t e CEM cademic and eneral Regulations for

Students and t e CEM Postgraduate Programme ssessment, Progression and ard

Regulations.

Career ros ects

is rogramme e ui s students it t e essential leaders i and management no ledge,

along it ostgraduate s ills and e ertise to enable t em to develo t eir careers it in

business management, focussing on t e real estate and construction sectors.

e rogramme rovides an academic ualification t at re ares students to rogress onto

members i of t e C artered Institute of uilding CIO , C artered Management Institute

CMI, Royal Institution of C artered Surveyors RICS and C artered ssociation of

uilding Engineers C E .

Students com leting t e M ill also ac ieve a C artered Management Institute CMI

evel 7 ualification in Strategic Management and eaders i sub ect to confirmation by t e

student t at t ey is to be registered for t is a ard. e CMI ualification ill rovide

M Construction and Real Estate Programme S ecification

students it t e o ortunity to ac ieve members i of t e leading rofessional body for managers and en anced access to C artered Manager status.

Programme Aims

Programme aims

e M in Construction and Real Estate is designed to educate individuals as managers. It

develo s t eir ability to reflect on t eir business e erience and solve com le business

issues it in t e conte t of t e Construction and Real Estate sectors. e rogramme is

designed to rovide as iring and current built environment rofessionals it an o ortunity

to develo leaders i and management s ills.

Mar et and internationalisation

is rogramme is aimed at a and broader international audience o ever, it as as its

basis, la and regulatory controls. e rogramme aims to utilise international case

studies to furt er understanding and ere ossible, international construction and

surveying is considered along it international codes and conventions.

Learning Outcomes

Having successfully com leted t e rogramme, you ill ave met t e follo ing learning outcomes.

M Construction and Real Estate Programme S ecification

Curriculum Map

is table indicates ic study units assume res onsibility for delivering and

summatively assessing articular Programme learning outcomes.

Module

Strategic Management of C ange

A X

lied International Mar eting X

A X

A X X

A X

Management inance and Science X

A X X

A X

Managing and eading Peo le X

A X

A X X

A X

Postgraduate Pro ect X

A X

Planning and Develo ment X X

A X

A X X

A X

Pro erty Management X X

A X X

A A XX X X

A X

Real Estate Investment X X

A X X

A X

A X X X

A X

Management of Construction X

A X X

A X

A A X X

X A X

A X

Procurement and endering X X

A X

A X X X

A X

M Construction and Real Estate Programme S ecification

Programme Structure

Module ist

Code Module evel Credits Core

/Elective

M 7SMC Strategic Management of C ange 7 20 Core

M 7M P Managing and eading Peo le 7 20 Core

P 7P D Planning and Develo ment 7 20 Core

M 7 IM lied International Mar eting 7 20 Core

M 7M I Management of inance and Science 7 20 Core

PR 7PR Postgraduate Pro ect 7 40 Core

P 7PRM Pro erty Management 7 20 Elective

I V7REV Real Estate Investment 7 20 Elective

CO 7CMC Management of Construction 7 20 Elective

M Construction and Real Estate Programme S ecification

Planning and Development

is module rovides an introduction to lanning la and t e lanning rocess relevant to ro erty develo ment. e ro erty develo ment rocess, site selection and financial

a raisal of develo ment sites and t eir funding are considered. e module blends t e

basic no ledge of lanning la it t e im lementation of a real estate develo ment

ro ect.

e module enables students to understand:

• t e conte t of lanning it in ro erty develo ment ro ects,

• rinci les of site layout and design, introducing met ods of a raisal of t e financial viability of develo ment sc emes,

• sources of funding for develo ment ro ects.

Postgraduate Project

is module re uires students to one t eir researc s ills ilst roviding t em it a

ve icle to conduct a self-directed researc ro ect t at reflects t e culmination of t eir studies in t e relevant rogramme.

is is a dual-use module serving MSc a ards and t e M . or t ose students studying t e

M rogramme, t e Postgraduate Pro ect s ould demonstrate leaders i t roug strategic management. or t ose students studying t e MSc rogrammes, t e Postgraduate Pro ect re uires students to demonstrate strong conce tual and t eoretical understanding as a lied it in a business conte t.

Strategic Management of Change

e rationale for t is module is to rovide an integrated a roac to cor orate strategy and

t e management of c ange and innovation in a com le and uncertain business environment

in construction and real estate. Part one considers frame or s to manage t e long-term

strategic direction of organisations it in a construction and real estate setting. Part t o

focuses on enhancing students’ understanding of, and response to organisational change.

e determination of a ro riate olicies and strategies to meet sta e older interests is

e lored it in different cultural conte ts.

ELECTIVE MODULES

Management of Construction

is module considers bot t e ersonnel and organisational as ects of construction management, it a focus on t e managers of construction ro ects. It rovides a com re ensive understanding of t e s ills re uired in managing, lanning and controlling t e safe im lementation of a construction ro ect.

e module enables students to understand :

• t e construction management environment,

• t e im lications of ealt and safety for construction,

• t e ersonnel s ills re uired of t e construction manager,

• construction management in ractice.

Procurement and Tendering

Students ill consider t e various ays in ic construction ro ects can be rocured and

t e conse uent effects of rocurement strategies on tendering. e module enables students

to understand:

• t e effects of ro ect ris allocation on t e rocurement rocess and o t e c oice of rocurement met od can im act on t e subse uent ases of t e ro ect cycle,

M Construction and Real Estate Programme S ecification

• rocurement t eories, toget er it e am les of ast and current ractices,

• o construction rocurement is affected by, and can be used to affect, t e ider economy.

Property Management

is module develo s t e rinci les of la and ractice a ro riate to t e effective and efficient management of commercial ro erty. e focus is rinci ally on t e landlord and

tenant relations i it in legal and regulatory frame or s, but also encom asses ro erty eld for o ner occu ation.

is module enables students to understand:

• t e significance of lease agreements from bot t e different landlord/investor and tenant/occu ier ers ectives,

• t e legal, regulatory and mar et constraints it in ic commercial ro erty management is underta en,

• t e andling of t e different interests of arties involved in ro erty management and t e referred solutions after consideration of all im lications.

Real Estate Investment

is module focuses on t e rationale and considerations for investing in ro erty as an asset class and ta es into account t e trade-off bet een t e resent and future use of resources by

organisations, it s ecial reference to t e returns and ris s of real estate investments. e module ill allo students to recognise ro erty as an investment asset it in t e overall

s ectrum of ot er investment media. Investment return, ris , a raisal, erformance measurement, modern ortfolio t eory and res onsible investment conce ts are considered and a lied in t e real estate investment conte t.

Learning, Teaching and Assessment

M Construction and Real Estate Programme S ecification

e resources it in t e Induction Module are available to students t roug out t e duration

of t eir study it CEM.

Student learning su ort:

e rogramme is delivered via t e Institution’s virtual learning environment V E and

academic teac ing and su ort is rovided online, giving students access to CEM tutors

and ot er students orld ide. e cademic Programme eader is t e first oint of contact

for academic ueries.

e cademic Programme eader ill act as t e main oint of contact for students

t roug out t e duration of t eir rogramme. e academic team ill guide and su ort

students’ learning. Ot er CEM teams rovide su ort for assessments, e ams and

tec nical issues including information and communication tec nologies IC . Eac student,

erever t eir location, ill ave access to a ealt of library and online materials to

su ort t eir studies.

Englis language su ort:

or t ose students ose first language is not Englis , or t ose students o is to

develo t eir Englis language s ills, additional su ort is rovided t roug online resources

on t e V E in t e resource ‘Developing Academic Writing’. The resource includes topics such as sentence structure, writing essays and guidance for writing at master’s level aimed at develo ing students’ study s ills.

Personal and rofessional develo ment:

Students are underta ing vocational courses t at are intrinsically lin ed to t e accrediting

rofessional bodies. Students are encouraged and su orted to understand t e need for t e

recognition of t ese bodies and guided as to o to meet t e rofessional members i

re uirements. More generally, CEM as a dedicated careers advisor to ensure students

ave a ro riate access to careers education, information, advice and guidance.

Programme s ecific su ort:

Eac rogramme as a Programme eader, Module eaders and Module utors to su ort

students t roug out t eir time it t e Programme. CEM staff are accessible during

normal or ing ours, during ic t ey also monitor t e 24/7 forums async ronously

and rovide encouragement, assistance and necessary tutor and student feedbac services.

ccess to t e CEM e- ibrary is on a 24/7 basis and CEM as a full -time librarian during

normal or ing ours.

M Construction and Real Estate Programme S ecification

Students are encouraged to develo and a ly t eir no ledge and understanding t roug a range of online activities and e ercises. ese re uire students to a ly researc and analysis it in com le business scenarios.

Subject practical skills

Students are encouraged to s are no ledge and ideas and to ta e t e initiative and demonstrate leaders i. Case studies and a range of online activities re uire students to analyse given information and ma e reasoned decisions.

Key/Transferable skills

e learning activities re uire students to underta e researc , evaluate t eir findings and

develo solutions. eac ing of module to ics requires students’ engagement with a range of

online activities t at develo researc and evaluation s ills and cultivate a systematic

a roac to roblem solving.

ssessment Strategy

Knowledge and understanding

ormative assessment o ortunities and feedbac are rovided t roug out t e rogramme. ese vary in format and may include self-assessment ui es and tutor guided discussion. ll are designed to motivate and su ort t e student.

Summative assessment met ods and formats vary across t e modules and are a ro riate to t e module and its stated learning outcomes.

Intellectual skills

Intellectual s ills are assessed t roug a range of course or artefacts, e aminations and a researc ro ect re ort.

Subject practical skills

range of formative assessment activities are utilised to el develo t e ability to formulate business strategies.

Summative assessment tests t at t e students ave formulated a ro riate business strategies using t e most relevant academic sources.

Key/Transferable skills

Formative assessment is used to develop students’ critical engagement with given scenarios, giving t em feedbac on t eir erformance.

Summative assessment tests t at t e students are able to a raise and a ly t eir researc to given scenarios.

ssessment Diet.

e niversity College of Estate Management su orted online-taug t ostgraduate rogrammes consist of a variety of assessment modes:

• assessed course or in essay, re ort, roblem or s ort uestion format ,

• ritten e amination a ers,

• ro ect or dissertation submissions,

• or -based learning ortfolios and ot er e-mediated submissions.

e e act combinations of assessment vary from rogramme to rogramme and from module to module.

M Construction and Real Estate Programme S ecification

Programme Assessment pattern CATS credits per module

MBA Programme 1 assessment

1 final assessment

2nd course or assignment or

e amination 20

MBA Programme 1 initial assessment

1 final assessment Dissertation 40

Proceedings of the 2016

IEEE International Symposium on

Workload Characterization

IEEE International Symposium on Workload Characterization

Copyright and Reprint Permissions: Abstracting is permitted with credit to the source. Libraries are permitted to photocopy beyond the limit of U.S. copyright law for private use of patrons those articles in this volume that carry a code at the bottom of the first page, provided the per-copy fee indicated in the code is paid through Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923.

***This publication is a representation of what appears in the IEEE Digital Libraries.

IEEE Catalog Number: CFP16236-USB

ISBN 13: 978-1-5090-3895-4

Additional Copies of This Publication Are Available From:

Curran Associates, Inc

57 Morehouse Lane

Red Hook, NY 12571 USA

Phone: (845) 758-0400

Fax: (845) 758-2633

E-mail: curran@proceedings.com

Table of Contents

2016 IEEE International Symposium on Workload Characterization

IISWC 2016

IEEE Copyright Notice ii

Table of Contents iii

Message from the General Chair vi

Message from the Program Co-Chairs vii

IISWC 2016 Organization viii

IISWC 2016 Sponsors and Supporters ix

Keynote Address I

Running on Empty: Getting Work Done on Battery-Free Energy Harvesting Platforms 1

Alanson Sample (Disney Research)

Keynote Address II

The Convergence of Physical/Digital Worlds: Implications on Workloads & Architecture 2

Ravishankar Iyer (Intel Corp.)

Session 1: Best Paper Nominees

TailBench: A Benchmark Suite and Evaluation Methodology for Latency-Critical Applications 3

Harshad Kasture, Daniel Sanchez (Massachusetts Institute of Technology)

Hetero-Mark, A Benchmark Suite for CPU-GPU Collaborative Computing 13

Yifan Sun, Xiang Gong, Amir Kavyan Ziabari, Leiming Yu, Xiangyu Li, Saoni Mukherjee, Carter McCardwell (Northeastern University), Alejandro Villegas (University of Málaga), David Kaeli (Northeastern University)

Measuring and Modeling On-Chip Interconnect Power on Real Hardware 23

Vignesh Adhinarayanan (Virginia Tech), Indrani Paul, Joseph L. Greathouse, Wei Huang (AMD Research), Ashutosh Pattnaik (Penn State University), Wu-chun Feng (Virginia Tech)

Session 2: Workload Characterization

Characterization of Quantum Workloads on SIMD Architectures 34

Robert Risque, Adwait Jog (College of William and Mary)

Characterizing the Workload of a Netflix Streaming Video Server 43

Jim Summers, Tim Brecht (University of Waterloo), Derek Eager (University of Saskatchewan), Alex Gutarin (Netflix)

Characterization and Mitigation of Power Contention across Multiprogrammed Workloads 55

Hiroshi Sasaki (Columbia University), Alper Buyuktosunoglu, Augusto Vega, Pradip Bose (IBM T. J. Watson Research Center)

Session 3: Operating Systems and Virtual Machines

Container Management as Emerging Workload for Operating Systems 65

Tatsushi Inagaki, Yohei Ueda, Moriyoshi Ohara (IBM Research - Tokyo)

iii

Overhead of Deoptimization Checks in the V8 JavaScript Engine 75

Gabriel Southern, Jose Renau (University of California, Santa Cruz)

Workload Characterization for Microservices 85

Takanori Ueda, Takuya Nakaike, Moriyoshi Ohara (IBM Research - Tokyo)

Session 4: Benchmark Formation and Suites

PBench: A Benchmark Suite for Characterizing 3D Printing Prefabrication 95

Fan Yang, Feng Lin, Chen Song, Chi Zhou (University at Buffalo, SUNY), Zhanpeng Jin (Binghamton University, SUNY), Wenyao Xu (University at Buffalo, SUNY)

ANMLZoo: A Benchmark Suite for Exploring Bottlenecks in Automata Processing Engines and

Architectures 105

Jack Wadden, Vinh Dang, Nathan Brunelle, Tommy Tracy II, Deyuan Guo, Elaheh Sadredini, Ke Wang,

Chunkun Bo, Gabriel Robins, Mircea Stan, Kevin Skadron (University of Virginia)

SPEC-AX and PARSEC-AX: Extracting Accelerator Benchmarks from Microprocessor Benchmarks 117

Snehasish Kumar, William N. Sumner, Arrvindh Shriraman (Simon Fraser University)

Session 5: Hardware-Software Codesign

Rebalancing the Core Front-End through HPC Code Analysis 128

Ugljesa Milic (Barcelona Supercomputing Center, Universitat Politècnica de Catalunya), Paul Carpenter (Barcelona Supercomputing Center), Alejandro Rico (ARM Inc.), Alex Ramirez (Nvidia Corp.)

Quantitative Characterization of the Software Layer of a HW/SW Co-Designed Processor 138

José Cano, Rakesh Kumar (University of Edinburgh), Aleksandar Brankovic (Intel), Demos Pavlou, Kyriakos Stavrou (11pets), Enric Gibert (Pharmacelera), Alejandro Martínez (ARM), Antonio González (Universitat Politècnica de Catalunya)

Fathom: Reference Workloads for Modern Deep Learning Methods 148

Robert Adolf, Saketh Rama, Brandon Reagen, Gu-Yeon Wei, David Brooks (Harvard University)

Session 6: GPGPUs and Heterogeneous Computing

ID-Cache: Instruction and Memory Divergence Based Cache Management for GPUs 158

Akhil Arunkumar, Shin-Ying Lee, Carole-Jean Wu (Arizona State University)

Evaluating the Effect of Last-Level Cache Sharing on Integrated GPU-CPU Systems with Heterogeneous

Applications 168

Victor García (Universitat Politècnica de Catalunya / Barcelona Supercomputing Center), Juan Gómez-Luna (Universidad de Córdoba), Thomas Grass (Universitat Politècnica de Catalunya / Barcelona Supercomputing Center), Alejandro Rico (ARM Inc.), Eduard Ayguade (Universitat Politècnica de Catalunya / Barcelona Supercomputing Center), Antonio J. Peña (Barcelona Supercomputing Center)

GPU Concurrency Choices in Graph Analytics 178

Masab Ahmad, Omer Khan (University of Connecticut)

Session 7: Memory and Storage

Memory Controller Design Under Cloud Workloads 188

Mostafa Mahmoud, Andreas Moshovos (University of Toronto)

A Simulation Analysis of Reliability in Primary Storage Deduplication 199

Min Fu (Huazhong University of Science and Technology), Patrick P. C. Lee (The Chinese University of Hong Kong), Dan Feng (Huazhong University of Science and Technology), Zuoning Chen (National Engineering Research Center for Parallel Computer), Yu Xiao (Huazhong University of Science and Technology)

Quantifying the Performance Impact of Large Pages on In-Memory Big-Data Workloads 209

Jinsu Park, Myeonggyun Han, Woongki Baek (UNIST)

Poster Session

Analyzing Power Consumption and Characterizing User Activities on Smartwatches: Summary 219

Emirhan Poyraz, Gokhan Memik (Northwestern University)

Resilience Characterization of a Vision Analytics Application Under Varying Degrees of Approximation ... 221

Radha Venkatagiri (University of Illinois at Urbana-Champaign), Karthik Swaminathan, Chung-Ching Lin

(IBM Research), Liang Wang (University of Virginia), Alper Buyuktosunoglu, Pradip Bose (IBM Research),

Sarita Adve (University of Illinois at Urbana-Champaign)

Identifying Representative Regions of Parallel HPC Applications: a Cross-architectural Evaluation 223

Alexandra Ferrerón (Universidad de Zaragoza), Radhika Jagtap, Roxana Ruşitoru (ARM Ltd., U.K.)

Power-Aware Characterization and Mapping of Workloads on CPU-GPU Processors 225

Kapil Dev, Xin Zhan, Sherief Reda (Brown University)

Treelogy: A Benchmark Suite for Tree Traversal Applications 227

Nikhil Hegde, Jianqiao Liu, Milind Kulkarni (Purdue University)

Characterizing Memory Bottlenecks in GPGPU Workloads 229

Saumay Dublish, Vijay Nagarajan, Nigel Topham (University of Edinburgh)

Author Index 231

www.ijraset.com Vol. 2 Issue IX, September 2014

ISSN: 2321-9653

INTERNATIONAL JOURNAL FOR RESEARCH IN APPLIED SCIENCE AND ENGINEERING TECHNOLOGY (IJRASET)

A Review Paper on Virtual Reality

Gajender Pal1, Kuldeep Kumar Barala2, Manish Kumar3

1,2,3Dronacharya College of Engineering, Gurgaon , Haryana(India)

Department of computer science and engineering

Abstract: This paper provides us a short survey on the topic virtual reality. Virtual reality some time also called virtual environments has drawn much attention in the last few years. Extensive media coverage causes this interest to grow rapidly. High lighting application domains, technological requirements, and currently available solutions. Section1 contain introduction, in section 2, we will discuss VR devices, section 3 contain VR application, conclusion in section 4 and 5 contain reference.

Keywords: virtual reality, virtual reality device virtual reality scope.

I. INTRODUCTION

The real implementation of virtual reality was done in 1989.but it was introduce Sutherland in 1965.there are four technology;

• the visual (and aural and haptic) displays that immerse the user in the virtual world and that block out contradictory sensory impressions from the real world;

• the graphics rendering system that generates, at 20 to 30 frames per second, the ever-changing images;

• the tracking system that continually reports the position and orientation of the user’s head and limbs; and

• the database construction and maintenance system for building and maintaining detailed and realistic models of the virtual world .

Sutherland’s 1965 Vision:

Display as a window into a virtual a real. Computer maintains world model in real time. User directly manipulates virtual objects Manipulated objects move realistically Immersion in virtual world via head-mounted display Virtual world also

sounds real, feels real.

Vehicle simulators were developed much earlier and independently of the VR vision. Although they today provide the best VR experiences available, that excellence did not arise from the development of VR technologies nor does it represent the state of VR in general, because of specialized properties of the application. Entertainment I exclude for two other reasons.

First, in entertainment the VR experience itself is the result sought rather than the insight or fruit resulting from the experience. Second, because entertainment exploits Coleridge’s “willing suspension of disbelief,”3 the fidelity demands are much lower than in other VR applications

2.VR devices:

• HMD

• Tracing devices

• VR glasses

• Data glove

• Cyber puck

Here we discuss these VR devices one by one

HMD:

The most noticeable advances in HMDs have occurred in resolution, although color saturation, brightness, and ergonomics have also improved considerably. In 1994, one had a choice of costly and cumbersome CRT HMDs, which had excellent resolution and color, or economical LCDs, which had coarse resolution and poor saturation. Today economical LCDs have acceptable resolution (640 × 480 tricolor pixels) and good color saturation. HMDs use separate displays mounted in a helmet for each eye. New versions of HMDs, still under development, are based on the creation of the image directly on

Page 404

www.ijraset.com Vol. 2 Issue IX, September 2014

ISSN: 2321-9653

INTERNATIONAL JOURNAL FOR RESEARCH IN APPLIED SCIENCE AND ENGINEERING TECHNOLOGY (IJRASET)

Cave:

Cave are also called VR system. Many major VR installations now use the surround-projection technology first introduced in the University of Illinois-Chicago Circle CAVE. From three to six faces of a rectangular solid are fitted with rear-projection screens, each driven by one of a set of coordinated image-generation systems

A cave image

III. APPLICATION OF VR

• In entertainment

• In medicine(like surgery)

• In arts

• In labs (VR labs)

• In aircraft training

IV. CONCLUSION

Virtual environment technology has been developing over a long period, and offering presence simulation to users as an interface metaphor to a synthesized world has become the research agenda for a growing community of researchers and industries. Considerable achievements have been obtained in the last few years, and we can finally say that virtual reality is here, and is here to stay. As the technology of VR will increases the application of VR become unlimited. From the hardware point of view, while full fidelity of sensory cues is still not achievable even with the most advanced and expensive devices, there exists

[1]. M. Akamatsu et al.: Multimodal Mouse: A Mouse-Type Device with Tactile and Force Display. Presence, Vol. 3, No. 1, pp. 73-80 (1994)

[2]. R. L. Anderson: A Real Experiment in Virtual Environments: A Virtual Batting Cage. Presence, Vol. 2, No. 1, pp. 16-33 (1993)

[3] Ascension: Ascension trackers technical data. http://www.world.std.com/~ascen (1995)

[4]Ascension: Ascension trackers technical data. ftp://ftp.std.com/ftp/vendors/Ascension/tecpaper.ps (1995)

[5] P. Astheimer: Acoustic Simulation for Visualization and Virtual Reality. EUROGRAPHICS’95 State Of The Art Reports, pp. 1 - 23 (1995)

[6]Atlantis: Atlantis VR Systems. http://vr

atlantis.com/vr_systems_guide/vr_systems_list2.html (1995) [7]BALAGUER, J.-F., AND GOBBETTI, E. i3D: A high speed 3D web browser. In VRML: Bringing Virtual Reality to the Interet, J. R. Vacca, Ed. AP Professional, Boston, MA, USA, 1996.

[8] BALAGUER, J.-F., AND MANGILI, A. Virtual environments. In New Trends in Animation and Visualization., D. Thalmann and N. Magnenat-Thalmann, Eds. Wiley, New York, NY, USA, 1992.

[9] BALET, O., LUGA, H., DUTHEN, Y., AND CAUBET, R. PROVIS: A platform for virtual prototyping and maintenance tests. In Proceedings IEEE Computer Animation (1997).

[10] BAYARRI, S., FERNANDEZ, M., AND PEREZ, M. Virtual Reality for

driving simulation. Communications of the ACM 39, 5 (May 1996),72–76.

[11] BIER, E. A., STONE, M. C., PIER, K., BUXTON, W., AND DEROSE, T. Toolglass and Magic Lenses: The see-through interface. In Computer Graphics (SIGGRAPH ’93 Proceedings) (Aug.1993), J. T. Kajiya, Ed., vol. 27, pp. 73–80. [12]R.L. Holloway, Registration Errors in Augmented Reality Systems,PhD dissertation, Department of Computer Science, University of North Carolina at Chapel Hill, 1995.

[13] M.R. Falvo et al., “Bending and Buckling of Carbon Nanotubes under Large Strain,” Nature, Vol. 389, Oct. 1997,pp. 582-584

Page 406

Performance and Efficiency:

Recent Advances in Supervised Learning

SHENG MA AND CHUANYI JI

This paper reviews recent advances in supervised learning with a focus on two most important issues: performance and efficiency. Performance addresses the generalization capability of a learning machine on randomly chosen samples that are not included in a training set. Efficiency deals with the complexity of a learn¬ing machine in both space and time. As these two issues are general to various learning machines and learning approaches, we focus on a special type of adaptive learning systems with a neural architecture. We discuss four types of learning approaches: training an individual model; combinations ofseveral well-trained models; combinations of many weak models; and evolutionary computation of models. We explore advantages and weaknesses of each approach and their interrelations, and we pose open questions for possible future research.

Keywords—Evolutionary computation, hybrid models, neural networks, supervised learning.

I. INTRODUCTION

In many important application areas such as signal pro¬cessing, pattern recognition, control, and communication, nonlinear adaptive systems are needed to approximate un¬derlying nonlinear mappings through learning from ex¬amples. In order for approximations to be sufficiently accurate, a good performance is required for nonlinear adaptive systems. Meanwhile, many applications, especially those in emerging areas of wireless communication and networking [24], [38], [79], [95], require the learning to be done in real time in order to adapt to a rapidly changing stochastic environment. Other applications such as data mining and searching the Web need to deal with very large data sets [37], [66], and thus the learning time must scale nicely with respect to the size of data sets. Since the size of learning machines determines the memory required for implementation, a learning machine with a compact structure is preferred. Therefore, a challenging problem is how to develop adaptive learning systems with a compact

Manuscript received October 30, 1998; revised April 2, 1999. This work was supported by the National Science Foundation under ECS-9312594 and (CAREER) IRI-95025 18.

S. Ma is with IBM T. J. Watson Research Center, Hawthorne, NY 10532 USA (e-mail: shengma@us.ibm.com).

C. Ji is with the Department of Electrical, Computer, and Systems Engineering, Rensselaer Polytechnic Institute, Troy, NY 12180 USA (e-mail: chuanyi@ecse.rpi.edu).

Publisher Item Identifier S 0018-9219(99)06910-8.

structure that can achieve good performance and be adapted in real time. The goal of this paper is to address the important issues of performance and real-time learning for nonlinear adaptive learning machines by reviewing recent work in the interdisciplinary areas of adaptive learning systems, statistics, and information theory.

There are two aspects when these issues are investigated: architecture of adaptive learning machines and learning scenarios (approaches). The learning scenario considered is the supervised learning for nonparametric nonlinear regres¬sion including classification as a special case. We discuss several general frameworks, such as the expectation-maximization (EM) framework, the combination scheme, weak learning, and evolutionary algorithms, all of which aim at improving the efficiency and performance of a learning machine. In order to be comprehensive, a neural network is used as a sample architecture to show how these general frameworks are applied. The reason why neural networks are chosen is that they have been shown to be universal approximators to a general class of nonlinear functions [9] and have become popular recently. The general framework can be applied to other learning machines. This, however, is not the focus here.

In supervised learning, performance addresses the prob¬lem of how to develop a learning machine to achieve optimal performance on samples that are not included in a training set. Efficiency deals with the complexity of a learning machine in both space and training time. Specifically, the space complexity of a neural network refers to its size, and the time complexity characterizes the computational time needed to develop such a neural network. These three issues are interrelated.

The performance of a supervised learning system is characterized by its generalization error, which measures the distance between the output function of a trained model and an underlying target function. Most existing methods for training neural networks in supervised learning suffer from an intrinsic problem in pattern recognition: the bias and variance dilemma [39], [47]. That is, if a neural network is too large,1 it may overfit a particular training set and

1For instance, feedforward neural networks with too many neurons.

0018–9219/99$10.00  1999 IEEE

PROCEEDINGS OF THE IEEE, VOL. 87, NO. 9, SEPTEMBER 1999 1519

thereby fail to maintain good generalization error. A small neural network, however, may be sufficient to approximate an optimal solution. In addition, one important algorith¬mic problem is how to deal with a complex optimization problem with possibly many local minima.

The size of a learning machine can be characterized by the space complexity, which is related to the number of free parameters (for instance, the number of weights of a neural network). A learning machine is considered to be efficient in space if its space-complexity scales as a polynomial function in terms of the dimension of feature vectors.2 It has been found that compact adaptive learning systems like multilayer feedforward neural networks are efficient approximators of a wide class of smooth functions. They possess a polynomial space complexity [10]. Other learning machines, which consist of localized models such as nearest neighbor classifiers [27], Parzen windows [35], and linear combinations of localized basis functions [10], [80], may suffer from the so-called “curse of dimensionality.” That is, their space-complexity scales exponentially with the dimension of the feature vectors [10], [65], [85].

Learning time can be characterized by the time complex¬ity as a scaling property with respect to the dimension of feature vectors. An adaptive learning system is considered efficient in time if the time complexity is polynomial. Training (nonlinear) feedforward neural networks and clas¬sifiers is slow, and in general, NP-complete [18], [59]. But training the aforementioned localized learning machines can be done quickly. For example, nearest-neighbor classifiers simply remember all training samples and therefore do not require training. This presents a dilemma that exists be¬tween performance, space complexity, and time complexity. Therefore, an important issue is whether or not it is possible and how to develop a neural network that can learn in real time while achieving a desired performance with a polynomial space complexity.

In this work, we address these three fundamental issues: performance; efficiency; and space-complexity. We discuss how they motivate the search for new solutions for adaptive learning machines with a neural architecture. To achieve our goal within a limited scope, we focus on reviewing recent work in three areas: training an individual neural network; combinations of well-trained models; and combinations of weak models. We also discuss hybrid methods, such as evolutionary computation in the emerging area of soft computing, and their role in tackling these fundamental issues.

This paper is organized as follows. Section II gives the background knowledge on performance and efficiency. Section III provides an overview to the approaches to be discussed in this paper. Section IV reviews neural networks that utilize a fixed structure to tackle the issues of perfor¬mance and efficiency. As finding an appropriate structure is very difficult, approaches have been developed to combine models. Section V reviews the research on combining well-trained models to improve the performance. Section VI

2Here the dimension of feature vectors is used as a measure of the complexity of a problem.

then discusses combinations of weak classifiers and how combined weak classifiers make a tradeoff among perfor¬mance, time complexity and space complexity. Section VII introduces hybrid methods, such as evolutionary computa¬tion, and their role in tackling issues of performance and efficiency. We conclude the paper in Section VIII. As there is a large volume of related work, our selections of material may be subjective, and we apologize for any significant omission.

II. BACKGROUND: PERFORMANCE AND EFFICIENCY

A. Notation

In a supervised learning environment, let

denote training samples pairs, where

is a -dimensional feature vector of the th

sample and is the corresponding target. For simplicity,

is assumed to be a scalar in this paper. Furthermore, we

assume that there is a function so that .3

For the function approximation problem, is a real

number. For classification, indicates the class to which

the th sample belongs.

Neural networks are a class of nonlinear models that consist of interconnected nonlinear processing nodes. As an example, a two-layer feedforward network with one linear

output unit, input units, and sigmoidal hidden units

is shown in Fig. 1. Let be a weight matrix at the

first layer and its th column

(1)

where is the sigmoidal transfer function. The output

of the network, , corresponding to an input can be

expressed as

(2)

B. Performance and Efficiency

1) Performance: Let be a model with a set

of parameters and trained on training set . The per

formance of can be measured in terms of the

difference between a function to be approximated and

its approximation through the squared norm

(3)

where is the probability density function of . If

training samples are drawn randomly, the expected squared

3 For simplicity, this paper does not address possible noise terms.

1520 PROCEEDINGS OF THE IEEE, VOL. 87, NO. 9, SEPTEMBER 1999

Fig. 1. The structure of a two-layer feedforward network.

norm is

(4)

2) Efficiency: The space complexity of is

the number of free parameters of when the

generalization error is no bigger than . For example,

if a two-layer feedforward neural network is considered, can be either the number of hidden units or the total number of independent weights of the network. The time-complexity of an adaptive learning system is defined as the expected training time needed to obtain a learning system when the generalization error is bounded by .

The scaling property of the time complexity and space complexity defines the efficiency of an adaptive learning machine (or algorithm) whose generalization error is no

bigger than a given quantity . is said to

be efficient in space if scales polynomially in terms of the dimension of feature vectors, , and other related parameters. If the time complexity scales polynomially in term of these quantities, the learning machine is said to be

efficient in time. If the efficiency can be achieved both in space and time, the learning machine is said to be efficient.

3) Relationships Between Performance and Efficiency: Performance and efficiency are interrelated. This can be understood through an example of a two-layer feedforward neural network.

Suppose a training algorithm exists that can obtain the weights of each hidden unit in polynomial time. The algorithm is not efficient if the number of hidden units needed is exponential in terms of the dimension of feature vectors. This is because the total training time can be estimated as the time needed to train a single hidden unit multiplied by the number of hidden units in the networks. In other words, a polynomial space complexity is needed to achieve the polynomial time complexity for training the entire network for this example. Together, space efficiency and time efficiency can define the efficiency of a learning algorithm for two-layer feedforward neural networks. This idea will be further explained in Section VI in the context of combinations of weak classifiers.

Efficiency has to be defined together with performance, i.e., the generalization error of a resulting two-layer network should be smaller than a desired quantity4 when the space complexity and time complexity are examined. This is because the concept of efficiency would be meaningless without a requirement on generalization performance. For example, a binary classifier with a generalization error of can certainly be obtained using an efficient (yet trivial) algorithm that only performs random guessing.

C. Performance Issues

There are two important factors affecting the perfor-mance: the bias and variance dilemma, and possibly a complex objective function with many local minima.

The first issue is intrinsic and is independent of algo-rithms used. The second issue is algorithm and problem dependent. In the following, we further discuss the first issue and its relationship with the second issue.

1) Bias and Variance Dilemma: The generalization error is affected by two factors: bias and variance. To define

the bias and variance, let be the best model

in the model space, that is,

. Therefore, the solution does not de

pend on the training data. Furthermore, the bias Bias

and variance Var for a model can be defined as

Bias (5)

where the expectation is on a randomly drawn , and

Var (6)

The inner expectation is on and the outer one is on ran-domly drawn training sets of the same size. Therefore, the generalization error can be decomposed into, and bounded by, a sum of the bias and variance [10], [22], [39]

Bias Var (7)

4 This is no smaller than the Bayes error of a given problem [35]. One could assume the Bayes error is small.

MA AND JI: PERFORMANCE AND EFFICIENCY 1521

The first term of the right-hand side is the error due to the bias that can be caused by an inappropriate choice of the size of a class of models when the number of training samples is assumed to be infinite. The latter is the error corresponding to the variance and is caused by the finite number of training samples.5

In a special case when a class of models was chosen to be two-layer feedforward neural networks with sigmoid hid¬den units,6 the bias and variance were bounded explicitly by

Barron [10]. In particular, and

are upper bounds for the bias and variance respectively,

where represents a quantity in the order of . The fact

that decreases with the number of hidden units sug

gests that a large neural network with many free parameters is less biased. Since it has been shown by Barron [10] that feedforward neural networks are capable of approximating a wide class of smooth functions when the number of hidden units is sufficiently large, a larger set of neural networks (with a larger ) certainly contains a better approximation

of than a smaller set of neural networks (with being

smaller). That is why the bias is smaller when is larger. However, when the number of training samples is finite, a network with an excessively large space complexity will overfit the training set. This can be understood from the concept of information capacity of neural networks [1], [13], [28], [52], [74] in information theory. The information capacity is the maximum number of training samples that can be memorized by a neural network given a certain space complexity. When the space complexity exceeds the information capacity, learning machines are so large that they only memorize training samples. Since there are many such neural networks with the same space complexity but different choices of weights which can memorize training samples, their (generalization) performance varies and thus leads to a large variance. That is, the average performance can decrease as gets larger. This can be observed from

the term , which increases with respect

to . Therefore, a tradeoff needs to be made between bias

and variance.

2) Bias Variance and Local Minima: An issue that needs to be clarified is whether the bias and variance are related to local minima, or the effectiveness of an algorithm at finding a good solution. In our view, the bias by definition only depends on the structure of a class of neural networks but not the choice of values of weights and is therefore independent of any training algorithm. To understand the variance as defined, we can consider the sample variance.

For any given (randomly drawn) training set of size

, a network corresponding to can be found,

which contributes to one sample in the sample variance. The true variance can be estimated through a sufficiently large number of such samples. In the definition given by Barron

[10], it was assumed that each corresponded to

the global minimum of the error function, i.e., it was the best network that can be obtained given information from a

5 This is equivalent to the space complexity versus the number of training samples.

6 The number of free parameters is thus approximately L(d + 1).

training set. To define the variance depending on a specific algorithm and applications, the theoretical result may lose its generality. This definition (given by Barron) simply suggests that the bias and variance dilemma is intrinsic for nonparametric regression (even if one had the most powerful algorithm to obtain the globally optimal network). Therefore, such a definition on the variance can also be considered to be algorithm independent.

In practice, for a given training set and the chosen error function, different algorithms result in different networks

corresponding to different ’s. If an algorithm (say,

an evolutionary algorithm to be introduced later) is always better at finding a globally optimal solution, and another algorithm (say, a gradient-descent algorithm) gets stuck at a local minimum more often, then the former would have a smaller sample variance than that of the latter.

III. OVERVIEW OF APPROACHES

Before we go into details, a quick overview on what we will soon discuss may elucidate the logical organization of this paper.

Focusing on performance and efficiency, we start our review from an individual model. For training an individual model, performance can be improved by balancing bias and variance through either finding a model with optimal size or adding a regulation term in an objective function to confine the complexity of a model. Efficiency in finding a model can be achieved through fast training algorithms based on the “divide-and-conquer” principle.

However, finding an optimal size is usually difficult if not impossible. To address the performance issue, a combination of well-trained models has been proposed. The essential idea is to pick a slightly oversized model as a base for combination. Since the base model is oversized, it has a low bias but a high variance. The algorithm, therefore, relies on combinations to diminish the overall variance, and thus the generalization error.

Although the combination of well-trained models relaxes the need to optimize the size of a model, its learning time increases with the number of models to be combined. The combination of weak models aims at improving the efficiency of the combination of well-trained models while preserving the nice performance property. In particular, this approach uses the similar combination scheme but it chooses a weak model as a base model. A weak model is a model whose performance is only slightly better than random guessing and thus has a large bias but a small variance. From the performance perspective, since different weak models are forced to learn different parts of a prob¬lem, combinations of many such weak models can diminish the overall bias and thus achieve good generalization. From the time-efficiency point of view, since the base model performs only slightly better than random guessing, it can be obtained efficiently. Furthermore, this approach uses an incremental learning procedure. The whole training process can be very efficient.

Since there is still a large body of related work in these three areas, we further narrow our scope in each area.

1522 PROCEEDINGS OF THE IEEE, VOL. 87, NO. 9, SEPTEMBER 1999

Specifically, for training an individual model, we focus on the issue of the time efficiency and illustrate how the “divide-and-conquer” principle is applied to the reviewed algorithms. For combinations of well-trained models, since there are papers with an extensive review on this subject [33], [77], we focus on one class of algorithms that hope¬fully serves as a bridge to the next topic on combinations of weak classifiers. More thorough review is given on combinations of weak models both on algorithms used and theoretical results.

As will soon be discussed, randomized algorithms play a key role in combinations of both well-trained and weak models. They are also one of a few approaches that can hopefully result in polynomial-time algorithms and achieve global optimization. This leads us to further discussing hybrid methods such as evolutionary computation in the emerging area of soft computing and illustrating the rela¬tionships with combinations of weak classifiers.

IV. PERFORMANCE AND EFFICIENCY:

TRAINING AN INDIVIDUAL MODEL

A. Performance

In order to make a tradeoff between bias and variance when a single network is used, we need to find the optimal architecture of a neural network. This is equivalent to determining an optimal space complexity, for example, the optimal number of hidden units and layers for multilayer neural networks. Tremendous efforts have been made on estimating and finding the optimal architecture of a learning system using a finite number of training samples. These approaches include computational learning theory [13], [39], [49], [53]–[55], [96]–[98], various statistical methods including cross validation and model selection [9]–[11], [81], [83], [100], [104], and stochastic exploration of an optimal structure based on evolutionary algorithms [73], [105]. Due to inherent difficulties in accurately estimating the (space) complexity of a nonlinear adaptive system like a neural network with limited training samples, finding the optimal architecture for a learning machine remains a very difficult task.

Another practical difficulty of training a neural network is that the error function (the objective function) to be minimized is usually very complex with many local minima and possibly flat areas [40]. Therefore, a deterministic gradient-based approach may be trapped by a local mini¬mum rather than a global minimum. To deal with this issue, stochastic algorithms have been developed. Since stochastic algorithms are also used in our next two subjects as a general framework for randomization, we will leave the further discussion on these algorithms to Section VII.

B. Space Complexity

Although an exact optimal architecture is difficult to obtain, nonlinear adaptive systems such as multilayer feed-forward neural networks have been shown to be efficient in space. In particular, Barron [10] showed that the space com¬

plexity of two-layer feedforward networks with sigmoidal hidden units is polynomial when used to approximate a wide class of smooth functions. The space complexity of linear combinations of localized basis functions, however, is exponential when used to approximate the same class of functions. Although feedforward two-layer neural networks can also be regarded as linear combinations of (sigmoidal) basis functions, they are efficient in space because of two major reasons: 1) sigmoidal functions are nonlocal and 2) sigmoidal functions at the hidden layer when regarded as basis functions are parameterized by their weights. Those weights effectively adapt the sigmoidal (basis) functions to the feature space and thus result in a good scaling property on the space complexity with respect to the dimension of feature vectors.

C. Fast Training Algorithms

Finding a nonlinear adaptive system like a neural network with an appropriate structure is a nonlinear optimization problem, which can be complex and slow. One of the most widely used algorithms for training neural networks is backpropagation (BP) based on gradient descent [90], [103]. Although this simple approach has been successfully used to train neural networks for a wide range of applications, the algorithm can be notoriously slow when used to train a large network in solving a complex problem. This is because neural networks with a complicated structure result in complex error functions that are difficult to minimize.

In fact, it has been shown that training a multilayer neural network of a fixed structure is NP complete [18], [59]. It is not clear whether an approximation of an op-timal solution can be found in polynomial training time either. Therefore, to tackle the issue of time complexity, efforts have been focused mostly on developing “fast” training algorithms that can improve upon the training time compared with, for instance, that of BP. There are two categories of fast training algorithms. The first one is motivated by the “divide-and-conquer” principle and tries to decompose the complicated problem into a set of relatively easy subproblems. The second class of algorithms is motivated by stochastic algorithms and tries to “escape” from local minima through randomized approaches. We discuss algorithms in the first category next and the second category in Section VII.

Three major types of fast training algorithms developed in the first category are: 1) decision-tree algorithms; 2) incremental learning algorithms for training multilayer neu¬ral networks; and 3) EM algorithms. Although they are in different forms, they were all developed based on the “divide-and-conquer” principle.

1) Decision Tree: A decision tree is a directed, acyclic graph in which each node is either a decision node with two or more successors or a leaf node. Every leaf is labeled as a class, while a decision node has a model for deciding to which successor a sample (feature vector) belongs. Usually, a model at a decision node can be as simple as a linear model (or a so-called perceptron) or just a threshold on an

MA AND JI: PERFORMANCE AND EFFICIENCY 1523

attribute. A decision tree can be considered as a special case of multilayer feedforward neural networks [4].

The decision tree algorithm can be considered as a typical example of “divide-and-conquer.” It is a recursive algorithm (see [19] and [86] for more elaborate review). The algorithm starts with a root node which utilizes all training data. First, a best model is computed according to a criterion for best separating the training data into two sets. Then the tree grows to the next level with two children so that the left child only takes care of one part of the training data, and the right child looks at the rest. The left and the right children have their own children in a similar fashion, and the tree grows until a stopping criterion is satisfied.

The decision tree algorithm is very efficient for two reasons. First, each node usually uses a simple decision model, (e.g., a perceptron or a threshold on one attribute). Secondly and more importantly, a complicated problem is divided by a decision node so that its successors only focus on parts of the problem instead of the entire problem.

2) Incremental Learning: Incremental learning trains one base model at a time using the residual error from the pre¬viously trained models. Once a base model is trained, it is fixed and combined with all previously trained models. The new residual error is calculated for training the next base model until a certain stopping criterion is met. Through this algorithm, training a large complicated model is reduced to training an individual base model sequentially. As an example, to train a two-layer feedforward neural network [36], a neuron is treated as a base model. Since incremental algorithms reduce the task of training an entire neural network into training individual neurons sequentially, they have been shown to be one or two orders of magnitude faster than gradient-descent-based algorithms [23], [36], [46].

3) EM Algorithms: As the previous two algorithms were developed based on heuristics, fast training algorithms based on EM algorithms were formulated through a better established theoretical framework [58], [69]–[71]. The EM algorithm was originated from statistics [31] for finding the maximum likelihood. It was introduced into the learning community in the past few years to speed up training. Specifically, EM algorithms were derived for training the stochastic Boltzmann machines [2], [25], the mixture of experts [16], [58], two-layer feedforward neural networks [71], and fully connected recurrent neural networks [69], [70].

In the following, we first review the general EM frame-work [31]. We discuss how to apply the EM framework in a supervised learning environment [58]. We then provide an EM algorithm for training a two-layer neural network [71]. a) EM framework: The EM algorithm can be viewed as a statistical framework for maximum likelihood estima

tion [31]. Let be a set of data observed directly, and let

be a set of parameters characterizing the corresponding distribution of the random variables. The maximum likeli¬hood problem is to find an optimal set of parameters

through maximizing the log likelihood of the data. That is

(8)

. The EM algorithm provides a framework to utilize such missing variables and simplify a complex optimization problem into several much simpler problems. The way to do so is through iterating between a so-called expectation(E)-step, and a maximization(M)-step. The E-step estimates the missing variables, and the M-step uses the estimated variables to find a locally optimal solution.

Specifically, in the E-step, the following conditional expectation can be computed (with respect to hidden vari¬ables) of the complete data ( ) log likelihood given

incomplete data and the previous parameters at

the th step:

(9)

. is called complete data

including both missing data and incomplete data. The no

tation indicates that the conditional expectation

is a function of and the expectation is evaluated using as parameters in the probability density function.

represents a probability density function. stands

for the expectation operation. In the M step, new param

eters are found by maximizing the expected log

likelihood . The algorithm alternates between

the E and M step and terminates when a certain con-vergence criterion is satisfied. Local convergence of the EM algorithm is guaranteed. That is, after every (E-M) iteration, the original log likelihood does not decrease, i.e.,

. As a result, the optimal

set of parameters can be obtained. In practice, this

implies convergence to a local maximum of the original

log likelihood .

b) EMfor supervised learning: Jordan et al. [58] in-troduced the EM framework into the supervised learning environment. Specifically, the observed data, i.e., incom-plete data, can be defined as

Moreover, if a set of hidden targets denoted as

can be chosen, the complete data corresponds to

. For notational simplicity, we will use

to represent . Therefore, the Q function (9) can

be rewritten as the following, when training samples are assumed independently drawn from the input domain:

1524 PROCEEDINGS OF THE IEEE, VOL. 87, NO. 9, SEPTEMBER 1999

Therefore, to apply the EM algorithm, we need to define a

proper set of hidden targets and the probability models

of and .

c) EMfor neural networks: To show a relatively simple example, we apply the EM framework to training a two-layer feedforward neural network. Our discussion is based mostly on [70]. More detailed treatment for more compli¬cated (recurrent) neural networks can be found in [69] and [71].

In [69] and [71], the hidden variables have been chosen to be the desired targets at the hidden layer denoted by for the th training sample. This is motivated by the fact [48], [64], [89], [91] that if the targets (missing variables) of each individual neuron could be obtained in addition to the input–output training examples, training a complicated neural network with many interconnected hidden neurons could be reduced to training a sequence of individual neurons.

To establish probability models, Gaussian models have been established in [69] and [71] based on previous work [67], [72], [104]. Through some tedious algebraic manipu¬lations, the following algorithm has been derived [71].

1) and initialize the weights at both layers (

and ) randomly.

2) E-step: Compute the expected hidden targets for the

hidden units as follows:

(11)

(12)

where stands for the new weight vector

at the first layer, for . This step is

equivalent to training individual sigmoidal (hidden) units in parallel

(13)

where are the new second-layer weights.

This can be accomplished by training individual sin¬gle neurons simultaneously.

4) Let , and go back to step 2) until certain

convergence conditions are satisfied.

The main idea of the derived EM algorithm for training a two-layer neural network can be summarized as follows. At each iteration, the expected hidden targets are evaluated first through the E-step (11). The expected hidden targets then serve as the targets at the hidden layer to train the

first-layer weights (12). Since the hidden targets for

the hidden units are completely separable, finding the first-layer weights can be done through training M individual neurons simultaneously. Furthermore, the hidden targets also serve as inputs to the second layer, and the quadratic training error between the outputs and the desired targets is minimized to obtain the new second-layer weights (13). This is equivalent to training a single linear neuron. Therefore, training the original two-layer (nonlinear) net¬work is decomposed into training a set of single neurons, which can be done much faster than training the original (nonlinear) layer network. The derived EM algorithms were tested extensively on various applications [70], [71] and were shown to reduce the training time by a factor of 10–20 as compared with BP.

4) Discussion: Three types of fast training algorithms have been reviewed in this section that apply the “divide-and-conquer” principle to improve training time.

Decision tree algorithms divide a problem (or a training set) recursively by the tree structure so that nodes (or base models) in the tree deal with successively easier problems moving down the tree. Therefore, a decision tree algorithm can be viewed as implementing “divide and conquer” in a “vertical” fashion. The incremental algorithms train a base model individually using residual error. Therefore, this scheme can be viewed as implementing the “divide and conquer” in a “horizontal” fashion. The EM algorithm works in a slightly different way. It divides a problem into small pieces iteratively using the “divide-and-conquer” procedure so that each base model can be trained in the maximization step. Once training is completed, it redivides the problem and reestimates information needed to train each small piece in the expectation step. Each piece is then trained again in the M step. The process continues until a convergence criterion is met.

D. Open Problems

Although fast training algorithms have been shown to improve the training time substantially, the improvement is usually measured heuristically. The theory on training time is still missing.

Many algorithms have been proposed to adapt the struc¬ture of a single model in order to find the optimal trade¬off between the bias variance. However, these algorithms are mostly developed through heuristics and lack a solid theoretical ground. The algorithmic complexity of these algorithms may also be higher. It has not been proved the¬oretically whether and how to develop an algorithm that is efficient for achieving good performance through adapting its structure. As a result, achieving a good generalization performance of an individual model is still considered to be a challenging open question.

V. PERFORMANCE: COMBINATIONS OF

WELL-TRAINED MODELS

To alleviate the difficulties in finding an optimal structure of a single nonlinear adaptive system (neural network), methods are proposed to combine different models. The

MA AND JI: PERFORMANCE AND EFFICIENCY 1525

main idea is to train multiple models, such as neural networks, decision trees, or classifiers individually, and then combine them as an ultimate model. The hope for using a combination to improve the generalization performance is that if individual models make mistakes differently, a combined model would be able to improve upon the performance of a single model.

Tremendous efforts have been made to investigate whether a combined model can indeed improve the performance, and how to combine models to achieve good performance [3], [88]. Specifically, combinations of experts have shown to be able to work at least as well as the best expert in the pool of experts on predicting binary strings [26], [68]. Combinations of neural networks [84], [102] and regressors [21] have also been used for both classification and function approximation problems. Somewhat similar ideas were also used for collective agents such as in classifier systems and genetic algorithms [43], [50], [51].

As a good review on this subject was offered by Di-etterich [33], in this section we do not intend to survey all aspects of combinations of well-trained models. Rather, we focus on the general issues that are relevant to the next subject we will discuss on combinations of weak classifiers.

A. Key Issues

A combined model can be represented in a simple form.

Let ’s be a set of total models defined on a feature

vector , where a can be a neural network

(or a single neuron). Let be the weighting factor for

the th model. The combination of these models can be represented as

(14)

is essentially a linear combination of the so-called

base models ’s.

When a (two-class) classification is considered, the com

bined model becomes a combined classifier , where

(15)

is an indicator function, where for , and

otherwise. is a label assigned to a feature

vector by a combined classifier. When ’s are also

classifiers, is a combination of classifiers.

There are two key issues in combinations of models.

2) Given a set of base models, how to choose an

optimal set of weighting factors so that

the generalization error of the combined model is minimized.

B. Obtaining Base Models

1) Requirements: Numerous algorithms have been devel-oped to explore the first question. These algorithms are distinct in two ways. One results from different architec-tures for base models and different algorithms for obtaining

base models. Another corresponds to different methods used to “perturb” the training process so that the base models obtained can have diverse error patterns. Having classifiers with diverse error patterns were shown to be crucial to effectively improving the performance through combinations of models [61], [63]. This can be understood through two (extreme) examples. At one extreme, if these classifiers are identical, there will be no gain from any combination. At the other extreme, if these classifiers make independent errors with a probability less than 0.5, the overall number of errors made after a combination can decrease exponentially as the number of such classifiers increases [61]. However, the “diversity” is not easy to achieve because all models are trained to do essentially similar tasks. Therefore, a certain randomness should be introduced to “perturb” the learning procedure so that models can learn different parts of a problem, and thereby make errors as differently as possible.

2) Diversification of Base Models: The common methods used to perturb training were categorized by Dietterich [33], [34] into the following cases.

1) Manipulating the input features, i.e., selecting differ¬ent features for training different base models (clas¬sifiers). One such example is Adaboot [45], where every individual classifier is trained on only one randomly chosen dimension of a feature vector.

2) Randomizing training procedures. For example, [84] sets randomized initial weights of neural networks; [57] generates hyperplanes randomly.

3) Manipulating the labels in a training set, i.e., adding the random noise to the labels [32].

4) Manipulating a training set, i.e., resampling or reweighting the original training set from a certain resampling probabilities (or reweighting schemes) [20], [30], [44].

A successful algorithm for generating base models may use one or several aforementioned techniques to diversify base models.

3) Algorithms: Among the above techniques, the most appealing is to randomize training data by selecting a re-sampling probability due to its effectiveness and simplicity.

Three algorithms to control resampling probabilities have been proposed and proven to be effective by extensive experiments. The first algorithm was originally proposed by Wolpert [102] and was extensively investigated by Breiman [20], [21] as the so-called bagging algorithm. This algorithm generates a training set for training a base model through resampling the original training set with replacement uniformly. Through this resampling algorithm, a newly generated training set is not exactly the same as the original one, although both data sets are drawn from the same distribution. Therefore, the resulting base models from different training sets are different.

The second algorithm was proposed by Freund et al. [45]. The algorithm starts from a uniform resampling probability

denoted as for a training sample . The th model is

1526 PROCEEDINGS OF THE IEEE, VOL. 87, NO. 9, SEPTEMBER 1999

trained on the th training set, which is resampled with replacement from the original training set according to

probability . changes dynamically as follows.

If the th sample is incorrectly classified by the current

combined model, is increased from the previous

resampling probability . Otherwise it remains the

same. Comparing with the above uniformly resampling al¬gorithm, the resampling probabilities here are dynamically changed based on the performance of all previous obtained models.

The third algorithm was proposed by Ji and Ma [30], [57]. This algorithm divides training samples into two classes: a set of “cares” and a set of “don’t cares.” The former consists of training samples that have not been classified correctly. The latter contains the training samples that have been classified correctly. A base model is then obtained using a set of “cares” alone. This algorithm can be viewed as a compromise between the first two algorithms, since it dynamically determines the set of “cares” and “don’t cares” but uniformly resamples only on all “cares.”7

C. Combination

The other important issue for combination is how to determine the weighting factors. Intuitively, the outputs

of base models can be used to “train” the

weighting factors . However, the straightforward

minimization approach often results in overfitting [75], [77] and is therefore not applicable. This is because the base classifiers are highly correlated since they are designed to solve similar tasks. One simplest algorithm called majority vote uses the equal weighting for base models, i.e.,

for all . Clearly, this scheme is not optimal because it does not take full consideration of differences among base models. But majority vote has been used widely due to its simplicity [20], [22], [57]. More elaborate algorithms have been investigated for combination. Wolpert [102] proposed combining base models through minimizing the squared error using cross validation. Breiman [21] later imposed

proper constraints on the weights ’s. Freund et al. [44],

[45] proposed using the confidence of a base model as the weight for that model. Principal component analysis [76] was also investigated to explore and thus discount correlation among individual base models.

D. Open Questions

Both theoretical and empirical results [20], [22] sug-gest that combinations of well-trained models can ease the bias-variance dilemma and improve the performance substantially. That is, we can select a base model with a relatively large size so that it has a small bias but a large variance. A combination scheme can be responsible for reducing the overall variance for a combined model [20], [22], [76], [84]. Therefore, finding an optimal structure of a model is no longer important. Very often, however, the price paid for the gain in performance is a larger space

7 Several popular incremental learning algorithms [23], [36], [46] can also be considered as special cases in this class of algorithms.

complexity. In addition, as several models are combined, each of which may take a long time to train, an even longer training time results in for a combination. Since training time is critical for real-time applications and is more difficult to tackle than the space complexity, a natural question to ask is whether combinations of models can be used to improve the time complexity at a reasonable cost of the space complexity. Combining weak models provides a promising answer to this question.

VI. PERFORMANCE AND EFFICIENCY: COMBINATIONS OF WEAK MODELS (CLASSIFIERS)

Although individual base models in a combination can be quite general, we focus on classifiers as base models in this section.

A. Three Approaches

There are three main approaches developed on combi¬nations of weak classifiers: the boosting algorithm called Adaboost (adaptive boosting algorithm) by Freund and Shapire [44], [45]; the stochastic discrimination (SD) by Kleinberg [61], [62]; and the combination of weak percep-trons (CWP) by Ji and Ma [30], [56], [57].

The concept of weak learning was first introduced by Kearns and Valiant [60] as a part of a theoretical question in the context of the probabilistic approximately correct (PAC) learning theory [13], [97]. The question can be informally stated as “Does the existence of a weak learner imply the existence of an efficient strong learner?” Schapire [93] first proved the existence with a “yes” through a recursive and constructive approach. He showed that if a classification problem is solvable in the PAC framework, the problem can be solved through combinations of weak classifiers that can do a little better than random guessing. Later, Freund [44] showed that in theory a combination of weak classifiers through the simple majority vote can combine the weak classifiers into a strong classifier. The Adaboost [44] was further proposed as a practical algorithm for combinations. Although this work [44], [45] illustrated that weak classifiers could be combined to achieve what a strong classifier could do, it did not further address whether combinations of weak classifiers had any advantage in performance or efficiency as compared with training an individual strong classifier or combining well-trained individual strong classifiers.

In an independent work, Kleinberg [61] proposed sto-chastic discrimination and showed that combining a large number of weak classifiers could improve the (training) performance monotonically. In addition, he also showed that the time-complexity of combined weak classifiers was polynomial. The derived theory, however, was built on an assumption that weak classifiers made independent classi¬fication errors. Such an assumption cannot be achieved in a real situation.

Ji and Ma [30], [56], [57] proposed combinations of weak perceptrons (or neurons). Through extensive exper¬iments, this work showed that a very simple algorithm for

MA AND JI: PERFORMANCE AND EFFICIENCY 1527

generating and combining weak classifiers may achieve better efficiency and performance than training a strong classifier or combining well-trained classifiers. Such a sim¬ple procedure is to generate randomly weak perceptrons from the perceptron space as weak classifiers and combine these weak perceptrons through a majority vote. Time-complexity and space-complexity of a combined classifier were formally defined and shown to be polynomial for a “special case” when the strength of weak classifiers was properly chosen. A tradeoff was explicitly shown to exist between time complexity and space complexity.

In the following, we further review these three algorithms and the corresponding theoretical results.

B. Weak Classifiers

1) Definition: The strength of a classifier can be

characterized by , the so-called weakness factor, where

. Let be the required (generalization)

error of classifier , i.e., Pr .

If , the classifier is considered to be a weak

classifier because it only performs a little better than random guessing. The larger the is, the weaker the weak classifier.

The time compelxity and space complexity of follow

the general definition given in Section II but should take into account the weakness factor.

If the space complexity and time complexity of a com-bined classifier scales polynomially with respect to both the dimension of feature vectors and parameters such as the weakness factor and a desired generalization error, the classifier is said to be efficient. Otherwise, if an exponential scaling is observed, they are considered to be inefficient.

A set of weak classifiers should satisfy the following two conditions: 1) each weak classifier should do better than random guessing and 2) the set of classifiers should have enough computational power to learn a problem. The first condition ensures that each weak classifier possesses a minimum computational power. The second condition sug¬gests that individual weak classifiers should learn different parts of a problem so that a collection of weak classifiers can learn an entire problem. If all the weak classifiers in a collection were to learn the same part of a problem, their combination would not do better than individual classifiers.

2) The Structure of Weak Classifiers: Both local and non-local classifiers have been proposed. Kleinberg et al. [62] proposed using a hypersphere as a weak classifier. Such a weak classifier classifies all the samples that fall into the sphere as one class and those outside as the other class. This classifier is local because samples close to the center of a sphere are in the same class. This choice of weak classifiers, however, may suffer from the “curse of dimensionality” and thereby is difficult to handle high-dimensional problems.

Kleinberg [62] and Freund et al. [45] independently proposed to use a half space as a weak classifier. A half space classifier classifies input features based on only one selected dimension (or feature) and ignores other dimensions. Therefore, the decision boundary of a half-space classifier is a hyperplane perpendicular to the selected dimension. Such a classifier is global and may be expected

to have a nice (polynomial) scaling property with re-spect to the dimension of feature vectors, thereby avoiding the curse of dimensionality for some cases. As shown by experimental results [44], the set of half-space weak classifiers is limited for representing an arbitrary decision boundary. As a result, the performance is not as good as that of combinations of stronger classifiers. To increase the representation power of the half-space classifiers, the union of two or more half spaces was also considered.

To generate a global classifier with a stronger represen¬tational power, Ji and Ma proposed using a perceptron (or a neuron) as a weak classifier [30], [57]. Similar to a half-space classifier, the decision boundary of this classifier is also a hyperplane, and therefore global. However, the orientation of its decision hyperplane can be arbitrary. Therefore, a hyperplane classifier is much more flexible than a half-space classifier. In fact, the combination of these perceptrons (or neurons) forms a two-layer neural network which has been shown to be able to approximate arbitrary functions in theory [10].

C. Algorithms for Combinations of Weak Classifiers

Once the structure of weak classifiers is determined, qualified weak classifiers can be generated and combined.

1) Generation of a Qualified Weak Classifier: There are two fundamentally different approaches to generate a qualified weak classifier. One is to apply a training algorithm to find the “best” weak classifier. The other is to randomly generate a weak classifier from the set of all weak classifiers until a qualified classifier is obtained.

Freund et al. [45] used the first scheme, where a fea-ture is first selected randomly and the best threshold is then obtained through exhaustive search. Both stochastic discrimination (SD) [62] and combinations of weak clas¬sifiers [57] generated a weak classifier through the second approach. The algorithm can be described as follows:

1) partition the training data into “cares” and “don’t cares”;

2) randomly generate a candidate classifier from the classifier space;

3) test classification error rate of the candidate classifier on the “cares”;

4) if the error rate on the “cared” samples is less than a

threshold , then keep this classifier and go to

step 1) to generate another weak classifier; otherwise

go to step 2), where is the weakness factor.8

In this algorithm, a random classifier is first chosen as a candidate classifier. The strength of this candidate classifier is then tested on a set of “cares,” which refer to those training samples incorrectly classified. If the candidate passes the threshold, it is accepted and combined with previous qualified classifier. Otherwise, it is rejected and the procedure is repeated until a qualified candidate is found.

8A typical value of v used in [56] is between 50 and 200. A typical number of combined weak classifiers is from 500 to 5000.

1528 PROCEEDINGS OF THE IEEE, VOL. 87, NO. 9, SEPTEMBER 1999

Therefore, this approach of generating a qualified classifier can also be called trial until qualified (TUQ).

To generate a perceptron classifier randomly [57], the direction of a hyperplane is first generated randomly and uniformly. Then a feature vector is picked randomly to place the hyperplane in the feature domain.

Three considerations motivate the choice of a randomized (TUQ) approach over the training classifier approach. First, if a qualified classifier is weak enough so that it can be obtained through only several trials, the TUQ may be computational cheaper than training a classifier. Second, the random sampling to obtain a weak classifier may reduce overfitting, which is the problem inherently associated with any training algorithm. Third, random sampling may facilitate theoretical analysis as we will discuss later.

2) Combinations of Qualified Weak Classifiers: In gen¬eral, the combination schemes developed to combine well-trained models can be used to combine weak classifiers. Specifically, Freund [45] used the weighted majority vote and chose the confidence of the corresponding classifier as the weighting factors. Both combinations of weak perceptrons [57] and stochastic discrimination [62] used the simple majority vote to combine weak classifiers.

3) Adaptively Reweighting and Combining (ARC): Several algorithms for combining either weak or strong models have been discussed. They are bagging [20], Adaboost [44], stochastic discrimination [61], [62], and a combination of weak hyperplanes [30], [57]. Although these algorithms were rooted from different theories and served different purposes, they share common characteristics. Recently, Breiman [22] proposed an ARC framework to capture the common characteristics in these algorithms, where ARC characterizes a combination algorithm as an iterative three

step procedure [assuming classifiers have

been obtained in the current combination].

1) Resampling (or reweighting) the original training set to obtain the th training data set. Bagging algo-rithm uniformly resamples the original training set. Boosting adaptively resamples the original training set. Stochastic discrimination and combinations of weak perceptrons resample uniformly on dynamically chosen “cares.”

2) Generating a qualified th classifier based on the

th training data. Bagging algorithm was designed to combine well-trained models. In practice, both neural networks and decision trees (CART) [20] have been used as base classifiers. Although Adaboost was originally designed to combine weak classifiers, it has been widely used effectively to boost the performance of strong classifiers [22], [87]. SD and CWP all used the TUQ algorithm to generate randomly either a hypersphere or a perceptron but only keep those which exceed a certain performance threshold.

3) Determining the weight for voting. Adaboost uses the confidence over the performance of a model as the corresponding weight for voting. All other algorithms use the simple majority vote combination scheme.

D. Performance of Combinations of Weak Classifiers

Since the performance of Adaboost [45] when used to combine a large number of weak classifiers was found to be either comparable or somewhat inferior to combining strong classifiers, Adaboost algorithms have been used mostly to combine a small number of well-trained classifiers [22], [45], [87]. Through extensive experimental comparisons [22], [45], [87], Adaboost for combining strong models has been shown to perform consistently better than training an individual model or a bagging combination algorithm.

Can combinations of weak classifiers do better in per-formance than training a strong classifier? Although the independent assumption used in the theory for stochastic discrimination is difficult to satisfy in reality, better perfor¬mance has been obtained consistently when a large number of weak classifiers are combined and used in handwritten digit recognition [62]. In particular, the resulting combined classifiers were shown to be less sensitive to overfitting. This means the performance of a combined classifier was usually improved when more and more weak classifiers were combined. Similar to combinations of well-trained classifiers, this in fact shows that combinations of weak classifiers reduce the variance when more and more weak classifiers are combined.

Combinations of weak perceptrons [57] were tested ex¬tensively on various synthetic and real data sets. The generalization performance of the combined weak percep-trons has been shown to be slightly better than combinations of well-trained classifiers and outperforms individual neu¬ral network classifiers and -nearest neighbor classifiers. Meanwhile, as will soon be shown, interesting phenomena have been observed that a tradeoff can be made between performance and efficiency by combinations of weak per-ceptrons.

E. Efficiency of Combinations of Weak Classifiers

As the performance of combinations of weak classi-fiers is comparable to that of combinations of well-trained classifiers, what really are the advantages of using weak classifiers?

1) SD: SD [61] first suggested that combinations of weak classifiers can be used to improve the training time. A theory was derived to show that when weak classifiers were assumed to make independent classification errors, the computational time for selecting a weak classifier was polynomial in terms of the dimension of feature vectors. As the space complexity was also shown to be polynomial, the resulting time complexity of a combined classifier was polynomial. Empirically, the training time needed to obtain a combination of weak classifiers was shown to be magnitudes faster than that of conventional training methods such as BP.

Three factors were not included when the time-complexity was derived: the structure of weak classifiers; the weakness factor; and the statistical dependence among outputs of weak classifiers. As discussed in Section V-A, the structure of weak classifiers relates directly to the space

MA AND JI: PERFORMANCE AND EFFICIENCY 1529

complexity of a combined classifier and thus determines its space efficiency. The weakness factor also affects the space complexity of a combined classifier, since the weaker the weak classifiers are, the more weak classifiers may be needed to learn a problem, and the larger the space complexity. This may in turn influence the time complexity, since the time complexity of a combined classifier can be regarded as the average training time needed to obtain a weak classifier multiplied by the space complexity.9

2) Combinations of Weak Perceptrons:

a) Empirical results: The efficiency of combinations of weak perceptrons was first tested empirically through various pattern classification problems [30], [57]. For the space complexity, when the -nearest neighbor classifiers suffer from the curse of dimensionality, a nice scaling prop¬erty in terms of the number of dimensions was observed for combined weak classifiers.

As for the training time, a qualified weak perceptron could usually be obtained within ten trials for all these experiments. Therefore, combinations of weak perceptrons are magnitudes faster than training a two-layered neural network by BP algorithms.

The empirical results also showed that the weaker the weak perceptrons, the more weak perceptrons (the larger the space complexity of a combined classifier) were needed to achieve a certain performance, and the less time was needed to obtain a weak perceptron. Therefore, a tradeoff may be made between the time and space complexity through a properly chosen weakness factor.

b) Theoretical results: To shed light on whether and why combinations of weak perceptrons are efficient, the-oretical analysis was carried out on a simple example when combinations of weak classifiers are used to learn underlying perceptrons [57].

The generalization error Pr of the

combined classifier with weak perceptrons

was derived and bounded above by a quantity in the order of

i.e.,

(16)

where is the weakness factor. stands for a quantity

in the order of .

Such a bound shows that the generalization error de-creases at a polynomial rate in terms of the number of weak perceptrons. By setting the upper bound to be equal

to , which is a bound on the desired generalization error,

the space complexity of a combined classifier10 can be obtained easily as

(17)

9 The number of weak classifiers by definition.

10S is the number of weak classifiers needed to achieve a certain

generalization error.

Therefore, is polynomial in both the weakness of factor and .

The time complexity of a combined classifier is defined as the average number of samplings needed to obtain a combined classifier when a desired generalization error at most . Such a time complexity was shown to satisfy

(18)

for and large but .

Since the larger the weakness factor , the larger the space complexity , but the smaller the time complexity , a tradeoff can be made between the space complexity and time complexity by finding an optimal . Specifically,

let and assume large, an optimal weakness

factor can be obtained as . Therefore, when

, the time complexity is polynomial

in the dimension of feature vectors; otherwise, is

an exponential function of . In the meantime, when this condition is satisfied, the space complexity

is also polynomial in . The existence of such a critical value for the weakness factor suggests that the polynomial time complexity may be obtained at a cost of a larger size classifier compared to that of a well-trained classifier with a fixed structure. The cost, however, is theoretically tolerable, since it scales polynomially in the dimension of feature vectors.

c) Discussion: Through analyzing the time complex-ity, an intuitive explanation can be drawn on when and why combinations of randomly selected weak perceptrons are efficient. If weak classifiers are weak enough, i.e., the weakness factor is large enough, there will be many such weak classifiers. Therefore, the chance of getting a weak classifier is high at each sampling. That is, the number of times needed to sample the classifier space until a qualified weak classifier is accepted is small. As soon as weak classifiers are not too weak to destroy the polynomial space complexity, the efficiency can be achieved both in time and space for combinations of weak perceptrons.

The theory provides a unique explicit relationship be-tween performance and efficiency but is limited to a special case for learning a linear decision boundary. There are no similar results derived so far on nonlinear classification problems.

F. Open Questions

Due to the intrinsic difficulties of tackling performance and efficiency of nonlinear classifiers, many open questions need to be investigated on combinations of weak classifiers. Some of these open questions are given below.

1) How to show the optimality of a (randomized) algo-rithm for choosing and combining weak classifiers. Assuming there were an infinite number of weak classifiers usable in a combination, this question essentially asks whether an algorithm is optimal in approximating a Bayes classifier. As an elegant anal-ysis showed that the nearest-neighbor classifiers [27]

1530 PROCEEDINGS OF THE IEEE, VOL. 87, NO. 9, SEPTEMBER 1999

are asymptotically Bayesian optimal, it remains open as to whether or not similar results could be obtained for a randomized algorithm using weak classifiers.

2) For what problems do a large number of weak clas-sifiers exist to achieve a desired performance?

Even when an optimal algorithm is used to obtain a combination of weak classifiers, it is not clear whether a large enough number of weak classifiers exist so that the combined classifier can achieve a desired performance for a given structure (perceptron, for example) and a weakness factor. If the classification problem is too difficult, there may not exist any weak classifier, since a set of weak classifiers with the chosen structure may not have enough capacity [13], [27], [52], [74] to do better than random guessing.

3) For what problems do a large number of weak clas-sifiers exist so that the time complexity can be poly¬nomial?

Not all problems can be solved in polynomial time, (e.g., training nonlinear neural classifiers is NP complete). A polynomial complexity for randomly selecting a weak classifier is obtained based on the assumption that there exist a large number of weak classifiers. For example, if there exists at least a polynomial fraction of classifiers that can do better than random guessing, the average number of tries required to get one weak classifier is polynomial. Otherwise, choosing one acceptable weak classifier from an exponentially small fraction of all classi¬fiers would take an exponential number of tries on the average. Since the number of acceptable weak classifiers depends on the problem, it is important to characterize problems for which a large number of weak classifiers do exist.

VII. PERFORMANCE AND EFFICIENCY THROUGH STOCHASTIC TRAINING: EVOLUTIONARY ALGORITHMS

As explained in the previous sections, randomization is the key for the success of the two combination approaches. In combining well-trained models, randomness is intro¬duced into different phases of the training process to obtain diverse base models. In combining weak classifiers, the randomized algorithm is crucial to generate different weak classifiers to achieve good performance and efficiency. In fact, randomization not only benefits the combination schemes but is also essential to exploring the global rather than local optimization (as deterministic gradient-based approaches do).

In this section, we further elaborate more general ran-domized algorithms based on evolutionary computation. Evolutionary computation has been used for solving com¬plicated problems in diverse areas of engineering and biology. Its general introduction and related applications can be found in [5], [40]–[42], and references therein. In what follows, we focus on specific applications of evolutionary algorithms to tackling the issues of training neural networks. Other applications can be found in [41],

which include applying evolutionary algorithms to finding appropriate network architecture [73], [105] and adaptively adjusting the learning rate [92]. We first illustrate how to apply evolutionary algorithms to train a neural network. We then discuss the similarities and differences between evolutionary algorithms and the algorithm for combinations of weak perceptrons.

A. Evolutionary Computation

A fundamental difficulty of using a deterministic gradient-descent-based algorithm for nonlinear optimization is that the algorithm can be easily trapped by a local minimum. As the error function of a complicated nonlinear system is itself usually very complex and may have many “tiny” local minima and flat areas, using a gradient-based algorithm to train a neural network is usually extremely slow and yields poor results. To ease this problem, evolutionary algorithms, which are stochastic algorithms, were proposed [41] as an alternative to deterministic approaches.

Evolutionary algorithms are a class of stochastic opti-mization and adaptation techniques that are inspired by natural evolution. They mainly consist of genetic algorithms [51], evolutionary programming [41], and evolution strate¬gies [94]. A comprehensive review of these three different methods can be found in [7]. Although each evolutionary algorithm is designed with a different methodology, all of them are population-based search procedures. When evolutionary algorithms are used to train neural networks, a typical algorithm can be described as follows.

Let be a real-valued -dimensional weight vector of a neural network with an input–output function

, where is the input vector, and is

the output. The goal is to find a vector so that a certain

criterion on , denoted as , can be minimized.

1) Initial step at , and randomly generate an

initial population with parent vectors for

2) Generate offspring from each parent, i.e.,

(19)

where . indicates a random varia

tion operation.

3) Select the th generation based on all offspring

’s. The fitness of offspring can be evaluated

through . Then based on the fitness measure,

a selection operator can select a set of offspring as

the new parents for the th generation. That is,

a set of the new parents (weight vectors) satisfies

(20)

where is a set of weight vectors consisting of

all ’s. Steps 2) and 3) can be combined into

one as

MA AND JI: PERFORMANCE AND EFFICIENCY 1531

At each iteration which is called a generation, an evo-lutionary algorithm first generates offspring (candidates) from each of parents through a randomization operator [Step 2)]. As an example, [40] describes a simple variation

operation as , where is a vector and

its elements are independent Gaussian random variables

with zero mean and variance proportional to . The

next key step [Step 3)] in an evolutionary algorithm is to select a set of offspring as a new generation (qualified

candidates). A simple way to select offspring as the

th generation is to pick the top offspring based

on their performance . , for instance, can be

the one that minimizes an error on a given training set while minimizing the complexity (the number of nonzero weights in ) of the network [10], [39]. More sophisticated variation and selection operators can be found in [41].

B. Efficiency: Relationships to Combinations of Weak Classifiers

To understand how evolutionary algorithms may con-tribute to tackling the issue of computational efficiency, it is beneficial to compare the similarities and differences between evolutionary algorithms and combinations of weak classifiers.

1) Comparison with Combinations of Weak Classifiers: There are two loops in the algorithm for CWP given in Section VI-C1. The inner loop is used to obtain a qualified weak classifier, and the outer loop is for growing the structure by combining more and more qualified weak classifiers. The algorithm at the inner loop, which is TUQ, can be viewed as a special case of the evolutionary algorithm described above. The outer loop can also be viewed as a special case of evolutionary architecture, where the training samples are evolving toward those which are misclassified.

When weak classifiers are weak perceptrons, corre

sponds to a weight vector that defines a hyperplane. The

fitness measure represents the (classification) error

rate. The selection threshold is , which is the

upper bound on the training error to ensure that candidate classifiers are not too weak. Corresponding to Step 1), a hyperplane ( ) is chosen initially from a uniform distribu

tion. Its strength is then tested using . The selection

operator simply keeps those that have a classification

error no larger than . To generate the offspring

(hyperplanes at the next step), the randomization operator simply generates another set of hyperplanes randomly. However, different from evolutionary algorithms, combi¬nations of weak classifiers combine hyperplanes (parents) at each stage in order to select new hyperplanes (offspring). The final solution for combinations of weak classifiers is a combination of selected parents and offspring at all stages, whereas an evolutionary algorithm intends to find a (single) optimal solution at the final stage. However, the com¬bination can also be included in evolutionary algorithms by revising either the section operator or the random variation operator .

Similar comparisons can be made between evolutionary algorithms and other related randomized algorithms given in Section V. Details will be omitted.

2) Discussion: To understand whether and when evo-lutionary algorithms could be efficient in time, we recall that combinations of weak classifiers use many randomly chosen (weak) classifiers whereas evolutionary algorithms (the current version) intend to find a (single) globally optimal solution. This leads to the question of whether and when it is possible to find a single solution through random sampling in hopefully a polynomial number of steps on the average.

Consider again the simple example on combinations of weak classifiers given in Section VI-E2. When random sampling was used to select a single weak classifier, weak classifiers needed to be weak enough so that the subset of all weak classifiers occupies a large enough fraction in the space of all classifiers defined by one perceptron. Then the average number of times needed to (uniformly) sample the classifier space until a qualified weak classifier was obtained would be polynomial. Since each classifier thus obtained was weak, many such classifiers would need to be combined so that the resulting classifier could have a low classification error. The number of weak classifiers needed in a combination would be polynomial if the classification error of the combined classifiers decreased at least at a polynomial rate with respect to the number of weak classifiers. This shows that the key ingredients for com¬bining randomized weak classifiers are 1) to have a large enough weak classifier space to sample in order to achieve polynomial time-complexity for each weak classifier, and 2) to have a polynomial number of weak classifiers in a combination (a polynomial space complexity) so that the time complexity for a combined classifier is polynomial.

Imagine evolutionary algorithms were used for this case. Then random sampling is used to obtain a single classifier with a low classification error. Since the number of such strong classifiers is an exponentially small fraction of all classifiers, sampling the classifier space randomly will not result in polynomial time complexity. That is, nonuniform sampling is needed in order to achieve polynomial time complexity. This needs to be done by carefully designing the random operator and the selection operator so that smaller and smaller classifier subspaces can be sampled to eventually obtain an optimal classifier with a high proba¬bility. Finding the operators and that can accomplish such a task is crucial to the resulting time complexity. This is equivalent to other problems of random sampling with different samplers [12]. In other words, as this problem is solvable in polynomial time, and extremely simple, it could be used as a simple example to see how an evolutionary algorithm can be designed to achieve the polynomial time complexity.

C. Open Questions

As evolutionary algorithms have shown potential in find¬ing optimal space complexity in a reasonable computational time, little has been done in investigating performance

1532 PROCEEDINGS OF THE IEEE, VOL. 87, NO. 9, SEPTEMBER 1999

and efficiency of evolutionary algorithms. This, in fact, poses several interesting open questions for possible future research through both empirical and theoretical studies.

1) How can one obtain experimental evidence on the quality of solutions (networks) obtained by evolution¬ary algorithms to achieve good generalization?

2) How does one rigorously define the time complexity and space complexity of evolutionary algorithms?

3) How does the time-complexity (once defined) scale with the complexity of a benchmark problem (char-acterized by the dimension of feature vectors)?

4) How should one analyze the performance and effi-ciency of evolutionary algorithms at least for simple cases?

Due to the complexity of nonlinear optimization, theo¬retical studies of randomized algorithms would probably be tractable for only simple problems and simplified algo¬rithms. As evolutionary algorithms were posed in a very general form in order to solve complex problems, for analytical tractability, it may be beneficial to consider very simple operators ( and ) for simple examples. In fact, this approach was taken by Karp et al., who introduced randomized algorithms in a general form but provided theoretical analysis on simple cases [15], [29]. He showed that using very simple random sampling, it was possible to obtain a single solution in polynomial time for some cases, which may help analyzing the time complexity of evolutionary algorithms.

For more complex problems such as finding (nonlin-ear) neural networks, empirical studies on performance and efficiency of evolutionary algorithms on benchmark problems should be possible. Either empirical or theoretical studies on the performance and efficiency of evolutionary algorithms will provide a better understanding of these emerging approaches.

VIII. CONCLUSION

We have reviewed several general techniques to improve efficiency and performance. Three different approaches have been discussed to improve the performance through making a tradeoff between the bias and variance: 1) search¬ing for an optimal structure for a single network—this was studied mainly for improving the performance of single model; 2) training several oversized models that have a low bias but a high variance, and then diminish the overall variance through combining these models—combinations of well-trained models improve the performance through this approach; 3) training a large set of weak models that have a large bias but a small variance, and then diminish the overall bias and thus the generalization error by combining these weak models—combinations of weak classifiers improve the performance through this scheme. Furthermore, to find global (rather than local) optimal, a stochastic training algorithm, such as an evolutionary algorithm, is essential.

To improve the efficiency for searching for a single optimal solution, the “divide-and-conquer” principle can be used vertically for decision tree algorithms, horizontally for incremental learning algorithms, or recursively for the EM algorithms.

Combinations of weak classifiers, which use an incre-mental combination scheme and a randomized algorithm, have shown the potential to achieve time efficiency as well as a good generalization performance at a cost of polynomial space complexity for benchmark problems. Explicit relationships have been provided to illustrate the interrelation and the tradeoff between performance and efficiency through combinations of weak classifiers.

As randomized algorithms play an important role for combinations of (weak) classifiers in tackling performance and efficiency, hybrid methods such as evolutionary compu¬tation are discussed as a class of more general randomized algorithms.

Although much progress has been made in performance and efficiency of nonlinear adaptive systems, many prob¬lems are still wide open for possible future research.

ACKNOWLEDGMENT

The authors would like to thank L. Breiman and A. Barron for pointing out the issue of the existence of weak classifiers, and for helpful discussions. They would also like to thank T. K. Ho, G. Kleinberg, G. Nagy, and G. Zhao for relevant references and helpful discussion. Special thanks are also due to D. B. Fogel for valuable and detailed comments.

REFERENCES

[1] Y. S. Abu-Mostafa, “Information theory, complexity and neural networks,” IEEE Commun. Mag., vol. 27, pp. 25–28, Nov. 1989.

[2] S. Atari, K. Urate, and H. Naga-oka, “Information geometry of Boltzmann machines,” Neural Networks, vol. 3, pp. 260–271, 1992.

[3] J. A. Benediktsson and P. H. Swain, “Consensus theoretic classification methods,” IEEE Trans. Syst., Man, Cybern., vol. 22, pp. 688–704, July–Aug. 1992.

[4] L. Atlas, R. Cole, L. Muthusamy, A. Lippman, J. Connor, D. Park, M. El-Sharkawai, R. J. Marks, “A performance compari-son of trained multilayer perceptrons and trained classification trees,” Proc. IEEE, vol. 78, pp. 1614–1619, Oct. 1990.

[5] W. Atmar, “Notes on the simulation of evolution,” IEEE Trans. Neural Networks, vol. 5, pp. 130–146, Jan. 1994.

[6] T. Back, Evolutionary Algorithms in Theory and Practice. New York: Oxford, 1996.

[7] T. Back, U. Hammel, and H. P. Schwefel, “Evolutionary computation: Comments on the history and current state,” IEEE Trans. Evolutionary Computation, vol. 1, pp. 1–23, Apr. 1997.

[8] P. Baldi and Y. Chauvin, “Temporal evaluation of generaliza-tion during learning in linear networks,” Neural Computation, vol. 3, pp. 589–603, 1991.

[9] A. Barron, “Approximation and estimation bounds for artificial neural networks,” Machine Learning, vol. 14, pp. 113–143, 1994.

[10] , “Universal approximation bounds for artificial neural networks,” IEEE Trans. Inform. Theory, vol. 39, pp. 930–944, May 1993.

[11] , “Predicted squared error: A criterion for automatic model selection,” in Self-Organizing Methods in Modeling, S. J. Far-low, Ed. New York: Marcel Dekker, 1984.

[12] , personal communication, 1997.

MA AND JI: PERFORMANCE AND EFFICIENCY 1533

[13] E. Baum and D. Haussler, “What size net gives valid gener-alization?” Neural Computation, vol. 1, no. 1, pp. 151–160, 1989.

[14] E. Baum,“Neural net algorithm that learn in polynomial time from examples and queries?” IEEE Trans. Neural Networks, vol. 2, pp. 5–19, Jan. 1991.

[15] S. Ben-David, A. Borodin, R. Karp, G. Tardos, and A. Wigder-son, “On the power of randomization in on-line algorithms,” Algorithmica, vol. 1, pp. 2–14, Jan. 1994.

[16] Y. Bengio, “Credit assignment through time: alternative to backpropagation,” NIPS, vol. 6, pp. 75–82, May 1994.

[17] D. P. Bertsekas and J. Tsitsiklis, Neuro-Dynamic Programming. Belmont, MA: Athena Scientific, 1996.

[18] A. L. Blum and R. L. Rivest, “Training a 3-node neural network is NP-complete,” Neural Networks, vol. 5, pp. 117–127, 1992.

[19] L. Breiman, J. Friedman, and C. Stone, Classification and Regression Trees. Belmont, CA: Wadsworth, 1984.

[20] L. Breiman, “Bagging predictors,” Machine Learning, vol. 24, no. 2, pp. 132–140, Aug. 1996.

[21] , “Stacked regressions,” Machine Learning, vol. 24, no. 1, pp. 49–64, July 1996.

[22] , “Arcing classifiers,” Ann. Statistics, vol. 26, no. 3, pp. 801–849, June 1998.

[23] L. E. Breiman and J. H. Friedman, “Function approximation us-ing ramps,” in Proc. Neural Networksfor Computing, Snowbird, UT, 1993.

[24] T. X Brown, H. Tong, and S. Singh, “Optimizing admission control while ensuring quality of service in multimedia net-works via reinforcement learning,” to be published.

[25] W. Byne, “Alternating minimization and boltzmann machine learning,” IEEE Trans. Neural Networks, vol. 3, pp. 612–620, July 1992.

[26] N. Cesa-Bianchi, Y. Freund, D. P. Helmbold, D. Haussler, R. E. Shapire, and M. K. Warmuth, “How to use expert advice,” in Proc. Foundation of Computer Science, 1993, pp. 382–391.

[27] T. M. Cover and P. E. Hart, “Nearest neighbor pattern classi-fication,” IEEE Trans. Inform. Theory, vol. IT-13, pp. 21–27, 1967.

[28] T. M. Cover, “Capacity problems for linear machines,” in Pattern Recognition, L. Kanal, Ed. New York: Thompson, 1968.

[29] P. Dagum, R. Karp, M. Luby, and S. Ross, “An optimal algo-rithm for Monte Carlo estimation,” in Proc. 36th Annu. Symp. Foundations of Computer Science, Oct. 1995, pp. 142–149.

[30] H. T. Demiral, S. Ma, and C. Ji, “Combined power of weak classifiers,” in Proc. World Congr. Neural Networks, 1995, pp. 591–595.

[31] A. P. Dempster, N. M. Laird, and D. B Rubin, “Maximum likelihood from incomplete data via EM algorithm,” J. Roy. Statistical Soc. B, vol. 39, pp. 1–33, 1977.

[32] T. G. Dietterich and G. Bakiri, “Solving multi-class learning problems via error-correcting output codes,” J. AI, vol. 2, pp. 263–286, 1995.

[33] T. G. Dietterich, “Machine-learning research,” AI Mag., vol. 18, no. 4, pp. 97–136, 1997.

[34] , “Discussion of Arcing classifiers,” Ann. Statistics, vol. 26, no. 3, pp. 838–841, June 1998.

[35] R. O. Duda and P. E. Hart, Pattern Classification and Scene Analysis. New York: Wiley, 1973.

[36] S. E. Fahlman, “The recurrent cascade-correlation architecture,” NIPS, vol. 3, pp. 190–196, May 1991.

[37] U. Fayyad, G. Piatetsky-Shapiro, and P. Smyth, “From data mining to knowledge discovery: An overview,” in Advances in Knowledge Discovery and Data Mining. Menlo Park, CA: AAAI/MIT Press, 1996.

[38] T. L. Fine, S. B. Wicker, T. Berger, and J. Halpern, “Sensor-assisted ALOHA for wireless networks,” in Proc. 1998 IEEE Int. Symp. Information Theory, Aug. 1999, pp. 161–162.

[39] T. L. Fine, Feedforward Neural Network Methodology. Berlin, Germany: Springer-Verlag, 1999.

[40] D. B. Fogel, L. J. Fogel, and V. W. Porto, “Evolving neural netowrks,” Bio. Cybern., vol. 63, pp. 487–493, 1990.

[41] , Evolutionary Computation: Toward A New Philosophy of Machine Intelligence. New York: IEEE Press, 1995.

[42] , “Practical advantages of evolutionary computation,” in Proc. SPIE, vol. 3165, pp. 14–22, 1997.

[43] S. Forest, Emergent Computation: Self-Organizing, Collective, and Cooperative Phenomena in Natural and Artificial Comput-ing Networks. Cambridge, MA: MIT Press, 1990.

[44] Y. Freund, “Boosting a weak learning algorithm by majority,” Inform. Computation, vol. 121, no. 2, pp. 256–285, 1995.

[45] Y. Freund and R. Schapire, “Experiments with a new boosting algorithm,” in Machine Learning: Proc. 13th Int. Conf., 1996, pp. 148–156.

[46] J. Friedman, T. Hastie, and R. Tibshirani, “Additive logistic regression: a statistical view of boosting,” Dept. of Statistics, Stanford Univ., Stanford, CA, Tech. Rep., 1998.

[47] S. Geman, E. Bienenstock, and R. Doursat, “Neural networks and the bias/variance dilemma,” Neural Computation, vol. 4, no. 1, pp. 1–58, 1992.

[48] R. Grossman, “Learning, capacity and implementation of neural network models,” Ph.D dissertation, Wizmann Inst. Sci., 1992.

[49] D. Haussler, M. Kearns, and R. Shapire, “Bounds on the sam¬ple complexity of Bayesian learning using information theory and the VC-dimension,” in Proc. 4th Workshop Computational Learning Theory, 1991.

[50] J. H. Holland, “A mathematical framework for studying learning classifier systems,” Physica, vol. 22D, pp. 307–317, 1986.

[51] ,“Genetic algorithms,” Scientific Amer., vol. 267, no. 1, pp. 66–72, 1992.

[52] C. Ji and D. Psaltis, “Capacity of two-layer networks with binary weights,” IEEE Trans. Inform. Theory, vol. 44, pp. 256–268, Jan. 1998.

[53] , “The VC-dimension versus the statistical capacity for two layer network with binary weights,” in Proc. 4th Workshop Computational Learning Theory, 1991.

[54] C. Ji, “Generalization error and the expected network complex-ity,” in Proc. Neural Information Processing Systems: Natural and Synthetic, 1993.

[55] C. Ji and D. Psaltis, “Data-driven network synthesis algorithm: addition and deletion,” Neural Networks, Aug. 1998.

[56] C. Ji and S. Ma. “Combined weak classifiers.” NIPS, pp. 494–500, May 1996.

[57] , “Combinations of weak classifiers,” IEEE Trans. Neural Networks, vol. 8, pp. 32–42, Jan. 1997.

[58] M. Jordan and R. A. Jacobs, “Hierarchical mixture of experts and the EM algorithm,” Neural Computation, vol. 6, no. 2, pp. 181–214, 1994.

[59] S. Judd, Neural Network Design and The Complexity ofLearn-ing. Cambridge, MA: MIT Press, 1990.

[60] M. Kearns and L. G. Valiant, “Cryptographic limitations on learning boolean formulae and finite automata,” in Proc. 21st Annu. ACM Symp. Theory of Computing, 1989, pp. 433–444.

[61] E. M. Kleinberg, “Stochastic discrimination,” Ann. Math. Arti-ficial Intell., vol. 1, pp. 207–239, 1990.

[62] E. M. Kleinberg and T. Ho, “Pattern recognition by stochastic modeling,” in Proc. 3rd Int. Workshop Frontiers in Handwriting Recognition, Buffalo, NY, May 1993, pp. 175–183.

[63] A. Krogh and J. Vedelsby, “Neural network ensembles, cross validation, and active learning,” NIPS, pp. 231–238, May 1995.

[64] A. Krogh, G. I. Thorberaso, and J. A. Hertz, “A cost function for internal representations.” NIPS, vol. 2, pp. 733–745, May 1990.

[65] S. R. Kulkarni, G. Lugosi, and S. S. Venkatesh, “Learning pattern classification-a survey,” IEEE Trans. Inform. Theory, vol. 44, pp. 2178–2206, Oct. 1998.

[66] S. Lawrence and C. L. Giles, “Searching the Web: general and scientific information access,” IEEE Commun. Mag., vol. 37, pp. 116–122, Jan. 1999.

[67] E. Levin, N. Tishby, and S. A. Solla, “A statistical approach to learning and generalization in layered neural network,” Proc. IEEE, vol. 78, pp. 1568–1574, Oct. 1990.

[68] N. Little and M. Warmuth, “The weighted majority algorithm,” in Proc. 3rd Workshop Computational Learning Theory, 1989.

[69] S. Ma and C. Ji, “Fast training of recurrent networks based on EM algorithm,” IEEE Trans. Neural Networks, vol. 9, pp. 11–26, Jan. 1998.

[70] , “A unified approach on fast training of feedforward and recurrent networks using EM algorithm,” IEEE Trans. Signal Processing, vol. 46, pp. 2270–2274, Aug. 1998.

[71] S. Ma, C. Ji, and J. Farmer, “An efficient EM-based training algorithm for feedforward neural networks,” Neural Networks, vol. 10, no. 2, pp. 243–256, 1997.

[72] D. J. Mackay, “A practical Bayesian framework for backprop-agation networks,” Neural Computation, vol. 4, pp. 448–472, 1992.

[73] V. Maniezzo, “Genetic evolution of the topology and weight distribution of neural networks,” IEEE Trans. Neural Networks,

1534 PROCEEDINGS OF THE IEEE, VOL. 87, NO. 9, SEPTEMBER 1999

vol. 5, pp. 39–53, Jan. 1994.

[74] R. J. McEliece, E. C. Posner, E. R. Rodemich, and S. S. Venkatesh, “The capacity of the Hopfield associative memory,” IEEE Trans. Inform. Theory, vol. IT-33, no. 4, pp. 461–482, July 1987.

[75] R. Meir, “Bias, variance and the combination of least squares estimators,” NIPS, vol. 7, pp. 295–302, 1995.

[76] C. Merz and M. J. Pazzani, “Handling redundancy in ensembles of learned models sing principal components.” in Proc. National Conf. Artificial Intell., 1996.

[77] , “Combining neural network regression estimates with regularized linear weights,” NIPS, vol. 9, 1997.

[78] Z. Michalewicz, Genetic Algorithms Data Structure = Evo-lution Programs, 3rd ed. Berlin, Germany: Springer, 1996.

[79] U. Mitra and H. V. Poor, “Neural network techniques for adap-tive multiuser demodulation,” IEEE J. Select. Areas Commun., vol. 12, pp. 1460–1470, Dec. 1994.

[80] J. Moody and C. Darken, “Fast learning in networks of locally-tuned processing units,” Neural Computation, vol. 1, pp. 281–294, 1989.

[81] J. Moody, “Generalization, weight decay, and architecture se-lection for nonlinear learning systems,” Neural Inform. Pro-cessing Syst., 1991.

[82] R. M. Neal, “Bayesian learning via stochastic dynamics,” Neural Inform. Processing Syst., 1993.

[83] S. J. Nowlan and G. E. Hinton, “Simplifying neural networks by soft weight sharing,” Neural Computation, vol. 4, pp. 473–493, 1992.

[84] M. P. Perrone and L. N. Cooper, “When networks disagree: ensemble method for neural networks,” in Artificial Neural Networks for Speech and Vision. London, U.K.: Chapman & Hall, 1993, ch. 10.

[85] D. Psaltis, R. R. Snap, and S. S. Venkatesh, “On the finite sample performance of nearest neighbor classifiers,” IEEE Trans. Inform. Theory, vol. IT-40, pp. 820–837, May 1994.

[86] J. R. Quinlan, C4.5: Programs for Machine Learning. San Mateo, CA: Morgan Kaufman, 1993.

[87] , “Bagging, boosting, and c4.5,” in Proc. 13th National Conf. Artificial Intelligence, 1996, pp. 725–730.

[88] S. Rangarajan, P. Jalote, and S. K. Tripathi, “Capacity of voting systems,” IEEE Trans. Software Eng., vol. 19, pp. 698–705, 1993.

[89] R. Rohwer, “The moving target training algorithm,” NIPS, vol. 2, pp. 558–565, May 1990.

[90] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, Learning Internal Representations By Error Propagation. Cambridge, MA: MIT Press, 1986.

[91] D. Saad and E. Marom, “Learning by choice of internal repre-sentations: An energy minimization approach,” Complex Syst., vol. 4, pp. 107–118, 1990.

[92] R. Salomon, “Evolutionary algorithms and gradient search: Similarities and differences,” IEEE Trans. Evolutionary Com-putation, vol. 2, pp. 45–55, July 1998.

[93] R. Schapire, “The strength of weak learnability,” Machine Learning, vol. 5, no. 2, pp. 197–227, 1990.

[94] H. P. Schwefel, Evolution and Optimum Seeking. New York: Wiley, 1995.

[95] M. Thottan and C. Ji, “Proactive anomaly detection using distributed intelligent agents,” IEEE Network Magazine (Special Issue on Network Management—Today and Tomorrow), Sept. 1998.

[96] N. Tishby, E. Levin, and S. Solla, “Consistent inference of probability in layered networks: prediction and generalization,” in Proc. Int. Joint Conf. Neural Networks, 1989.

[97] L. G. Valient, “A theory of learnable,” in Commun. ACM, vol. 27, no. 11, pp. 1134–1142, 1984.

[98] V. Vapnik, Estimation of Dependences Based on Empirical Data. New York: Springer-Verlag, 1982.

[99] , The Nature of Statistical Learning Theory. New York:

Springer Verlag, 1995.

[100] A. Weigend and D. E. Rumelhart, “Generalization by weight elimination with application to forecasting,” in Proc. Int. Joint Conf. Neural Networks, 1991, pp. 2374–2379.

[101] A. S. Weigend and D. E. Rumelhart, “The effective dimension of the space of hidden units,” in Proc. Int. Joint Conf. Neural Networks, 1991, pp. 2069–2074.

[102] D. Wolpert, “Stacked generalization,” Neural Networks, vol. 5, no. 2, pp. 241–259, 1992.

[103] P. J. Werbos, “Supervised learning: can it escape local mini-mum?” in Proc. Int. Conf. Neural Networks, 1993.

[104] H. White, “Parametric statistical estimation with artificial neural networks,” Univ. California, San Diego, Tech. Rep., Mar. 1992.

[105] X. Yao and Y. Liu, “A new evolutionary system for evolving artificial neural networks,” IEEE Trans. Neural Networks, vol. 8, pp. 694–713, May 1997.

Sheng Ma received the B.S. degree in electrical engineering from Tsinghua University, Beijing, China, in 1987. He received the M.S. and Ph.D degrees in electrical engineering from Rensse¬laer Polytechnic Institute, Troy, NY, in 1995 and 1998, respectively.

Since 1998, he has been with the IBM T. J. Watson Research Center, Hawthorne, NY, as a Research Staff Member. His current research in¬terests are network traffic modeling and control, network/system management, and data mining.

Chuanyi Ji received the B.S. degree (with hon¬ors) from Tshinghua University, Beijing, China, in 1983, the M.S. degree from the Univer¬sity of Pennsylvania, Philadelphia, in 1986, and the Ph.D. degree from California Institute of Technology (Caltech), Pasadena, in 1992, all in electrical engineering.

In November 1991, she joined the faculty of Rensselaer Polytechnic Institute, Troy, NY, as an Assistant Professor. She is now an Associate Professor of Electrical Computer and Systems Engineering at the same institution. Currently, she is on sabbatical at Bell Laboratories, Lucent Technologies, Murray Hill, NJ. In the past, she has done reseach in the area of adaptive learning systems, where she has investigated capacity and generalization capability of nonlinear classfiers and neural networks, randomized algorithms for combinations of weak classifiers, and expectation and maximization algorithms for training static and dynamic neural networks. In recent years, her interests have also been in the areas of computer communication networks. She has been involved in modeling and analysis of heterogeneous network traffic using wavelets, network fault and performance management using adaptive systems, and admission control for wireless networks.

Dr. Ji was a reciepient of the Ming-Li Scholarship at Caltech in 1989 and the NSF Early Career Development (CAREER) Award in 1995.

MA AND JI: PERFORMANCE AND EFFICIENCY 1535

KRONECKER GRAPHICAL LASSO

Theodoros Tsiligkaridis*, Alfred O. Hero III*, and Shuheng Zhou,*

University of Michigan, *EECS Dept. and  Dept. Statistics, Ann Arbor, USA

{ttsili,hero,shuhengz}@umich.edu

ABSTRACT

We consider high-dimensional estimation of a (possibly sparse) Kronecker-decomposable covariance matrix given i.i.d. Gaussian samples. We propose a sparse covariance esti¬mation algorithm, Kronecker Graphical Lasso (KGlasso), for the high dimensional setting that takes advantage of structure and sparsity. Convergence and limit point characterization of this iterative algorithm is established. Compared to stan¬dard Glasso, KGlasso has low computational complexity as the dimension of the covariance matrix increases. We de¬rive a tight MSE convergence rate for KGlasso and show it strictly outperforms standard Glasso and FF. Simulations val-idate these results and shows that KGlasso outperforms the maximum-likelihood solution (FF), in the high-dimensional small-sample regime.

Index Terms— sparsity, structured covariance estima¬tion, penalized maximum likelihood, graphical lasso

1. INTRODUCTION

Covariance estimation is a problem of great interest in many different disciplines, including machine learning, signal pro¬cessing, economics and bioinformatics. In this paper we consider covariance estimation in the multivariate Gaussian model under the separable positive definite pf x pf covari¬ance matrix assumption:

Σ0 = A0 ® B0 (1)

where A0 is a p x p positive definite matrix and B0 is an f x f positive definite matrix. Model (1) arises in channel modeling for MIMO wireless communications, where A0 is a transmit covariance matrix and B0 is a receive covariance matrix, and in other applications, see [1]. Let Θ0 := Σ1

0 de-note the inverse covariance, or precision matrix. As compared to the standard saturated (unstructured) model, the number of independent parameters in (1) is reduced from Θ(p2f2) to Θ(p2) + Θ(f2). Furthermore, as shown [1], factorization (1) results in a significant reduction in estimation mean squared error and in estimator computational complexity. In this paper

The research reported here was partially supported by a grant from ARO, W911NF-11-1-0391, and a Rackham Merit fellowship to the first author.

we propose estimation of a sparse version of the Kronecker product model (1) resulting in even more significant perfor¬mance improvements than for the saturated model studied in [1].

Under model (1), the joint probability distribution of the measurements can be represented by an undirected graph 9 = (V, £), where V is the vertex set (each vertex corresponding to a variable) and £ is the edge set. If, given all the other vari¬ables, the ith variable is conditionally independent of the jth variable, then (i, j) / £ [2]. Estimating an undirected Gaus¬sian graphical model is equivalent to estimating the inverse covariance matrix. Penalized likelihood estimators for Gaus¬sian graphical models, such as the graphical lasso (Glasso) have been proposed [3, 4, 5]. The maximum-likelihood (ML) estimator of the Kronecker product (1) has been studied in [1, 6]. While the ML estimator has no known closed-form solution, an approximation to the solution can be iteratively computed via an alternating algorithm: the flip-flop (FF) al¬gorithm [1, 6].

To our knowledge, ML estimation for the situation where the Kronecker component matrices are themselves sparse has not been studied. In addition to the Kronecker factorization, we exploit sparsity in order to derive better estimators, espe¬cially for the large-dimension small-sample regime. In this paper, we propose an B1-penalized likelihood estimator for the sparse Kronecker product case.

Statistical consistency is guaranteed, i.e., the estimator converges in probability to the true inverse covariance matrix Θ0 asymptotically as the number of samples and dimensions of Kronecker factor matrices grows to infinity. The main con¬tribution is the derivation of the high-dimensional MSE con¬vergence rates for KGlasso. When both Kronecker factors are sparse, it is shown that KGlasso strictly outperforms FF and naive Glasso in MSE, and the performance improvement can be very significant. Simulations show that KGlasso exhibits superior empirical performance.

2. NOTATION

For a square matrix M, define |M|1 = IIvec(M)II1 and |M|. = IIvec(M)II., where vec(M) denotes the vectorized form of M (concatenation of columns into a vector). IIMII2 is the spectral norm of M. Mi,j and [M]i,j are the (i, j)th el

The Glasso mapping (3) is written as G(•) : dd,

 

G(T) = arg min tr(ΘT)  log det(Θ) + Θ1 

ΘSd ++

(10) As compared to the (44) computational complexity of Glasso [8], KGlasso has a computational complexity of only (4 + 4).

Assuming ˆSn is p.d., KGlasso converges to a critical point of the objective function [10]. Under a mild assumption on the starting point, KGlasso can be shown to converge to a local minimum [10].

5. HIGH DIMENSIONAL CONSISTENCY OF FF

In this section, we show that the flip-flop (FF) algorithm achieves the optimal (non-sparse) statistical convergence rate

p2+f2

of P (up to a log-factor). This result (Thm.

1) allows us to establish that the proposed KGlasso has sig¬nificantly improved MSE convergence rate (Thm 2). We make the following standard assumption on the spectra of the Kronecker factors.

Assumption 1. Uniformly Bounded Spectra

There exist absolute constants AABBAinitAinit

such that:

1a. 0 Amin(A0) max(A0) A 

1b. 0 Bmin(B0) max(B0) B 

2. 0 Ainit min(Ainit) max(Ainit) 

Ainit 

Let RFF(3) := ˆA( ˆB(Ainit))ˆB( ˆA( ˆB(Ainit))) denote the 3-step (noniterative) version of the flip-flop algorithm [1]. More generally, let ˆRFF () denote the -step version of the flip-flop algorithm. Let ΘFF (k) = (RFF (k ))1.

Theorem 1. Let A0 B0, and Ainit satisfy Assumption 1 and define  = max(). Assume  2 and  log  for some finite constant  0. Finally, assume pf + 1. Then, for  2 finite,

 

(2 + 2) log 

ΘFF ()  Θ0F = P (11)



as .

Proof. Due to space limitations the proof is given in [10].

The bound (11) specifies the rate of reduction of the es¬timation error for the multi-iteration FF algorithm, which in¬cludes the three step FF algorithm ( = 3) [1] as a special case. The error reduction decreases as long as  and  do not increase too quickly in .

Note that (11) specifies a faster rate than that of the naive sample covariance matrix estimator (5). Furthemore, since

the computational complexity for FF is (2 + 2) which is less than the (22) complexity of SCM, by exploiting Kro-necker structure FF simultaneously achieves improved MSE performance and reduced computational complexity.

6. HIGH DIMENSIONAL CONSISTENCY OF

KGLASSO

In this section, high dimensional consistency is established for KGlasso as .

Define ΘKGlasso(k) as the output of the kth KGlasso it¬eration.

Theorem 2. Let A0 B0 Ainit satisfy Assumption 1. Let

 = max(). Let ¯(1) log M

Y  np and ¯(k)

 1 log M  1

p + 1 f n  ¯(k~) log M

Y p + 1 f n as

 for all  1 and  2. Assume sparse X0 and Y0, i.e. X0 = ()Y0 = (). Assume

max f f log  = (). Then, for  2 finite, we

have

ΘKGlasso()  Θ0F = P  

(12) 

( + ) log

as .

Proof. The proof uses results from large deviation theory. See

[10].

Thm. 2 generalizes Thm. 1 to the case of sparse Kro-necker structure. Comparison between the error expressions (4), (11) and (12) show that, by exploiting both Kronecker structure and sparsity, KGlasso can attain significantly lower estimation error than standard Glasso [3] and FF [1].

7. SIMULATION RESULTS

In this section, we compare the KGlasso algorithm with the flip-flop (FF) algorithm [1] that iteratively computes the ML solution. The Glasso algorithm implementation used was based on [8] with a stopping criterion determined by when the duality gap falls below a threshold of 105.

To empirically evaluate performance, Monte Carlo simu¬lations were used. The true matrices X0 := A1

0 and Y0 := B1

0 were unstructured randomly generated positive definite matrices based on an Erd¨os-R´enyi graph model. Performance assessment was based on normalized Frobenius norm error in the covariance and precision matrix estimates. The normal



NMC

ized error was calculated using 1 Σ0 ˆΣ(i)k2

NMC i=1 F ,

Σ02

where MC is the number of Monte Carlo runs and ˆΣ(i) is the covariance output from the ith trial run. 1

1The same formula can be adapted to calculate the normalized error in the precision matrix ˆΘ0.

In our implementation of KGlasso, the regularization parameters were chosen as follows. The initialization was Xinit = Ip. The regularization parameters were selected as

λ(1) /V log M /V log M

Y = cy np , λ(2)

X = cx nf + λ(1)

Y , λ(2)Y = λ(2)

X ,

λ(3)

X = λ(2)

X , etc. We set cx = cy = 0.4.

We consider the setting where X0 and Y0 are large sparse matrices of dimension p = f = 60 (see Fig. 1). The dimen¬sion of Θ0 is d = 3,600, which is too large for standard Glasso to handle. Thus, it is not shown.

Figure 2 compares the root-mean squared error (RMSE) performance in precision and covariance matrices as a func¬tion of n. As expected, KGlasso outperforms FF [1] over the exhibited range of n for both the covariance and the inverse covariance estimation problem.

Fig. 1. Sparse Kronecker matrix representation. Left panel: left Kronecker factor. Right panel: right Kronecker factor.

Fig. 2. Normalized RMSE performance for precision matrix (top) and covariance matrix (bottom) as a function of n. For n = 10, there is a 69% RMSE reduction for the precision matrix and 36% RMSE reduction for the covariance matrix when using KGlasso instead of FF.

8. CONCLUSION

We considered high-dimensional estimation of a Kronecker-decomposable covariance matrix given i.i.d. Gaussian sam¬

ples. A B1-penalized likelihood approach was proposed for estimating the covariance matrix when the kronecker factors are sparse. This led to an iterative algorithm (KGlasso) that takes advantage of structure and sparsity. A tight MSE con¬vergence rate was derived for KGlasso, showing significantly better MSE performance than standard Glasso and FF [1]. Simulations validated our theoretical predictions.

9. ACKNOWLEDGEMENT

The authors thank Prof. Mark Rudelson for very helpful dis¬cussions on large deviation theory.

10. REFERENCES

[1] K. Werner, M. Jansson, and P. Stoica, “On estimation of covariance matrices with Kronecker product structure,” IEEE Trans. Sig. Proc., vol. 56, no. 2, Feb. 2008.

[2] S. L. Lauritzen, Graphical Models, Oxford University Press, 1996.

[3] A. Rothman, P. Bickel, E. Levina, and J. Zhu, “Sparse permutation invariant covariance estimation,” Elec¬tronic Journal of Statistics, vol. 2, pp. 494–515, 2008.

[4] S. Zhou, J. Lafferty, and L. Wasserman, “Time varying undirected graphs,” Machine Learning Journal, vol. 80, pp. 295–319, 2010.

[5] P. Ravikumar, M. J. Wainwright, G. Raskutti, and B. Yu, “High-dimensional covariance estimation by minimiz¬ing i1-penalized log-determinant divergence,” Elec¬tronic Journal of Statistics, vol. 5, pp. 935–980, 2011.

[6] N. Lu and D. Zimmerman, “On likelihood-based in-ference for a separable covariance matrix,” Tech. Rep., Statistics and Actuarial Sc. Dept., Univ. Iowa, IA, 2004.

[7] O. Banerjee, L. El Ghaoui, and A. d’Aspremont, “Model selection through sparse maximum likelihood estimation for multivariate gaussian or binary data,” Journal of Machine Learning Research, March 2008.

[8] J. Friedman, T. Hastie, and R. Tibshirani, “Sparse in-verse covariance estimation with the graphical lasso,” Biostatistics, vol. 9, no. 3, pp. 432–441, 2008.

[9] G. I. Allen and R. Tibshirani, “Transposable regularized covariance models with an application to missing data imputation,” The Annals of Applied Statistics, vol. 4, no. 2, pp. 764–790, 2010.

[10] T. Tsiligkaridis, A. O. Hero III, and S. Zhou, “Con¬vergence properties of kronecker graphical lasso algo¬rithms,” arXiv:1204.0585v1, 2012.

Children, Memory,

and Family Identity

in Roman Culture

Edited by

VE´ RONIQUE DASEN

and

THOMAS SPA

Great Clarendon Street, Oxford ox2 6dp

Oxford University Press is a department of the University of Oxford.

It furthers the University’s objective of excellence in research, scholarship,

and education by publishing worldwide in

Oxford New York

Auckland Cape Town Dar es Salaam Hong Kong Karachi

Kuala Lumpur Madrid Melbourne Mexico City Nairobi

New Delhi Shanghai Taipei Toronto

With offices in

Argentina Austria Brazil Chile Czech Republic France Greece

Guatemala Hungary Italy Japan Poland Portugal Singapore

South Korea Switzerland Thailand Turkey Ukraine Vietnam

Oxford is a registered trade mark of Oxford University Press

in the UK and in certain other countries

Published in the United States

by Oxford University Press Inc., New York

# Oxford University Press 2010

The moral rights of the author have been asserted

Database right Oxford University Press (maker)

First published 2010

stored in a retrieval system, or transmitted, in any form or by any means,

without the prior permission in writing of Oxford University Press,

or as expressly permitted by law, or under terms agreed with the appropriate

reprographics rights organization. Enquiries concerning reproduction

outside the scope of the above should be sent to the Rights Department,

Oxford University Press, at the address above

You must not circulate this book in any other binding or cover

and you must impose the same condition on any acquirer

British Library Cataloguing in Publication Data

Data available

Library of Congress Cataloging in Publication Data

Library of Congress Control Number: 2010933158

Typeset by SPI Publisher Services, Pondicherry, India

Printed in Great Britain

on acid-free paper by

MPG Books Group, Bodmin and King’s Lynn

ISBN 978–0–19–958257–0

13579108642

Preface and Acknowledgements

This volume presents a selection of the papers delivered at the Fifth Roman Family Conference, ‘Secret Families, Family Secrets’, which took place in June 2007 in Fribourg (Switzerland). The conference, held for the first time in Europe, assembled specialists from different academic and cultural traditions: American, Australian, Belgian, Finnish, French, German, Italian, and Swiss.

Thanks are due to a number of institutions for their generous financial support which made the conference possible: the Swiss National Sciences Foundation, the Swiss Academy of Humanities and Social Sciences, and the De´partement des Sciences de l’Antiquite´ of the University of Fribourg. We are also very grateful for the editorial assistance of Oxford University Press, for the judicious comments of OUP’s readers, and to Lukas Grossmann and Diana Valaperta who contributed to the preparation of the final version of the volume.

Our thanks go above all to Beryl Rawson, for the outstanding impulse she gave to Roman family studies and for her trust and continuous encouragement.

V.D. and T.S.

Contents

List of Contributors ix

List of Figures x

Abbreviations xiii

Introduction 1

Ve´ronique Dasen and Thomas Spa¨th

I. FAMILY IDENTITIES AND TRADITIONS

1. Remembering one’s Ancestors, following in their Footsteps, being like them: The Role and Forms of Family Memory

in the Building of Identity 19

Catherine Baroin

2. Roman Patchwork Families: Surrogate Parenting,

Socialization, and the Shaping of Tradition 49

Ann-Cathrin Harders

3. Children and the Transmission of Religious Knowledge 73

Francesca Prescendi

4. Women and Children in Ancient Landscape 95

Michel E. Fuchs

5. Wax and Plaster Memories: Children in Elite and

non-Elite Strategies 109

Ve´ronique Dasen

6. Cicero, Tullia, and Marcus: Gender-Specific Concerns

for Family Tradition? 147

Thomas Spa¨th

7. Children and the Memory of Parents in the Late Roman

World 173

Ville Vuolanto

viii Contents

II. CHILDREN ON THE MARGINS?

8. Degrees of Freedom: Vernae and Junian Latins in the

Roman familia 195

Beryl Rawson

9. Modestia vs. licentia: Seneca on Childhood and Status

in the Roman Family 223

Francesca Mencacci

10. Delicia-Children Revisited: The Evidence of Statius’

Silvae 245

Christian Laes

11. The Sick Child in his Family: A Risk for the Family

Tradition 273

Danielle Gourevitch

12. Hidden in Plain Sight: Expositi in the Community 293

Judith Evans Grubbs

13. Rome: The Invisible Children of Incest 311

Philippe Moreau

References 331

General Index 363

Index of Ancient Authors and People 367

Index of Latin and Greek Terms 372

List of Contributors

Catherine Baroin is Maıˆtre de confe´rences of Latin Language and Literature at the University of Rouen (France).

Ve´ronique Dasen is Professor in Classical Archaeology at the University of Fribourg (Switzerland).

Judith Evans Grubbs is Professor of Roman History at Emory University (USA).

Michel Fuchs is Professor in Roman Provincial Archaeology at the University of Lausanne (Switzerland).

Danielle Gourevitch is Directeur d’e´tudes at the E´cole pratique des Hautes E´tudes in Paris (France).

Ann-Cathrin Harders is Lecturer in Ancient History at the University of Mu¨nster (Germany).

Christian Laes is Assistant Professor of Ancient History at the Uni¬versity of Antwerp, and Assistant Professor of Latin and Ancient History at the Free University of Brussels (Belgium).

Francesca Mencacci is Professor of Classical Philology at the Uni-versity of Siena (Italy).

Philippe Moreau is Professor of Latin Language and Literature at the University Paris Est Cre´teil (France).

Francesca Prescendi is Professor in History of Religions at the University of Geneva (Switzerland).

Beryl Rawson is Professor Emerita and Adjunct Professor in Classics at the Australian National University, Canberra (Australia).

Thomas Spa¨th is Professor in Ancient Cultures and Constructions of Antiquity at the University of Bern (Switzerland).

Ville Vuolanto is Lecturer at the Department of History and Philo-sophy, University of Tampere (Finland).

Cicero, Tullia, and Marcus

Gender-Specific Concerns for Family Tradition?*

Thomas Spa¨th

Various recent studies have suggested that the love of fathers and mothers for their children was not only one of the ideals of Roman culture but also a natural everyday occurrence. This perspective can be seen as a reaction to Philippe Arie`s’s opposite claim, advanced in his landmark study of children and families in the Ancien Re´gime (1960). Contrary to recent views, Arie`s argued that a proper notion of childhood emerged only in the modern period, and that the high child mortality rate in pre-modern societies thwarted the develop¬ment of an emotional bond between parents and children, especially infants and small children.1 Over the past few decades, such claims have met with increasing criticism. In her contribution to the second Roman Family Conference (1988), Suzanne Dixon, for instance, as¬serts that a ‘sentimental ideal of Roman family life’ had already arisen in the Late Republic. She compares this emotional ideal without further hesitation with the modern ideal of the family as a haven of

* I began developing the ideas set forth in this essay in a seminar on ‘Gender Relations in Cicero’s Letters’, co-taught with Leonhard Burckhardt at the University of Basel in the summer semester 2002. I am indebted to discussions with both him and the seminar participants, and gratefully acknowledge Dagmar Bargetzi’s seminar paper (2002) in particular. I am very grateful to Mark Kyburz for rendering my German thoughts into English prose.

1 Arie`s 1973; for details, see the Introduction to this volume.

148 Thomas Spa¨th

peace and refuge against a hostile outside world.2 In the same collection of essays, edited by Beryl Rawson, Emiel Eyben observes that sons would have likewise benefited from increasing paternal ‘warmth and tenderness’, and dates this development to the turn of the third century BCE.3

Seen against this background, the present essay discusses Marcus Tullius Cicero’s relationship with his children—Tullia and Marcus— on the basis of his letters.4 A glance at earlier studies would lead one to expect support for Eyben’s notion of ‘paternal love’5 and also for Dixon’s ‘sentimental ideal’. The relevant literature does indeed men¬tion Cicero’s almost obsessive love for Tullia,6 his ‘unbounded affec¬tion for Tullia’,7 or at the very least observes that he loved his daughter more than his son.8 Can our modern notions of parental love, however, actually grasp Cicero’s ‘paternal love’ for his daughter?

2 Dixon 1991: 113. Likewise, Judith Hallett (1984) emphasizes the emotional bond between father and daughter (see also Suzanne Dixon’s critical review of Hallett’s essay in American Journal of Philology 107 (1986), 125–30); see further Pomeroy 1976.

3 Eyben 1991: 142 seems to believe in ‘natural’ feelings when he speaks of ‘ma-ternal instinct’ (117), which he would probably set alongside paternal ‘instinct’. In her contribution to the fourth Roman Family Conference, Susan Treggiari (2005: 18) also uses the term ‘paternal instinct’, but places it in its historical and cultural context, defining it as ‘defence of hearth, home, fortunes, household gods, wives and children’ rather than as a purported universal. Her essay analyses the Roman perception of ‘natural affection’ in family relations in terms of their rhetorical use in Cicero’s political and court speeches.

4 Hereinafter, I refer to M. Tullius Cicero (106–43) as ‘Cicero’, and to his epon-ymous son (65–after 25) by his first name ‘Marcus’; ‘Q. Cicero’ designates Cicero’s younger brother (c.102–43), and ‘Quintus’ his son (67/66–43). All dates refer to the period BCE.

5 The second section of Eyben’s essays bears this title; see Eyben 1991: 116–21.

6 Hallett 1984: 134.

7 Carp 1981: 351.

8 Eyben 1991: 139. Elizabeth Rawson (1979: 197) also emphasizes how close Cicero was to his daughter, who showed greater understanding for him than his wife Terentia, for whom their son Marcus possessed ‘his mother’s practical outlook and abilities’, thus explaining his father’s lack of continuous interest in him (Rawson 1979: 223). See also Treggiari 2007: 161–2, especially n. 49, which cites the relevant literature, and n. 51, which collates the diminutives and other epithets used to refer to Tullia (see further Ermete 2003: 232 n. 1311). For a general discussion of paternal affection, specifically of fathers favouring daughters over sons, see especially Hallett 1984: 62 ff.; Pomeroy (1976: 215) refers to Plutarch’s coniugalia praecepta 36, mor. 143 B, in which he observes that fathers would love their daughters more than their sons because they felt more needed by the former.

Cicero, Tullia, and Marcus 149

Or does his behaviour instead reveal a specifically Roman type of parental affection? Did Cicero’s endeavour as paterfamilias to secure and perpetuate both the tradition of a consular domus and the name of the Tullii Cicerones, which he had founded as a homo novus, determine his relationship with his children? If so, however, how would this dovetail with Cicero favouring his daughter over his son, who could after all hand down his name from one generation to the next? Was not a male child,9 specifically, obliged to follow in the ‘footsteps of his ancestors’, as Catherine Baroin describes the prevail¬ing social norms and expectations about male descendants? Would not Marcus therefore, by virtue of imitatio patris, have been obliged to imitate his father’s founding of the family’s consular standing?10

These questions raise two issues: first, Cicero’s specific concerns for his children; and, second, the gender-specific differences between Cicero’s treatment of his daughter and his son. In what follows, I will first compare Cicero’s comments in his letters about his children’s education and schooling. Secondly, I will discuss his views on pro¬moting his children’s careers (that is, grooming Marcus for a career in politics and preparing Tullia for her various marriages—and thus for her integration into society and occupation of a particular social rank). This comparative approach to Cicero’s treatment of his chil¬dren serves to illuminate the everyday practices involved in establish¬ing a tradition in a ‘new family’,11 how such a tradition was passed on to its children, and how it thus became established as a tradition.

9 Compared to the numerous studies on the relationship between Cicero and Tullia, his relationship with his son has hitherto received scant attention; see, how-ever, Hall 2005. In the early 1930s, James Stinchcomb (1932/3) compared the biographical facts about Quintus and Marcus with the corresponding passages in the letters. He claims that Marcus received Cicero’s continuous support, even though he hands down the negative image of Marcus still evident in Syme (1939: 303, 498) and even in Fu¨ndling’s more recent DNP essay (‘[I 10] Tullius Cicero, M.’, in DNP 12 (2002), 902–3).

10 See Baroin’s contribution to this volume; for a discussion of imitatio patris as a social norm in Roman culture, see Scholz 2006.

11 Even though the relevant literature and scholarship customarily speaks of homo novus—and not, for instance, of domus nova—I would like to coin the term ‘new family’ to identify the family of the novus homo, because individual social ascendency obviously also implies that of the narrower and wider kinship. On the close relation¬ship between the domus, that is, the domestic sphere and socio-political status and prestige, see Burckhardt 2003.

150 Thomas Spa¨th

Finally, I will consider the nature of Cicero’s ‘paternal love’, and discuss the similarities and contrasts with present-day conceptions.

1. SCHOOLING AND EDUCATING CHILDREN

This first area of paternal care reveals a radical difference between daughter and son: Cicero utterly ignores his daughter’s education in his letters, unlike his son’s. It would be mistaken, however, to assume that Tullia remained uneducated. In the aristocratic domus, daugh¬ters quite obviously took part in social events, such as invitations to cultivate friendships and the conversations held on such occasions. Cicero’s earliest surviving letters to Atticus, written in 68 and 67, attest to his daughter’s involvement. For instance, Tullia, who was about 10 years old at the time,12 asks her father to be remembered to Atticus, whom she admonishes for not yet giving her the small gift that he had promised her.13 What Beryl Rawson has observed about the Roman aristocracy in general thus also applies to Tullia’s acquisi-tion of knowledge and social customs:14 in an intellectual milieu in which Roman poets and Greek intellectuals moved freely, the every¬day communication about philosophy, literature, and politics in the aristocratic domus contributed to the socialization of daughters as well as sons.15 Accordingly, Tullia would have had access to Cicero’s library, and we can assume that father and daughter would have discussed books.16 Later letters reveal that Cicero looked upon his daughter as an educated woman: for not only does she write to

12 Tullia’s precise year of birth is unknown; she must have been born between 79 and 76. See further Hallett 1984: 140 n. 77.

13 Cic. Att. 1.5.8 [SB 1], 1.8.3 [SB 4], 1.10.6 [SB 2]; here and in the following, the first numeral refers to the Vulgata edition of Cicero’s Letters, the second to Shackleton Bailey’s edition. For the English translation and the dating of the letters, I follow Shackleton Bailey, unless otherwise mentioned.

14 Rawson 2003: 153–7.

15 See also Peter Scholz (2006: 128), who considers the parental home the site of ‘primary socialization’. As Ann-Cathrin Harders mentions in the introduction to her contribution to this volume, socialization as the unintentional acquisition of knowl-edge must be distinguished from education as the intentional transfer of knowledge.

16 Rawson 2003: 156.

Cicero, Tullia, and Marcus 151

him,17 but she also reads the letters addressed to him over his shoulder,18 and indeed shares her assessment of the critical political situation in 49 with him.19 But the education she receives at home prior to her first marriage—somewhere between the age of 13 and 16—seems to have been so self-evident that it goes unnoticed in Cicero’s letters.

By contrast, Cicero’s correspondence contains many explicit refer¬ences to his son’s education. In a letter to Atticus written in April 59, Cicero conveys his 6-year-old son’s request to ‘give Aristodemus the same answer about him as you gave about his cousin, your nephew’. Shackleton Bailey suspects that Aristodemus, referred to only once in Cicero’s letters, was the boys’ private tutor, and that they were obliged to send their apologies for missing a grammar lesson.20 Two weeks later, Cicero wrote two further letters to Atticus, on 16 (or 17) and 20 April respectively, whose final salutations each contain a line of Greek. In the first letter, this reads a KtKEpwv o µtps dcma';ETat TiTov Tov ABqvaiov—‘Le petit Cice´ron salue Tite l’Athe´-nien’, as Shackleton Bailey translates the line according to his inge¬nious convention to render the Greek in French. He refers here to Wieland’s nice idea, for which there is obviously no evidence, that Marcus, who had begun to learn writing Greek, had appended the line.21

Three aspects of this first reference to Marcus’ education are worth noting: first, Cicero’s paternal interest in his son’s schooling; second, the sons of the two Cicero brothers are educated together; and third, the interesting reference to the essentially bilingual education of the Roman elite.22 A survey of the other passages in the letters indicating

17 Even though the correspondence contains no letter written by Tullia, explicit reference to such letters is made; see, for instance, Cic. Att. 10.2.2 [SB 192] (dated 5 or 6 Apr. 49) or Att. 10.8.1 [SB 199] (2 May 45).

18 Cic. Att. 10.13.1 [SB 205].

19 Cic. Att. 10.8.1 [SB 199].

20 Cic. Att. 2.7.5 [SB 27]; see Shackleton Bailey 1965–70: vol. 1, 367.

21 Cic. Att. 2.9.4 [SB 29]; see also Att. 2.12.4 [SB 30] and 2.15.4 [SB 35]. See also Shackleton Bailey’s comments on the passages.

22 Unlike along tradition of German classics scholarship, which has postulated the fundamental difference between Greek and Roman culture since the nineteenth century, a view that is also held in this volume by Ann-Catherin Harders (with reference especially to the work of Ulrich Gotter), I assume the indistinguishable

152 Thomas Spa¨th

Cicero’s efforts on behalf of his son’s education confirms the first two aspects mentioned above: Cicero’s concern for his son’s education and for good tutoring, and his mutual concern for his son and nephew. Another aspect is also evident, namely Cicero’s paternal endeavour to oversee and ensure his son’s progress and studiousness.

One of the letters written to his brother, Quintus, establishes quite clearly that Cicero was actively involved in the education of both his son and nephew. Should Quintus raise no objections, Cicero writes in 54, he would tutor his nephew himself, since he had now gained quite some practice through teaching Marcus during his enforced political inactivity.23 Thereupon, his brother, Quintus Cicero, writes to his son, instructing him to now regard his uncle as his tutor.24 Cicero sees his tuition as meaningfully complementing Paeonius’ rhetoric lessons. He informs his brother, furthermore, that he will introduce Quintus to the declamation exercises himself.25

One important tutor for both Marcus and Quintus was Dionysius, a freedman of Atticus. In July 54, Cicero writes to Atticus, requesting his earliest possible visit so that Dionysus could teach both him and his son.26 He reiterates his demand a few months later.27 Three years later, Dionysius is in fact present in Cicero’s household, who com¬mends him in his letters to Atticus.28 Even though he remarks that the two boys complain about Dionysius’ fits of violent temper, Cicero defends their tutor to the utmost: never had a man been more learned, more virtuous, and more loving of Atticus and himself than Dionysius.29 It is thus quite surprising that two years later, in 49, Cicero describes this once highly reputable tutor as lacking the

Graeco-Roman blending of educational ideals among the Roman elite; see further Spa¨th 2007 (for a discussion of bilingual education, see 163–4 and the references to Dubuisson and Dupont/Vallette-Cagnac).

23 Cic. ad. Q. fr. 2.13.2 [SB 17]; see also Cic. ad Q. fr. 3.4.6 [SB 24]: Cicero takes Marcus to the Tusculanum, not for recreational but instead for educational purposes.

24 Cic. ad. Q. fr. 3.1.19 [SB 21].

25 Cic. ad Q. fr. 3.3.4 [SB 23]; for a general discussion of rhetorical education, see Rawson 2003: 147–53; for a detailed discussion of declamation, see Kaster 2001.

26 Cic. Att. 4.15.10 [SB 90], letter dated 27 July 54.

27 Cic. Att. 4.18.5 [SB 92], written between 24 Oct. and 2 Nov.

28 Cic. Att. 5.9.3 [SB 102].

29 Cic. Att. 6.1.12 [SB 115]: pueri autem aiunt eum furenter irasci; sed homo nec doctior nec sanctior fieri potest nec tui meique amantior.

Cicero, Tullia, and Marcus 153

gift of teaching. He would therefore rather teach his son and his nephew himself. In a letter to Atticus written two days later, he reports the dismissal of Dionysius.30

These references suggest that, as a father, Cicero was intensely concerned with his son’s progress and tutoring, just as he was as a patruus (a paternal uncle) with his nephew’s.31This concern involved strict supervision. Various letters concerning Marcus’ study visits to Athens in 45 and 44 attest to his father’s surveillance: Marcus, on the one hand, writes his father letters that demonstrate that his writing style and knowledge of literature were progresssing—in two letters to Atticus, Cicero praises his son’s letters, written ‘in a good archaic style indeed and pretty long’.32 Cicero, however, is not content to let the matter rest there. In several letters to Atticus written in the spring of 44, he expresses his intention to travel to Athens to observe his son’s progress for himself.33 This intention never materialized. Fail¬ing his own inspections, Cicero commissioned various tutors, in-cluding Leonidas and Herodes, to send regular progress reports.34 C. Trebonius, one of Cicero’s fellow senators, visits Marcus in Athens on his journey to assuming office as proconsul of the province of Asia. In a letter dated May 44, he congratulates Cicero on his

30 Att. 8.4.1–2 [SB 156], dated 22 Feb. 49; see, however, 8.5.1 [SB 157] (written on the same day) where Cicero mentions reconciliation with Dionysius and demands a sharp letter from him, addressed to Atticus, in return; Att. 8.10 [SB 159], dated

25 Feb. 49, reports the dismissal of Dionysius; in this letter, Cicero mentions that while he is reluctant to see him leave as the boys’ tutor, he is pleased to see the back of an homo ingratus, an ‘ungrateful fellow’.

31 On Cicero’s conduct as patruus, see Bettini 1986: 47–9.

32 Cic. Att. 14.7.2 [SB 361; my translation is slightly modified]: litterae sane nentvope vat et bene longae; see also Cic. Att. 15.16 [SB 391].

33 Att. 14.16.3 [SB 370], dated 2 May 44, and beforehand 14.13.4 [SB 367], dated

26 Apr. 44.

34 Cic. Att. 14.16.3–4 [SB 370]. Incidentally, Cicero quite as a matter of course also informs his brother, Q. Cicero, about his son’s progress during his absence from Rome—Quintus is obviously quite often at his uncle’s house, for instance in March 56 when Cicero reports his nephew’s good progress, since he was being taught by Theophrastus of Amisus (referred to as Tyrannio in the letter); see Cic. ad Q. fr. 2.4.2 [SB 8]. See further Cic. ad Q. fr. 3.1.14 [SB 21], where Quintus’ studiousness is praised; in the same letter, Cicero assures his brother that he forgives his ‘continual enquiries’ about young Quintus; the letter also requests his brother’s wife Pomponia to come to Arpinum, since Cicero would like to have Quintus’ company during the otium (3.1.7 [SB 21]).

154 Thomas Spa¨th

son: ‘I came to Athens [ ... ] and there saw what I most desired to see, your son devoting himself to liberal studies and bearing an extra¬ordinary reputation on account of his modesty.’35 He also mentions that following Marcus’ interest in becoming acquainted with Asia, he had invited Cicero’s son to visit him during his governorship. Trebo-nius assures Cicero that Marcus would be accompanied by Cratippus, his tutor, so that his education would suffer no interruption.36

Cicero was not afraid to impose decisions concerning tutors upon his son, even against his will. Thus Plutarch mentions that Cicero suspected one tutor, Gorgias, of inciting Marcus to indulge in mer¬rymaking and excessive drinking. He had therefore forbidden his son from attending any more lessons with Gorgias.37 With Cicero’s Letters to Friends, a parallel body of correspondence has survived, allowing us to verify Plutarch’s statement. In a letter to Tiro, Cicero’s secretary, dating from the summer of 44, Marcus mentions not his father’s reasons but indeed his directive, observing that his ‘kindest and dearest father’ (humanissimus et carissimus pater)38 had imposed upon him the dismissal of Gorgias, his teacher of rhetoric. While Marcus found the latter’s lessons useful, he realizes that he would have been ‘taking a lot upon myself in judging my father’s judge-ment’ (grave esse me de iudicio patris iudicare).39

These references to Cicero’s paternal concern for his son’s studies suggest a notion of education that differs markedly from modern educational goals: in ancient Rome, sons were not meant to develop their individual abilities and interests, but instead lessons were aimed

35 Cic. fam. 12.16.1 [SB 328]: Athenas veni [ ... ] atque ibi, quod maxime optabam, vidi filium tuum deditum optimis studiis summaque modestiae fama (I adapt Shackle-ton Bailey’s translation in order to retain the literal sense of the Latin modestia).

36 Cic. fam. 12.16.2 [SB 328].

37 Plut. Cic. 24.8.

38 Cic. fam. 16.21.1 [SB 337]; Shackleton Bailey dates this letter ‘August (?) 44’, Kasten (1997) suspects a date ‘between the end of July and the end of October 44’. In addition, we can refer to Marcus’ remark in fam. 14.25 [SB 338], where he writes in a further letter to Tiro: de sua in me voluntate semper ad me perscribit pater (‘my father constantly writes to me about his kind feelings towards me’—on the phrase voluntas in aliquem, see Shackleton Bailey’s commentary on this passage); while scholars have often referred to Cicero’s apparently less obvious sympathy for his son (see nn. 8–9 above), such passages advise caution.

39 Cic. fam. 16.21.6 [SB 337]. For an excellent analysis of this letter, see Jon Hall’s essay (2005, especially 260–3).

Cicero, Tullia, and Marcus 155

at imparting skills designed to enable male children to further pursue the social and political prestige established by their fathers. The purpose of paternal control was to maintain a certain social standing for male offspring, and to thus safeguard the family name. Cicero’s funding of Marcus’ education plainly reveals this underlying inten¬tion: his repeated enquiries to Atticus about the safe receipt of monies,40 and his frequent requests that Marcus be well endowed indicate his concern about his son possessing sufficient freely dis¬posable assets to afford a lifestyle commensurate with his status. Marcus must be amply provided for (honestissime copiosissimeque), not simply as a matter of paternal duty but also as a matter of safeguarding his father’s social standing and dignity.41 Cicero com¬pares Marcus to the sons of other senators, and observes that his expenditure should not be lower than that of Bibulus, Acidinus, or Messalla, who were all staying in Athens at the same time.42 Conse¬quently, Marcus’ lifestyle in Athens must at the very least match but preferably surpass that of his peers. He thus becomes his father’s alter ego, whose political and social standing he must display outwardly. What Cicero has achieved for his domus, Marcus must show before the Greeks and Romans in Athens.

Various other aspects of the education of Marcus and his cousin Quintus illustrate how the Tullii Cicerones sought to establish a joint family tradition. Cicero’s comments to his brother, Quintus, about Paeonius’ rhetoric lessons are remarkable in this respect. He informs his brother that he will teach his son and his nephew additional lessons.43 He asserts that his own rhetorical training is more learned and more abstract than Paeonius’. He therefore intends to introduce their sons to a declamatory technique that both fathers ‘have been

40 See, for instance, Cic. Att. 12.24.1 [SB 263], 12.32.2 [SB 271].

41 Cic. Att. 14.7.2 [SB 361]: id cum ad officium nostrum pertinet tum ad existima-tionem et dignitatem (‘That is for me a matter both of duty and of reputation and prestige’); see also 14.11.2 [SB 365].

42 Cic. Att. 12.32.2 [SB 271]; see further the comparison with the sons of other senators in Cic. Att. 12.7.1 [SB 244]. Here, Dixon (1984: 94) also acknowledges: ‘It was partly affection which prompted the wish, but also a matter of Cicero’s own standing.’

43 See above, p. 152 and n. 25.

156 Thomas Spa¨th

through [ ... ] ourselves’.44 In so doing, he resorts to the fathers’ own youth. Is the joint education of Marcus and Quintus not aimed precisely at establishing a tradition, or indeed at continuing a tradi¬tion based on the bond between their fraternal fathers?45 On account of its only recently achieved upward social mobility, and on the basis of unique historical sources,46 I would argue that the case of the Tullii Cicerones allows us to explore how family identity was established in Roman culture. Whereas established aristocratic families could model themselves upon a more or less long line of ancestors selected on the basis of their success, thereby allowing descendants to ‘follow in their footsteps’, as Catherine Baroin’s contribution to the present volume suggests, or whose images function as continuous reminders or admonishment, as Ve´ronique Dasen and Ann-Cathrin Harders demonstrate, the Tullii Cicerones must first establish their connec¬tion with a glorious past. One constituent element of this endeavour is Cicero’s active involvement in the education of his son and ne¬phew. He thus assumes the task of both pater and patruus severus— even though he later reproaches himself for not having been strict enough with either boy. Marcus’ letter to Tiro allows us a glimpse of the Roman notion of severitas: even at the age of 20, Marcus would never dream of questioning his father’s judgement, at least not in a letter to his secretary which ran the risk of being seen by Cicero. Evidently, the relationship between father and son was such a funda¬mental part of Roman culture that a son’s obligation to exercise pietas and dutiful respect of his father’s will made open criticism incon-ceivable.47

44 Cic. ad Q. fr. 3.3.4 [SB 23].

45 See Bradley 1991d: 169 and nn. 56–8.

46 With regard to Andrew Lintott’s (2008) warning about placing too much faith in the accuracy of facts in Cicero’s speeches and letters (on the need to situate the texts in their pragmatic context of utterance, see Hall 2005), I would argue that the relevance of family tradition and identity concepts have nothing to do with Lintott’s concern for positivist facts but rather with conceptions and meanings; Cicero’s representation of matters is thus singly decisive in historical terms, irrespective of whether it coincides with extratextual reality. For a more detailed discussion of the relationship between text and reality, see Spa¨th 2006.

47 See also Scholz 2006: 128–36; for a discussion of the social and especially legal dimensions of the relationship between fathers and sons, see Thomas 1983.

Cicero, Tullia, and Marcus 157

Cicero’s endeavour to establish a family tradition along the lines described above becomes evident in areas other than schooling and education; these concern his daughter as much as his son.

2. CICERO’S SUPPORT FOR HIS CHILDREN’S

CAREERS IN POLITICS AND SOCIETY

In the aristocratic domus, a son’s cursus honorum corresponded to a daughter’s marital career.48 In a culture in which her father’s and her husband’s positions in society determined a woman’s social standing, marriage represented a crucial decision for daughters. In the 1970s and 1980s, women’s studies heavily criticized the fact that women served solely to secure relationships between men, and that their personal interests were ignored as a result.49 Does such an assessment of marriage as a means of abusing or exploiting daughters not amount to a rather simple projection of current notions onto an¬other, foreign society such as ancient Rome? Closer examination of Tullia’s marriages suggests that our modern concept hardly corre¬sponds to Tullia’s and Terentia’s perception. While a Roman father by all means ‘instrumentalized’ his daughter by marrying her off to political friends, and thereafter dissolved the marriage depending upon political and financial developments, to subsequently remarry her, he instrumentalized his son in exactly the same fashion. Instru-mentalization must here be conceived as a descriptive rather than as an evaluative term.

Tullia’s ‘Marital Career’

Tullia’s first engagement was to C. Calpurnius Piso Frugi in Decem¬ber 67. Cicero conveys this news to Atticus in a somewhat terse comment: Tulliolam C. Pisoni L.f. Frugi despondimus—‘We betrothed

48 See Clark 1991: 28 for a discussion of Tullia’s ‘marital career’.

49 See, for instance, Teresa Carp’s observation (1981: 352): ‘[ ... ] Cicero, no less than other Roman aristocrats, did not fail to exploit her [scil. his daughter’s] political value’.

158 Thomas Spa¨th

Tullia to C. Piso Frugi, son of Lucius.’ In his letters, Cicero routinely uses the diminutive Tulliola, which carries an affective connotation, to refer to his daughter. Tullia, however, was also ‘small’ with regard to her age: she was aged between as little as 9 and no more than 12, by no means an unusual age for sponsalia in the Roman elite.50 Her fiance´, aged about ten years older,51 was the son of a praetor and descended from a consular family with whom Cicero entertained friendly relations on a political level.52 There are no records of the exact date of the marriage. In 63, Cicero refers to Piso as gener in his fourth Catiline Oration; the designation could apply not only to an actual son-in-law, but in a broader sense to a man only engaged to be married to his daughter.53 Calpurnius Piso was appointed quaestor in 58, and died either while holding office or shortly thereafter. He had stood up for his father-in-law during Cicero’s exile.54

Tullia became a widow as early as 58 or 57, that is, when she was about 20. Her next marriage was instigated in 56 when she became engaged to Furius Crassipes, a rich patrician,55 who attained the quaestorship in 51. Cicero mentions the engagement in various

50 Cic. Att. 1.3.3 [SB 8]; I replace Shackleton Bailey’s impersonal translation (‘Tullia is engaged...’) with the exact pluralis maiestatis employed by Cicero in the original text. On the age of marriage, see, among others Hopkins 1965; Shaw 1987b; Lelis, Percy, and Verstraete 2003.

51 Presuming that he became quaestor in 58, at the earliest at the age of 30 (in accordance with Sulla’s lex Cornelia de magistratibus of 82), he would have been born in 88. See Treggiari 2007: 42.

52 The father of Tullia’s fiance´, L. Calpurnius Piso Frugi, was made tribune of the plebs in 90 and appointed praetor in 74, together with Verres; his grandfather, who also served as praetor, was killed in the Hispania ulterior in 112; his great-grand¬father, the annalist C. Calpurnius Piso Frugi, was appointed tribune of the plebs in 149, held the office of consul in 133, and probably became censor in 120. See Shackleton Bailey’s commentary on Att. 1.3.3 [SB 8].

53 See Beryl Rawson (2003: 247; nn. 105–6) and Patricia Clark (1991: 33–4); all that is certain is the terminus ante quem: as Rawson notes, the marriage must have occurred before 58. For a reference to Piso as gener, see Cic. Catil. 4.3. Treggiari 2007 (43, 47) suggests dating the marriage to the end of Cicero’s consulship or immediately thereafter in 62.

54 Cicero mentions that he had accompanied him to the unsuccessful discussion with L. Calpurnius Piso Caesoninus, who was consul in 58, and whose assistance he sought against Clodius, the tribune of the plebs; see also Cic. Pis. 12–3.

55 Clark (1991: 28 n. 2) refers to Treggiari 1984: 441. Treggiari 2007: 75 describes Furius Crassipes as ‘young, wealthy, a patrician and a prospective senator’, and who possessed ‘the great attraction of a house and park on the outskirts of Rome’.

Cicero, Tullia, and Marcus 159

letters written to his brother, Quintus, in April 56. One letter, for instance, refers to an engagement dinner hosted by Cicero to cele-brate the occasion.56 In a letter to Lentulus57 dating from July 56, he mentions the engagement and closes by thanking Atticus for extend¬ing his congratulations.58 Scholars have on the whole assumed that Tullia and Crassipes married shortly afterwards, and that they were divorced in 51. No direct records of either occasion have survived.59 In an essay published in 1991, Patricia A. Clark asks whether the marriage ever occurred.60 Her interesting reasoning contrasts with Susan Treggiari’s hardly disputable view that it is difficult to imagine a young widow remaining unmarried for six years, and that Cicero would have waited for five years to urgently pursue his objective to remarry his daughter shortly before leaving Rome to assume the governorship of the province of Cilicia.61

Cicero was indeed looking for a new husband for Tullia in 51, the year in which he left Rome to assume his duties as proconsul. Much has been written about Tullia’s third marriage. This episode offers a striking example of how in Roman culture the mater familias con¬ducted the affairs of the domus in the absence of the paterfamilias, and also took decisions independently of her husband.62 Not only did Terentia, Cicero’s wife, and Tullia, his daughter, choose the

56 The engagement is first mentioned in Cic. ad. Q. fr. 2.4.2 [SB 8] (mid-March 56), thereafter 2.6.1 [SB 10] (9 Apr. 56), which gives the date of the engagement as 4 Apr. and of the banquet as 6 Apr.

57 In Cic. fam. 1.7.11 [SB 18], Cicero thanks P. Cornelius Lentulus Spinther, consul in 57, for his congratulations on Tullia’s engagement.

58 Att. 4.4a.2 [SB 78]; see also Clark 1991: 31.

59 This might be connected to the fact that Cicero wrote only infrequently to Atticus between 56 and 55, and not at all from November 54 to May 51 when they were both in Rome. About 50 letters ad familiares have survived from the same period, but these are scarcely concerned with family matters. From Cic. Att. 4.4a.2 [SB 78], we can infer that in June 56 (Shackleton Bailey dates letter 78 to ‘circa 20 June (?) 56’), Tullia is still staying at Cicero’s country estate at Antium. In a letter to P. Cornelius Lentulus Spinther written in December 54 (fam. 1.9.20 [SB 20]), Crassipes is referred to as gener, which need not, however, as suggested above in n. 53, imply a formal marriage.

60 Clark 1991.

61 Treggiari 2007: 76.

62 See especially the letter to Appius Claudius Pulcher dating from either 3 or 4 Aug. 50 (Cic. fam. 3.12.2 [SB 75]): Quibus ego ita mandaram ut, cum tam longe afuturus essem, ad me ne referrent, agerent quod probassent (‘I had told them [i.e. my family] not to consult me since I should be so far away, but to act as they thought best.’)

160 Thomas Spa¨th

latter’s new husband during his absence but they also proceeded with the engagement and shortly thereafter with the marriage between May and early June 50.63 This placed Cicero in a delicate political situation: shortly before the marriage, his future son-in-law, Dola-bella, had accused Appius Claudius Pulcher, Cicero’s predecessor as proconsul of Cilicia, of a breach of official duties, thereby pre¬venting his triumph. Shortly thereafter, Dolabella instigated legal proceedings for electoral bribery.64 Immediately upon his appointment as Appius’ successor (in the first months of 51), however, Cicero sought to establish amicable relations with Appius Claudius Pulcher.65 In a letter to this influential politician,66

63 In a letter written at the beginning of June 50, Caelius Rufus congratulates Cicero on the marriage alliance with Dolabella (Cic. fam. 8.13.1 [SB 94]); for a discussion of the possible assumptions about the date of the marriage, for which no historical evidence exists, see Treggiari 2007: 97 f. At the time of the marriage, Tullia was between 26 and 29 years old; while Dolabella’s age has remained uncertain, there is good reason to believe that he was born around 74 (see Shackleton Bailey 1965–70, vol. 3, 269; Treggiari 2007: 93), thus making him two to five years younger. The marriage lasted almost four years. Various letters contain references to the fact that payment of the dos in three instalments in 49, 48, and 47 meant a considerable financial burden (see Ioannatou 2006: 225, 426–9).

64 See Cicero’s letter to Appius Pulcher, Cic. fam. 3.10.1 [SB 73]; in the same letter, dating from the the first half of April 50, Cicero also mentions that he had defended Dolabella twice against a capital charge (capitis iudicia), and therefore considered Dolabella’s action against Appius Pulcher as a breach of the obligation to friends (fam. 3.10.5 [SB 73]). Moreover, he assures Appius Pulcher that Dolabella’s ‘silly, childish talk’ (sermo stultus et puerilis)—that is, his allusions to a prospective mar¬riage with Tullia—should under no circumstances be taken seriously. This clearly suggests that in April 50 Cicero was utterly unaware of the marriage negotiations between Tullia and Dolabella, or that he at least followed Caelius Rufus’ advice to postpone such a possibility (see Caelius Rufus’ letter to Cicero, fam. 8.6.1–2 [SB 88], written in February 50—which Cicero probably received in April). See, moreover, Cicero’s direct congratulations on Appius Pulcher’s acquittal from both the charge de maiestate (fam. 3.11.1–2 [SB 74], 26 June (?) 50) and de ambitu (fam. 3.12.1 [SB 75], 3 or 4 Aug. 50).

65 See his letters to Appius Pulcher, Cic. fam. 3.2–13 [SB 65–76], written between February/March 51 and August 50. The letters written between February and August 51 reveal that before leaving for the province he did his utmost to arrange a meeting with Appius Pulcher, who was on the return journey; the meeting, however, never took place (which Cicero complains about politely but none the less assertively; see fam. 3.6.3–4 [SB 69]).

66 One indication of his eminent network of friends is that Appius Pulcher’s daughters were married to the eldest son of Pompeius and to M. Junius Brutus, the son of Servilia and M. Junius Brutus (tr. pl. in the year 83).

Cicero, Tullia, and Marcus 161

written in August 50, Cicero emphasizes his embarrassment about the ‘arrangement made by my family without my knowledge’ (ea quae me insciente facta sunt a mei),67 meaning Tullia’s marriage with Dolabella. Cicero thanks Appius Pulcher for conveying his tactful good wishes for the couple’s marriage fortunes.

Even though the reasons prompting Terentia and Tullia to choose Dolabella can be discerned no more than hypothetically through a web of conjectures, closer scrutiny of this affair is worthwhile—since it allows us to at least make substantiated assumptions about what determined the choice of husband from a female viewpoint, and about the scope of action available to female Roman aristocrats. Based on a reading of three letters written by Cicero to Atticus,68 John H. Collins suggested some time ago that there were originally three marriage candidates.69 From the letter that Cicero wrote to Atticus on 12 May 51 from Beneventum,70 Collins deduces the identity of an unnamed candidate, whom he calls ‘B’; Cicero rules out this candidate because he would not be acceptable to Tullia. Marriage negotiations with the second candidate (‘C’), whom Cicero refers to as ‘Servius’, could be conducted, as he writes, through Servilia, the mother of Brutus, acting as an intermediary. This can¬didate can be identified as Servius Sulpicius Rufus, the son of Pos-tumia and Servius Sulpicius Rufus, a jurist, who stood for election in 52 to become consul in 51. The third candidate (‘A’), finally, remains

67 Cic. fam. 3.12.2 [SB 75]; see also the reiteration two sentences later: in quo unum non vereor, ne tu parum perspicias ea quae gesta sint ab aliis esse gesta, ‘On one point, though, my mind is easy—you will not fail to realize that what has been done has been done by others’; this follows on from the apology cited above (see n. 62) concerning the assignment of decision-making powers to his family during his absence—these passages clearly show how Cicero is desperate to maintain good political relations and pulls out all the stops of epistolary-rhetorical courtesy.

68 Cic. Att. 5.4.1 [SB 97] (Beneventum, 12 May 51), 5.21.14 [SB 114] (Laodicea, 13 Feb. 50), and 6.1.10 [SB 115] (Laodicea, 20 Feb. 50). Cicero also admonishes Atticus—without being more specific—in Att. 5.13.3 [SB 106], 5.14.3 [SB 107], and 5.17.4 [SB 110] (written between July and August 51) to attend to what he considers an important ‘domestic affair’ (domesticus scrupulus, vSoµvXov, mea domestica).

69 Collins 1951: 164.

70 Cic. Att. 5.4.1 [SB 97].

162 Thomas Spa¨th

unnamed, but Cicero mentions that he comes on the recommenda-tion of a woman known as Pontidia.71 Writing to Atticus on 13 February 50 from Laodicea in the province of Cilicia, Cicero ap-proves of the advice given by Atticus in an earlier letter which has not survived: Atticus had evidently recommended Sulpicius Rufus (can¬didate ‘C’), Postumia’s son, ‘since Pontidia is trifling’ (quoniam Pontidia nugatur), thus ruling out candidate ‘A’.72 One week later, however, in a letter to Atticus written on 20 February, he returns to the matter on the basis of newly received letters (from either Terentia or Atticus, or indeed from both). He writes, ‘I much prefer Pontidia’s candidate [‘A’] to Servilia’s [‘C’]’, citing one of Atticus’ previous letters: ‘you had written to me “but I wish you had gone back to your old gang”’.73 Thus Cicero had made up his mind, and a letter containing a corresponding instruction was duly dispatched to Ter-entia and Tullia. Given six to eight weeks conveyance time,74 the letter would have reached his wife and daughter in mid-April. Who was ‘Pontidia’s candidate’, however, who would return Cicero to his ‘old gang’—whether socially in terms of his knightly status, or indeed in terms of his Arpinate origins.75

Based on Cic. Brut. 70.246, which mentions a Marcus Pontidius as municeps noster, Collins suggests that candidate ‘A’, who meets with the unanimous approval of Atticus and Cicero, must have belonged to a gens Pontidia and would have stemmed from an Arpinate family of equestrian rank. He argues that it ‘seems clear that Pontidia had proposed some good, solid eques, perhaps an Arpinate, but certainly

71 She is mentioned in Cic. Att. 5.21.14 [SB 114] and 6.1.10 [SB 115].

72 Cic. Att. 5.21.14 [SB 114].

73 Cic. Att. 6.1.10 [SB 115]: De Tullia mea tibi adsentior scripsique ad eam et ad Terentiam mihi placere; tu enim ad me iam ante scripseras ‘ac vellem te in tuum veterem gregem rettulisses’ [ ... ] multo enim malo hunc a Pontidia quam illum a Servilia: ‘I agree with what you say about my Tullia, and have written to her and to Terentia to say that I approve; you had already written to me “but I wish you had gone back to your old gang” [ ... ] for I much prefer Pontidia’s candidate to Servilia’s.’

74 Collins 1951: 167.

75 With regard to vetus grex, Shackleton Bailey observes: ‘the implication here is doubtless partly social’ (Shackleton-Bailey 1965–70: vol. 3, 244).

Cicero, Tullia, and Marcus 163

not a member of a patrician family active in politics’.76 Consequently, in December 51 or at the beginning of January 50, Terentia and Tullia found themselves in the following situation: the negotiations with Pontidia, whose respectability had at first been doubted, had resulted in a tangible outcome, since Atticus supported Pontidia’s candidate; Tullia and Terentia obviously knew, moreover, that Atticus would write to Cicero along these lines. Apparently, however, rumours about a possible marriage between Dolabella und Cicero’s daughter77 were afloat in Rome as early as February 50, suggesting that discus¬sions between Tullia, Terentia, and Dolabella had already taken place—despite, or as I would suggest, because Atticus, Cicero’s ad¬viser, had begun to express his preference for the Arpinate eques. As mentioned above, Cicero’s wife and daughter did not receive news of his explicit approval for Pontidia’s marriage candidate until April. Cicero, moreover, changed his mind again a few weeks later—pre-sumably in April 5078—after Tiberius Claudius Nero had conducted negotiations with him about the marriage in the province. He now dispatched ‘reliable persons’ to Rome—but these evidently only arrived after Tullia’s engagement to Dolabella. In a letter to Atticus written in early August, Cicero observes: ‘Here am I in my province paying Appius all manner of compliments, when out of the blue I find his prosecutor becoming my son-in-law!’79

The circumstances surrounding Tullia’s third marriage point to various interesting aspects: first, women quite evidently performed a decisive role in proposing possible marriage candidates; negotiations

76 Collins 1951: 166. Shackleton-Bailey (1965–70, vol. 3, 195) thus speaks of ‘a mere eques’, and Susan Treggiari (2007: 87 f.) suspects that the man in question could have been either a young eques at the beginning of his political career, or that fathers in Cicero’s situation ‘might lower their sights to a husband not active in public service’—for instance opting for a ‘cultured, wealthy, well-born eques, rather like Atticus himself’.

77 See Caelius Rufus’ letter to Cicero written in February 50, fam. 8.6.1–2 [SB 88].

78 Shackleton Bailey assigns ‘April (?) 50’ as a date to Cicero’s letter of recommen¬dation concerning Tiberius Claudius Nero, Cic. fam. 13.64 [SB 138], addressed to Minucius Thermus, propraetor of the province of Asia (Shackleton Bailey amends the traditional name of the addressee ‘Publius Silius’ thus).

79 Cic. Att. 6.6.1 [SB 121], dated 3 Aug. 50: ego dum in provincia omnibus rebus Appium orno, subito sum factus accusatoris eius socer; on the charges brought by Dolabella against Appius Claudius Pulcher, see n. 64 above.

164 Thomas Spa¨th

were conducted with Servilia and Pontidia, and the correspondence between Cicero and Atticus specifies marriage candidates under reference to their female brokers. Such arrangements constitute a remarkable form of identifying suitable men in a society in which individuals were routinely named after their father and grandfather. Secondly, we can infer from Cicero’s first letter to Atticus, in which the subject is broached, that a daughter’s views were much heeded: Cicero, as observed, rules out candidate ‘B’ because he doubts whether Tullia ‘could be brought to consent’.80 Thirdly, the episode provides a concrete example of how marriages served to establish and cultivate political and amicable relations (which amounts to the same) in the Roman elite. Tullia’s first two husbands, whom Cicero helped choose, belonged to a highly prestigious and considerably affluent domus. The engagements and marriages enable Cicero to secure amicable relations. It is thus neither accidental that Calpur-nius Piso committed himself to Cicero’s return from exile,81 nor that Crassipes visited Cicero at his country estate in 49 to convey news of the situation in Pompeius’ camp.82

While securing advantages for himself, these marriages also en¬abled Cicero to establish circumstances for Tullia commensurate with her social standing.83 In this respect, the episode surrounding the third marriage reveals another important aspect: Tullia and her mother quite obviously proceeded in full compliance with the cri¬teria applicable to marriages among the Roman elite, aimed at enhancing the social prestige of one’s own family—and, by further implication, of one’s daughter. Not only did they take decisions in Cicero’s absence but they in effect took advantage of it to avert a mistake on his part. If Cicero and Atticus were actually seriously considering someone born into an Arpinate equestrian family as

80 Cic. Att. 5.4.1 [SB 97]: vereor adduci ut nostra possit. We must consider, however, that Tullia was between 25 and 28 years old in 51; she obviously had less influence on her first engagement when she was aged no more than between 9 and 12.

81 See above, p. 158 and n. 54.

82 Cic. Att. 9.11.3 [SB 178].

83 See Servius Sulpicius Rufus’ reference to Tullia’s husband as belonging to ‘young men of distinction’ in his letter to Cicero, Cic. fam. 4.5.5 [SB 248].

Cicero, Tullia, and Marcus 165

Tullia’s third husband, this would have entailed her clear social demise. Just like her father, this marriage would have returned Tullia to what Atticus calls the ‘old gang’ (vetus grex), which Cicero’s brilliant career had catapulted him out of, and which her first two marriages had elevated her above. I would argue, possibly somewhat daringly, that faced with the impending decision for wedlock with a man from a rural equestrian family, mother and daughter forged ahead with negotiations with Dolabella. Choosing him contributed decisively to safeguarding Tullia’s social prestige but also brought about a better decision for the domus Tullia than the pater familias would have reached himself. The episode demonstrates that Terentia and Tullia made an essential contribution to protecting and continu¬ing the family tradition. Given that Dolabella held the promise of good connections to Caesar’s camp, we can perhaps even surmise that Cicero’s wife and daughter made a very conscious political choice based on their assessment of the circumstances prevailing in Rome at the time.84

Along these lines, the episode discussed above exemplifies my introductory remarks about so-called ‘instrumentalization’.85 It is by no means impossible that women were themselves actively in-volved in employing marriage and matrimony to serve political ends. Given the opportunity, for instance in the absence of the pater familias, as shown, they took matters into their own hands or brokered engagements. Hence, we can justifiably speak of a female marriage career, one which by all means compares with the male political career in terms of social standing.

84 By no means is this the only passage revealing the independent political deliberations of the women in Cicero’s domus: see, for instance, the decision at the beginning of 49 about whether to remain in Rome or to flee the city in Cic. Att. 7.14.3 [SB 137], 7.16.3 [SB 140], 7.17.5 [SB 141], or Cicero’s complaint that while he had been mindful of his family, they were now reproaching him for his indecisive wavering between Caesar and the Pompeians, Cic. Att. 9.6.4 [SB 172]. As mentioned above (n. 19), in Cic. Att. 10.8.1 [SB 199], Cicero alludes to letters written by Tullia in which she advises her father on how to assess the political situation in the spring of 49.

85 See above, p. 157.

166 Thomas Spa¨th

Preparing Marcus’ Career

Cicero’s efforts to advance his son’s career are at first the subject of great concern in his letters. In the letters to his wife and family, and to Atticus and his brother, during his exile in 58, Cicero incessantly reproaches himself for what he has brought upon his family, and in particular upon his son.86 In November 58, he tells Atticus how unfortunate his son is for having a father who has passed on nothing other than resentment and an ignominious name: invidia and igno-minia nominis mei.87 Cicero’s complaint reveals how much it mat¬tered to him as a father to hand down a good name to his son.88 Almost ten years later, this concern has dissipated and self-assured¬ness has returned. In a letter to M. Caelius, Cicero writes that his son Marcus will be ideally placed should the res publica somehow survive the Civil War, for his name would ensure a ‘grand heritage’, amplum patrimonium.89

Cicero’s governorship of Cilicia clearly illustrates how he paves the way for Marcus’ future career. He takes his son and his nephew, Quintus, along on his journey east in early summer 51. Their slow progress and numerous stopovers en route to the province, among others in Athens, Delos, and Ephesus, undoubtedly served to intro¬duce the two boys to the Greek Orient.90 During the military cam¬paign in the summer and autumn of the same year, Cicero entrusts the boys to the younger Deiotaros, the son of the King of the Galatians, upon whom the Senate had already conferred the royal title during his father’s lifetime.91 Deiotaros returned Marcus and Quintus to Cicero in Laodicea in February 50. Cicero confers upon

86 See the letters written between April and November 58, for instance Cic. fam. 14.1 [SB 8], 14.2 [SB 7], 14.3 [SB 9]; Cic. Att. 3.19 [SB 64], 3.23 [SB 68], ad Q. fr. 1.3 [SB 3]; in Cic. fam. 14.3.1, he laments the acerbissimos dolores miseriasque (‘bitter sorrow and suffering’) that his guilt has brought upon his son.

87 Cic. Att. 3.23.5 [SB 68] (29 Nov. 58).

88 For a detailed investigation, see Catherine Baroin’s Chapter 1 in this volume.

89 Cic. fam. 2.16.5 [SB 154] (2 or 3 May 49).

90 See Stinchcomb 1932/3: 443, who has gone to great lengths to collate the passages in the letters concerning the course of the journey.

91 Cic. Att. 5.17.3 [SB 110], 5.18.4 [SB 111], 5.20.9 [SB 113].

Cicero, Tullia, and Marcus 167

Quintus the toga pura92 while Marcus receives the adult’s toga a year later at Arpinum.93 Cicero and the two boys thereafter travelled to Rhodes where Poseidonios and Molon had taught him philosophy and rhetoric twenty-five years before.94 From Rhodes they continued their return journey to Athens and then to Rome.95

Following the outbreak of the Civil War, Cicero is undecided whether to take the boys along to Greece or, should he hold out in Italy, to send them there on their own.96 He also considers fleeing to Malta, but does not wish to create the impression that he has no stomach for danger. Cicero commends Marcus, whose chief concern is his father’s dignity, for being more courageous than himself.97 In 49, Cicero finally decided to join Pompeius in Greece, accompanied by his son, his brother, Quintus Cicero, and his nephew. Soon after¬wards, in 47, he returns to Brundisium, and contemplates dispatch¬ing Marcus as an envoy to Caesar.98 A year later, Marcus himself considers joining Caesar in Spain, thus following his uncle and cousin who had already broken away from Cicero following the defeat of the Pompeians at Pharsalos in August 48 to join Caesar’s forces. Cicero, however, calls this decision into question: Atticus had

92 Cic. Att. 5.20.9 [SB 113], 6.1.12 [SB 115].

93 In Cic. Att. 9.6.1 [SB 172], Cicero also contemplates whether Caesar might take offence if he did not perform the ritual in Rome; on conferring upon Marcus the toga virilis, see also 9.17.1 [SB 186], 9.19.1 [SB 189]; the letters were written between 11 Mar. and 2 Apr. 49, thus before and after the traditional dates of the ritual, the festival of Liber and Libera on 17 Mar. On the ritual, see Dolansky 2008 (including further references in n. 2, 59 f.)

94 See Plut. Cic. 4.5–7.

95 Cic. Att. 6.7.2 [SB 120], fam. 14.5.1 [SB 119].

96 See Cicero’s deliberations in the letters written to Atticus in January and February 49: Att. 7.13.3 [SB 136], 7.17.1, 4 [SB 141], 8.2.4 [SB 152], 8.3.5 [SB 153].

97 Cic. Att. 10.9.1–2 [SB 200].

98 See the two letters written to Terentia on 14 and 19 June 47, fam. 14.11 [SB 166] und 14.15 [SB 167]. While Cicero does not ask for his wife’s opinion, he promises to inform her should he send their son to Caesar. In the second letter, he informs her that he has decided against sending him to Caesar. Written shortly before his divorce from Terentia, scholars have often interpreted these terse letters as a clear indication of the alienation of affection and breakdown of relations between Cicero and his wife (see, for instance, Claassen 1996: 217, or the unacceptable simplification in Dixon 1984: 88). Notwithstanding these circumstances, Cicero none the less keeps his wife abreast of the questions and plans he is turning over in his mind.

168 Thomas Spa¨th

evidently raised the subject,99 and Cicero recalls his ‘very open’ (liberalissime) discussion with his son, in which he held out the prospect of financial assistance compatible with the funds available to the sons of other dignified aristocrats. Cicero, however, bids his son remember that his father had faced a volley of reproaches for leaving the optimates to return to Italy; Marcus’ prospective journey to Spain, moreover, would be interpreted as Cicero himself switching allegiances. This is a further indication of how self-evidently father and son were considered one and the same. Besides, Cicero warns Marcus that he could feel excluded once he realized how much more popular and well-connected his cousin Quintus, who had been stay¬ing with Caesar in Spain for quite some time, was. The letters reveal no more about the further deliberations between Marcus, Cicero, and Atticus.100 Eventually, however, Marcus travelled to Athens in 45 rather than to Spain.101

The available evidence shows how Marcus was groomed for aris¬tocratic duties through accompanying his father, thus providing him with first-hand experience, through acquiring the necessary intellec¬tual skills like the art of conversation and rhetoric, and especially through his introduction to vital social networks. The correspon¬dence, however, provides only one example of Cicero actively boost¬ing his son’s political career, namely when he has Marcus, his nephew

99 The wording in Cic. Att. 12.7.1 [SB 244]—[ ... ] de Cicerone, cuius quidem cogitationis initium tu mihi attulisti (‘[ ... ] about Marcus, you started me thinking about this’)—leads one to suspect that Marcus had first conferred with Atticus before disclosing his intentions to his father. Does this perhaps point to the downside of Marcus’ above-mentioned respect for his father, namely that he does not in the first instance seek to conduct an open discussion with his father?

100 It is difficult to interpret the following remark in Cic. Att. 12.8.1 [SB 245]: de Cicerone multis res placet (‘About Marcus, many people approve the plan’). The res has been seen mostly as a reference to the decision to travel to Athens; as Shackleton Bailey has quite rightly suggested, however, it could also refer to joining Caesar in Spain. Presumably, this letter was written towards the end of the second intercalary month (that is, the month of November according to the actual calendar) in 46 while Att. 12.7.1 [SB 244], mentioned in the previous note, was written at the beginning of this intercalary month, so that the res could very well refer to the preceding letter.

101 See pp. 153–5 above.

Cicero, Tullia, and Marcus 169

Quintus, and his friend M. Caesius elected aediles in Arpinum to manage municipal finances.102

Cicero died too soon to properly lead his son, aged only 22 at his father’s death at 43, to the cursus honorum. The above elements suggest, however, that Cicero actively laid the foundations for Mar¬cus’ political career in a manner characteristic of the Roman aris¬tocracy, specifically through involving his son in his own activities. The above division between ‘Schooling and Education’, on the one hand, and ‘A Career in Politics’ on the other, would therefore have made little sense from a Roman perspective. Other than grammar and rhetoric, a son’s education would have included practical in¬struction alongside his father: male children were assigned duties and tasks serving their father’s ends. Such instrumentalizing, however, enabled sons to obtain the necessary career qualifications. This per¬mitted Marcus, the sole male family member surviving proscription, to become pontifex, without his father’s support but due to Octa-vian’s patronage; ultimately, he even became suffect consul in 30.103

Comparing Cicero’s advancement of his children’s respective social careers shows that Tullia and Marcus received equal attention. The gender-specific difference in his attention lies simply in the distinc¬tion between socially determined female and male careers. The chil-dren’s welfare and social standing, on the one hand, and their father’s prestige and position, on the other, are all inextricably bound up in his paternal concern. ‘Instrumentalizing’ one’s children thus befitted a Roman citizen and a paterfamilias who instrumentalized himself to safeguard for the longer term his family’s social esteem and name. Both Tullia and Terentia’s decision concerning her marriage and Marcus’ pious obedience demonstrate that Cicero’s children and wife adopted and internalized Cicero’s conception of self. Far from being mere ‘victims’ or ‘objects’ of his paternalistic decisions, they

102 In a letter to M. Brutus, who is based in Gaul at the time, Cicero requests him to assist the Arpinate delegation of knights to collect tributes so that municipal finances can be rehabilitated; he refers to the office held by his son, nephew, and friend. See Cic. fam. 13.11.3 [SB 278], presumably written in the first half of 46.

103 The letters make no reference to a marriage involving Marcus—which would have also been part of his political career—with the exception of an allusion to a marriage offer in connection with a discussion about financial matters in a letter to Atticus written on 8 July 44 (Att. 16.1.5 [SB 409]).

170 Thomas Spa¨th

instead contributed as active subjects to attaining social prestige, if necessary redressing his indecision; they thus became an integral part of family tradition.

3. EMOTIONAL AFFECTION OR FAMILY TRADITION?

Examining Tullia and Marcus’ education and how Cicero sought to advance their careers (and the expenses he incurred in pursuing both objectives) has revealed his very rational motives, aimed at construct¬ing and maintaining family tradition in social and political respects. Obviously, such motives do not exclude emotional affection. Closer scrutiny of Cicero’s expressions of love and affection,104 however, reminds us to exercise caution about all too readily assuming ‘pa¬ternal love’ or ‘unbounded affection’ on his part. Contrary to claims for an ahistorical, universal ‘paternal love,’ Cicero’s letters reveal a specifically Roman paternal love, situated within a concrete historical context in which family tradition is a decisive element.

Cicero, as observed, conveys the prestige of the domus onto his children through advancing their respective careers, thus establishing a family tradition in the first place; doing so obviously presupposed his own ascendency, through which he acquired the necessary social and political capital for himself and for the Tullii Cicerones. The love and affection that he expresses for his children fits into this context. Cicero protests his love through recourse to prevailing norms of expressing esteem and appreciation. Rather than valuing particu¬lar individual attributes, Cicero commends Tullia, ‘the most loving, modest, and clever daughter a man ever had’, for her pietas, modestia, ingenium.105 In a letter to Terentia written in 47 while anxiously awaiting Caesar’s benevolence at Brundisium, he praises

104 On the emotional connotations of the vocabulary used with regard to Tullia, see Treggiari 2007: 161 f. and Ermete 2003: 232 n. 1311.

105 Cic. ad Q. fr. 1.3.3 [SB 3]; Cicero wrote this letter to his brother from exile, on 13 June 58, and it becomes generally apparent that his exuberant expressions of love and affection occur especially in letters written in extreme circumstances.

Cicero, Tullia, and Marcus 171

his daughter’s virtus,106 humanitas, and dignitas,107 for she is ‘so wonderfully brave and kind’. Likewise, Marcus must embody his father’s dignitas, and Cicero praises his fortitudo and modestia, ‘cour¬age and modesty’.108 Apparently, Cicero’s love for his children focuses not on specific behaviour or abilities but instead on their highest possible conformity with those aristocratic values that he expects every vir bonus to possess, and for which he commends both his children and his friends.

Notwithstanding that his esteem for his children has nothing to do with their individual personality but instead with conforming to overarching social norms and values, we need not refute claims to Cicero’s ‘paternal love’. His esteem, however, allows us to recognize the historical and cultural contingency of this particular ‘love’. Cicero expresses his affection for his children because this corresponds to the requirements and moral rules that he expects from the boni and that serve as the yardstick with which he measures his own life. For it is precisely this yardstick that determines whether a name will find the approval of the senatorial aristocracy in Rome or not.109 In Cicero’s case, the essential condition of Roman ‘paternal love’ is the successful adoption of a family tradition and its continuation. Cicero loves his children as images of the Tullii Cicerones family. Tullia and Marcus follow in the footsteps of their father, each according to their gender, in what amounts to the transmission of tradition through

106 In this context, virtus designates ‘virtue’ or ‘reaching one’s highest human potential’ (in the sense of Greek arete¯). Even though it is undisputed that the etymology of the Roman notion of virtue refers to male virtue in the first instance (and thus also to Greek andreia), I would argue that attempts to reduce the re-emergence of the word in texts written in Latin during the Republican period to an exclusive ‘manliness’—see, for instance, McDonnell 2006, who postulates a semantic field with exclusively military connotations—fail to capture actual usage. While virtus is attributed much more frequently to male figures, assigning this quality to Tullia is by no means an isolated case.

107 Cic. fam. 14.11 [SB 166] (14 June 47).

108 See the above-cited references to dignitas (Cic. Att. 14.7.2 [SB 361], 14.11.2 [SB 365]), to fortitudo (Cic. Att. 10.9.1–2 [SB 200]), and to the son’s modestia mentioned in Trebonius’ letter (Cic. fam. 12.16.1 [SB 328]).

109 Cicero refers to those values which conventionally decide on social ascendency; during his lifetime and amid the turmoil of several civil wars, however, these values were severely called into question.

172 Thomas Spa¨th

their metonymization: each part of the family refers to the family as a whole.

In a letter to his brother written from exile in June 58, Cicero observes that Tullia is ‘the image of my face and speech and mind’,110 thus lavishing the highest praise on her. This might also account for the fanum111 that Cicero intended to erect following his daughter’s death. Cicero’s unbounded grief and his obsession with such a ‘monument’ or ‘shrine’ puzzled his contemporaries—Atticus showed little enthusiasm for the idea—as much as it has continued to mystify scholarship.112 If we, however, assume paternal love along specifically Roman historical and social lines, that is, as love for one’s own image, one’s own name and its prestige, Cicero mourning his daughter’s death could be conceived as the loss of part of his own identity. Identity was considered to be neither autonomous nor individual in Roman culture, but instead it was strictly aligned with generational lines of transmission. Cicero’s endeavour to erect a fanum as a means of solace thus amounts not merely to an emotionally upset father’s overstrung reaction to his daughter’s death. Erecting such an edifice would have served not only to render eternal homage to his virtuous daughter but also to immortalize the Tullii Cicerones.113 If ever erected, it would have acted as a substitute for the loss of Tullia as a bearer of family tradition.

110 Cic. ad Q. fr. 1.3.3 [SB 3]: effigiem oris sermonis animi mei; thereafter follows the remark addressed to his brother tuum filium imaginem tuam, ‘Likewise your son, your image [ ... ]’. Catherine Baroin’s chapter in this collection examines in depth the physical and moral similarities between father and son, which also serve to recall the father; see especially Sect. 8, pp. 37–47, of her chapter. Baroin’s reflections on the similarity not only between sons and fathers but also between daughters and fathers (see pp. 41–2) place Cicero’s comments on Tullia in their social context.

111 See Treggiari 1998: 16 and n. 58 on the relevant passages.

112 See, for instance, Pierre Boyance´’s critical reflections (1944) on the possible philosophical backgrounds of the construction of the fanum. Boyance´ takes issue with what he considers to be Pierre Grimal’s untenable claims about the neo-Pythagorean foundations and mystical religious beliefs involved. Treggiari 1998: 14–23 analyses the various phases of Cicero mourning Tullia’s death based on a (modern) three-phase psychological model of grief, and thus postulates an ahisto-rical-universal emotionality of human ‘nature’: she suggests that Cicero’s ‘bitter grief [ ... ] was entirely natural’ (16, my emphasis).

113 The significance of the monument built to commemorate Tullia would be comparable to the inscriptions for deliciae, which Christian Laes refers to as ‘monu-ments of self-representation in a status-conscious society’ in the conclusions to his contribution to this volume.

References

AASGAARD, R. (2006), ‘Children in Antiquity and Early Christianity: Research History and Central Issues’, Familia. Revista de ciencias y orientacio´n familiar, 33: 23–46.

ADRIANI, A. (1970), ‘Ritratti dell’egitto greco-romano’, Mitteilungen des

Deutschen Archa¨ologischen Instituts, Ro¨mische Abteilung, 77: 72–109. ALDHOUSE GREEN, M. J. (1999), Pilgrims in Stone: Stone Images from the

Gallo-Roman Sanctuary of Fontes Sequanae, Oxford.

ALFONSO, G., and BLAIZOT, F. (eds.) (2004), La villa gallo-romaine de Champ Madame a` Beaumont (Puy-de-Doˆme): Habitat et ensemble fune´raire de nourrissons (Documents d’Arche´ologie en Rhoˆne-Alpes et en Auvergne 27), Lyon.

ALLMER, A., and DISSARD, P. (1885), Trion, Antiquite´s de´couvertes en 1885, 1886 et ante´rieurement: Au quartier de Lyon dit de Trion, Lyon.

AMERIO, A. (1968), ‘Galeno “puericultore”’, Pagine di storia della medicina, 12/5: 24–7.

ANDREAU, J., and BRUHNS, H. (eds.) (1990), Parente´ et strate´gies familiales dans l’Antiquite´ romaine. Actes de la table ronde des 2–4 octobre 1986, Paris (MSH), Rome.

ANKUM, H. (2005), ‘D.40.4.29 Scaevola L.23 Digestorum: Un drame matrimonial romain ayant des conse´quences juridiques complexes’, in Z. Słuzewska and J. Urbanik (eds.), Marriage, Ideal—Law—Practice, Proceedings of a Conference Held in Memory of Henryk Kupiszewski in Warsaw on the 24th of April 2004, Warsaw, 1–5.

ARGENTO, R. (1972), ‘Due epicedii di Papinio Statio. Silvae II.2, V.5’, Rivista di Studi Classici, 20: 331–62.

ARIE`S, P. (1973 [1st edn. Paris 1960]), L’Enfant et la vie familiale sous l’Ancien Re´gime, Paris.

ARJAVA, A. (1996), Women and Law in Late Antiquity, Oxford.

ASCENZI, A. (1994), La mummia romana di Grottarossa, Corso di perfezionamento in paleopatologia, Pisa, private document for students.

——BIANCO, P., NICOLETTI, R., et al. (1996), ‘The Roman Mummy of Grottarossa’, in K. Spindler (ed.), Human Mummies: A Global Survey of their Status and the Techniques of Conservation, Vienna and New York, 205–18.

332 References

AssMANN, A. (1991a), ‘Fest und flu¨ssig: Anmerkungen zu einer Denkfigur’, in A. Assmann and D. Harth (eds.), Kultur als Lebenswelt und Monument, Frankfurt a.M., 181–99.

——(1991b), ‘Kultur als Lebenswelt und Monument’, in A. Assmann and D. Harth (eds.), Kultur als Lebenswelt und Monument, Frankfurt a.M., 11–25.

——and AssMANN, J. (1988), ‘Schrift, Tradition und Kultur’, in W. Raible (ed.), Zwischen Festtag und Alltag (ScriptOralia 6), Tu¨bingen, 25–49.

AssMANN, J. (1988), ‘Kollektives Geda¨chtnis und kulturelle Identita¨t’, in J. Assmann and T. Ho¨lscher (eds.), Kultur und Geda¨chtnis, Frankfurt a, M., 9–19.

——(19972 [1st edn. 1992]), Das kulturelle Geda¨chtnis: Schrift, Erinnerung und politische Identita¨t in fru¨ hen Hochkulturen, Munich.

AsTIN, A. E. (1967), Scipio Aemilianus, Oxford.

ATHERToN, C. (1998), ‘Children, Animals, Slaves and Grammar’, in Y. Lee

Too and N. Livingstone (eds.), Pedagogy and Power, Cambridge, 214–44. AUBERT, M.-F., and CoRTopAssI, R. (2004), Portraits fune´raires de l’Egypte

romaine, I, Masques en stuc, Paris.

AUFFARTH, Ch. (2005), ‘How to Sacrifice Correctly—without a Manual?’, in R. Ha¨gg and B. Alroth, Greek Sacrificial Ritual, Olympian and Chthonian, Stockholm, 11–21.

BACKE-DAHMEN, A. (2006), Innocentissima aetas—Ro¨mische Kindheit im Spiegel literarischer, rechtlicher und archa¨ologischer Quellen des 1. bis 4. Jahrhunderts n. Chr., Mainz.

BADEL, C. (2005), La Noblesse de l’Empire romain: Les Masques et la vertu, Seyssel.

BADIAN, E. (1957), ‘Caepio and Norbanus: Notes on the decade 100–90 B.C.’, Historia, 6: 318–46.

——(1965), ‘M. Porcius Cato and the Annexation and Early Administration of Cyprus’, JRS 55: 110–21.

BAGNALL, R. (1997), ‘Missing Females in Roman Egypt’, Scripta Classica Israelica, 16: 121–38.

——and FRIER, B. (1994), The Demography of Roman Egypt, Cambridge.

BAILLs-TALBI, N., and BLANCHARD, P. (2006), ‘Se´pultures de nouveau-ne´s et de nourrissons du 1er aˆge de Fer au haut Moyen Aˆ ge de´couvertes hors des contextes fune´raires traditionnels sur les territoires carnute, turon et bituriges cube: Inventaire, synthe`se et interpretations’, Revue arche´ologique de Centre, 157–205.

BAKKE, O. M. (2005), When Children Became People: The Birth of Childhood in Early Christianity, Minneapolis.

References 333

BALCH, D., and OSIEK, L. (eds.) (2003), Early Christian Families in Context: An Interdisciplinary Dialogue, Cambridge and Grand Rapids.

BALDASSARRE, I., PONTRANDOLFO, A., ROUVERET, A., and SALVADORI, M. (2006), La Peinture romaine de l’e´poque helle´nistique a` l’Antiquite´ tardive, Arles and Milan.

BARBET, A., and COMPOINT, S. (1999), Les Cite´s enfouies du Ve´suve: Pompe´i, Herculanum, Stabies et autres lieux, Paris.

BARCLAY, W. (1959), Educational Ideals in the Ancient World, London. BARDON, H. (1972), ‘Le Concept de similitude a` Rome’, in Aufstieg und Niedergang der ro¨mischen Welt, I.2: 857–68.

BARGETZI, D. (2002), Cicero und Tullia: Einblicke in eine Beziehung zwischen Vater und Tochter im alten Rom (Seminararbeit, Seminar fu¨r Alte Geschichte Basel), Basel.

BAUMAN, R. (1992), Women and Politics in Ancient Rome, London.

BE´AL, J.-C. (1983), Catalogue des objets de tabletterie du Muse´e de la civilisation gallo-romaine de Lyon, Lyon and Paris.

BEARD, M. (1993), ‘Looking (Harder) for Roman Myth: Dume´zil, Declamation and the Problems of Definition’, in F. Graf (ed.), Mythos in mythenloser Gesellschaft: Das Paradigma Roms, Stuttgart–Leipzig, Teubner, 44–64 (trans. by B. Mezzadri, in Mythe et mythologie dans l’Antiquite´ gre´co-romaine, Europe 904–905, 2004: 190–218).

BELLEMORE, J., and RAwSON, B. (1990), ‘Alumni: The Italian Evidence’, Zeitschrift fu¨r Papyrologie und Epigraphik, 83: 1–19.

BELTRAMI, L. (1998), Il sangue degli antenati: Stirpe, adulterio e figli senza padre nella cultura romana, Bari.

BERGER, P. L., and LUCKMANN, T. (1967), The Social Construction of Reality: A Treatise in the Sociology of Knowledge, New York.

BERNSTEIN, N. W. (2005), ‘Mourning the Puer Delicatus: Status Inconsistency and the Ethical Value of Fostering in Statius, Silvae 2.1’, American Journal of Philology, 126: 257–80.

——(2008), ‘Each Man’s Father Served as his Teacher: Constructing Relatedness in Pliny’s Letters’, Classical Antiquity, 27/2: 203–30.

BERTIER, J. (1991), ‘La Me´decine des enfants dans l’Oribase latin’, in P. Mudry and J. Pigeaud (eds.), Les E´coles me´dicales a` Rome, Geneva, 269–84.

——(1996), ‘La Me´decine des enfants a` l’e´poque impe´riale’, in Aufstieg und Niedergang der ro¨mischen Welt, II.37.3: 2147–227.

BETTINI, M. (1986), Antropologia e cultura romana: Parentela, tempo, immagini dell’anima, Rome.

334 References

BETTINI, M. (1989), ‘Le riscritture del mito’, in G. Cavallo, P. Fedeli, and A. Giardina (eds.), Lo spazio letterario di Roma antica, I, La produzione del testo, Rome, 15–35.

——(1992), Familie und Verwandtschaft im antiken Rom, Frankfurt/M. and New York.

——(1999 [1st edn. 1986]), Antropologia e cultura romana, Rome (Eng. trans. Baltimore 1991).

——(2005), ‘Death and its Double: Imagines, Ridiculum and Honos in the Aristocratic Funeral’, in Mustakallio, Hanska, Sainio, and Vuolanto 2005, 191–202.

——(2008a), Voci, Turin.

——(2008b), ‘Weighty Words, Suspect Speech: fari in Roman Culture’, Arethusa, 41: 313–75.

BIANCHI BANDINELLI, R. (1980), La pittura antica (F. Coarelli and L. Franchi dell’Orto eds.), Biblioteca di storia antica, 11, Rome.

BIEZUNSKA-MALOWIST, I. M. (1971), ‘Die Expositio von Kindern als Quelle der Sklavenbeschaffung im griechisch-ro¨mischen Aegypten’, Jahrbuch fu¨r Wirtschaftgeschichte, 2: 129–33.

——(1977), L’Esclavage dans l’Egypte gre´co-romaine, II, Pe´riode romaine, Wrocław.

BINDER, G. (1985), ‘Lusus Troiae: L’Ene´ide de Virgile comme source

arche´ologique’, Bulletin de l’Association Guillaume Bude´, 44: 349–56. BIONDI, B. (1954), Il diritto romano cristiano, III, La famiglia: Rapporti

patrimoniali: Diritto pubblico, Milan.

BIRLEY, A. (1997), Hadrian: The Restless Emperor, New York.

BLANC, N., and MARTINEZ, J.-L. (1998), ‘La Tombe de Patron a` Rome’, in N. Blanc (ed.), Au royaume des ombres: La Peinture fune´raire antique, Catalogue d’exposition, Paris and Saint-Romain-en-Gal-Vienne, 82–95.

BLU¨MNER, H. (1879), Technologie und Terminologie der Gewerbe und Ku¨nste bei Griechen und Ro¨mern, Leipzig, II.

BODEL, J. (1999), ‘Death on Display: Looking at Roman Funerals’, in B. Bergmann and C. Kondoleon (eds.), The Art of Ancient Spectacle (Studies in the History of Art 56), Washington, DC, 259–81.

BOE¨LS-JANSSEN, N. (1993), La Vie religieuse des matrones dans la Rome archaı¨que, Rome.

BOETHIUS, A. (1942), ‘On the Ancestral Masks of the Romans’, Acta Archaeologica, 13: 226–35.

BONNARD, J.-B. (2006), ‘Il paraıˆt en effet que les fils ressemblent aux pe`res’, in F. Prost and J. Wilgaux (eds.), Penser et repre´senter le corps dans l’Antiquite´, Rennes, 307–18.

References 335

BONNER, S. F. (1977), Education in Ancient Rome: From Cato the Elder to Pliny the Younger, Berkeley and Los Angeles.

BOOTH, A. D. (1979), ‘The Schooling of Slaves in First-Century Rome’,

Transactions of the American Philological Association, 109: 11–19. BORDENACHE BATTAGLIA, G. (1983), Corredi funerari di eta` imperiale

e barbarica nel Museo nazionale romano, Rome.

BORGEAUD, P. (ed.) (1988), Me´moire des religions, Geneva.

——(2004), ‘L’Enfance au miel dans les re´cits antiques’, in Dasen 2004, 113–26.

BOSWELL, J. (1988), The Kindness of Strangers: The Abandonment of Children

in Western Europe from Late Antiquity to the Renaissance, New York. BOURDIEU, P. (1972), Esquisse d’une the´orie de la pratique, Geneva. ——(1982), Ce que parler veut dire: L’E´conomie des e´changes linguistiques,

Paris.

BOYANCE´, P. (1944), ‘L’Apothe´ose de Tullia’, Revue des E´tudes Anciennes, 46: 179–84.

BRADLEY, K. R. (1986a), ‘Wet-Nursing at Rome: A Study in Social Relations’, in Rawson 1986, 201–29

——(1986b), ‘Seneca and Slavery’, Classica et Mediaevalia, 37: 161–72. ——(1991a), Discovering the Roman Family, New York and Oxford. ——(1991b): ‘Child Care at Rome—The Role of Men’, in Bradley 1991a,

37–75.

——(1991c), ‘Dislocation in the Roman family’, in Bradley 1991a, 125–55. ——(1991d), ‘Remarriage and the Structure of the Upper-Class Family at Rome’, in Bradley 1991a, 156–76.

——(1991e), [Review of] ‘Thomas Wiedemann, Adults and Children in the Roman Empire, London 1989’, Classical Philology, 86: 258–63.

——(1998), ‘The Roman Family at Dinner’, in I. Nielsen and H. Sigismund Nielsen (eds.), Meals in a Social Context: Aspects of the Communal Meal in the Hellenistic and Roman World, Aarhus, 36–55.

——(2005), ‘The Roman Child in Sickness and Health’, in George 2005, 67–92.

BRAUN, M., HALTENHOFF, A., and MUTSCHLER, F.-H. (eds.) (2000), Moribus antiquis res stat Romana: Ro¨mische Werte und ro¨mische Literatur im 3. und 2. Jh. v. Chr., Munich and Leipzig.

BRECHTER, S. (1941), Die Quellen zur Angelsachsenmission Gregors des Grossen, Mu¨nster.

BRELICH, A. (1969), Paides e partenoi, Rome.

BREMMER, J. A. (1995), ‘The Family and other Centres of Religious Learning in Antiquity’, in J. W. Drijvers and A. A. MacDonald (eds.), Centres of

336 References

Learning: Learning and Location in Pre-Modern Europe and the Near East, Leiden and New York, 29–38.

——(1997), ‘Jokes, Jokers and Jokebooks in Ancient Greece’, in J. Bremmer and H. Roodenburg (eds.), A Cultural History of Humour from Antiquity to the Present Day, Cambridge, 11–28.

BRENNAN, T. C. (2000), The Praetorship in the Roman Republic, Oxford.

BRETIN-CHABROL, M. (2007), La Naissance et l’origine: Me´taphores ve´ge´tales de la filiation dans les textes romains de Caton a` Gaius (stirps, propago, suboles, semen, satus, inserere), Doctoral Dissertation, under the supervision of P. Moreau, Universite´ Paris XII.

BRISSON, L. (1994), Platon, les mots et les mythes, Paris.

BROMMER, F. (1953–4), ‘Zu den ro¨mischen Ahnenbildern’, Mitteilungen des Deutschen Archa¨ologischen Instituts, Ro¨mische Abteilung, 60–1: 163–71.

BROWNING, R. (2000), ‘Education in the Roman Empire’, in A. Cameron et al. (eds.), Cambridge Ancient History XIV. Late Antiquity: Empire and Successors, A.D. 425–600, Cambridge, 855–83.

BRUHNS, H. (1990), ‘Parente´ et alliances politiques a` la fin de la Re´publique romaine’, in Andreau and Bruhns 1990, 571–94.

——(1994), ‘Verwandtschaftsstrukturen, Geschlechterverha¨ltnisse und Max Webers Theorie der alten Stadt’, in C. Meier (ed.), Die okzidentale Stadt nach Max Weber, Munich, 59–94.

BRULE´, P. (1987), La Fille d’Athe`nes: La Religion des filles a` Athe`nes a` l’e´poque classique: Mythes, cultes et socie´te´, Paris.

——(2007), La Gre`ce d’a` coˆte´: Re´el et imaginaire en miroir en Gre`ce antique, Rennes.

BUCKLAND, W. W. (1908), The Roman Law of Slavery, Cambridge.

BU¨LL, R. (1959), Vom Wachs: Hoechster Beitra¨ge zur Kenntnis der Wachse, I, Beitra¨ge 10/11, Frankfurt a.M.

——and MOSER, E. (1973), s.v. Wachs, RE, Suppl. Bd. XIII: 1347–416. BURCKHARDT L. (2003), ‘“Zu Hause geht Alles, wie wir wu¨nschen ...”— Privates und Politisches in den Briefen Ciceros’, Klio, 85: 94–113.

——and UNGERN-STERNBERG, J. VON (1994), ‘Cornelia, Mutter der Gracchen’, in M. H. Dettenhofer (ed.), Reine Ma¨nnersache? Frauen in den Ma¨nnerdoma¨nen der antiken Welt, Cologne, Weimar, and Vienna, 97–132.

BURGUIE`RE, A., KLAPISCH-ZUBER, C., and SEGALEN, M., et al. (eds.) (1986), Histoire de la famille, I–II, Paris.

BUTLER, S. (1998), ‘Notes on a “membrum disiectum”’, in S. R. Joshel and S. Murnaghan (eds.), Women and Slaves in Greco-Roman Culture: Differential Equations, London and New York, 236–55.

References 337

BUTRICA, J. (2005), ‘Some Myths and Anomalies in the Study of Roman Sexuality’, Journal of Homosexuality, 49/3–4: 209–69.

BUTTERWORTH, A., and LAURENCE, R. (2005), Pompeii: The Living City, London.

BYL, S., and BARGIBANT, A. (1991), ‘L’Enfant chez Galien’, in J. A. Lo´pez

Fe´rez (ed.), Galeno: Obra, pensiamento e influencia, Madrid, 107–17. CALAME, C. (1996 [1st edn. 1990]), The´se´e et l’imaginaire athe´nien,

Lausanne.

CAMERON, A. (1939), ‘Threptos and Related Terms in the Inscriptions of Asia Minor’, in W. M. Calder and J. Keil (eds.), Anatolian Studies Presented to William Hepburn Butler, Manchester, 27–62.

CANCIK, H. (1973), ‘Ro¨mischer Religionsunterricht in apostolischer Zeit: Ein pastoralgeschichtlicher Versuch zu Statius Silvae 5, 3, 176–184’, in H. Feld and J. Nolte (eds.), Wort Gottes in der Zeit, Du¨sseldorf, 181–97.

CANFORA, L. (2001), Caesar: Der demokratische Diktator, Eine Biographie, Munich.

CANTARELLA, E. (2003), ‘Fathers and Sons in Rome’, Classical World, 96: 281–98.

CAPASSO, L. (2000), ‘Indoor Pollution and Respiratory Diseases in Ancient Rome’, The Lancet, 356, Nov.: 1774.

CARP, T. (1981), ‘Two Matrons of the Late Republic’, in H. P. Foley (ed.), Reflexions of Women in Antiquity, New York, London, and Paris, 343–54 [1st publ. in Women’s Studies 8, 1981: 189–200].

CATALANO, P., et al. (2003), ‘Studio del microclima per la conservazione della Mummia di Grottarossa’, Materiali e strutture: Problemi di conservazione, 2: 135–61.

CENNINI, C. (1933 [1st edn. 1390]), Libro dell’Arte (Eng. trans. by D. V. Thompson, The Craftsman’s Handbook: The Italian ‘Il Libro dell’ Arte’, New York).

CERULLI IRELLI, G., AOYAGI, M., DE CARO, S., and PAPPALARDO, U. (eds.) (1993), La Peinture de Pompe´i: Te´moignages de l’art romain dans la zone ensevelie par (le) Ve´suve en 79 ap. J.-C., I–II, Paris.

CHAMPEAUX, J. (1982), Fortuna: Recherches sur le culte de la Fortune a` Rome et dans le monde romain des origines a` la mort de Ce´sar, I, Fortuna dans la religion archaı¨que, Rome.

CHAMPLIN, E. (1991), Final Judgments: Duty and Emotion in Roman Wills 200 B.C.–A.D. 250, Berkeley.

CHARLES-LAFORGE M.O. (2007), ‘“Imagines Maiorum” et portraits d’anceˆtres a` Pompe´i’, in La Norme a` Pompe´i (Ier sie`cle avant – Ier sie`cle apre`s J.-C.), Colloque organise´ par M.-O. Charles-Laforge, Centre Jacob

338 References

Spon, Romanitas, Universite´ Lyon 2—le 17 novembre 2004 (Studi della Soprintendenza archeologica di Pompei 21), Rome, 158–71.

CHARLIER, P. (2003), ‘Les Maladies infantiles: Cas pale´opathologiques en Gaule romaine’, in Gourevitch, Moirin, and Rouquet 2003, 184–8.

CICHORIUS, C. (1921), ‘Ein Heiratsprojekt im Hause Caesars’, in Festgabe Friedrich Bezold. Dargebracht zum 70. Geburtstag von seinen Schu¨lern, Kollegen und Freunden, Bonn and Leipzig, 59–80.

CLAASSEN, J.-M. (1996), ‘Documents of a Crumbling Marriage: The Case of Cicero and Terentia’, Phoenix, 50: 208–32.

CLARK, G. (1993), Women in Late Antiquity: Pagan and Christian Lifestyles, Oxford.

——(1994), ‘The Fathers and the Children’, in D. Wood (ed.), The Church and Childhood (Studies in Church History 31), Cambridge, 1–28. CLARK, P. A. (1991), ‘Tullia and Crassipes’, Phoenix, 45: 28–38.

CLARKE, J. R. (1991), The Houses of Roman Italy, 100 B.C.–A.D. 250: Ritual, Space, and Decoration, Oxford.

——(1995), ‘Landscape Paintings in the Villa of Oplontis’, Journal of Roman Archaeology, 9: 81–107.

COARELLI, F. (1969), ‘Le Tyrannoctone du Capitole et la mort de Tibe´rius

Gracchus’, Me´langes de l’E´cole franc¸aise de Rome. Antiquite´, 81: 137–60. COHEN, A., and RUTTER, J. (eds.) (2007), Constructions of Childhood in

Ancient Greece and Italy (Hesperia suppl.).

COLGRAVE, B., and MYNORS, R. A. B. (1969), Bede’s Ecclesiastical History of the English People, Oxford.

COLLINS, J. H. (1951), ‘Tullia’s Engagement and Marriage to Dolabella’, Classical Journal, 47: 164–8.

CORBEAUD, E. (1900), s.v. imago, in Daremberg and Saglio, III: 412–14. CORBEILL, T. (1996), Controlling Laughter: Political Humour in the Late Republic, Princeton.

——(2001), ‘Education in the Roman Republic: Creating Traditions’, in Y. L. Too (ed.), Education in Greek and Roman Antiquity, Leiden, Boston, and Cologne, 261–87.

CORBETT, P. B. (1986), The Scurra, Edinburgh.

CORBIER, M. (1987), ‘Les Comportements familiaux de l’aristocratie romaine (IIe`me sie`cle avant J.-C. – IIIe`me sie`cle apre`s J.-C.)’, Annales: E´conomies, socie´te´s, civilisations, 42: 1267–85.

——(1991), ‘Divorce and Adoption as Roman Familial Strategies’, in Rawson 1991, 47–78.

——(1999), ‘La Petite Enfance a` Rome: Lois, normes, pratiques individuelles et collectives’, Annales: Histoire, sciences sociales, 6: 1257–90. ——(2001), ‘Child Exposure and Abandonment’, in Dixon 2001b, 52–73.

References 339

——(2006), ‘Parente´ et alliance dans le monde romain (IIe sie`cle a.C. – VIe sie`cle p.C.)’, in A. Bresson (ed.), Parente´ et socie´te´ dans le monde grec de l’antiquite´ a` l’aˆge moderne, Bordeaux, 199–208.

CORCORAN, S. (2000), ‘The Sins of the Fathers: A Neglected Constitution of Diocletian on Incest’, Journal of Legal History, 21: 1–34.

COUDRY, M., and SPA¨TH, T. (2001), L’Invention des grands hommes de la Rome antique/Die Konstruktion der grossen Ma¨nner Altroms. Actes du colloque du Collegium Beatus Rhenanus, Augst 16–18 septembre 1999 (Collections de l’Universite´ Marc Bloch—E´tudes d’arche´ologie et d’histoire ancienne), Paris.

COULON, G. (1994), L’Enfant en Gaule romaine, Paris (2nd revised edn. 2004).

CROISILLE, J.-M. (2005), La Peinture romaine, Paris.

——(2010), Paysages dans la peinture romaine, Paris.

CROOK, J. A. (1967), Law and Life of Rome, 90 B.C. – A.D. 212, Ithaca, NY. D’AMBRA, E. (2002), ‘Acquiring an Ancestor: The Importance of Funerary

Statuary among the Non-E´lite Orders of Rome’, in J. M. Hojte (ed.),

Images of Ancestors, Aarhus, 223–46.

DAMON, C. (1997), The Mask of the Parasite: A Pathology of Roman Patronage, Ann Arbor.

D’ARMS, J. (1999), ‘Performing Culture: Roman Spectacle and the Banquets of the Powerful’, in B. Bergman and C. Kondoleon (eds.), The Art of Ancient Spectacle, New Haven, 301–19.

DASEN, V. (2003), ‘L’Enfant malade’, in Gourevitch, Moirin, and Rouquet 2003, 178–82.

——(ed.) (2004), Naissance et petite enfance dans l’Antiquite´, Actes du colloque de Fribourg, 28 novembre – 1er de´cembre 2001, Fribourg and Go¨ttingen.

——(2005), Jumeaux, jumelles dans l’Antiquite´ grecque et romaine, Kilchberg.

——(2006), ‘La Mort des enfants a` Rome: L’Impossible chagrin?’, in L’enfant, La Vouivre, 15: 29–37.

——(2009a), ‘Empreintes maternelles’, in La madre/The Mother, Micrologus XVII, Florence, 35–54.

——(2009b), ‘Roman Birth Rites of Passage Revisited’, Journal of Roman Archaeology, 22: 199–214.

——(2010), ‘Naissance et petite enfance, grec’, in ThesCRA VI.

——LETT, D., MOREL, M.-F., and ROLLET, C. (2001), ‘Dix ans de travaux sur l’enfance’, in Enfances: Bilan d’une de´cennie de recherche, Annales de de´mographie historique, 2: 5–100.

340 References

DAUT, R. (1975), Imago: Untersuchungen zum Bildbegriff der Ro¨mer, Heidelberg.

DAVID, J.-M. (1980), ‘Maiorum exempla sequi: l’exemplum historique dans les discours judiciaires de Cice´ron’, Me´langes de l’E´cole franc¸aise de Rome. Moyen-aˆge, 92/1: 67–86.

——(1992), Le Patronat judiciaire au dernier sie`cle de la Re´publique romaine, Rome and Paris.

——(1998a), ‘De l’action exemplaire a` la mise en sce`ne des actes a` Rome aux deux derniers sie`cles avant notre e`re’, in Travail sur la figure, travail de la me´moire, Strasbourg, 5–16.

——(ed.) (1998b), Valeurs et me´moire a` Rome. Vale`re Maxime ou la vertu recompose´e (Coll. de l’Universite´ des Sciences Humaines de Strasbourg. E´tudes d’arche´ologie et d’histoire ancienne), Paris.

——and DONDIN, M. (1980), ‘Dion Cassius, XXXVI, 41, 1–2: Conduites symboliques et comportements exemplaires de Lucullus, Acilius Glabrio et Papirius Carbo (78 et 67 a. C.)’, Me´langes de l’E´cole franc¸aise de Rome. Antiquite´, 92/1: 199–213.

DEANESLY, M., and GROSJEAN, P. A. (1959), ‘The Canterbury Edition of the Answers of Pope Gregory I to St. Augustine’, Journal of Ecclesiastical History, 10: 1–49.

DIDI-HUBERMANN, G. (1994), ‘Ressemblance mythifie´e et ressemblance oublie´e chez Vasari: La Le´gende du portrait “sur le vif”’, Me´langes de l’E´cole franc¸aise de Rome. Moyen-aˆge, 106: 383–432.

——(1999), ‘Wax Flesh, Vicious Circles’, in Encyclopaedia Anatomica: A Complete Collection of Anatomical Waxes, Cologne, 64–74.

DIXON, S. (1984), ‘Family Finances: Tullia and Terentia’, Antichthon, 18: 78–101 [= (abridged) in Rawson 1986, 93–120].

——(1991), ‘The Sentimental Ideal of the Roman Family’, in Rawson 1991, 99–113.

——(1992), The Roman Family, Baltimore.

——(1999), ‘The Circulation of Children in Roman Society’, in M. Corbier (ed.), Adoption et fosterage, Paris, 217–30.

——(2001a), ‘Rape in Roman Law and Myth’, in S. Dixon (ed.), Reading Roman Women: Sources, Genres and Real Life, London, 45–55.

——(ed.) (2001b), Childhood, Class and Kin in the Roman World, London and New York.

——(2007), Cornelia: Mother of the Gracchi, London and New York. DOLANSKY, F. (2008), ‘Togam Virilem Sumere: Coming of Age in the Roman

World’, in J. Edmondson and A. Keith (eds.), Roman Dress and the Fabrics

of Roman Culture, Toronto, 47–70.

DONATI, F. (2005), ‘Pueri nell’arena: Fonti per una iconografia’, Archeologia classica, 56, NS 6: 80–129.

References 341

DONDIN-PAYRE, M. (1993), Exercice du pouvoir et continuite´ gentilice: Les Acilii Glabriones du 3e`me sie`cle ap. J.-C. au 5e`me sie`cle ap. J.-C., Rome.

DRERUP, H. (1980), ‘Totenmaske und Ahnenbild bei den Ro¨mern’, Mitteilungen des Deutschen Archa¨ologischen Instituts, Ro¨mische Abteilung, 87: 81–129.

DUDAY, H., LAUBENHEIMER F., and TILLIER, A.-M. (1995), Salle`les d’Aude: Nouveau-ne´s et nourrissons gallo-romains (Centre de recherche d’histoire ancienne 144), Paris.

DUNAND, F., and Lichtenberg, R. (1998), Les Momies et la mort en E´gypte, Paris.

DUNBABIN, K. M. D. (1999), Mosaics of the Greek and Roman World, Cambridge.

DUPONT, F. (1998), ‘Imago identique et imago identitaire: Le Jeu du double dans la come´die romaine’, in C. Auvray-Assayas (ed.), Images romaines, Paris, 247–60.

DURAND, R. (2003), ‘La Mortalite´ infantile en Gaule romaine’; ‘Les se´pultures d’enfants’, in Gourevitch, Moirin, and Rouquet 2003, 94–97, 112–15.

DURLING, R. J. (1993), A Dictionary of Medical Terms in Galen, Leiden. DWYER, E. J. (1982), Pompeian Domestic Sculpture: A Study of Five Pompeian Houses and their Contents, Rome.

ECK, W. (1986), ‘Inschriften aus der Vatikanischen Nekropole unter St. Peter’, Zeitschrift fu¨r Papyrologie und Epigraphik, 65: 245–93.

——(1993), ‘Das s. c. de Cn. Pisone patre und seine Publikation in der Baetica’, Cahiers du Centre Glotz, 54: 189–208.

EHRHARDT, W. (1998), Casa di Paquius Proculus (I 7,1.20) (Ha¨ user in Pompeji 9), Munich.

EIJK, T. VAN (1972), ‘Marriage and Virginity, Death and Immortality’, in J. Fontaine and C. Kannengiesser (eds.), Epektasis: Me´langes patristiques offerts au Cardinal Jean Danie´lou, Paris, 220–35.

ERMETE, K. (2003), Terentia und Tullia—Frauen der senatorischen Oberschicht, Frankfurt a.M.

ESTIENNE, S. (2005), ‘Saliens’, in ThesCRA V: 85–7.

——PRESCENDI, F., SCHEID, J., and VAN HAEPEREN, F. (2005), ‘Les Enfants (pueri patrimi matrimique, camilli)’, in ThesCRA V: 116.

EVANS GRUBBS, J. (1993a), ‘Constantine and Imperial Legislation on the Family’, in J. Harries and I. Wood (eds.), The Theodosian Code: Studies in the Imperial Law of Late Antiquity, London and Ithaca, 120–42.

——(1993b), ‘“Marriage more Shameful than Adultery”: Slave–Mistress Relationships, “Mixed Marriages”, and Late Roman Law’, Phoenix, 47: 125–54.

342 References

EVANS GRUBBS, J. (1995), Law and Family in Late Antiquity: The Emperor Constantine’s Marriage Legislation, Oxford.

——(2000), ‘The Slave who Avenged her Master’s Death’, Ancient History Bulletin, 14: 81–8.

——(2005), ‘Parent–Child Conflict in the Roman Family: The Evidence of Justinian’, in George 2005, 93–128.

——(2009), ‘Church, State, and Children: Christian and Imperial Attitudes toward Infant Exposure in Late Antiquity’, in A. Cain and N. Lenski (eds.), The Power of Religion in Late Antiquity, Aldershot, 119–31.

——(forthcoming), Children without Fathers in Roman Imperial Law: Paternity, Legitimacy, and Freedom, to be published by Oxford University Press.

EWALD, P., and HARTMANN, L. E. (1899), Gregorii I Papae registrum epistolarum, II, Berlin.

EYBEN, E. (1986), ‘Sozialgeschichte des Kindes im ro¨mischen Altertum’, in J. Martin and A. Nitschke (eds.), Zur Sozialgeschichte des Kindes, Freiburg and Munich, 317–64.

——(1991), ‘Fathers and Sons’, in Rawson 1991, 114–43.

——(1993), RestlessYouth in Ancient Rome, London and New York. FANTHAM, E. (1999), ‘Chironis Exemplum: On Teachers and Surrogate Fathers in Achilleid and Silvae’, Hermathena, 167: 59–70.

FARMER, H. (1959), ‘St Gregory’s Answers to St Augustine of Canterbury’, Studia Monastica, 1: 419–22.

FAYER, C. (1994), La familia romana, I, Rome.

FEHRLE, R. (1983), Cato Uticensis, Darmstadt.

FENTRESS, J., and WICKHAM, C. (1992), Social Memory (New Perspectives on the Past), Oxford.

FERAUDI-GRUE´NAIS, F. (2001), Ubi diutius nobis habitandum est: Die Innendekoration der kaiserzeitlichen Gra¨ber Roms, Wiesbaden.

FERRARY, J.-L. (1988), Philhelle´nisme et impe´rialisme: Aspects ide´ologiques de la conqueˆte romaine du monde helle´nistique, de la seconde guerre de Mace´doine a` la guerre contre Mithridate, Rome.

FLAIg, E. (1995), ‘Die Pompa Funebris: Adlige Konkurrenz und annalistische Erinnerung in der Ro¨mischen Republik’, in O. G. Oexle (ed.), Memoria als Kultur, Go¨ttingen, 115–48.

FLESS, F. (1995), Opferdiener und Kultmusiker auf stadtro¨mischen historischen Reliefs, Mainz.

FLEURY, J. (1933), Recherches sur les empeˆchements de parente´ dans le mariage canonique des origines aux fausses de´cre´tales, Paris.

References 343

FLORY, M. B. (1993), ‘Livia and the History of Public Honorific Statues for Women in Rome’, Transactions of the American Philological Association, 123: 287–308.

FLOWER, H. I. (1996), Ancestor Masks and Aristocratic Power in Roman Culture, Oxford.

——(2002), ‘Were Women ever Ancestors in Republican Rome?’, in J. M. Hojte (ed.), Images of Ancestors, Aarhus, 159–84.

FORBES, C. A. (1955), ‘The Education and Training of Slaves in Antiquity’,

Transactions of the American Philological Association, 86: 321–360. FORTGENS, H. W. (1959), ‘Publius Papinius Statius, de Latijnse dichter van

het kinderleven’, Hermeneus, 31/1: 52–9.

FOSSATI VANZETTI, M. B. (1983), ‘Vendita ed esposizione degli infanti da Costantino a Giustiniano’, Studia et Documenta Historiae et Iuris, 49: 179–224.

FRANCIOSI, E. (1988), ‘Il regime delle nozze incestuose nelle Novelle giustinianee’, Estudios en homenaje al Profesor Juan Iglesias, II, Madrid, 727–46.

FRANC¸OIS, E., and SCHULZE, H. (2001), Deutsche Erinnerungsorte, I–III, Munich.

FRASCHETTI, A. ([1st Ital. edn. 1994] 1996), ‘Jeunesses romaines’, in G. Levi and J.-C. Schmitt (eds.), Histoire des jeunes en Occident, I, De l’Antiquite´ a` l’e´poque moderne, Paris, 63–100.

FRIESEN, B. (2006), ‘Answers and Echoes: The Libellus Responsionum and the Hagiography of North-Western European Mission’, Early Medieval Europe, 4/2: 153–73.

FUCHS, M. (2007), ‘Peintures de paysages au nord des Alpes romaines’, in C. Robotti (ed.), Disegni e progetti di citta` e paesaggi fortificati. Atti del Quinto Colloquio internazionale di Studi Castello di Carlo V, Capua 3–4 dicembre 2005, Lecce, 188–99.

——(2008), ‘Petite vieille ou noble dame: Portraits de femmes aˆge´es sous l’Empire romain’, in V. Dasen and J. Wilgaux (eds.), Langages et me´taphores du corps, Rennes, 73–89.

GABORIT, J.-R., and LIGOT, J. (eds.) (1987), Sculptures en cire de l’ancienne Egypte a` l’art abstrait, Paris.

GARDNER, J. (1986), Women in Roman Law and Society, London and Sydney. ——(1989), ‘The Adoption of Roman Freedmen’, Phoenix, 43: 236–57. ——(1998), The Family and Familia in Roman Law and Life, Oxford. GARTHWAITE, J. (1984), ‘Statius, Silvae 3.4: On the Fate of Earinus’, Aufstieg

und Niedergang der ro¨mischen Welt, II.32.1: 111–24.

GAUDEMET, J. (1993), Les Sources du droit canonique. VIIIe–XXe s., Paris.

344 References

GEHRKE, H.-J. (1994), ‘Ro¨mischer mos und griechische Ethik:

U¨ berlegungen zum Zusammenhang von Akkulturation und politischer

Ordnung im Hellenismus’, Historische Zeitschrift, 258: 593–622. GEIGER, J. (1972), ‘Canidius or Caninius?’, Classical Quarterly, 22: 130–4.

——(1973), ‘The Last Servilii Caepiones of the Republic’, Ancient Society, 4:

143–56.

GELZER, M. (1918), s.v. Iunius (Brutus) No. 53, RE X.1: 973–1020. GENNEP, A. VAN (1981 [1st edn. 1909]), Les Rites de passage, Paris. GEORGE, M. (ed.) (2005), The Roman Family in the Empire: Rome, Italy and

Beyond, Oxford.

GIBSON, N. J. (2006), Statius Silvae 5. Edited with Introduction, Translation and Commentary, Oxford.

GIPPERT, J. (1997), Index Galenicus: Wortformenindex zu den Schriften Galens, Dettelbach, 2 vols.

GIULIANI, L. (1986), Bildnis und Botschaft, Hermeneutische Untersuchungen zur Bildniskunst der ro¨mischen Republik, Frankfurt a.M.

——(2008), ‘Imagines maiorum, o¨ffentliche Bildnisstatuen und die

Extensivierung der res publica’, in M. Bernett, W. Nippel, and A.

Winterling (eds.), Christian Meier zur Diskussion, Stuttgart, 143–59. GLADIGOw, B. (1997), ‘Erwerb religio¨ser Kompetenz: Kult und

¨ O ffentlichkeit in den klassischen Religionen’, in G. Binder and K. Ehlich

(eds.), Religio¨se Kommunikation: Formen und Praxis vor der Neuzeit, Trier, 103–18.

GOETTE, H. R. (1989), ‘Ro¨mische Kinderbildnisse mit Jugendlocken’, Mitteilungen des Deutschen Archa¨ologischen Instituts, Athenische Abteilung, 104: 203–17.

GOFFMAN, E. (1963), Behavior in Public Places: Notes on the Social Organization of Gatherings, Glencoe, Ill.

GOLDEN, M. (1988), ‘Did the Ancients Care when their Children Died?’, Greece & Rome, 35: 152–63.

——(2004), ‘Mortality, Mourning and Mothers’, in Dasen 2004, 145–57.

GOMBRICH, E. H. (1972), ‘The Mask and the Face: The Perception of Physiognomic Likeness in Life and Art’, in E. H. Gombrich, J. Hochberg, and M. Black, Art, Perception and Reality, Baltimore and London, 1–46.

GONZENBACH, V. VON (1957), Untersuchungen zu den Knabenweihen im Isiskult der ro¨mischen Kaiserzeit, Bonn.

GOTTER, U. (2001), Griechenland in Rom? Die ro¨mische Rede u¨ber Hellas und ihre Kontexte (3.–1. Jh. v.Chr.) (unpub. Habilitation), Freiburg.

References 345

GOUREVITCH, D. (1975), ‘Apropos de la simulation dans l’Antiquite´: Galien et sa monographie princeps, Quomodo morbum simulantes sint deprehendendi libellus’, Me´decine le´gale et expertise me´dicale: 13–18.

——(1982), ‘La gestatio the´rapeutique a` Rome’, Centre Jean Palerne. Me´moires, 3: 55–65.

——(1987), ‘L’Esthe´tique me´dicale de Galien’, Les E´tudes classiques, 55: 267–90.

——(1988), ‘Galien, me´decin de´cideur’, Prospective et sante´: La Clinique demain, 46: 81–8.

——(1990), ‘La Malade, sa maladie et ses the´rapeutes dans la Rome impe´riale’, Actions et recherches sociales, Mar. 1990: 42–7.

——(1991), ‘Un enfant muet de naissance s’exprime par le dessin : A propos d’un cas rapporte´ par Pline l’Ancien’, L’E´volution psychiatrique, 56: 889–93.

——(1992), ‘Is it Beautiful, is it Ugly? Considerations on the Aesthetics of Newborn Babies during the Roman Empire’, Forum, 2, History of medicine: 5–11.

——(1994), ‘Comment rendre a` sa ve´ritable nature le petit monstre humain?’, in P. J. van der Eijk, H. F. J. Horstmanshoff, and P. J. Schrijvers (eds.), Ancient Medicine in its Socio-Cultural Context, II, Amsterdam, 239–60.

——(1995), ‘Les Noms de la maladie, continuite´s et nouveaute´s’, in D. Gourevitch (ed.), Histoire de la me´decine: Lec¸ons me´thodologiques, Paris, 52–7.

——(1998), ‘Les Socie´te´s antiques connaissaient-elles l’addiction?’, in Les Passions dangereuses, Paris, Dossiers de l’IEPS, 9–11.

——(2001), I giovani pazienti di Galeno, Rome and Bari.

——(2006), ‘L’Enfant handicappe´ a` Rome: Mise au point et perspectives’, Medicina nei Secoli, 18/2: 459–77.

——and GOUREVITCH, M. (1983a), ‘Chronique anachronique, XI, Un enfant boulimique?’, L’E´volution psychiatrique, 48: 853–7.

————(1983b), ‘Chronique anachronique, IX, Marc-Aure`le devint-il toxico-de´pendant?’, L’E´volution psychiatrique, 48: 253–6.

——and GRMEk, M. (1986), ‘Medice, cura te ipsum: Les Maladies de Galien’, E´tudes de lettres: 45–64.

——MOIRIN, A., and ROUQUET, N. (eds.) (2003), Maternite´ et petite enfance dans l’Antiquite´ romaine. Catalogue de l’exposition, Bourges, Muse´um d’histoire naturelle, 6 novembre 2003 – 28 mars 2004, Bourges.

GOYON, J.-C. (ed.) (1972), Rituels fune´raires de l’ancienne Egypte: Le Rituel de l’embaumement, le rituel de l’ouverture de la bouche, les livres des respirations, Paris.

346 References

GRIFFIN, M. T. (1976), Seneca: A Philosopher in Politics, Oxford.

GRIMM, G. (1974), Die ro¨mischen Mumienmasken aus A¨gypten, Wiesbaden. GRMEK, M. D. (1983), Les Maladies a` l’aube de la civilisation occidentale, Paris (2nd edn. 1994; Eng. trans. Baltimore, 1989).

——and GOUREVITCH, D. (1998), Les Maladies dans l’art antique, Paris. GRUNDMANN, M. (2006), Sozialisation: Skizze einer allgemeinen Theorie, Konstanz.

GUASTELLA, G. (1985), ‘La rete del sangue: Simbologia delle relazioni e modelli dell’identita` nella cultura romana’, Materiali e discussioni per l’analisi dei testi classici,15: 49–123.

GUITTARD, C. (1988), ‘Naissance et de´veloppement d’une le´gende: Les “Decii”’, in D. Porte and J.-P. Ne´raudau (eds.), Hommages a` H. Le Bonniec. Res sacrae (Latomus 201), Brussels, 256–66.

GUROIAN, V. (2001), ‘The Ecclesial Family: John Chrysostom on Parenthood’, in M. J. Bunge (ed.), The Child in Christian Thought, Grand Rapids, 61–77.

GWYNN, A. (1964), Roman Education from Cicero to Quintilian, New York. HAAKE, M. (2007), Der Philosoph in der Stadt: Untersuchungen zur

o¨ffentlichen Rede u¨ber Philosophen und Philosophie in den hellenistischen

Poleis, Munich.

HACKWORTH PETERSEN, L. (2006), The Freedman in Roman Art and History, Cambridge.

HALBWACHS, M. (1968 [1st edn. 1950]), La Me´moire collective, Paris (Eng. trans. New York, 1980).

——(1971 [1st edn. 1941]), La Topographie le´gendaire des E´vangiles en Terre Sainte: E´tude de me´moire collective, Paris.

——(1994 [1st edn. 1925]), Les Cadres sociaux de la me´moire, Paris.

HALL, J. (2005), ‘Cicero Fam. 16.21, Roman Politeness, and the Socialization of Marcus Cicero the Younger’, in K. Welch and T. W. Hillard (eds.), Roman Crossings: Theory and Practice in the Roman Republic, Swansea, 259–78.

HALLETT, J. P. (1984), Fathers and Daughters in Roman Society: Women and the Elite Family, Princeton.

——(2004), ‘Matriot Games? Cornelia, Mother of the Gracchi, and the Forging of Family-Oriented Political Values’, in F. McHardy and E. Marshall (eds.), Women’s Influence on Classical Civilization, London and New York, 26–39.

HALLIWELL, S. (1991), ‘The Uses of Laughter in Greek Culture’, Classical Quarterly, 41: 279–96.

——(2008), Greek Laughter, Cambridge.

References 347

HALTENHOFF, A., HEIL, A., and MUTSCHLER, F.-H. (eds.) (2005), Ro¨mische Werte als Gegenstand der Altertumswissenschaft, Munich.

HANDEL, G. (ed.) (2006 [1st edn. New York, 1988]), Childhood Socialization, New Brunswick and London.

HANKINSON, R. J. (ed.) (1998), Galen on Antecedent Causes, Cambridge. HARDERS, A.-C. (2007a), ‘Die verwandtschaftlichen Beziehungen der

Servilia, Ehefrau des L. Licinius Lucullus: Schwester oder Nichte des

Cato Uticensis?’, Historia, 56: 453–61.

——(2007b), ‘Die Identita¨t der Iunia, De´kmou thuga´te¯r und Ehefrau des P. Servilius Isauricus (cos. 48): U¨ berlegungen zu BMusImp 3:3,1, PH 206 und IG XII Suppl. 60’, Klio, 89: 403–16.

——(2008), Suavissima Soror: Untersuchungen zu den Bruder-Schwester-Beziehungen in der ro¨mischen Republik, Munich.

HARLOW, M., LAURENCE, R., and VUOLANTO, V. (2007), ‘Past, Present and Future in the Study of Roman Childhood’, in S. Crawford and G. Shepherd (eds.), Children, Childhood and Society (IAA Interdisciplinary Series 1/BAR International Series 1697), 1–14.

HARRIS, W. (1994), ‘Child-Exposure in the Roman Empire’, Journal of Roman Studies, 84: 1–22.

——(1999), ‘Demography, Geography, and the Sources of Roman Slaves’, Journal of Roman Studies, 89: 62–75.

HEDRICK, C. W. (2000), History and Silence: Purge and Rehabilitation of Memory in Late Antiquity, Austin.

HELBIG, W. (1963), Fu¨hrer durch die o¨ffentlichen Sammlungen klassischer Altertu¨mer in Rom, I, Die Pa¨pstlichen Sammlungen im Vatikan und Lateran, Tu¨bingen.

HENRIKSE´N, C. (1997), ‘Earinus: An Imperial Eunuch in the Light of the Poems of Martial and Statius’, Mnemosyne, 50/3: 281–94.

HERRMANN-OTTO, E. (1994), Ex ancilla natus. Untersuchungen zu den ‘hausgeborenen’ Sklaven und Sklavinnen im Westen des ro¨mischen Kaiserreiches, Stuttgart.

HO¨LKESKAMP, K.-J. (1996), ‘Exempla und mos maiorum: U¨ berlegungen zum kollektiven Geda¨chtnis der Nobilita¨t’, in H.-J. Gehrke and A. Mo¨ller (eds.), Vergangenheit und Lebenswelt: Soziale Kommunikation, Traditionsbildung und historisches Bewusstsein, Tu¨bingen, 302–38.

HO¨LSCHER, T. (1978), ‘Die Anfa¨nge ro¨mischer Repra¨sentationskunst’, Mitteilungen des Deutschen Archa¨ologischen Instituts, Ro¨mische Abteilung, 85: 315–57.

HONORE´, T. (1994; 1st edn. London, 1981), Emperors and Lawyers, Oxford. HOPKINS, K. (1965), ‘The Age of Roman Girls at Marriage’, Population Studies, 19: 309–27.

348 References

HOPKINS, K. (1983), Death and Renewal, Cambridge.

HORN, C. B. (2005), ‘Children’s Play as Social Ritual’, in V. Burrus (ed.), A People’s History of Christianity, II, Christianity in the Roman World, Minneapolis, 95–116.

——(2007), ‘Children as Pilgrims and the Cult of Holy Children in the Early Syriac Tradition: The Cases of Theodoret of Cyrrhus and the Child-Martyrs Behna¯m, Sarah, and Cyriacus’, ARAM Periodical, 19/1–2: 439–62.

HORSFALL, N. (1989), ‘Virgil and Marcellus’ Education’, Classical Quarterly, 39: 266–7.

——(1996), La cultura della plebs romana, Barcelona.

HU¨BNER, S. R., and RATZAN, D. M. (eds.) (2009a), Growing up Fatherless in Antiquity, Cambridge.

————(2009b), ‘Fatherless Antiquity? Perspectives on “Fatherlessness”

in the Ancient Mediterranean’, in Hu¨bner and Ratzan 2009a, 3–28. HUET, V., and WYLER, S. (2005), ‘“Copies romaines d’un original grec”, ou les

arts grecs revisite´s par les Romains’, Me`tis, NS 3: 151–77.

HUMPHREYS, S. C. (1993 [1st edn. 1983]), The Family, Women and Death: Comparative Studies, Ann Arbor.

HUNECKE, V. (1994), ‘The Abandonment of Legitimate Children in Nineteenth-Century Milan and the European Context’, in J. Henderson and R. Wall (eds.), Poor Women and Children in the European Past, London and New York, 117–35.

HUNZIKER, E. (1887), ‘Camilli, camillae’, in Daremberg and Saglio, I.2: 858–9.

HUYS, M. (1995), ‘The Spartan Practice of Selective Infanticide and its Parallels in Ancient Utopian Tradition’, Ancient Society, 27: 47–74.

ILBERG, J. (1905), ‘Aus Galens Praxis: Ein Kulturbild aus der ro¨mischen Kaiserzeit’, Neue Jahrbu¨cher fu¨r das klassische Altertum, 15: 276–312 (republished in H. Flashar (ed.), Antike Medizin, Darmstadt 1971, 361–416).

IOANNATOU, M. (2006), Affaires d’argent dans la correspondance de Cice´ron: L’Aristocratie se´natoriale face a` ses dettes, Paris.

ISKANDAR, A. Z. (1988), De optimo medico cognoscendo: On Examinations by

which the Best Physicians are Recognized, edn. of the Arabic version, with

English translation and commentary (CMG Suppl. Or. IV), Berlin. ISNENGHI, M. (ed.) (1996–8), I luoghi della memoria, I–III, Rome.

JACQUES, F. (1987), ‘L’E´thique et la statistique a` propos du renouvellement du se´nat romain (Ier–IIIe sie`cles de l’Empire)’, Annales: E´conomies, socie´te´s, civilisations, 42: 1287–1303.

KASTEN, H. (19975), Marcus Tullius Cicero, An seine Freunde, lt.–Dt., Du¨sseldorf and Zurich.

References 349

KASTER, R. A. (1988), Guardians of Language: The Grammarian and Society in Late Antiquity, Berkeley, Los Angeles, and London.

——(1995), C. Suetonius Tranquillus De Grammaticis et Rhetoribus, Oxford. ——(2001), ‘Controlling Reason: Declamation in Rhetorical Education at

Rome’, in Y. L. Too (ed.), Education in Greek and Roman Antiquity, Leiden,

317–37.

——(2005), Emotion, Restraint and Community in Ancient Rome, Oxford. KERTZER, D. (1993), Sacrificed for Honor: Italian Infant Abandonment and the Politics of Reproductive Control, Boston.

KIERNER, C. (2004), Scandal at Bizarre: Rumor and Reputation in Jefferson’s America, Charlottesville.

KING, M. (2000), ‘Commemoration of Infants on Roman Funerary Inscriptions’, in G. J. Oliver (ed.), The Epigraphy of Death, Liverpool, 117–54.

KoCH, C. (1949), ‘Patrimi et matrimi’, in RE XVIII.4: 2250–2.

KoTIlA, H. (1992), Memoria mortuorum: Commemoration of the Departed in Augustine (Studia Ephemeridis Augustinianum 38), Rome.

KRASSER, H. (2002), ‘Poeten, Papageien und Patrone: Statius Silve 2, 4 als Beispiel einer kulturwissenschaftlichen Textinterpretation’, in J. P. Schwindt (ed.), Klassische Philologie inter disciplinas: Aktuelle Konzepte zu Gegenstand und Methode eines Grundlagenfaches, Heidelberg, 151–68.

KUNST, C. (2005), Ro¨mische Adoption: Zur Strategie einer Familienorganisation, Hennef.

LA MoNACA, V. (2007), ‘Festius: Un caso emblematico di delicatus?’, Epigraphica, 69: 169–80.

——(2008), ‘I Delicati nella Cisalpina’, in P. Basso, A. Buonopane, A. Cavarzere, and S. Pesavento Mattioli (eds.), Est enim ille Italiae flos... Vita economica e sociale nella Cisalpina romana. Atti delle giornate di studio in onore di Ezio Buchi, Verona 30 novembre – 1 dicembre 2006, Verona, 211–19.

LAES, C. (2003), ‘Desperately Different? Delicia Children in the Roman Household’, in Balch and Osiek 2003, 298–326.

——(2004a), ‘High Hopes, Bitter Grief: Children in Latin Literary Inscriptions’, in G. Partoens, G. Roskam, and T. Van Houdt (eds.), Virtutis Imago. Studies on the Conceptualisation and Transformation of an Ancient Ideal, Leuven, Paris, and Dudley, Mass., 43–77.

——(2004b), ‘Romeinse “paedagogi” tussen minachting en waardering: Een voorbeeld van “differential equations”’, Handelingen van de Koninklijke Zuid-Nederlandse Maatschappij voor Taal-en Letterkunde en Geschiedenis, 58: 161–80.

350 References

LAES, C. (2004c), ‘Children and Accidents in the Roman Empire’, Ancient Society, 34: 153–70.

——(2006), Kinderen bij de Romeinen: Zes eeuwen dagelijks leven, Leuven. ——(2010a), ‘When Classicists Need to Speak Up: Antiquity and Present

Day Pedophilia’, in V. Sofronievski (ed.), Aeternitas Antiquitas. Proceedings

of the Symposium Held in Skopje (Ziva Antika), Skopje, 30–59.

——(2010b, forthcoming), ‘The Educated Midwife in the Roman Empire: An Example of Differential Equations’, in H. F. J. Horsmanshoff (ed.), Hippocrates and Medical Education. Selected Papers read at the XIIth International Hippocrates Colloquium, Universiteit Leiden, 24–26 August 2005, Leiden.

LAGUNA, G. (1992), Estacio, Silvas III. Introduccion, edicion critica, traduccion e commentario, Madrid.

LAHUSEN, G. (1982), ‘Statuae et imagines’, in B. von Freytag et al. (eds.), Praestant Interna. Festschrift fu¨r Ulrich Hausman, Tu¨bingen, 101–9.

——(1985), ‘Zur Funktion und Rezeption des ro¨mischen Ahnenbildes’, Mitteilungen des Deutschen Archa¨ologischen Instituts, Ro¨mische Abteilung, 92: 261–89.

——(1995), ‘Sur l’origine et la terminologie des portraits romains’, in Le Regard de Rome: Portraits romains des muse´es de Me´rida, Toulouse et Tarragona, Toulouse, 246–59.

LAIRD, M. L. (2008), ‘Interpreting Children in Roman Imperial Art’ [Review of J. D. Uzzi, Children in the Visual Arts of Imperial Rome, Cambridge 2005], Journal of Roman Archaeology, 21: 509–14.

LANCIANI, R. (1892), Pagan and Christian Rome, Boston and New York. LASTEYRIE, R. DE (1878), ‘Notice sur un cimetie`re romain’, Revue Arche´ologique, 19/35: 371–83.

LAUBENHEIMER, F. (2004), ‘La Mort des tout petits dans l’Occident romain’, in Dasen 2004, 293–315.

LAVAGNE, H., BALANDA, E. DE, and URIBE ECHEVERRIA, A. (eds.) (2001), Jeunesse de la beaute´: La Peinture romaine antique, Paris.

LAWRENCE, R., and PATERSON, J. (1999), ‘Power and Laughter: Imperial Dicta’, Papers of the British School at Rome, 67: 183–97.

LAZZARONI, S. (1991), Sepulcra familiaria: Un’indagine epigrafico-giuridica, Padua.

LEACH, E. W. (1988), The Rhetoric of Space, Literary and Artistic Representations of Landscape in Republican and Augustan Rome, Princeton.

——(1990), ‘The Politics of Self-Presentation: Pliny’s Letters and Roman Portrait Sculpture’, Classical Antiquity, 9: 14–37.

References 351

LE GOFF, J. (1986), Storia e memoria, Turin [1st published in Enciclopedia Einaudi, Turin, 1977; French edn.: Paris, 1986; German trans. Berlin, 1999].

LEHNHERR, Y., and SCHUSTER CORDONE, C. (eds.) (2003), Au-dela` du visible: Reliquaires et travaux de couvents, Fribourg.

LEITH, D. (2008), ‘The Diatritus and Therapy in Graceo-Roman Medicine’, Classical Quarterly, 58: 581–600.

LELIS, A. A., PERCY, W. A., and VERSTRAETE, B. C. (2003), The Age of Marriage in Ancient Rome, Lewiston [etc.].

LENTANO, M. (1995), ‘Nupta verba—Praetextata verba: Considerazioni sul frammento 68 Lindsay di Plauto’, Aufidus, 25: 7–17.

LEO, F. (1901), Die griechisch-ro¨mische Biographie nach ihrer literarischen Form, Leipzig.

LE´VI-STRAUSS, C. (1967), Les Structures e´le´mentaires de la parente´, Paris. LEYERLE, B. (1997), ‘Appealing to Children’, Journal of Early Christian Studies, 5/2: 243–70.

LIFTON, R. J. (1977 [1st edn. 1973]), ‘The Sense of Immortality: On Death and the Continuity of Life’, in H. Feifel (ed.), New Meanings of Death, New York, 273–90.

LING, R. (1977), ‘Studius and the Beginnings of Roman Landscape Painting’, Journal of Roman Studies, 67: 1–16.

——(1991), Roman Painting, Cambridge.

——(1993), ‘The Paintings of the Columbarium of Villa Doria Pamphili in Rome’, in E. M. Moormann (ed.), Functional and Spatial Analysis of Wall Painting. Proceedings of the Fifth International Congress on Ancient Wall Painting (Amsterdam, 8–12 September 1992) (Babesch Suppl. 3), Leiden, 127–35.

——(1998), ‘Le Columbarium de la villa Doria Pamphili a` Rome’, in N. Blanc (ed.), Au royaume des ombres: La Peinture fune´raire antique, Catalogue d’exposition, Paris, 45–7.

LINKE, B., and STEMMLER, M. (eds.) (2000), Mos Maiorum: Untersuchungen zu den Formen der Identita¨tsstiftung und Stabilisierung in der ro¨mischen Republik, Stuttgart.

LINTOTT, A. (2008), Cicero as Evidence: A Historian’s Companion, Oxford. MCDONNELL, M. (2006), Roman Manliness: Virtus and the Roman Republic, Cambridge.

MAIURI, A. (1933), La casa del Menandro e il suo tesoro di argenteria, Rome. MALAISE, M. (1972), Les Conditions de pe´ne´tration et de diffusion des cultes e´gyptiens en Italie, Leiden.

MANCA MASCIADRI, M., and MONTEVECCHI, O. (1984), I contratti di baliatico (Corpus Papyrorum Graecorum 1), Milan.

352 References

MANTLE, I. C. (2002), ‘The Roles of Children in Roman Religion’, Greece & Rome, 49: 85–106.

MARGANNE, M.-H. (1991), ‘L’E´gypte me´dicale de Pline l’Ancien’, in Le Latin me´dical (Centre Jean Palerne, Me´moires X), Saint-E´tienne, Universite´ de Saint-E´tienne, 155–71.

MARINESCU, C. A., COx, S. E., and WACHTER, R. (2005), ‘Walking and Talking among Us: Personifications in a Group of Late Antique Mosaics’, in H. Morlier (ed.), La Mosaı¨que gre´co-romaine IX, II, Rome, 1269–77.

——————(2007), ‘Paideia’s Children: Childhood Education on a

Group of Late Antique Mosaics’, in Cohen and Rutter 2007, 101–14. MARROU, H.-I. (1957), Geschichte der Erziehung im klassischen Altertum,

Freiburg and Munich.

MARTIN, J. (1984), ‘Zur Stellung des Vaters in antiken Gesellschaften’, in H. Su¨ssmuth (ed.), Historische Anthropologie: Der Mensch in der Geschichte, Go¨ttingen, 84–109.

——(1997), ‘Zwei Alte Geschichten: Vergleichende historisch-anthropologische Betrachtungen zu Griechenland und Rom’, Saeculum, 48: 1–20.

——(2002), ‘Familie, Verwandtschaft und Staat in der ro¨mischen Republik’, in J. Spielvogel (ed.), Res Publica reperta: Zur Verfassung und Gesellschaft der ro¨mischen Republik und des fru¨ hen Prinzipats, Festschrift fu¨r Jochen Bleicken zum 75. Geburtstag, Stuttgart, 13–24.

——and NITSCHKE, A. (eds.) (1986), Zur Sozialgeschichte der Kindheit (Kindheit, Jugend, Familie 2), Freiburg and Munich.

——and ZOEPFFEL, R. (eds.) (1989), Aufgaben, Rollen und Ra¨ume von Frau

und Mann, I–II (Kindheit, Jugend, Familie 3), Freiburg and Munich. MARTINEZ-PINNA, J. (1995), ‘L’Oenoche´ de Tragliatella: Conside´rations sur

la socie´te´ e´trusque archaı¨que’, Studi Etruschi, 60: 79–92.

MAVROUDIS, A. D. (1986), ‘Critical Notes on a Medical Papyrus (PSI III 252)’, Aristotelian University of Thessalonica, Annual of the Faculty of Medicine, 16: 187–209.

MENCACCI, F. (1996), I fratelli amici: La rappresentazione dei gemelli nella cultura romana, Venice.

——(2001), ‘Genealogia metaforica e maiores colletivi (prospettive anthropologiche sulla costruzione dei Viri illustres)’, in Coudry and Spa¨th 2001, 421–37.

——(2005), Parentela e quasi parentela: Pedagoghi, nutrici e altri ruoli ‘vicari’ nella letteratura latina, Pisa.

MEYVAERT, P. (1959), ‘Les Responsiones de S. Gre´goire le Grand a` S. Augustin de Cantorbe´ry’, Revue d’Histoire Eccle´siastique, 54: 879–94.

References 353

——(1971), ‘Bede’s Text of the Libellus responsionum of Gregory the Great to Augustine of Canterbury’,in England before the Conquest: Studies in Primary Sources presented to Dorothy Whitelock, Cambridge, 15–33.

——(1986), ‘Le Libellus responsionum a` Augustin de Cantorbe´ry: Une oeuvre authentique de saint Gre´goire le Grand’, in J. Fontaine, R. Gillet, and S. Pellistrandi (eds.), Gre´goire le Grand, Paris, 543–50.

MICHEL, D. (1990), Casa dei Cei (I 6, 15) (Ha¨ user in Pompeji 3), Munich. MIELSCH, H. (1975), Roemische Stuckreliefs, Heidelberg.

——(2001), Ro¨mische Wandmalerei, Darmstadt.

——and VON HESbERG, H. (1995), Die heidnische Nekropole unter St. Peter in Rom, die Mausoleen E-I und Z-PSI, Rome.

MINARD, P. (1991), Gre´goire le Grand: Registre des lettres, I (Livres I et II), Paris.

MOMIGLIANO, A. (1971), ‘La liberta` di parola nel mondo antico’, Rivista storica italiana, 83: 499–544.

——(1973), ‘Freedom of Speech in Antiquity’, in Dictionary of the History of Ideas, II, New York, 252–63.

MOREAU, P. (1990), ‘Adfinitas: La Parente´ par alliance dans la socie´te´ romaine (Ier sie`cle av. J.-C. – IIe sie`cle ap. J.-C.)’, in Andreau and Bruhns 1990, 3–26.

——(2002), Incestus et prohibitae nuptiae: Conception romaine de l’inceste et histoire des prohibitions matrimoniales pour cause de parente´ dans la Rome antique, Paris.

MORGAN, T. (1998), Literate Education in the Hellenistic and Roman Worlds, Cambridge.

MOTOMURA, R. (1988), ‘The Practice of Exposing Infants and its Effects on the Development of Slavery in the Ancient World’, in T. Yuge and M. Doi (eds.), Forms of Control and Subordination in Antiquity, Leiden, 410–15.

MUDRY, P. (2004), ‘Non pueri sicut viri: Petit apercu de pe´diatrie romaine’, in Dasen 2004, 451–62.

MU¨LLER, E. W. (ed.) (1985), Geschlechtsreife und Legitimation zur Zeugung (Kindheit, Jugend, Familie 1), Freiburg and Munich.

MU¨LLER, R. J. (1993), ‘Tradierung religio¨sen Wissen in den Mysterienkulten am Beispiel von Andania’, in W. Kullmann and J. Althoff (eds.), Vermittlung und Tradierung von Wissen in der griechischen Kultur, Tu¨bingen, 307–16.

MU¨NZER, F. (1918a), s.v. Iunius (Brutus) No. 52, RE X.1: 972–3.

——(1918b), s.v. Iunius (Silanus) No. 163, RE X.1: 1090–1.

——(1920), Ro¨mische Adelsparteien und Adelsfamilien, Stuttgart.

354 References

MUSTAKALLIO, K., HANSKA, J., SAINIO, H.-L., and VUOLANTO, V. (eds.) (2005), Hoping for Continuity: Childhood, Education and Death in Antiquity and the Middle Ages, Rome.

NAGLE, B. R. (2004), The Silvae of Statius. Translated with Notes and Introduction, Bloomington, Indianapolis.

NANI, T. G. (1943–4), ‘Threptoi’, Epigraphica, 5–6: 45–84.

NATHAN, G. (2000), The Family in Late Antiquity: The Rise of Christianity and the Endurance of Tradition, London and New York.

NAUTA, R. R. (2002), Poetry for Patrons: Literary Communication in the Age of Domitian, Leyden.

NAvA, M. L., PARIS, R., and FRIGGERI, R. (eds.) (2007), Rosso Pompeiano: La

decorazione pittorica nelle collezioni del Museo di Napoli e a Pompei, Milan. NEILS, J., and OAKLEY, J. (eds.) (2004), Coming of Age in Ancient Greece:

Images of Childhood from the Classical Past, Hanover and New Haven. NE´RAUDAU, J.-P. (2008 [1st edn. 1984]), Eˆtre enfant a` Rome, Paris.

NEWLANDS, C. (2002), Statius’ Silvae and the Poetics of Empire, Cambridge. NILSSON, M. P. (1955 [1st edn. 1941]), Geschichte der griechischen Religion, Munich.

NORA, P. (1997 [1st edn. 1984–92]), Les Lieux de me´moire, I–III, Paris. NORBERG, D. (1982), S. Gregorii Magni registrum epistularum. Libri I–VII et

Libri VIII–XIV, I–II, Turnhout (Corpus Christianorum. Series Latina, CXL

et CXL A).

NUTTON, V. (1990), ‘The Patient’s Choice: A New Treatise by Galen’, Classical Quarterly, 40: 236–57.

——(1997), ‘Galen on Theriac: Problems of Authenticity’, in A. Debru (ed.),

Galen on Pharmacology: Philosophy, History and Medicine, Leiden, 133–51. OEPKE, A. (1934), ‘Amphithaleis in griechischen und hellenistischen Kult’,

Archiv fu¨r Religionswissenschaft, 31: 42–56.

OOST, S. I. (1955), ‘Cato Uticensis and the Annexation of Cyprus’, CPh 50: 98–112.

O’ROARK, D. (1999), ‘Parenthood in Late Antiquity: The Evidence of Chrysostom’, Greek, Roman and Byzantine Studies, 40: 53–81.

PARKER, R. (1983), Miasma: Pollution and Purification in Early Greek Religion, Oxford.

PATTERSON, C. (1985), ‘“Not Worth the Rearing”: The Causes of Infant Exposure in Ancient Greece’, Transactions of the American Philological Association, 115: 103–23.

PEDERZANI, O. (1995), Il talamo, l’albero e lo specchio: Saggio di commento a Stat. Silv. I, 2, II 3, III 4, Bari.

PEDICINI, R., and PEDICINI, L. (eds.) (1986), Le Collezioni del Museo Nazionale di Napoli, Naples.

References 355

PELLING, C. B. J. (1980), ‘Plutarch’s Adaptation of his Source Material’, Journal of Hellenic Studies, 100: 127–40.

PERTz, G. H. (1837), Monumenta Germaniae historica. Legum II, Pars altera.

Capitularia spuria. Canones ecclesiastici. Bullae pontificum, Hanovre. PETERS, W. J. T. (1963), Landscape in Romano-Campanian Mural Painting,

Assen.

——(1993), La Casa di Marcus Lucretius Fronto a Pompei e le sue pitture. Scrinium V, Amsterdam.

PETERSEN, J. M. (1994), ‘The Education of Girls in Fourth-Century Rome’, in D. Wood (ed.), The Church and Childhood (Studies in Church History 31), Cambridge, 29–38.

PETERSEN, L. H. (2006), The Freedman in Roman Art and Art History, Cambridge.

PIGEAUD, J. (1978), ‘Du rythme dans le corps: Quelques notes sur l’interpre´tation du pouls par le me´decin He´rophile’, Bulletin de l’Association Guillaume Bude´, 258–67.

POLY, J.-P. (2003), Le Chemin des amours barbares: Gene`se me´die´vale de la sexualite´ europe´enne, Paris.

POMEROY, A. G. (1992), ‘Trimalchio as Deliciae’, Phoenix, 46: 45–53. POMEROY, S. B. (1976), ‘The Relationship of the Married Woman to her Blood Relatives in Rome’, Ancient Society, 7: 215–27.

PRESCENDI, F. (1996), ‘Il lutto dei padri nella cultura romana’, in F. Hinard (ed.), La Mort au quotidien dans le monde romain, Actes du colloque organise´ par l’Universite´ de Paris IV (Paris, Sorbonne 7–9 octobre 1993), Paris, 147–54.

PROSPERI VALENTI, G. (1985), ‘Attori-bambini del mondo romano attraverso le testimonianze epigrafiche’, Epigraphica, 47: 71–82.

PULIATTI, S. (2001), Incesti crimina: Regime giuridico da Augusto a Giustiniano, Milan.

RAwSON, B. (1966), ‘Family Life among the Lower Classes at Rome in the

First Two Centuries of the Empire’, Classical Philology, 61: 71–83. ——(1986) (ed.), The Family in Ancient Rome: New Perspectives, London

and Sydney.

——(1989), ‘Spurii and the Roman View of Illegitimacy’, Antichthon, 23: 10–41.

——(1991) (ed.), Marriage, Divorce and Children in Ancient Rome, Oxford. ——(1997a), ‘Representations of Roman Children and Childhood’, Antichthon, 31: 74–95.

——(1997b), ‘The Iconography of Roman Childhood’, in Rawson and Weaver 1997, 205–32.

356 References

RAWSON, (1999), ‘Education, the Romans and Us’, Antichthon, 33: 81–98. ——(2003), Children and Childhood in Roman Italy, Oxford.

——(2005), ‘Circulation of Staff between Roman Households’, Zeitschrift fu¨r Papyrologie und Epigraphik, 151: 223–5.

——and WEAVER, P. (eds.) (1997), The Roman Family in Italy: Status, Sentiment, Space, Oxford and New York.

RAWSON, E. (1979), Cicero: A Portrait, London.

RICHLIN, A. (1992), The Gardens of Priapus: Sexuality and Aggression in Roman Humor, New York and Oxford.

RIETH, A. (1973), ‘Antike Goldmasken’, Antike Welt, 4: 28–34.

RIPOSATI, B. (1939), M. Terenti Varronis: De vita populi Romani. Fonti, esegesi, edizione critica dei frammenti, Milan.

ROLLER, M. B. (2006), Dining Postures in Ancient Rome, Princeton. ROMANO, G. (2005), ‘Luperques’, ThesCRA V: 89–91.

ROSENBAUM, E. (1960), A Catalogue of Cyrenaican Portrait Sculpture, London.

ROSSINI, O. (2006), Ara Pacis, Roma, Musei in Comune, Milan. ROSTOWZEW, M. (1904), ‘Pompejanische Landschaften und ro¨mische Villen’,

Jahrbuch des Deutschen Archa¨ologischen Instituts, 19: 103–26. ——(1911), ‘Die hellenistisch-ro¨mische Architekturlandschaft’, Mitteilungen

des Deutschen Archa¨ologischen Instituts, Ro¨mische Abteilung, 26: 1–186. ROUSSELLE, A. (1986), ‘Gestes et signes de la famille dans l’Empire romain’,

in Burguie`re et al. 1986, 231–69.

RUDHARDT, J. (1988), ‘Mne´mosyne et les Muses’, in P. Borgeaud (ed.), Me´moire des religions, Geneva, 37–62.

RU¨HL, M. (2003), ‘Confer gemitus pariterque fleamus! Die Epikedien in den Silven des Statius’, Hyperboreus, 9/1: 114–26.

——(2006), Literatur gewordener Augenblick: Die Silven des Statius im

Kontext literarischer und sozialer Bedingungen, Berlin and New York. SAGLIO, E. (1887), s.v. cera, in Daremberg and Saglio, I: 1019–20.

SALLER, R. P. (1984), ‘Familia, Domus, and the Roman Conception of the

Family’, Phoenix, 38: 336–55.

——(1994), Patriarchy, Property and Death in the Roman Family, Cambridge.

——(1997), ‘Roman Kinship: Structure and Sentiment’, in Rawson and Weaver 1997, 7–34.

SALZMAN, M. R. (2000), ‘Elite Realities and Mentalite´s: The Making of a Western Christian Aristocracy’, Arethusa, 33: 347–62.

——(2001), ‘Competing Claims to Nobilitas in the Western Empire of the Fourth and Fifth Centuries’, Journal of Early Christian Studies, 9/3: 359–85.

References 357

SCAMuZZI, U. (1964), ‘Studio sulla mummia di bambina cosidetta “Mummia di Grottarossa” rinvenuta a Roma, sulla via Cassia, il 5–2–1964’, Rivista di studi classici, 12: 264–80.

——(1965), ‘La “Mummia di Grottarossa”’, Aegyptus, 45: 74–83.

SCARpAT, G. (1964), Parrhesia: Storia del termine e delle sue traduzioni in latino, Brescia.

SCHA¨FER, T. (1989), Imperii insignia: Sella curulis und Fasces: zur Repra¨sentation ro¨mischer Magistrate, Mainz.

SCHEID, J. (1990), Romulus et ses fre`res: Le Colle`ge des fre`res arvales, mode`le du culte public dans la Rome des empereurs, Rome.

——(1998), La Religion des Romains, Paris.

——(2003), ‘Religion romaine et spiritualite´’, Archiv fu¨r Religionsgeschichte, 5: 198–209.

——(2007), ‘Le Sens des rites: L’Exemple romain’, in J. Scheid (ed.), Rites et croyances dans les religions du monde romain (Entretiens sur l’Antiquite´ classique, Fondation Hardt 53), 39–63.

——and SVENbRO, J. (2003), Le Me´tier de Zeus: Mythe du tissage et du tissu dans le monde gre´co-romain, Paris.

SCHEIDEL, W. (1997), ‘Quantifying the Sources of Slaves in the Early Roman Empire’, Journal of Roman Studies, 87: 156–89.

——(2005), ‘Human Mobility in Roman Italy, II: The Slave Population’, Journal of Roman Studies, 95: 64–79.

——(2009), ‘The Demographic Background’, in Hu¨bner and Ratzan 2009, 31–40.

SCHLANGE-SCHO¨NINGEN, H. (2003), Die ro¨mische Gesellschaft bei Galen: Biographie und Sozialgeschichte, Berlin and New York.

SCHLOSSER, J. VON (1910–11), ‘Tote Blicke: Geschichte der Portra¨tbildnerei in Wachs: ein Versuch’, Jahrbuch der Kunsthistorischen Sammlungen des Allerho¨chsten Kaiserhauses, 29: 171–258 (French trans. Histoire du portrait en cire, Paris, 1997).

SCHMITZ, T. A. (2002), Moderne Literaturtheorie und antike Texte: Eine Einfu¨hrung, Darmstadt.

SCHNEGG-KO¨HLER, B. (2002), Die augusteichen Sa¨kularspiele, Munich.

SCHOLZ, P. (2004), ‘Der Senat und die Intellektualisierung der Politik— Einige Bemerkungen zur Krise der traditionellen Erziehung in der spa¨ ten ro¨mischen Republik’, in C. Kretschmann, H. Pahl, and P. Scholz (eds.), Wissen in der Krise: Institutionen des Wissens im gesellschaftlichen Wandel, Berlin, 17–27.

——(2005), ‘Zur o¨ffentlichen Repra¨sentation ro¨mischer Senatoren und Magistrate: Einige U¨ b erlegungen zur (verlorenen) materiellen Kultur

358 References

der republikanischen Senatsaristokratie’, in T. L. Kienlin (ed.), Die Dinge als Zeichen: Kulturelles Wissen und materielle Kultur, Bonn, 409–31.

——(2006), ‘Imitatio Patris statt griechischer Pa¨dagogik: U¨ berlegungen zur Sozialisation und Erziehung der republikanischen Senatsaristokratie’, Jahrbuch des Historischen Kollegs 2005: 121–48.

——(2007), ‘Senatorisches Ma¨zenatentum: U¨ berlegungen zum Verha¨ltnis von Dichtern, Gelehrten und ro¨mischen nobiles in republikanischer Zeit’, in U. Oevermann, J. Su¨ßmann, and C. Tauber (eds.), Die Kunst der Ma¨chtigen und die Macht der Kunst: Untersuchungen zu Ma¨zenatentum und Kulturpatronage, Berlin, 25–46.

SCHOLZ, U. (1992), ‘Von der Bildung eines ro¨mischen Politikers’, Humanistische Bildung, 16: 25–37.

SHACKLETON BAILEY, D. R. (1965–70), Cicero’s Letters to Atticus, Lat.–Engl., 7 vols. (Cambridge Classical Texts and Commentaries 3–9), Cambridge. ——(1968), Cicero’s Letters to Atticus, III, CUP, Cambridge.

——(1980), Epistulae ad Quintum fratrem et M. Brutum, Lat.–Eng., (Cambridge Classical Texts and Commentaries 22), Cambridge.

——(1991 [1st edn. 1976]), Two Studies in Roman Nomenclature, New York. SHAW, B. D. (1987a), ‘The Family in Late Antiquity: The Experience of Augustine’, Past and Present, 115: 3–51.

——(1987b), ‘The Age of Roman Girls at Marriage: Some Re-considerations’, Journal of Roman Studies, 77: 30–46.

——(2001), ‘Raising and Killing Children: Two Roman Myths’, Mnemosyne, 54: 31–77.

SHERWIN-WHITE, A. N. (1966), The Letters of Pliny: A Historical and Social Commentary, Oxford.

SHRIMPTON, G. S. (1997), History and Memory in Ancient Greece, Montreal. SIGISMUND NIELSEN, H. (1990), ‘Delicia in Roman Literature and in the

Urban Inscriptions’, Analecta Romana Instituti Danici, 19: 79–88. ——(1991), ‘Ditis examen domus? On the Use of the Term “Verna” in the

Roman Epigraphical and Literary Sources’, Classica et Mediaevalia, 42:

221–40.

——(1998), ‘Roman Children at Mealtimes’, in I. Nielsen and H. Sigismund Nielsen, Meals in a Social Context: Aspects of the Communal Meal in the Hellenistic and Roman World, Aarhus, 56–66.

SLATER, W. J. (1974), ‘Pueri, turba minuta’, Bulletin of the Institute of Classical Studies, 21: 133–40.

SLIM, H. (1976), ‘Masques mortuaires d’El-Jem (Thysdrus)’, Antiquite´s Africaines, 10: 79–92.

——(1996), El Jem: L’Antique Thysdrus, Tunis.

——(2001), La Tunisie antique: De Hannibal a` Augustin, Paris.

References 359

SLUITER, I., and ROSEN, R. M. (2004), Free Speech in Classical Antiquity, Leiden.

SMITH, C. J. (2006), The Roman Clan. The Gens from Ancient Ideology to Modern Anthropology, Cambridge.

SMITH, R. R. R. (1981), ‘Greeks, Foreigners and Roman Republican Portraits’, Journal of Roman Studies, 71: 24–38.

SPA¨TH, T. (1998), ‘Faits de mots et d’images: Les Grands Hommes de la Rome ancienne’, traverse, 5/1: 35–56.

——(2006), ‘Geschlechter—Texte—Wirklichkeiten’, in C. Ulf and R. Rollinger (eds.), Frauen und Geschlechter: Bilder—Rollen—Realita¨ ten in den Texten antiker Autoren der ro¨mischen Kaiserzeit, Vienna, 39–76.

——(2007), ‘Blick auf Helden statt Blick auf Rom: Plutarchs Rezepte fu¨r ein globales Bankett der Moral’, in M.-L. Freyburger and D. Meyer (eds.), Visions grecques de Rome/Griechische Blicke auf Rom, Paris, 143–70.

SPINAZZOLA, V. (1953), Pompei alla luce degli scavi nuovi di Via dell’Abbondanza, I, Rome.

SPITZBART, G. (1997), Beda der Ehrwu¨rdige: Kirchengeschichte des englischen Volkes, Darmstadt.

STEIN-HO¨LKESKAMP, E. (ed.) (2006), Erinnerungsorte der Antike: Die ro¨mische Welt, Munich.

STINCHCOMB, J. (1932/3), ‘The Two Younger Tullii’, Classical Journal, 28: 441–8.

STOCKTON, D. (1979), The Gracchi, Oxford.

STONE, L. (1990), Road to Divorce: England 1530–1987, Oxford.

STUVERAS, R. (1969), Le putto dans l’art romain, Brussels.

SYME, R. (1939), The Roman Revolution, Oxford.

——(1960), ‘Bastards in the Roman Aristocracy’, Proceedings of the American Philosophical Society, 104: 323–7 (= Roman Papers, II, Oxford 1979, 510–17).

——(1980), ‘No Son for Caesar?’, Historia, 29: 422–37.

——(1986), The Roman Aristocracy, Oxford.

TANNER, J. (2000), ‘Portraits, Power, and Patronage in the Late Roman Republic’, Journal of Roman Studies, 90: 18–50.

TAYLOR, L. R. (1960), The Voting Districts of the Roman Republic: The Thirty-Five Urban and Rural Tribes, Rome.

TEMKIN, O. (1934), ‘Galen’s Advice for an Epileptic Boy’, Bulletin of the History of Medicine, 2: 179–89.

——(1971 [1st edn. 1945]), The Falling Sickness: A History of Epilepsy from the Greeks to the Beginnings of Modern Neurology, Baltimore.

THOMAS, R. (1989), Oral Tradition and Written Record in Classical Athens (Cambridge Studies in Oral and Literate Culture 18), Cambridge.

360 References

THOMAS, Y. (1983), ‘Paura dei padri e violenza dei figli: Immagini retoriche e norme di diritto’, in E. Pellizer and N. Zorzetti (eds.), La paura dei padri nella societa` antica e medievale, Rome and Bari, 113–40.

——(1984a), ‘Vitae Necisque Potestas: Le Pe`re, la cite´, la mort’, in Du chaˆtiment dans la Cite´: Supplices corporels et peine de mort dans le monde antique. Table ronde organise´e par l’E´cole franc¸aise de Rome avec le concours du Centre national de la recherche scientifique (Rome 9.–11. Novembre 1982), Rome, 499–548.

——(1984b), ‘Se venger au Forum. Solidarite´ familiale et proce`s criminel a` Rome’, in R. Verdier and J.-P. Ply (eds.), La vengeance. Vengeance, pouvoirs et ide´ologies dans quelques civilisations de l’Antiquite´, Paris, 65–100.

——(1986), ‘Rome: A Rome, pe`res citoyens et cite´ des pe`res (IIe sie`cle avant J.-C. – IIe sie`cle apre`s J.-C.)’, in Burguie`re et al. 1986, 195–229.

——(1988), ‘Le Droit romain: Le Traite´ des computs du juristeconsulte Paul’, in P. Legendre, Lec¸ons IV, Le Dossier occidental de la parente´: Textes juridiques inde´sirables sur la ge´ne´alogie, Paris, 27–119.

TILLY, L., FUCHS, R., KERTZER, D., and RANSEL, D. (1992), ‘Child Abandonment in European History: A Symposium’, Journal of Family History, 17: 1–23.

TIMMER, J. (2008), Altersgrenzen politischer Partizipation in antiken Gesellschaften, Berlin.

TREGGIARI, S. (1969), Roman Freedmen during the Late Republic, Oxford. ——(1975), ‘Family Life among the Staff of the Volusii’, Transactions of the American Philological Association, 105: 393–401.

——(1979), ‘Questions on Women Domestics in the Roman West’, in Schiavitu`, manomissione e classi dipendenti nel mondo antico (pubblicazioni dell’Istituto di storia antica, Universita` degli studi di Podova 13), Rome, 185–201.

——(1984), ‘Digna condicio: Betrothals in the Roman Upper Class’, E´chos du Monde Classique, 3: 419–51.

——(1991), Roman Marriage: Iusti Coniuges from the Time of Cicero to the Time of Ulpian, Oxford.

——(1998), ‘Home and Forum: Cicero between “Public” and “Private”’, Transactions of the American Philological Association, 128: 1–23.

——(2003), ‘Ancestral Virtues and Vices: Cicero on Nature, Nurture and Presentation’, in D. Braund and C. Gill (eds.), Myth, History and Culture in Republican Rome: Studies in Honour of T. P. Wiseman, Exeter, 139–64.

——(2005), ‘Putting the Family Across: Cicero on Natural Affection’, in George 2005, 9–35.

——(2007), Terentia, Tullia and Publilia: The Women of Cicero’s Family, London.

References 361

TURCAN, R. (1995), L’Art romain dans l’histoire: Six sie`cles d’expressions de la romanite´, Paris.

TYRRELL, R. Y., and PURSER, L. C. (1969 [1st edn. Dublin and London, 1914]), The Correspondence of M. Tullius Cicero, III, Hildesheim.

UZZI, J. D. (2005), Children in the Visual Arts of Imperial Rome, Cambridge. VALETTE-CAGNAC, E. (2003), ‘Eˆtre enfant a` Rome: Le Dur Apprentissage de la vie civique’, Terrain, 40: 49–64.

——(2004), ‘La Figure de la reticentia dans les traite´s de rhe´torique’, in L. Louvel and C. Rannoux (eds.), La Re´ticence, Rennes, 101–21.

VAN DAM, H.-J. (1984), Papinius Statius. Silvae Book II. A Commentary, Leiden.

VANSINA, J. (1985), Oral Tradition as History, Madison [= revised edn. of De la tradition orale, Tervuren 1961].

VATTUONE, R. (2004), Il mostro e il sapiente: Studi sull’ erotica greca, Bologna. VERSTRAETE, B. (1989), ‘Panegyric and Candour in Statius, Silvae, 3, 4’, in

C. Deroux (ed.), Studies in Latin Literature and Roman History, Brussels,

405–13.

VESSBERG, O. (1941), Studien zur Kunstgeschichte der ro¨mischen Republik, Lund and Leipzig.

VEYNE, P. (1991), La Socie´te´ romaine, Paris.

VIAZZO, P. P. (1994), ‘Family Structures and the Early Phase in the Individual Life Cycle: A Southern European Perspective’, in J. Henderson and R. Wall (eds.), Poor Women and Children in the European Past, London and New York, 32–50.

VOLLMER, F. (1898), P. Papinii Statii, Silvarum libri: herausgegeben und erkla¨rt, Leipzig.

VOLTERRA, E. (1939), ‘L’efficacia delle costituzioni imperiali emanate per le provincie e l’instituto dell’expositio’, in Studi di storia e diritto in onore di Enrico Besta, Milan, 449–77.

VO¨SSING, K. (2004), Mensa Regia: Das Bankett beim hellenistischen Ko¨nig und beim ro¨mischen Kaiser, Munich and Leipzig.

VOUT, C. (2007), Power and Eroticism in Imperial Rome, Cambridge. VUOLANTO, V. (2002), ‘Women and the Property of Fatherless Children in

the Roman Empire’, in P. Seta¨la¨ et al. (eds.), Women, Wealth and Power in

the Roman Empire, Rome, 203–43.

——(2003), ‘Selling a Freeborn Child: Rhetoric and Social Realities in the Late Roman World’, Ancient Society, 33: 169–207.

——(2009), ‘Choosing Asceticism: Children and Parents, Vows and Conflicts’, in C. Horn and R. Phenix (eds.), Children in Late Ancient Christianity (Studien und Texte zu Antike und Christentum 58), Tu¨bingen, 255–91.

362 References

WALKER, S. (ed.) (2000), Ancient Faces: Mummy Portraits from Roman Egypt, London.

WALTER, U. (2002), ‘Geschichte als Lebensmacht im republikanischen Rom’, Geschichte in Wissenschaft und Unterricht, 53: 326–39.

——(2003), ‘AHN MACHT SINN: Familientradition und Familienprofil im republikanischen Rom’, in K.-J. Ho¨lkeskamp et al. (eds.), Sinn (in) der Antike: Orientierungssysteme, Leitbilder und Wertkonzepte im Altertum, Mainz, 255–75.

——(2004), ‘Ein Ebenbild des Vaters: Familiale Wiederholungen in der historiographischen Traditionsbildung der ro¨mischen Republik’, Hermes, 132: 406–25.

WARBURG, A. M. (20083 [1st edn. 2000]), Der Bilderatlas Mnemosyne

(Gesammelte Schriften, II.1, M. Warnke and C. Brink, eds.), Berlin. WATSON, A. (1985), trans., The Digest of Justinian, Philadelphia, 4 vols. WATSON, P. (1992), ‘Erotion: Puella delicata?’, Classical Quaterly, 42: 253–68. WEAVER, P. R. C. (1972), Familia Caesaris, Cambridge.

——(1990), ‘Where Have All the Junian Latins Gone? Nomenclature and Status in the Roman Empire’, Chiron, 20: 275–305.

——(2001), ‘Reconstructing Lower-Class Roman Families’, in Dixon 2001b, 101–14.

WHITEHEAD, A. (2009), Memory, London.

WIEDEMANN, T. (1989), Adults and Children in the Roman Empire, London. WILLIAMS, W. (1990), Pliny the Younger: Correspondence with Trajan from Bithynia (Epistles X). Translation and Commentary, Warminster.

WLOSOK, A. (1978), ‘Vater und Vatervorstellungen in der ro¨mischen Kultur’, in H. Tellenbach (ed.), Das Vaterbild im Abendland, I, Rom, fru¨hes Christentum, Mittelalter, Neuzeit, Stuttgart, Berlin, and Cologne, 18–54.

WREDE, H. (1981), Consecratio in formam deorum: Vergo¨ttlichte Privatpersonen in der ro¨mischen Kaiserzeit, Mainz am Rhein.

YATES, F. A. (1966), The Art of Memory, London.

ZADOKS-JOSEPHUS JITTA, A. N. (1932), Ancestral Portraiture in Rome and the Art of the Last Century of the Republic, Amsterdam.

ZAMBON, S. (2005), Una voce tu sei e null’altro: L’usignolo nella tradizione culturale del mondo antico (Tesi di dottorato), Siena.

ZANKER, P. (1975), ‘Grabreliefs ro¨mischer Freigelassener’, Jahrbuch des Deutschen Archa¨ologischen Instituts, 90: 267–315.

ZIOLKOWSKI, J. M. (1998), ‘Obscenity in the Latin Grammatical and Rhetorical Tradition’, in J. M. Ziolkowski (ed.), Obscenity: Social Control and Artistic Creation in the European Middle Ages, Leiden, Boston, and Cologne.

ZOz, M. G. (1970), ‘In tema di obbligazioni alimentari’, Bullettino dell’Istituto di Diritto Romano ‘Vittorio Scialoja’, 73: 323–55.

MASTERS THESIS/PHD DISSERTATION FORMAT PROCEDURES SHEET

PLEASE READ CAREFULLY PRIOR TO SUBMITTING MASTERS THESIS/PHD DISSERTATION

I. FORMAT REQUIREMENTS

The body of the thesis must be formatted to fit these specific requirements:

• Paper must be of good quality and size 8.5”x11” or 21.5cmx28cm (for paper copies)

• Print must be of high quality with legible font (for paper copies)

• Text must be double spaced on single sided pages

• Left margin must measure 1.5” or 3.8cm

• Right, top and bottom margins must measure 1” or 2.5cm

• Page numbers, headers and footers must be within the margin measurements

• Ensure charts do not exceed page margin measurements

• Electronic version should be in a single PDF file.

The title page must include the following information:

• Title of thesis

• Name of author

• Degree to be awarded

• The words “Saint Mary’s University”

• Date of submission (title page date and signature page date must be the same)

• Names of Examining Committee members (do not include signatures)

The abstract must:

• Not exceed 150 words (300 words for PHD dissertations)

• Be single spaced

• Include the word “Abstract”

• Include author’s name

• Include thesis title

• Include date of submission (title page date and signature page date must be the same)

• Have print of good quality

Examples of title pages and abstracts may be viewed on the University Archives website (smu.ca/archives)

MASTERS THESIS/PHD DISSERTATION FORMAT PROCEDURES SHEET

page 1 of 4

TH 023 April 2018

II. COPYRIGHTED MATERIAL – What to consider when creating/submitting your thesis

• According to Library and Archives Canada “Students should ensure that the use of copyrighted material from other sources in their theses meets the requirements of the Copyright Act. Some written permissions from copyright holder(s) may be required”. Educational exceptions (such as fair dealing) may apply if certain conditions are met.

• Using a greater amount than what is considered an ‘insubstantial amount’ of a work requires a written letter of permission from the copyright holder. This letter must be addressed to you and submitted with your thesis.

• Full attribution must always be given to the original creator of the work, regardless of how much you are including in your theses.

• The Copyright Guide for Students (libguides.smu.ca/studentcopyright) on the library website goes into more detail on copyright procedures.

• If you are unsure about the copyright guidelines of the work you are using you can contact the library’s copyright office at copyright@smu.ca

III. LICENCE TO REPRODUCE

University regulations require all Masters theses/PHD dissertations be submitted to the National Library of Canada. Students completing a Masters/PHD degree are required to complete a “Non-Exclusive Licence to Reproduce Thesis” form. This form grants permission to the National Library of Canada to reproduce and sell the thesis/dissertation. This form and the accompanying “UMI” form must be submitted with the paper copies of the thesis.

IV. RESEARCH ETHICS BOARD

If you used human participants as part of your research (for example, conducted interviews or surveys), you will have obtained a “Certificate of Ethical Acceptability for Research Involving Humans” from the University’s Research Ethics Board. A copy of the certificate should be submitted with your project. This is in accordance with the Tri-Council Policy Statement: Ethical Conduct for Research Involving Humans (TCPS 2), article 2.1.

V. RESTRICTIONS

If your thesis contains confidential or sensitive material, you may wish to have your thesis restricted. This request must be submitted in writing in a letter addressed to the University Librarian, and the letter should be included with your theses submission. The request should specify the length of the restriction (up to a maximum of five years). Restricted theses will not be published for the length of the restriction.

VI. SIGNATURES OF EXAMINING COMMITTEE

It is required that the student submit one page with the names of the examining committee and their original signatures (not photocopied). This page must include all the elements that are on the title page (see section one above), as well as the signatures. This page will not be bound with the rest of your manuscript, and due to Canadian privacy law the title page that is bound with your manuscript cannot have signatures on it. See the University Archives web site (smu.ca/archives) for

Some departments have a “certification page”, which is separate from the title page and contains the names and positions of the examining committee. This is fine, as long as the certification page that is bound with your manuscript does not include signatures and you do include a separate, signed certification page.

MASTERS THESIS/PHD DISSERTATION FORMAT PROCEDURES SHEET

page 2 of 4

TH 023 April 2018

VII. COPYING PROCEDURES

When the thesis has been approved and signed by your advisor and you have complied with the formatting requirements:

1. Make paper copy

• submit one paper copy of the thesis/dissertation (for the Archives collection)

• submit a PDF version of your thesis/dissertation to fgsr@smu.ca

2. Fill out the Master’s Thesis/PHD Dissertation Information Sheet indicating where you can be reached in case of problems. Current and forwarding addresses are required. Be sure to review and sign the Thesis/Dissertation Checklist.

3. At the Graduate Studies Office (Atrium 210), submit:

• One paper copy (one PDF version of your thesis/dissertation is submitted online: see step one).

• the completed Master’s Thesis/PHD Dissertation Information Sheet, the completed Non-Exclusive Licence to Reproduce, and the UMI form

• Copyright letters, Research Ethics Certificates or Restriction Letters if applicable.

Graduate Studies forwards the thesis manuscripts to the Patrick Power Library which in turn processes and authorizes the manuscripts.

VIII. DISTRIBUTION OF THESES/DISSERTATIONS

• one copy is catalogued and added to the University Archives Collection (this will not be available immediately)

• Digital copies will be added to the Institutional Repository and to the National Library theses program

QUESTIONS?

If you have any questions about the thesis process, drop in or telephone the staff of the Faculty of Graduate Studies and Research at 420-5089 or University Archives at 496-8750.

FREQUENTLY ASKED QUESTIONS

1. Why do I need copyright permission if I’ve referenced the material in my thesis and bibliography? If you have taken material directly from an existing source you are required by University Senate regulation to obtain permission from the copyright holder in order to reproduce the material in your thesis. Referencing the material is not sufficient. If the material is public domain or open source, proof of this must be included. Only short quotations of another person’s work do not require copyright permission.

2. When will my thesis be available on online or in the library? You will be notified when the online copy is available. The Library copy of your thesis will be given a call number by the Cataloguing Department in the order in which it is received. This may take several months due to the schedule of the bindery and volume of items to be catalogued.

MASTERS THESIS/PHD DISSERTATION FORMAT PROCEDURES SHEET

page 3 of 4

TH 023 April 2018

Once you have read and understood the Format Procedures,

fill out and submit the Master’s Thesis/PHD Dissertation Information Sheet with your

manuscripts.

Be sure to review and sign the Thesis Checklist!

If you ave any uestions about t e formatting of your t esis /dissertation or t e submission rocess, lease

contact t e niversity rc ives 496-8750 or visit t e ebsite at tt :// .smu.ca/academics/arc ives. tml

urt er assistance in riting your t esis, including an in-de t t esis or s o , is available at t e riting Centre riting smu.ca, 491-6202 .

MASTERS THESIS/PHD DISSERTATION FORMAT PROCEDURES SHEET

page 4 of 4

TH 023 April 2018

Guoshen YU, 1/3

PROFESSIONAL EXPERIENCE

- Quantitative researcher - systematic trading

BlueCrest Capital Management LLP, Geneva, Switzerland 12/2010 – present

- Postdoctoral research associate

University of Minnesota, Electrical and Computer Engineering Dept., USA 09/2009 – 11/2010

Supervisor: Prof. Guillermo Sapiro

EDUCATION

- Ph.D., image and signal processing 09/2006 – 06/2009

Ecole Polytechnique, Center for Applied Mathematics, Palaiseau, France

“Sparse Grouping and Invariant Representations for Estimation and Recognition”

Advisor: Prof. Stéphane Mallat

Part of work under the supervision of Prof. Jean-Michel Morel (ENS Cachan) and Prof. Jean

Jacques Slotine (MIT).

Defended on June 30, 2009, with highest honors (mention très honorable).

Jury : Emmanuel Bacry, Michael Elad, Henri Maître, Stéphane Mallat, Jean-Michel Morel, Jean

Ponce, Guillermo Sapiro, Jean-Jacques Slotine.

- Visiting graduate student 02/2008 – 05/2008

MIT, Mechanical Engineering Dept, Cambridge, MA, U.S.A.

- M.Sc., applied mathematics and image processing 09/2005 – 07/2006

Ecole Normale Supérieure de Cachan, Cachan, France

With highest honors (mention très bien)

- Engineering Degree, signal and image processing 09/2003 – 07/2006

Ecole Nationale Supérieure des Télécommunications (Telecom Paris), Paris, France

- B.Sc., electronic engineering 09/1999 – 07/2003

Fudan University, Shanghai, China

INDUSTRY INTERN EXPERIENCE

• Let It Wave, Paris, France 01/2005 – 08/2005 Full-time research intern, with Prof. Stéphane Mallat

- Video deinterlacement and super-resolution (algorithmic research)

- Microscopic image restoration (kernel calibration and deconvolution)

• STMicroelectronics, Advanced System Technology Lab, Agrate, Italy 07/2004 – 12/2004 Full-time engineering intern, with Dr. Daniele Bagni

- AVS Video Coding Standard (Chinese version of H.264) research and its VLIW implementation

Guoshen YU, 2/3

PUBLICATIONS

Journal Papers

- G. Yu and G. Sapiro, Statistical Compressive Sensing of Gaussian Mixture Models, submitted,

arxiv.org/abs/1101.5785, Jan. 2011.

- G.Yu, G. Sapiro, and S. Mallat, Solving Inverse Problems with Piecewise Linear Estimators: From Gaussian Mixture Models to Structured Sparsity, submitted, arxiv.org/abs/1006.3056, June, 2010.

- J.M. Morel and G.Yu, Is SIFT Scale Invariant?, Inverse Problems and Imaging, vol. 5, no. 1, Feb., 2011

- S. Mallat and G. Yu, Super-Resolution with Sparse Mixing Estimators, IEEE Trans. on Image Processing, vol.19, issue 11, pp. 2889 - 2900, 2010.

- J.M. Morel and G. Yu, ASIFT: A New Framework for Fully Affine Invariant Image Comparison, SIAM Journal on Imaging Sciences, vol.2, issue 2, pp. 438-469, 2009.

- G. Yu and J.J. Slotine, Visual Grouping by Neural Oscillator Networks, IEEE Trans. on Neural Networks, vol.20, issue 12, pp. 1871-1884, 2009.

- G. Yu, S. Mallat, and E. Bacry, Audio Denoising by Time-Frequency Block Thresholding, IEEE Trans. on Signal Processing, vol 56, no. 5, pp. 1830-1839, May 2008.

Conference Proceedings

- G. Yu and G. Sapiro, Statistical Compressive Sensing of Gaussian Mixture Models, accepted to IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2011.

- F. Léger, G. Yu and G. Sapiro, Efficient Matrix Completion with Gaussian Models, accepted to IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2011.

- G. Yu, G. Sapiro, and S. Mallat, Image Modeling and Enhancement via Structured Sparse Model Selection, Proc. IEEE International Conference on Image Processing (ICIP), Hong Kong, 2010.

- S. Mallat and G. Yu, Structured Space Pursuits for Geometric Super-Resolution, invited paper, IEEE International Conference on Image Processing (ICIP), Cairo, 2009.

- G. Yu and S. Mallat, Sparse Super-Resolution with Space Matching Pursuits, Proc.SPARS'09, Saint-Malo, 2009.

- G. Yu and J.M. Morel, A Fully Affine Invariant Image Comparison Method, Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Taipei, 2009.

- G. Yu and J.J. Slotine, Audio Classification from Time-Frequency Texture, Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Taipei, 2009.

- G. Yu and J.J. Slotine, Visual Grouping by Neural Oscillators, Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Taipei, 2009.

- G. Yu and J.J. Slotine, Fast Wavelet-based Visual Classification, Proc. IEEE International Conference on Pattern Recognition (ICPR), Tampa, 2008.

- G. Yu, E. Bacry, and S. Mallat, Audio Signal Denoising with Complex Wavelets and Adaptive Block Attenuation, Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Hawaii, pp. 869-872, 2007.

Book Chapters

- G. Yu and G. Sapiro, Image Enhancement and Restoration, «Encyclopedia of Computer Vision», in press, Springer, 2011.

- A. Castrodad, I. Ramirez, G. Sapiro, P. Sprechmann, and G. Yu, Second Generation Sparse Modeling: Structured and Collaborative Signal Analysis, «Compressed Sensing: Theory and Applications», in press, Cambridge Univ. Press, 2011.

PATENTS

- S. Mallat and G. Yu, Video Enhancement Using Recursive Bandlets, PCT/IB2008/051770, pending, 2008.

- J.M. Morel and G. Yu, Viewpoint Invariant Object and Shape Recognition in Digital Images, FR 08/53244 pending, 2008.

- J.J. Slotine and G. Yu, Fast Pattern Classification Based on a Sparse Transform, US/PCT 13191, pending, 2009.

ONLINE DEMOS

- ASIFT demo: an online demo that allows you to try the ASIFT algorithm with your own images. (Until Sept 2010, the demo has received more than 8600 experiments from anonymous researchers and engineers all over the world. See [Morel and Yu 09], [Yu and Morel 09].) http://www.cmap.polytechnique.fr/~yu/research/ASIFT/demo.html

- Online lecture: Sparse geometric super-resolution. (Speaker Stéphane Mallat. See [Yu and Mallat, 09], [Mallat and Yu, 09].) http://videolectures.net/etvc08_mallat_sgsr/

Guoshen YU, 3/3

TALKS

- Solving Inverse Problems with Piecewise Linear Estimators, IEEE International Conference on Image Processing (ICIP), Hong Kong, Sept., 2010.

- Solving Inverse Problems with Gaussian Mixture Models, The Hong Kong Polytechnic University (invited by Professor Lei Zhang), Sept., 2010.

- Solving Inverse Problems with Piecewise Linear Estimators, IAS/Park City Mathematics Program (PCMI), June, 2010.

- Image Modeling and Enhancement with Structured Sparse Mixing Estimators, SIAM conference on imaging science (invited by Professors Michael Elad, Peyman Milanfar and Gabriel Peyré), Chicago, IL, U.S.A, March, 2010.

- Image Modeling and Enhancement with Structured Sparse Model Selection, Johns Hopkins University (invited by Professor Laurent Younes), Baltimore, MD, U.S.A., March, 2010.

- Structured Sparse Image Super-Resolution, University of Minnesota (invited by Professor Guillermo Sapiro), Minneapolis, MN, U.S.A., September, 2009.

- ASIFT: A New Framework for Fully Affine Invariant Image Comparison, EPFL (invited by Professor Martin Vetterli), Lausanne, Switzerland, June, 2009.

- ASIFT: A New Framework for Fully Affine Invariant Image Comparison, Jiaotong University (invited by Professor Yuncai Liu), Shanghai, China, April, 2009.

- Fully Affine Invariant Image Comparison, Fudan University (invited by Professor Yuanyuan Wang), Shanghai, China, August, 2008.

- Fully Affine Invariant Image Comparison, East China Normal University (invited by Professor Chunli Shen), Shanghai, China, July, 2008.

- Viewpoint Invariant Image Comparison, MIT (invited by Professor Jean-Jacques Slotine), Cambridge, MA, U.S.A., May, 2008.

- Audio Denoising by Time-Frequency Block Thresholding, IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Hawaii, U.S.A., April, 2007.

JOURNAL REVIEWER

IEEE Transactions on Image Processing, SIAM Journal on Imaging Sciences, IEEE Signal Processing Letters, IEEE Journal of Selected Topics in Signal Processing, IEEE Transactions on Circuits and Systems for Video Technology, Journal of Mathematical Imaging and Vision, Signal Processing, Applied and Computational Harmonic Analysis, Neurocomputing, Journal of Computer Science and Technology

HONORS

- ParisTech Best Thesis Prize 2010 (3 doctoral theses out of 550 defended in 2009 in ParisTech,

comprised of the twelve most prestigious French Grandes Ecoles including Ecole Polytechnique,

Mine ParisTech, Ecole des Ponts ParisTech, etc., were selected to award the prize.)

- CNRS (French National Center for Scientific Research) national recruitment competitions

2010

1st ex-aequo (199 candidates), information and communication sciences and technologies

2nd (66 candidates), applied mathematics

- Innovation Prize 2010, Ecole Polytechnique, finalist, 2nd ex-aequo

- GASPARD MONGE International Doctoral Fellowship, Ecole Polytechnique 2006 – 2008

- Best Poster Prize 2008, Xdoc, Ecole Polytechnique

LANGUAGE SKILLS

Chinese native language.

English fluent (TOEFL 277/300, Writing 6/6, living in the U.S.A. since 2009).

French fluent (living in France from 2003 to 2009).

Italian notation.

COMPUTER SKILLS

Languages C/C++, Matlab, VHDL, Java, Yorick, Assembly languages Intel 8086, Intel 8051 Operating systems Windows, Unix/Linux, Mac

INTERESTS AND SOCIAL ACTIVITIES

- Trainer of the Ping-pong Club of France Telecom R&D, Issy-les-Moulineaux 2007 – 2008

- Trombonist in the orchestra of Fudan University 2000 – 2003

- Vice chair, Students’ International Affairs Study Association, Fudan University 2000 – 2001

Don’t Fear the Bit Flips: Optimized Coding

Strategies for Binary Classification

Frederic Sala, Shahroze Kabir, Guy Van den Broeck, and Lara Dolecek

University of California, Los Angeles

fredsala, shkabir@ucla.edu, guyvdb@cs.ucla.edu, dolecek@ee.ucla.edu

Abstract

After being trained, classifiers must often operate on data that has been corrupted by noise. In this paper, we consider the impact of such noise on the features of binary classifiers. Inspired by tools for classifier robustness, we introduce the same classification probability (SCP) to measure the resulting distortion on the classifier outputs. We introduce a low-complexity estimate of the SCP based on quantization and polynomial multiplication. We also study channel coding techniques based on replication error-correcting codes. In contrast to the traditional channel coding approach, where error-correction is meant to preserve the data and is agnostic to the application, our schemes specifically aim to maximize the SCP (equivalently minimizing the distortion of the classifier output) for the same redundancy overhead.

I. INTRODUCTION

Noise is an enduring component of all computing and communication systems. Information is corrupted when transmitted over noisy channels [1], stored in unreliable memories [2], or processed by noisy or low-quality circuits [3]. Moreover, techniques aimed at saving power or increasing efficiency can increase noise. For example, voltage scaling further increases the probability of data corruption [4]. The traditional approach to handling noisy storage media is to implement strategies to detect and correct errors. For example, many memories and drives implement error-correcting codes [5].

The typical policy of such systems is to ensure that information read back is the same as it was when written. Such a policy represents a very strong constraint; implementations tend to be expensive, with significant storage overhead dedicated to redundant data. In this paper, we ask whether we may relax this constraint in the context of feature data for classifiers. Our goal is to reduce the effect of noise on the output of the algorithm; in other words, we consider the distortion of the algorithm outputs, rather than the inputs.

The basic setup is shown in Figure 1. We have a na¨ıve Bayes classifier with class variable C and binary features X1, X2, .. . , Xn. Learning is performed on noiseless training data. Afterwards, the classifier operates on noisy data; rather than the true features X1,..., Xn, only the corrupted features X1, X2, ... , Xn are used. Feature Xi’s value is flipped from 0 to 1 or from 1 to 0 with probability ei. A fundamental question is to determine the same classification probability (SCP), that is, the probability that the classifier output is the same for X1,..., Xn and for X1, ... , Xn.

Features can be protected by using an error-correction strategy. A set of redundant bits is computed when the data is written; before the data is used, both the data and redundant bits are decoded, with the goal of correcting some of the errors. This procedure is known as channel coding. The traditional approach is to protect all bits equally, uniformly reducing the eis for all n features. As we will see, this is not an efficient approach. Some features are more valuable and more worthy of protection than others. In fact, given a budget of redundancy, our major goal is an optimal resulting set of e1, ... , en to minimize the classifier distortion.

Fig. 1. Noisy classifier process model. The noiseless classification uses the features X1, X2,. . . , Xn. Noisy feature Xi is produced from Xi with the binary symmetric channel BSC(ei): Xi is equal to Xi with probability 1 ei and equal to ¯Xi with probability ei. Classification is performed on the noisy features X1, ... , Xn, without observing the true values X1, ... , Xn.

TABLE I

SCPS FOR VARIOUS REDUNDANCY ALLOCATIONS

Parameters

p0(x1) p0(x2) p1(x1) p1(x2) SCP by protection strategy Uniform None All x1 All x2

0.9 0.89 0.1 0.11 0.963 0.900 0.991 0.900

0.8 0.81 0.3 0.3 0.964 0.905 0.946 0.946

We illustrate this idea with a simple example. Take a classifier with n = 2 features X1, X2, noisy versions X1, X2, and uniform prior. The noise parameter for both binary symmetric channels (BSCs) is e = 0.1 by default. Let pi(xj) denote the conditional probability p(xj = 0C = i). We show how the SCP varies for different values of the pi(xj)’s and noise parameters. We also demonstrate how protecting features uniformly or non-uniformly affects the SCP. “Protection” refers to an error-correction scheme that effectively reduces the e noise parameter for a particular feature. Here, we allocate 4 additional redundancy bits for protection; uniform protection gives 2 such bits to each feature, while protection on Xi alone grants all 4 bits to Xi. (More precise definitions for this terminology are provided later on in the paper; the result of protection is a reduction of the Ei noise parameter for feature i.)

In the first row of Table I, X1 contains more information about C compared to X2. Observe that allocating all bits to X1 yields a better SCP than equal protection or protection on X2. Conversely, if we look at the last row we see that even though X2 contains more information about the class variable, the SCP is maximized by a uniform allocation of redundancy bits. Thus we cannot simply examine the conditional probabilities to decide how to allocate redundancy. Such results motivate us to seek an informed redundancy allocation strategy.

Our contributions are

• A framework for measuring the impact of noise by the same-classification probability (SCP),

• A low-complexity approximation for the SCP based on quantization and polynomial multiplication,

• A study of coding strategies based on allocating redundant bits in a way that minimizes the SCP (and thus the distortion on the classifier output). We show that, surprisingly, the empirical SDP is non-monotonic with respect to error protection. Afterwards, we give an optimization for optimal redundancy allocation based on a greedy approach.

II. PRIOR WORK

Channel coding for data protection is a vast field. Error-correction has been proposed for disk drives [6], write-once memories [7], and for RAID architectures [8]. More recently, techniques have been developed to take advantage of the specific properties of modern non-volatile memories such as flash [9], [10]. Distributed storage systems can be protected through replication or erasure coding [11], [12]. Solid state drives can also benefit from sophisticated error-correction [13]. A common aspect of such research, in contrast to our work, is that the goal is to preserve the data being stored without considering the application. That is, error-correction is found in a different abstraction layer.

The research in [14] introduces an informed channel coding scheme for linear regression. Here, the error model was also bit flips applied to a finite-width binary representations of integers. The goal was to minimize the distortion on the algorithm outputs. An approximation scheme was introduced by considering the contribution to the distortion by a flip in a single bit independently of the others.

There are a number of papers that have examined issues related to learning algorithms dealing with uncertainty. Recent work on information dropout, a technique introduced in [15] to prevent neural networks from overfitting, has considered adding noise to the activations of deep neural networks [16]; unlike our work, here noise is used during learning.

The robustness of various algorithms under noise exposure has been considered in a series of papers. [17] performed an experimental study of algorithms when data sets were corrupted by synthetic noise. Deleted features were tackled in [18] and [19]. [18], which is perhaps the most closely related work to the present paper, also considers corrupted features, and introduces two techniques to tackle noise, one based on linear programming, and the other using an online batching scheme. However, unlike our work, this paper considers only adversarial (not probabilistic) noise and does not propose optimized channel coding. [19] uses a game-theoretic approach to avoid over-reliance on a feature that could be deleted. Another work tackling the case of missing data in Bayes classifiers is [20].

Noise and coding are often important issues in distributed learning. For example, [21] introduced optimality guarantees for distributed estimators in a setting where nodes can be isolated, but limited communication between them is allowed. [22] considered problems including hypothesis testing and parameter estimation in the case of multiterminal algorithms where each terminal has a data compression constraint.

A related area is known as value of information [23], [24]. The idea is to maximize an expected reward on some observed features. One example is to observe those features that maximize the information gain [25].

Decision robustness in the case of hidden variables can be measured by the same decision probability (SDP) [26], [27]. SDP has been applied to the evaluation of adaptive testing [28] and of mammography-based diagnosis [29] The same classification probability we employ in this paper bears some similarity to the SDP. The problem of handling label noise is considered in

[30]. Label noise in crowdsourcing is tackled through coding-theoretic means in [31]. Additionally, our work can be viewed as a form of feature selection [32], although we do not remove features, but rather allow them to be noisier than certain more critical features.

III. NOISE IN BINARY CLASSIFICATION

We use the standard notation of upper-case letters for random variables and lower-case letters for their instantiations. Boldface denotes vectors of variables.

Throughout this paper we consider binary classification problems with binary features. For some test point x = (x1, x2, ... , xn ) with n binary features, we denote the conditional probabilities by αi = p(xi|C = 0) and βi = p(xi|C = 1) for 1 <, i <,n.

Throughout this paper, we employ a na¨ıve Bayes classifier such that x is classified to

( )

n n

c(x) = arg max p(C = 0)  αi, p(C = 1)  βi.

i=1 i=1

Taking logarithms, x is classified according to to the value of

log p(C = 0)

c(x) = sgn p(C = 1) + n



i=1 !log(αi/βi) . (1)

A. Impact of Noise

In this work, we employ a simple noise model: we define an noise parameter vector e = (1, 2, ... , n) with 0 <, i < 1/2 for 1 <, i <, n. The binary feature Xi is flipped to its opposite value ¯Xi with probability i and stays unchanged with probability 1  i. Using information theory terminology, we view this operation as placing a binary symmetric channel on each of the features, as shown in Figure 1. We express the resulting error vector as E = (E1, E2, ... , En), where Ei = 1 if an error has occurred for feature i and 0 otherwise. As a special case, we consider the vector e = (, , ... , ), where the error probability is identical on all features. In this setting, the probability of some error instantiation e is a function of its Hamming weight

wt(e): Pr(e) = wt(e)(1  )nwt(e).

B. Same Classification Probability

Now we examine the effect of the feature bit errors on the output of the classification algorithm. Our goal is to determine when noise does not impact the algorithm output. That is, we wish to compute the probability that the noiseless point X and the noisy version X(E) = (x01, ... , x0n) have the same classification:

SCP(X,E) = Pr(c(X) = c(X(E))). (2)

We call this quantity the (real) same classification probability. We are mainly concerned with an empirical version of the SCP for a feature realization x,

SCP(x, E) =  Pr(c(x) = c(x(e)))Pr(e). (3)

e{0 , 1}n

The SCP resembles the same decision probability (SDP) introduced in [26], defined as the probability of producing a decision (using evidence e) confirmed by Pr(d|e > T) as when some hidden variable H is revealed, h Pr(d|e, h)Pr(h|e). A key difference from SDP is that SCP considers two separate distributions: the augmented distribution that captures the data generation process (Figure 1), and the original na¨ıve Bayes network where classification is performed. Instead, SDP is calculated a single Bayesian classifier that is assumed to capture both processes.

If ej = 1 and thus there is a feature bit flip at position j, the corresponding term in (1) changes as

αj! 1 αj!

log  log .

βj 1  βj

αj 1αj

We denote Aj := log βj and Bj := log 1βj . Let us write Dj for the difference

Dj = Bj  Aj = log 1 αj 1  βj ! αj !

 log

βj

An error ej = 1 replaces Aj in (1) with Bj; equivalently, Dj is added to this sum. Therefore, for some error vector e with ` errors given by the components ei1, ei2, ... , ei` = 1, the noisy classification is given by

Producing the same classification requires c(x(e)) = c(x), or, equivalently,

Without loss of generality, fix the noiseless classifier output to c = 0, so that the sign function value is positive1. Next, write

log p(c = 0)

T :=  p(c = 1) +

Here, T represents a target value. Then, the equality (5) is equivalent to



j= 1

Recall that the Dij correspond to all the components in e that are equal to 1. Then, to sum over all e in (3), we must determine all the subsets of D1, D2, ... , Dn with sum greater than the target T. A subset of size f corresponds to an error vector with f feature bit errors that does not change the classification relative to the noiseless version. Such a subset contributes a term (given by (3)) to the same classification probability.

Example: Consider a model with n = 3 binary features and equal error probability c = (e, e, e). Take D1 + D2 > T. This implies that the error vector e = (1, 1, 0) does not change the classification output, contributing a probability term of e2( 1  e) to the SCP.

IV. SCP APPROXIMATION

In this section, we consider issues surrounding the computation of the SCP. Our main result is an efficient SCP approximation. A. Exact Empirical SCP

The problem described by (6) is a variant of the subset sum problem with real numbers. The na¨ıve approach is to enumerate

the 2n subsets of D1, D2, ... , Dn and compute each of their sums. This task is not tractable for large n.

A compact representation of the SCP is the function

We define [> T]G(z) to be the sum of the coefficients of terms in G(z) with exponent larger than T. It is easy to see that

SCP(x,E) = [> T]G(z).

Any subset Di1, Di2, ..., Dij  with sum greater than T produces a term in G(z) with exponent Di1 + ... + Dij > T and coefficient

ei1ei2 ••• eij 

i=i1,...,ij (1  ei).

The sum of all such coefficients is indeed the SCP. This may lead us to attempt to expand the function G(z) and examine the resulting coefficients. However, the fact that the exponents are real-valued prevents us from doing so efficiently.

B. Quantization

Observe that if the exponents of terms in G(z) were non-negative integers, G(z) would be a polynomial, allowing for fast multiplication. In fact, G(z) would become a generating function [33]. Inspired by this notion, we introduce the following quantization scheme.

The key idea is to quantize Di into k buckets, for k a constant. By performing the quantization in a clever way, we induce a structure that enables us to approximate the SCP in no more than O(n2 log n) operations. The approximation is described in Algorithm 1; the concept is detailed below.

Our quantization scheme only relies on the minimal and maximal values of D1, ... , Dn. Let Dmin = minD1, ... , Dnand Dmax = maxD1, ... , Dn. Let DI = Dmax  Dmin. Consider the family of k intervals

1If the noiseless classifier output is c = 1, the outcome is that we seek subsets whose sum is less than the target value T rather than greater. All of the arguments we present in this paper are unchanged. For simplicity of notation, we do not add the case to our setup.

Algorithm 1 SCP Approximation

Input: Difference terms D1, D2, ... , Dn, Error probabilities e1, e2, ... , en, Target T, Number of buckets in quantization

scheme k

Output: SCP approximation SCPapp

Initialize S1, ... , Sk to 0, SCPapp to 0

for j = 1 ton do

if Dj  [Ebegin i , Eend

i ) then

Si  Si + 1

ei Si  ej

end if

end for

expand G[i, j] = ki= 1 Si

j= 1(( 1  ei j) + ei j yzi)

for i = 0 to n do ( ( 11

Tk1

DI T  i Dmin  DI

k 1

for j > 0 do

if G[i, j] > T then

SCPapp  SCPapp + G[i, j]

end if

end for

1J [ 1J

DI DI

2(k  1), Dmin + , Dmin + DI

2(k  1), Dmin + 3DI , ... ,

2(k  1) 2(k  1)

[ 1J [ 1J

Dmin + (2k  5)DI

2(k  1) , Dmin + (2k  3) DI , Dmin + (2k  3)DI

2(k  1) , Dmax + (2k  1)DI . (7)

2(k  1) 2(k  1) Note that each bucket has width Di/ (k  1). For compactness, we also write the ith interval as [Ebegin

i , Eend

i ).

We quantize any value of Dj that falls into a bucket with the midpoint of the bucket interval. We compute Si, the number

of Dj in each bucket:

{ li [ 1J

Si = Dj : Dj  Dmin + (2i  3)DI

2(k  1) , Dmin + (2i  1)DI }, ,

, 1 <- j <- n

2(k  1)

for 1 <- i <- k. In addition, we re-label the noise parameters such that ei1, ei2, ... , eiSi are the noise parameters corresponding

to the features Dj that fall into the ith bucket.

Recall that we seek to determine the subsets of  D1, D2, ... , Dn that have sum greater than T. We approximate any Dj

that falls into the ith interval with the midpoint of that interval:

i  1

Dmin + k  1 DI. (8) Additionally, observe that Dmin and Dmax falls are represented by their own values.

We approximate the SCP by computing the SCP on the quantized versions of the Di; we show that this can be performed efficiently by Algorithm 1.

Consider the multivariate generating function

We write [f, > R]G(y, z) for the sum of all coefficients of terms with exponent f in y and exponent greater than R in z. Note that in the case where e = (e, . . . , e), G(y, z) reduces to

G(y, z) =  ((1  e) + eyzi)Si.

i= 1

We show that the appropriate coefficients in the generating function yield the SCP for the quantized versions of the Di:

Theorem 1. Under the quantization scheme given by Algorithm 1, the resulting SCP is given by

SCPapp(x,E) = n

f=0 [f,> T(f)]G(y, z),

where

( ( / /

k  1

T(f) := T  ~ Dmin  DI .

DI k  1

Proof: Consider some subset with size f and sum greater than the target T in the quantization scheme given by (7). We distribute our f choices of Dj into the k buckets, where we may take up to Si values from bucket i. Let (a1, a2, ... , ak) be such a choice; ai, for 0 <, ai <, Si represents the number of Dj falling into the ith bucket. A subset of size f requires a1 + a2 + ... + ak = f. The total sum for the subset is given by

(a1Dmin + a2 Dmin + / DI

k  1 + ... + akDmax.

Here we used the fact that terms in each bucket are quantized to the midpoint of that bucket (i.e., expression (8)). Since the sum of the subset is greater than T, we may write

( / ( /

(a1 +a2+...+ak) DminDI DI 2DI kDI

+ a1

k  1 k  1 + a2 k  1 + ... + ak > T

k  1

( /

= f Dmin  DI + DI

k  1 (a1 + 2a2 + ... + kak) > T

k  1

= a1 + 2a2 + . .. + kak >

Now we relate the generating function given in (9) to the equation a1 + 2a2 + ... + kak > T. The generating function is the product of n binomials. To generate a term in the expansion, we select from each binomial either 1  ei j or ei j yzi. The subsets of size f satisfying a1 + 2a2 + ... + kak > T correspond to choosing a1 terms e1 jyz, a2 terms e2 jyz2, and so on. The products of such terms have degree f in y and degree a1 + 2a2 + ... + kak > T in z. Considering the subsets of size 0, 1, ... , n, we indeed have that SCPapp(x, E) = En~=0[ f, > T]G(y, z).

C. Computation

To compute SCPapp(x, E), we must perform the expansion of the generating function G(y, z) = j jki= 1 j jSi

j=1((1 ei j) + ei jyzi) and examine the coefficients. Polynomial multiplication is equivalent to convolution (of vectors or arrays, depending on whether the polynomials are univariate or multivariate); this effort can be sped up by using the Fast Fourier Transform (fft) [34]. In the Fourier domain, convolution is equivalent to multiplication. That is, for two polynomials q(x), r(x),

q(x)r(x) = ifft(fft(q)  fft(r)),

where the multiplication on the right side is performed term by term and ifft denotes the inverse transform. If the maximum degree of q(x) and r(x) is d, the number of operations required is O(d log d).

First, we expand the terms j jSi

j=1((1  ei j) + ei j yzi). Replace yzi by v; then, we are multiplying Si polynomials of degree 1. This can be performed in O(Si log log Si) operations. This fact is not difficult to check: our Si monomials can be paired up and the pairs multiplied; the Si/ 2 resulting degree-2 polynomials are paired up, and so on. The total time is O(S1 log log S1) + O(S2log log S2) + ... + O(Sk log log Sk) <, O(n log log n). We note that each expansion can be performed in parallel.

Note that in the case of identical e, the polynomial to be expanded is ((1 e) +eyzi)Si, and the expansion can be performed in O(Si) operations with the binomial theorem.

Next, each of the resulting expansions must be multiplied. The expansion of j jSi

j=1((1  ei j) + ei j yzi) can be represented as a two-dimensional array, where rows represent the exponent of y and columns the exponent of z. In our case, we have k arrays. The maximal degree for y is n, while the maximal degree for z is kn, so that the product of all of the arrays (and thus any intermediate multiplication) has size at most n  kn. Each multiplication has cost at most O(kn2log(kn)); the total cost is thus no more than O(k2n2log(kn)). (Again, it is possible to speed to perform the multiplications in parallel for speedup.)

Number of buckets k

Fig. 2. SCP approximation error (against the real SCP) for the two approximation schemes. The parameters here are e = 102 and n = 16. 50 data points were selected randomly; the true SCP was computed and the average absolute error was measured for various k in the two schemes.

D. Approximation Quality

Next we comment on the quality of the approximation. Clearly, the quality is a function of the number of buckets k; the larger the k, the finer the approximation. We note, however, that the approximation error is not a monotonic function, since there are edge effects for different k, as can be seen in Figure 2.

We introduce two improvements to optimize the worst-case error of the SCP approximation. We describe these errors as a function of the noise parameter e, and for simplicity take the case where all the components equal to e. The first improvement comes at no extra cost and reduces the worst-case error to O(e2), while the second improvement requires O(nr) time but reduces the worst-case error to O(er+ 1). We adopt the first improvement for our experiments.

Improvement 1. For the first improvement, consider the intervals defined in (7). If our target value T lines somewhere in [Dmin, Dmax], we can translate (shift) all of our intervals over by some constant Tshift that is no larger than half an interval width, DI/ 2 (k  1), such that T is now precisely the edge between two intervals. Then, any Di < T is mapped to a bucket center smaller than T, while any Di > T is mapped to a bucket center larger than T. We conclude that any single-element subset of D1, . . . , Dn that contributes to the SCP also contributes to the SCP approximation; thus, all SCP approximation errors must be for two- or more element subsets, reducing this error to O(e2).

Improvement 2. We refer to the second improvement as a hybrid SCP approximation. The idea here is simple: we use the approximation only for the subsets of size greater than r for some constant 2 < r < n. For those subsets that are of size r or smaller, we check the sums using the actual values of Di, requiring at most O(nr) time. This strategy is most suitable for very small values of r, such as r = 2.

Example. An evaluation was performed on the small data set house-votes-84 from the UCI repository [35]. This set has n = 16 binary features corresponding to ‘yes’ or ‘no’ votes on various congressional proposals; the binary class represents democrat or republican. The small size of n enables us to compute the real SCP value (though much slower than the approximate SCP). Figure 2 shows the error versus the true SCP for the two approximation schemes and various k averaged over 50 random test points; the noise parameter was e = 102.

V. OPTIMIZED CHANNEL CODING

Thus far we have only considered the impact of noise without taking any action to protect the algorithm output from distortion. In this section, we consider a simple but effective form of error correction. Our goal is to tailor this strategy to the characteristics of the features in order to minimize the classifier distortion. This is a departure from traditional error protection, which is typically present in a different abstraction layer and is agnostic to the algorithm.

Consider storing multiple copies of a certain feature bit; these copies are produced when the data is written and before the addition of noise. Afterwards, when the bit is ready to be used, the repeated values of the bit (some of which may now be corrupted) are subject to a majority vote (equivalent to maximum-likelihood decoding) and the decoded value is placed back into the algorithm.

What is the effect of repetition on the error probability e? If the jth feature xj, corrupted by a bit flip with probability e, is repeated 2r + 1 times for some r > 0, the probability e(r) that the feature is decoded incorrectly is equal to the probability

Algorithm 2 Channel Code Optimization

Input: Test points x1, x2 ... , xt, Noise vector e = (E, ... , E), redundancy budget r

Output: Optimized redundancy allocation vector rr

Initialize r(0) to 0

for j = 0 to r  1 do

i  arg mini 1t ti= 1 SCPapp(xt, Er(j)+1(i)) )

r(j+ 1)  r(j) +1(i)

end for

that the majority of votes are corrupted:

For example, if r = 1, the smallest meaningful repetition, E( 1 ) = E3 + 3E2(1  E). Note that we always add pairs of repetitions, since an even number of repeated bits can produce ties. In general, employing r pairs of repeated bits drives the error probability E to O(E1 +r).

A. Monotonicity

Intuitively, we expect that reducing the error probability on features will increase the SCP. Clearly, reducing all of the Ei to 0 yields an SCP of 1, as the only non-zero term in the SCP is (1  E1) (1  E2) × ... × (1  En) = 1.

Nevertheless, the SCP does not necessarily increase when we reduce noise parameters E. The following result is applicable to both the SCP and the SCP approximation:

Theorem 2. Let x be a feature vector and e = (E1, ... , Ei1, Ei, Ei+ 1, . . . , En) and E' = (E1, . . . , Ei1, E'i, Ei+ 1, . .. ,En) be noise parameter vectors with Ei > E'i Let the associated error vectors be E, E', respectively. If Di > 0, then SCP(x, E) >SCP(x, E'), while if Di <, 0, then SCP(x, E) <, SCP(x, E').

Proof: Every summand in SCP(x, E) is either a multiple of Ei (if it corresponds to a subset of the Djs that includes Di) or a multiple of (1  Ei) (if it corresponds to a subset of the Dj that does not include Di). We write

SCP(x,E) = EiC1 + (1  Ei)C2,

where C1, C2 are non-negative.

Case 1) Di > 0. Consider the subsets of {D1, ... , Dn} \{Di}. corresponding to C1. These subsets must have sum larger than T  Di, since including Di, their sum must exceed T. In the case of C2, these subsets have sum larger than T. Since Di is positive, any subset with sum larger than T has sum larger than T  Di as well, so all subsets corresponding to C2 also correspond to terms in C1. We may write C1 = C2 + C3 for some C3 > 0. Then,

SCP(x, E) = EiC1 + (1  Ei)C2

= Ei(C2 +C3) + (1 Ei)C2

= C2 + EiC3.

Since C3 > 0, reducing Ei to E'i reduces the SCP as well.

Case 2) Di <, 0. Then, T  Di > T, so all subsets corresponding to C1 also correspond to C2. Thus, we can write

SCP(x, E) = EiC1 + (1  Ei)C2

= EiC1 + (1  Ei) (C1 + C3)

=C1 + (1 Ei)C3.

Reducing Ei to E'i increases (1  Ei)C3, so the SCP is increased.

In words, noise helps the SCP when applied to those features that disagree with the classification, since flipping these bits increases confidence (and can cover up for other bit flips that reduce it). However, noise hurts the SCP when applied to features that agree with the classification; the sign of Di is a direct consequence of this idea.

The fact that reducing the error probability (by adding redundancy, etc.) is not always helpful leads us to seek an optimized solution. Clearly a uniform allocation of protection for all features is not always a good idea.

TABLE II

EXPERIMENTAL RESULTS FOR CLASSIFICATION CHANGE PROBABILITY (1-SCP)

Movie dataset, feature set 1 Movie dataset, feature set 2 Voting dataset NLTCS dataset

R Uniform Optimal Ratio Uniform Optimal Ratio Uniform Optimal Ratio Rb Optimal No protection

1 0.10112 0.08664 1.1670 0.1263 0.10245 1.2328 0.05635 0.04135 1.3628 6 0.0131 0.0218

2 0.03484 0.03024 1.1524 0.04419 0.03028 1.4595 0.03182 0.02377 1.3390 7 0.0120 0.0218

3 0.01154 0.01002 1.1513 0.01469 0.00879 1.6711 0.01903 0.01429 1.3319 8 0.0108 0.0218

4 0.00382 0.00331 1.1535 0.00486 0.00256 1.9028 0.01154 0.0085 1.3555 9 0.0097 0.0218

5 0.00127 0.00109 1.1576 0.00162 0.00077 2.1019 0.00702 0.00504 1.3911 10 0.0086 0.0218

6 0.00043 0.00037 1.1602 0.00055 0.00024 2.2549 0.00423 0.00298 1.4328 11 0.0074 0.0218

7 0.00015 0.00013 1.1622 0.00019 7.70E-05 2.3983 0.00261 0.00177 1.4742 12 0.0062 0.0218

8 4.94E-05 4.24E-05 1.1636 6.30E-05 2.48E-05 2.539 0.0016 0.00105 1.5182 13 0.0052 0.0218

9 1.69E-05 1.45E-05 1.1646 2.16E-05 8.08E-06 2.6720 0.00098 0.00063 1.5632 14 0.0044 0.0218

10 5.83E-06 5.00E-06 1.1654 7.43E-06 2.64E-06 2.8109 0.00060 0.00038 1.6107 15 0.0037 0.0218

Total redundancy budget / n

Fig. 3. Classification change probability (1  SCP) for a movie reviews data set with n = 20 features and a test set of size of t = 250. Comparison of the results of optimized versus uniform redundancy allocation.

B. Greedy Optimization

In the remainder of this work, we introduce an algorithm to optimize the allocation of a redundancy budget for coded feature protection. We use a budget of 2r redundancy bits, which will be used to protect our n features. Feature i is then represented by i+ 2ri copies, and has error probability e(ri), for 1 <_ i <_ n. The noise vector for all n features C(r) = (e(r1),e(r2), ... ,e(rn)). The corresponding error vector is written E(r). The values r = (r1, r2, ... , rn) are constrained by r1 + r2 + ... + rn = r. To minimize the distortion due to noise on point x, we must maximize the SCP with respect to r and the resulting e vector. We have the following optimization

Our first step is to perform a tractable version of this optimization using our SCP approximation:

Although we can now efficiently check each value of r, doing so is still computationally expensive. The number of solutions to r1 + r2 + ... + rn = r with ri > 0 is given by (n+r1

r ). We turn to a greedy (myopic) optimization approach to further reduce the complexity. This approach enables us to perform the optimization one redundancy unit (two repeated bits) at a time. After the jth step, we write the redundancy vector as r(j) = (r1(j), r2(j), ... , rn(j)). In the (j + 1)st step, one of the ri(j) terms is selected and increased by 1. For ease of notation, let us write 1(i) for the vector (0, 0,. . . , 0, 1, 0,. . . , 0) with a 1 in the ith position and 0s elsewhere. Then, the (j + 1)st step is given by

arg max

i SCPapp(x, Er(j)+1(i))) s.t. 1 <_ i <_ n.

This reduces our complexity to that of performing n SCP computations per each of the r steps. Of course, we can also perform the optimization over t test data points. The procedure is given in Algorithm 2

2.5

1.5

1 0 2 4 6 8 10

Total redundancy budget / n

Fig. 4. The classification change probability (uniform to optimized allocation) ratio reaches 3. The setup is the same as that of Figure 2.

We demonstrate our approach with SCP experiments using the voting dataset [35], the movie review dataset from [36], and the NLTCS dataset in [37]. In all cases, we used k = 50 for the number of buckets in our SCP approximations. For the movie reviews set we used two sets of n = 20 features and 250 test data points. The e parameter was 0.2 for the first 2 datasets (a very high amount of noise that can nevertheless be handled through channel coding). Table II shows the classification change (1  SCP) probability for uniform and optimized assignment of redundancy bits given a budget of 2nR bits. We report the ratio of the classification change probability (uniform versus optimized); this ratio reaches up to 3.

In the movie review dataset, the first feature set included features with similar dependence on the class variable. The second feature set included a mixture of features (some more informative of the class variable than others). In this case the optimized assignment has a clear benefit over the uniform assignment. The results on the voting dataset (using n = 16 features and t = 454 test data points) are in between those of the two movie review sets. A plot of the SCP and the improvement ratio for the more beneficial assignment is shown in Figures 3 and 4.

For the NLTCS dataset, we tested the case of a small amount of redundancy Rb < 15, so that a uniform allocation is impossible, and we must rely on the optimization. Here, n = 15 features and t = 3236 test data points were used. The results are compared to the unprotected version.

VI. CONCLUSION

We studied binary na¨ıve Bayes classifiers operating on noisy test data. First, we characterized the impact of bit flips due to noise on the classifier output with the same classification probability (SCP). We introduced a low-complexity approximation for the SCP based on quantization and polynomial multiplication. Next, we considered minimizing the classifier distortion by allocating redundant bits among the features. Our informed approached for redundancy allocation is among the first principled methods combining coding theory with machine learning. We demonstrated the results of this idea with experiments.

REFERENCES

[1] C. E. Shannon, “Communication in the presence of noise,” Proceedings of the IEEE, vol. 86, no. 2, pp. 447–457, 1998.

[2] L. Dilillo, P. Girard, S. Pravossoudovitch, A. Virazel, and M. B. Hage-Hassan, “Data retention fault in SRAM memories: analysis and detection procedures,” in Proceedings of the 23rd IEEE VLSI Test Symposium (VTS), 2005.

[3] S. Mittal, “A survey of techniques for approximate computing,” ACM Computing Surveys, vol. 48, no. 4, 2016.

[4] V. Chandra and R. Aitken, “Impact of technology and voltage scaling on the soft error susceptibility in nanoscale CMOS,” in Proc. Defect and Fault Tolerance of VLSI Systems (DFTVS ‘08), (Boston, MA), 2008.

[5] V. Sridharan and D. Liberty, “A field study of DRAM errors,” in Proc. International Conference for High Performance Computing, Networking, Storage and Analysis (SC), (Salt Lake City, UT), 2012.

[6] C. M. Riggle and S. G. McCarthy, “Design of error correction systems for disk drives,” IEEE Transactions on Magnetics, vol. 34, no. 4, pp. 2362–2371, 1998.

[7] R. L. Rivest and A. Shamir, “How to reuse a “write-once memory”,” Information and Control, vol. 55, no. 1-3, pp. 1–19, 1982.

[8] M. Blaum, J. Brady, J. Bruck, and J. Menon, “EVENODD: an efficient scheme for tolerating double disk failures in RAID architectures,” IEEE Transactions on Computers, vol. 44, no. 2, pp. 192–202, 1995.

[9] Y. Cassuto, M. Schwartz, V. Bohossian, and J. Bruck, “Codes for asymmetric limited-magnitude errors with application to multilevel flash memories,” IEEE Transactions on Information Theory, vol. 56, no. 4, pp. 1582–1595, 2010.

[10] L. Dolecek and F. Sala, “Channel coding methods for non-volatile memories,” Foundations and Trends in Communications and Information Theory, vol. 13, no. 1, pp. 1–136, 2016.

[11] H. Weatherspoon and J. Kubiatowicz, “Erasure coding vs. replication: A quantitative comparison,” in Proceedings of First International Workshop on Peer-to-Peer Systems (IPTPS ’01), 2002.

[12] A. G. Dimakis, P. B. Godfrey, Y. Wu, M. J. Wainwright, and K. Ramchandran, “Network coding for distributed storage systems,” IEEE Transactions on Information Theory, vol. 56, no. 9, pp. 4539–4551, 2010.

[13] K. Zhao, W. Zhao, H. Sun, T. Zhang, X. Zhang, and N. Zheng, “LDPC-in-SSD: Making advanced error correction codes work effectively in solid state drives,” in Proceedings of Conference on File and Storage Technologies (FAST ‘13), (San Jose, CA), 2013.

[14] K. Mazooji, F. Sala, G. Van den Broeck, and L. Dolecek, “Robust channel coding strategies for machine learning data,” in Proc. Annual Allerton Conference on Communication, Control, and Computing (Allerton), pp. 609–616, 2016.

[15] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: a simple way to prevent neural networks from overfitting,” Journal of Machine Learning Research, vol. 15, no. 1, pp. 1929–1958, 2014.

[16] A. Achille and S. Soatto, “Information dropout: Learning optimal representations through noisy computation,” tech. rep., University of California, Los Angeles, 2016.

[17] E. Kalapanidas, N. Avouris, M. Craciun, and D. Neagu, “Machine learning algorithms: a study on noise sensitivity,” in Proc. 1st Balcan Conference in Informatics, (Thessaloniki, Greece), pp. 356–365, 2003.

[18] O. Dekel and O. Shamir, “Learning to classify with missing and corrupted features,” in Proc. 25th International Conference on Machine Learning (ICML 2008), (Helsinki, Finland), 2008.

[19] A. Globerson and S. Roweis, “Nightmare at test time: robust learning by feature deletion,” in Proc. 25th International Conference on Machine Learning (ICML 2006), (Pittsburgh, PA), 2006.

[20] M. Ramoni and P. Sebastiani, “Robust bayes classifiers,” Artificial Intelligence, vol. 125, no. 1-2, pp. 209–226, 2001.

[21] J. C. Duchi, M. I. Jordan, M. J. Wainwright, and Y. Zhang, “Optimality guarantees for distributed statistical estimation,” tech. rep., University of California, Berkeley, 2014.

[22] T. S. Han and S. Amari, “Statistical inference under multiterminal data compression,” IEEE Transactions on Information Theory, vol. 44, no. 6, pp. 2300–2324, 1998.

[23] D. Heckerman, E. Horvitz, and B. Middleton, “An approximate nonmyopic computation for value of information,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 15, no. 3, pp. 292–298, 1993.

[24] M. Bilgic and L. Getoor, “Value of information lattice: Exploiting probabilistic independence for effective feature subset acquisition,” Journal of Artificial Intelligence Research (JAIR), vol. 41, pp. 69–95, 2011.

[25] A. Krause and C. Guestrin, “Optimal value of information in graphical models,” Journal of Artificial Intelligence Research (JAIR), vol. 35, pp. 557–591, 2009.

[26] A. Choi, Y. Xue, and A. Darwiche, “Same-Decision Probability: A confidence measure for threshold-based decisions,” International Journal of Approximate Reasoning (IJAR), vol. 2, pp. 1415–1428, 2012.

[27] S. Chen, A. Choi, and A. Darwiche, “Algorithms and applications for the Same-Decision Probability,” Journal of Artificial Intelligence Research (JAIR), vol. 49, pp. 601–633, 2014.

[28] S. Chen, A. Choi, and A. Darwiche, “Computer adaptive testing using the same-decision probability,” in Proceedings of the Twelfth UAI Conference on Bayesian Modeling Applications Workshop, pp. 34–43, 2015.

[29] F. J. Gimenez, Y. Wu, E. S. Burnside, and D. L. Rubin, “A novel method to assess incompleteness of mammography reports,” in AMIA Annual Symposium Proceedings, vol. 2014, p. 1758, American Medical Informatics Association, 2014.

[30] B. Frenay and A. Kaban, “A comprehensive introduction to label noise,” in Proc. European Symposium on Artificial Neural Networks, Comp. Intelligence and Machine Learning (ESANN), (Bruges, Belgium), 2014.

[31] A. Vempaty, L. R. Varshney, and P. K. Varshney, “Reliable crowdsourcing for multi-class labeling using coding theory,” IEEE Journal of Selected Topics in Signal Processing, vol. 8, no. 4, pp. 667–679, 2014.

[32] I. Guyon and A. Elisseeff, “An introduction to variable and feature selection,” Journal of Machine Learning Research, vol. 3, pp. 1157–1182, 2003.

[33] H. S. Wilf, Generatingfunctionology. A K Peters/CRC Press, 3rd ed., 2005.

[34] K. R. Rao, D. N. Kim, and J.-J. Hwang, Fast Fourier Transform - Algorithms and Applications. Springer Publishing Company, 1st ed., 2010.

[35] M. Lichman, “UCI machine learning repository,” 2013.

[36] J. Bekker, J. Davis, A. Choi, A. Darwiche, and G. Van den Broeck, “Tractable learning for complex probability queries,” in Proceedings of Conference on Neural Information Processing Systems (NIPS), (Montreal, Canada), 2015.

[37] D. Lowd and J. Davis, “Learning markov network structure with decision trees,” in Proc. IEEE International Conference on Data Mining (ICDM), (Sydney, Australia), 2010.

International Journal of Strategic Property Management

ISSN: 1648-715X (Print) 1648-9179 (Online) Journal homepage: http://www.tandfonline.com/loi/tspm20

Valuation of properties in close proximity to waste

dumps sites: The Nigeria experience

Victoria Amietsenwu Bello & Mustapha Oyewole Bello

To cite this article: Victoria Amietsenwu Bello & Mustapha Oyewole Bello (2009) Valuation of properties in close proximity to waste dumps sites: The Nigeria experience, International Journal of Strategic Property Management, 13:4, 309-317

To link to this article: https://doi.org/10.3846/1648-715X.2009.13.309-317

Published online: 18 Oct 2010.

Article views: 646

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=tspm20

International Journal of Strategic Property Management (2009) 13, 309–317

VALUATION OF PROPERTIES IN CLOSE PROXIMITY TO WASTE DUMPS SITES: THE NIGERIA EXPERIENCE

Victoria Amietsenwu BELLO 1 and Mustapha Oyewole BELLO 2 

1 Department of Estate Management, Federal University of Technology, Akure, Nigeria E-mail: vicbellofuta@yahoo.com

2 Department of Estate Management, Federal University of Technology, Akure, Nigeria E-mail: oyewolebello@yahoo.com

Received 4 August 2009; accepted 28 September 2009

1. INTRODUCTION

The assessment of a property usually de-pends on the property’s unique characteristics, each of which provides utilities or disutility to individuals. These characteristics are gener¬ally classified into those which are external and those that are internal to the property (Mackmin, 1985). External influences relate to the general state of the economy, population, employment, immigration, finance, location, transportation and environmental attributes, while the internal influences essentially con-stitute the specific details of the property such

as size, accommodation, condition, design, lay¬out, age, type and plot size (Adair et al., 1996; Bello, 2000). These value determining factors can be attributable to the nature of property as a package of goods and services (Bello and Bello, 2008). Therefore, property extends be¬yond shelter (the physical attributes of a prop¬erty) to include environmental characteristics or attributes which refers to the quality of the neighborhood and location within the neigh-borhood. Environmental characteristics are manifested in the form of pull and push effect of the neighbourhood. The push effects could be characterized by proliferation of squatter

International Journal of Strategic Property Management

http://www.ijspm.vgtu.lt

DOT: 10.3846/1648-715X.2009.13.309-317

310 V. A. Bello and M. O. Bello

settlements, air and water pollution, squalid condition of environmental sanitation, and breakdown of waste disposal arrangement (Bello and Bello, 2008) while the pull effect could be good roads, schools etc. When prop-erties are in close proximity to environmental factors which could lead to push effects like waste dump sites, they demand special as¬sessment on the part of the Estate Surveyors and Valuers. The question therefore, is what techniques do the Nigerian Valuers adopt in valuing properties close to waste dump sites? Are the techniques different from those used in properties not close to dump site or any en-vironmental hazards?

The remainder of the paper is organized as follows. The next section (section 2) deals with previous research on the subject matter. Section (3) three presents a description of the methodology employed for the study, while sec¬tion (4) four focuses on the empirical results. The concluding remarks and policy implication are contained in section (5) five.

2. REVIEW OF PREVIOUS

LITERATURE

The pricing of properties close to waste dump sites have been studied extensively since the early 1980s using variety of methods (Liz-ieri et al., 1995). The methods range from the traditional market Comparison Approaches, the direct application of Cost through Yield Decomposition Techniques to Explicit Dis-counted Cash Flow Scenario Modeling Ap-proach. In Nigeria, the traditional Valuation methodologies are used for pricing properties. The methods ranged from the Direct Compari¬son Method, the Cost Method to the Income Approach (Ogunba, 1997; Ogunba et al., 2005).

The Direct Capital Comparison Method al¬lows valuers to analyze specifically the results of buyers and sellers interaction and to use market evidence (Sale price) to value other properties. While the method may be appro¬

priate in assessing the value of property under normal circumstances, it however, presents some difficulties with respect to valuation of properties close to waste dump sites, as Ped¬erson (2002) noted that it would be difficult to make a meaningful comparative analysis since contamination can be specific to a property.

The Cost Method on its own is based upon the premise that the value of property is ap-proximated by the investment necessary to re¬place that property (Acks, 1995). Replacement Cost typically includes land acquisition, the cost of site and building improvements, and an allowance for the developers’ profit, less ac¬crued depreciation (Acks, 1995). Wilson (1994) offered a cost based methodology for estimat¬ing the effects of waste (contamination) on val¬ue. He stressed that the estimation of the val¬ue of properties contaminated by waste must consider the “negative impact of intangible factors.” Among these factors are the general demand for the subject in the market place, the level of confidence in remediation and its cleaning, the availability of financing for the contaminated property and the possibility of third-party liabilities. In Wilson’s framework, the quantified effect of these intangible factors is deducted from an unimpaired value together with remediation costs and any quantifiable effect of use restrictions. In spite of Wilson (1994) argument in favour of Cost Method, Marchitelli (1992) questioned the use of the method and suggested that the method should be abandoned in the valuation process in most situations. This is because depreciation, which is one of the components used in Cost Meth¬od, is difficult to measure especially when the property is close to environmental hazards (waste dump sites in this case). If therefore, depreciation is to be accounted for explicitly, it must be on a realistic basis. To argue that depreciation has been taken into account by an implied means as in a “traditional” valuation, when there is no consensus as to the correct methods or amount is farcical.

Valuation of Properties in Close Proximity to Waste Dumps Sites: the Nigeria Experience 311

The Income Approach is based on the prin¬ciple that annual values and capital values are related to each other and that given the in¬come a property produces or its annual value, the capital value can be found (Kinnard, 1970; Millington 1990 and Baum et al., 1997). Vari¬ants of this method have been developed over the years, as Real Value Approach propounded by Wood (1972), which has been amended and reconciled with the Equated Yield Approach to produce a Real Value/Equated Yield Hybrid by Crosby (1985 and 1991). Also, included in the model are the Rational Approach based on the earlier work of Greaves (1972), and further developed by Sykes (1984) and the Discounted Cash Flow Technique etc. These techniques (variants of income approach) incorporate year-by-year projections of income and expenses and discounts the cash flows by an appropriate rate, and add the present value of the property at some future date. The methodology is ami¬cable to income streams that do not conform to a pattern. When used for contaminated prop¬erties, the cost of remediation is central. For instance, Mundy (1992a) and Chalmers and Roehr (1993) opted for a Discounted Cash Flow Technique with multiple discount rates corre¬sponding to the contamination related stigma risk for each year of the presumed holding pe¬riod. The underlying concept echoes the com¬ments of Patchin (1994) and Mundy (1992b) in that risk is not constant over the holding period. Indeed the authors suggested that the effect of stigma and its risk premium may de¬cline over the holding period and subsequent to remediation. Furthermore, Neustein (1992) demonstrated a technique whereby simple in¬come ratio can be used in a direct capitaliza¬tion formula rather than actual difference in net operating income. This allowed Neustein (1992) to make a graduated set of comparisons of different capitalization rate premium and income ratios in terms of their effect on value. This simplifies the problem, but may not allow for final adjustments based on other variables.

In spite of these approaches, several ana-lytical approaches are currently being used to measure environmental externalities and the extent to which they are internalized into real estate values (Furby et al., 1988). Well known to property appraisers are the Paired Sales Analysis used by Jackson (2001), Contingent Valuation Analysis (CVA) used by Carson, (2000); Clinch and Murphy, (2001), Option Pricing Model by Lentz and Tse (1995) and the Hedonic Model developed by Rosen (1974). These techniques have been used to estimate the benefits of things such as increased air and water quality as a result of contaminants like waste, increased risk from drinking water and groundwater contaminants, outdoor rec¬reation, and protecting wetlands, wilderness areas and endangered species on property val¬ues at one point or the other (Carson, 2000; Clinch and Murphy, 2001). However, some of these techniques have been criticized by prop¬erty valuation experts. For instance Des Ros-iers et al. (1999) noted that the Paired Sales Analysis is somewhat speculative because the inherent heterogeneity of property market makes it difficult to isolate the price impact of a particular attribute. In this regard, Lan¬caster (1966), Rosen (1974), Damodaran (2006) and Des Rosiers and Theriault (2006) noted that the Hedonic Approach remains the most reliable and acceptable tool for valuing proper¬ties that are affected by environmental exter¬nalities as it reveals buyers perceptions of any potential environmental hazard through their actual pricing behaviour. Hedonic Model is a statistical technique used to isolate the effect and contribution of various housing attributes to real estate prices (Rosen, 1974). However, Rosen (1974) approach has been applied to a wide range of discipline among which are hous¬ing and environmental economics (Bajari and Kahn, 2005). In the hedonic model, a house is seen as a bundle of characteristics, and its price is given as a function of those character¬istics. Homebuyers maximize their utility by

312 V. A. Bello and M. O. Bello

purchasing the house that falls within their budget and contains the most characteristics in monetary terms (Vanderford et al., 2005). Theoretically, a homebuyer purchases a set of characteristics, not a house per say. Hence it’s growing popularity among urban economist and property appraisers (Des Rosiers et al., 1999). The hedonic technique has been used in numerous studies to determine the effects of various characteristics on house prices and to compare the prices of homes that differ on key characteristics. The value that a characteristic adds to the price of a house can be thought of as an implicit price – the expected benefits over time – for the characteristic, where it is assumed that the expected benefits of a given characteristic are the same for all potential consumers (Vanderford et al., 2005). Such im¬plicit price determination is particularly help¬ful for characteristics that are not priced in-dependently, as is the case for most housing characteristics.

The general consensus in the literature is that the traditional method of valuation is in-adequate in the valuation of environmentally contaminated property. With this in view, the study identified and evaluated the predomi¬nant methods in use among Nigerian practi¬tioners.

3. DATA AND METHODOLOGY EMPLOYED

The target population for the study is the registered Estate Surveyors and Valuers in Lagos whose names are contained in an up-dated 2002 Directory published by the Estate Surveyors and Valuers Registration Board of Nigeria (ESVRBON). The total number of reg¬istered practicing Estate Surveyors and Valu¬ers whose names are in the directory are 228 of which a sample size of 107 was chosen us¬ing the simple random technique. 107 Ques¬tionnaires were administered on these Estate Surveyors and Valuers out of which 99 were

filled and returned for analysis using percen-tiles and mean score.

Using the methodology of Finlay and Tyler (1991) and Bello (2002) the respondents were asked of their awareness, usage and under¬standing of the methods of valuation used in practice.

In evaluating the techniques used in valu-ing contaminated properties, the fifty one (51) Estate Surveyors and Valuers (those who are registered) were engaged to value eight recent¬ly sold properties, two located in close proxim¬ity to each of the waste dump site at Oluso-sun, Solous and Abule-Egba respectively and the other two located at Ketu where there is no identifiable waste dump site in sight. Their valuations were compared with the actual sales price of the properties. The dispersion between these valuations and the sales price were analyzed and tested using t - statistic.

4. EMPIRICAL ANALYSIS AND DISCUSSION OF RESULTS

Tables 1 and 2 show the qualification and status of the respondents. In Table 1, 51 out of the 99 respondents are registered or deemed to be registered Estate Surveyors and Valu¬ers. This category was considered for further analysis. This was to ensure that the views of only those qualified to assess property values according to Decree No 24 of 1975 were con¬sidered.

Also, the survey according to Table 2 shows that over 70 percent of the respondents held key position in their respective organisations. This lays credence to the information collected for this analysis.

Table 3 shows that majority of the Estate Surveyors and Valuers have heard, used and understood mainly the five traditional methods (Comparative, Cost, Investment, Residual and Profit) of valuation. Although, 52.94 percent of the Estate Surveyors and Valuers have heard of the Discounted Cashflow Techniques, only

Valuation of Properties in Close Proximity to Waste Dumps Sites: the Nigeria Experience 313

Table 1. Distribution of respondents according to qualification

Qualification No of respondents Percentage no of respondents

HND only 14 14.14

B.SC only 29 29.30

B.SC/ HND with ANIVS/Fellow* 44 44.44

M.SC with ANIVS* 7 7.07

Others 5 5.05

Total 99 100

Source: Field data (2008)

* Registered Estate Surveyors and Valuers (44 + 7 = 51)

Table 2. Distribution of respondents according to status

Status No of respondents Percentage no of respondents

Principal partners 22 22.22

Branch managers 16 16.16

Senior estate surveyors and valuers 42 42.42

Estate surveyors and valuers 15 15.15

Others 4 4.05

Total 99 100

Source: Field data (2008)

Table 3. Awareness of the specific techniques of valuing contaminated properties

Techniques

Awareness of methodology Usage of methodology Understanding of methodology

No Percentage No Percentage No Percentage

Paired sales Analysis 6 11.76 1 1.96 3 5.88

Contingent Valuation Analysis (CVA) 5 9.80 0 0.00 3 5.88

Option Pricing Model 6 11.76 4 7.84 5 9.80

Discounted Cash Flow Techniques 27 52.94 14 27.45 18 35.29

Hedonic approach 10 19.61 0 0.00 1 1.96

Comparative method 51 100 51 100 51 100

Cost method 51 100 51 100 51 100

Investment method 51 100 47 92.16 51 100

Residual method 51 100 28 54.90 30 58.82

Profit method 51 100 15 29.41 38 74.51

Source: Field data (2008)

27.45 have actually used the method. Also, other available techniques are the Paired Sales Analysis, the Contingent Valuation Analysis, Option Pricing Model and the Hedonic Model. For the Paired Sales Analysis and the Option

Pricing Model, 11.76 percent of the Estate Sur¬veyors and Valuers have heard of the method¬ology respectively. 1.96 percent of the Estate Surveyors and Valuers have actually used the Paired Sales Analysis while 7.84 percent of the

314 V. A. Bello and M. O. Bello

Valuers have used the Option Pricing Model. Also, 9.80 percent and 19.61 percent of the Estate Surveyors and Valuers have heard of the Contingent Valuation Analysis (CVA) and Hedonic Model respectively while none of the Estate Surveyors and Valuers has used the Methodologies.

Table 4 shows Comparative Method as the most frequently used techniques by the re-spondents (mean rating 3.81). This is followed by Cost method with mean rating of 3.70; In¬vestment Method with mean rating of 3.35.

These three methods are the methods the Es¬tate Surveyors and Valuers normally used in the valuation of properties not contaminated. The question is whether they use these meth¬ods in the manner that will reflect the effect of contamination on the affected properties.

Table 5 gives the description of the proper¬ties that were recently sold in the four neigh¬bourhoods (Olusosun, Solous, Abule – Egba and Ketu). These were the properties the Valu¬ers were engaged to value without disclosing their sales price to them.

Table 4. Methods used in the valuation of properties that are contaminated by waste dump sites

Methods

SCALE

1 2 3 4 Mean score Ranking

Pared sales Analysis 97.3 2.70 0.00 0.00 1.02 8

Contingent Valuation Analysis (CVA) 100 0.00 0.00 0.00 1.00 9

Option Pricing Model 89.20 8.10 2.70 0.00 1.13 7

Discounted Cash Flow Techniques 19.00 27.00 24.30 29.70 2.64 4

Hedonic Approach 100 0.00 0.00 0.00 1.00 10

Comparative method 0.00 0.00 18.90 81.10 3.81 1

Cost method 0.00 5.40 18.90 75.70 3.70 2

Investment method 0.00 18.90 27.00 54.10 3.35 3

Residual method 81.10 5.40 5.40 8.10 1.40 5

Profit method 78.40 10.80 5.40 5.40 1.37 6

Source: Field data (2008)

Table 5. Description of property

Property type

Description Location Sales price

A 14 rooms tenement building Olusosun 5 400 000

B 3 bedroom detached house Olusosun 8 500 000

C 13 rooms tenement building Solous 5 800 000

D 3 bedroom detached house Solous 9 800 000

E 10 rooms tenement building Abule – Egba 6 000 000

F 3 bedroom detached house Abule – Egba 12 000 000

G 12 rooms tenement building Ketu 5 800 000

H 3 bedroom detached house Ketu 9 000 000

Source: Field data (2008)

Valuation of Properties in Close Proximity to Waste Dumps Sites: the Nigeria Experience 315

Table 6. One sample t – test

Property type Test value

(Sales price) T Df Sig (2 – tailed) Mean diff 95% Confidence interval

Upper limit Lower limit

A 5400000 8.26 50 .000 828431.37 626980.60 1029882

B 8500000 9.43 50 .000 676470.59 532883.30 820057.90

C 5800000 7.09 50 .000 831372.55 595823.60 1066922

D 9800000 5.05 50 .000 633333.33 381289.90 885376.70

E 6000000 8.15 50 .000 747058.82 563027 931090.70

F 12000000 7.08 50 .000 652941.18 467795.10 838087.30

G 5800000 0.89 50 0.377 98039.22 –122955 319033.70

H 9000000 0.55 50 0.585 64705.88 –171938 301349.30

Source: Field data (2008)

Table 6 shows one sample t – test for the valuation of the 51 Estate Surveyors and Valu¬ers together with their Sales Prices. This pro¬cedure tests whether the mean of the valuation figures estimated by the Estate Surveyors and Valuers for each of the property differs from the Sales Price of the property (i.e. Valuation Accuracy).

For properties located in Olusosun (proper¬ty A and B), Solous (property C and D) and Ab-ule – Egba (property E and F), they have their confidence intervals entirely above 0.00, which implies that the Estate Surveyors and Valuers produce valuation figures that are significantly higher than the sales price, while this is not so for the two properties in Ketu (property G and H). The results obviously come from the fact that none of the Estate Surveyors and Valuers took the effect of the waste dump site on the properties into consideration, hence they tend to over value the properties in close proximity to the waste dump sites.

5. CONCLUSION AND POLICY IMPLICATION

It is evident from this study that the Es-tate Surveyors and Valuers are not aware and do not understand the techniques of valuation

that has been developed for the valuation of properties that are contaminated especially by waste. The few of the techniques they are aware of and understood are the traditional valuation techniques commonly used in prac-tice whose reliability has been questioned by valuation experts.

It is therefore, not uncommon that the in-ability of the majority of the Estate Survey-ors and Valuers to understand the theoretical basis underlying these techniques stems from the nature and content of their undergradu¬ate curricula. In this wise, the undergraduate curriculum should be broadened and extended beyond the confines of the traditional Methods of valuation. Those in academics should also brace up to the challenges in keeping them¬selves abreast with what obtains in the world in the field of valuation as the world is fast becoming a global village.

The Nigerian Institution of Estate Survey¬ors and Valuers should not be left out of this. The Institution should make sure that the Continuous Professional Development (CPD) programme addresses this area of deficiency. This will help to enlighten the practicing Es¬tate Surveyors and Valuers who are ignorant of most of the available techniques in carrying out valuation of contaminated properties.

316 V. A. Bello and M. O. Bello

REFERENCES

Acks, K. (1995) Valuation of Environmental Dam-ages to Real Estate. [Online] Environmental Valuation & Cost Benefit Group. Available at: http://www.costbenefitgroup.com/text/html/ valredi4.htm

Adair, A., Berry, J. and McGreal, S. (1996) Valua-tion of residential property: analysis of partici-pant behaviour, Journal of Property Valuation and Investment, 14(1), pp. 20–35.

doi:10.1108/14635789610107453

Bajari, P. and Kahn, M. (2005) Estimating hous¬ing demand with an application to explaining racial segregation in cities, Journal of Busi-ness & Economic Statistics, 23(1), pp. 20–33. doi:10.1198/073500104000000334

Baum, A.E., Mackmin, D. and Nunnington, N. (1997) Income Approach to Property Valua¬tion. London: International Thompson Busi¬ness Press.

Bello, O.M. (2002) A study on the residential proper¬ty investment performance and its measurement in Lagos metropolis. A PhD thesis submitted to the Department of Estate Management, Fed¬eral University of Technology, Akure.

Bello, M.O. and Bello, V.A. (2008) Willingness to pay for better environmental services: evidence from the Nigerian real estate market, Journal of African Real Estate Research, 1(1), pp. 19–27.

Bello, V.A. (2000) An empirical analysis of the de-terminants of residential property values in Lagos metropolis. M.Tech thesis, Department of Estate Management, Federal University of Technology, Akure.

Carson, R.T. (2000) Contingent valuation: A users guide, Environmental Science & Technology, 34(8), pp. 1413–1418. doi:10.1021/es990728j

Chalmers, J.A., and Roehr, S.A. (1993) Issue in the valuation of contaminated property, The Ap-praisal Journal, 61(1), pp. 28–40.

Clinch, J. and Murphy, A. (2001) Modeling win¬ners and losers in contingent valuation of public goods: appropriate welfare measures and econometric analysis, Economic Journal, 111(470), pp. 420–443.

doi:10.1111/1468-0297.00614

Crosby, N. (1985) The Application of Equated Yield and Real Value Approaches to the Market Valu¬ation of Commercial Property Investment. PhD thesis, University of Reading.

Crosby, N. (1991) The practice of property invest-ment appraisal: reversionary freeholds in the UK, Journal of Property Valuation and Invest-ment, 9(2), pp. 109–122.

doi:10.1108/14635789110030840

Damodaran, A. (2006) Valuation approaches and metrics: A survey of the theory and evidence. [Online] Stern School of Business. Available at: http://pages.stern.nyu.edu/~adamodar/pdfiles/ papers/valuesurvey.pdf

Des Rosiers, F., Bolduc, A. and Theriault, M. (1999) Environment and value: Does drinking water quality affect house prices? Journal of Property Investment and Finance, 17(5), pp. 444–463. doi:10.1108/14635789910294877

Des Rosiers, F. and Theriault, M. (2006) Mass ap-praisal, hedonic price modelling and urban externalities: understanding property value shaping processes. [Online] Discussion Paper presented at the Advances in Mass Appraisal Methods Seminar, Delft University of Tech-nology, October 30-31, 2006. Available at: http://www.tudelft.nl/live/binaries/ac23443b-6d94-4513-a9ea-d312362c0b8b/doc/Paper%20 Des%20Rosiers%20revised.pdf

Finlay, P.N. and Tyler, S.B. (1991) The performance measurement of property investment, Journal of Property Valuation and Investment, 9(4), pp. 295–312. doi:10.1108/14635789110030985

Furby, L., Slovic, P., Fischhoff, B. and Gregory, R. (1988) Public perception of electric power transmission lines, Journal of Environmental Psychology, 8, pp. 19–43.

doi:10.1016/S0272-4944(88)80021-5

Greaves, M.J. (1972) The investment method of property valuation and analysis: An examina-tion of some of its problems. Ph.D thesis, Uni-versity of Reading.

Jackson, T.O. (2001) The effects of previous envi-ronmental contamination on industrial real estate prices, The Appraisal Journal, 69(2), pp. 200–210.

Kinnard, W.N., Jr. (1970) Income Property Valua-tion. Lexington, MA: Health and Company.

Lancaster, K.J. (1966) A new approach to consum¬er theory, Journal of Political Economy, 74, pp. 132–156. doi:10.1086/259131

Lentz, G.H. and Tse, K.S.M. (1995) An option pric-ing approach to the valuation of real estate contaminated with hazardous materials, Jour-nal of Real Estate Finance and Economics, 10(2), pp. 121–144. doi:10.1007/BF01096985

Lizieri, C., Palmer, S., Finlay, L. and Charlton, M. (1995) Valuation methodology and environ-mental legislation: A research project for the RICS Education Trust, City University Busi-ness School discussion paper series, City Uni-versity: London.

Mackmin, D. (1985) Is there a residential valuer in the house? Journal of Property Valuation and

Valuation of Properties in Close Proximity to Waste Dumps Sites: the Nigeria Experience 317

Investment, 3(4), pp. 384–390. doi:10.1108/eb007982

Marchitelli, R. (1992) Rethinking the cost approach,

The Appraisal Journal, 60(3), pp. 424–426. Millington, A. (1990) An Introduction to Property

Valuation. London: Estates Gazette.

Mundy, B. (1992a) The impact of hazardous materi-als on property value, The Appraisal Journal, 60(2), pp. 156–162.

Mundy, B. (1992b) Stigma and value, The Appraisal Journal, 60(1), pp. 7–13.

Neustein, R.A. (1992) Estimating value diminution by the income approach, The Appraisal Jour-nal, 60(2), pp. 283–285.

Ogunba, O.A. (1997) A study of valuation and pric-ing practices in the residential property mar¬ket in Lagos metropolis. M.Sc thesis, Obafemi Awolowo University, Ile-Ife.

Ogunba, O.A., Ajayi, C.A. and Aluko, B.T. (2005) Substitution versus anticipation: A resolution of the Nigerian valuation practice paradox, Journal of Land Use and Development Stud¬ies, 1(1), pp. 1–12.

Patchin, P. (1994) Contaminated properties and the sales comparison approach, The Appraisal Journal, 62(3), 402–409.

Pederson, P. (2002) Contaminated Property Ap-praisal: A Practical Guide to Minimizing Li-ability. Appraisal Institute of Canada Resource Center, pp. 1–17.

Rosen, S. (1974) Hedonic prices and implicit mar-kets: product differentiation in pure compe-tition, Journal of Political Economy, 82(1), pp. 34–55. doi:10.1086/260169

Sykes, S. (1984) Property valuation: A rational mod¬el, Journal of Property Valuation and Invest¬ment, 2(3), pp. 258–270. doi:10.1108/eb007956

The Estate Surveyors and Valuers Registration Board of Nigerian (2002) Membership Directo-ry, Sixth Edition, Lagos, Donald Rowland’s Ltd.

Vanderford, S.E., Mimura, Y. and Sweaney, A.L. (2005) A hedonic price comparison of manu-factured and site-built homes in the Non-MSA U.S. Journal of Real Estate Research, 27(1), pp. 83–104.

Wilson, A. R. (1994) The environmental opinion basis for an impaired value opinion, The Ap-praisal Journal, 62(3), pp. 410–423.

Wood, E. (1972) Property investment – A real value approach. Ph.D thesis, University of Reading.

Jasa Konsultasi Disertasi Keuangan (1) dan Perbankan S3, hubungi WA 0821 2230 7021

Selasa, 09 Agustus 2022

CONTOH PAPER MENGENAI ALGORITHM, INTELLIGENT SYSTEM, COMPUTATIONAL

Tidak ada komentar:

Posting Komentar