Contoh riset mengenai Burst Buffer Storage Architecture.

 

Jitter-Free Co-Processing on a Prototype Exascale Storage Stack

John Bent Sorin Faibish Jim Ahrens Gary Grider

john.bent@emc.com sfaibish@emc.com ahrens@lanl.gov ggrider@lanl.gov

John Patchett Percy Tzelnic Jon Woodring

patchett@lanl.gov tzelnic@emc.com woodring@lanl.gov

 

Abstract

In the petascale era, the storage stack used by the ex¬treme scale high performance computing community is fairly homogeneous across sites. On the compute edge of the stack, file system clients or IO forwarding services direct IO over an interconnect network to a relatively small set of IO nodes. These nodes forward the requests over a secondary storage network to a spindle-based parallel file system. Unfortunately, this architecture will become unviable in the exascale era.

As the density growth of disks continues to out-pace increases in their rotational speeds, disks are be¬coming increasingly cost-effective for capacity but de¬creasingly so for bandwidth. Fortunately, new storage media such as solid state devices are filling this gap; although not cost-effective for capacity, they are so for performance. This suggests that the storage stack at exascale will incorporate solid state storage between the compute nodes and the parallel file systems. There are three natural places into which to position this new storage layer: within the compute nodes, the IO nodes, or the parallel file system. In this paper, we argue that the IO nodes are the appropriate location for HPC workloads and show results from a proto¬type system that we have built accordingly. Running a pipeline of computational simulation and visualiza¬tion, we show that our prototype system reduces total time to completion by up to 30%.

1 Introduction

The current storage stack used in high performance computing (HPC) is recently imperiled by trends in disk technology and harsh economic realities. The storage requirements in HPC are bursty and pre¬dictable. Although there is a small amount of other traffic, the vast majority of IO is comprised of fre¬quent checkpoints and infrequent restarts. Although sites and workloads vary, there are several general

978-1-4673-1747-4/12/$31.00 c⃝2012 IEEE

 

rules of thumb. One, checkpoint and restart work-loads are highly concurrent due to the tight coupling of the running jobs. Two, checkpoint frequency is on the order of a small number of hours. Three, HPC sites have a minimum required bandwidth in order to store a checkpoint quickly enough to ensure suffi¬cient forward progress and a minimum required ca¬pacity to store a sufficient number of checkpoints for a variety of reasons such as time-series analysis and back-tracking due to computational steering.

Until recently, the number of disks required for ca¬pacity has been larger than the number required for bandwidth. In other words, buying the number of disks required for capacity has provided excess band¬width essentially for free. However, disk capacity is increasing much faster than disk performance. New technologies such as shingled disks [4] are only ex¬acerbating this trend. The result is that the num¬ber of disks required for capacity has now become fewer than the number required for bandwidth. Un¬fortunately, purchasing disks for bandwidth is cost-prohibitive [9]. Solid-state drives (SSDs) however are cost-effective for bandwidth but cost-prohibitive for capacity. Clearly, a hybrid hierarchy in which SSDs are used for checkpoint bursts and disks are used for checkpoint storage is the answer.

However, it is not immediately obvious whether to place SSDs inside the compute nodes, inside the stor¬age system, or somewhere in the middle. Although each placement may be appropriate for some work¬loads, we believe that placement in the middle is best for HPC. Placement on the compute nodes is problematic due to computational jitter [17] as any perturbation caused by asynchronously copying data from the SSDs to the disks can ripple devastatingly through the tightly-coupled computation [11]. Place¬ment inside the storage system is also problematic as it would require a more expensive storage network matched to the higher bandwidth of the SSDs instead of being matched to the lower bandwidth of the disks.

Therefore, in this paper we examine a prototype exascale storage stack built with SSDs placed in spe 

 

Figure 1: A Burst Buffer Storage Architecture. Eight processes spread across four compute nodes intersperse writes into a shared file. By using the PLFS middleware layer, the illusion of a single file is preserved in a manner completely transparent to the application and the user. Physically however, PLFS transforms the IO to leverage both the global visibility of the parallel file system as well as the faster performance of the SSDs in the burst buffers.

 

cialized burst buffers stitched transparently into the storage stack using a modified version of the PLFS middleware [6]. Evaluations with a workload of simu¬lation and visualization show a speedup in total time to completion of up to thirty percent. The remainder of the paper is as follows: we describe our architec¬ture in Section 2, our results in Section 3, related work in Section 4, and our conclusions in Section 5.

2 Design and Implementation

Our design consists of both a new hardware archi-tecture and a modified software stack. The hard-ware architecture is relatively straight-forward. In order to transparently weave the SSDs into the stor¬age stack, we extend the notion of IO nodes [3, 10, 20] to add buffering to quickly absorb the bursty check¬point workloads; we call these augmented IO nodes burst buffers. As is typical of IO nodes, we place the burst buffers within the computational fabric and also attach them to the secondary storage network.

The modified software stack is a bit more complex. We extend the PLFS middleware layer [6] to incorpo¬rate the burst buffers. PLFS is a middleware virtual file system which transparently interposes on applica¬tion IO. This interposition can dramatically improve write bandwidth by rearranging the workload into one requiring less locking within the parallel file sys¬tem (PFS). This is primarily achieved by decoupling writes to a shared logical file into writes to multi¬ 

 

ple physical files; each physical file is then written by only a single process. PLFS also maintains metadata sufficient to reconstruct the logical file for reads.

PLFS stores file data and its metadata into con-tainers. Containers are implemented as physical di-rectories on the underlying file system which PLFS uses as its store. Subdirs within the container dis-tribute the large number of files within the container into multiple directories to avoid performance prob¬lems with too many files in a single directory [15]. Since version 2.0 [5], PLFS has used a new abstrac¬tion called metalinks which allow the subdirs to phys¬ically exist in shadow containers. This allows sites with multiple metadata servers to configure PLFS such that its load is distributed across them.

By default, PLFS randomly places subdirs; we modified it to allow each compute node to specify a lo¬cation for its subdirs. In this way, the relatively small amount of PLFS metadata is stored on the PFS but the larger amount of file data is stored in the burst buffers as shown in Figure 1. We then augmented the PLFS API by adding two functions to allow the user to control the management of the burst buffers: one to start an asynchronous copy of the file data from the shadow subdirs into into the container on the PFS and a second to remove the shadow subdirs and their contents.

An important feature of PLFS is that it is transpar¬ent to users and applications and runs with unmodi¬fied programs. Although we preserved this as much as possible, existing interfaces such as POSIX and MPI 

 

8 Compute Nodes 2 Burst Buffers Storage 2 File System Blades

•

• 40 GB RAM

Dual socket, quad core, 2.8

GHz Intel Nehalem •

• 128 GB RAM

Quad socket, 10 core, 3.1

GHz Intel Westmere •

•

• EMC VNX 7500

24 GB RAM

10 Near-Line 2TB 7200 •

•

• Lustre 1.8.0

16 GB RAM

Dual socket, quad core, 2.8

• Mellanox QDR Infiniband • 512 core nVidia Tesla 2090 RPM SATA drives GHz Intel Nehalem

port • GPGPU

4 Mellanox QDR Infiniband

ports • 4 8gb fiber channel ports •

• Mellanox QDR Infiniband

port

2 8gb fiber channel ports

• 16 Samsung 200 GB SSDs

Table 1: Evaluation System Specifications. The SSDs on the burst buffers were grouped into four RAID-0 arrays, mounted using ext4, and exported via NFS. In all cases, the operating system was 64-bit CentOS 6.0. Both file system blades ran a Lustre Object Storage Server; one also ran a metadata server which used a local Samsung 200 GB SSD. The fiber channels were directly connected; the other ports used a Mellanox QDR Infiniband switch.

 

IO do not support burst buffer management: thus we were forced to modify our simulation to initiate the asyncronous copy and our visualization to remove the shadows. Aside from these modifications, the burst buffers were transparent to the user and the appli¬cation. The files were always accessed via the same path regardless of the data’s physical location; the only difference perceivable is a faster bandwidth when the data is moving to or from the burst buffers.

3 Evaluation

To evaluate our design, we built a small proto¬type burst buffer system consisting of eight compute nodes, two burst buffers, and a PFS. On the compute nodes, we ran the PLFS client through which the sim¬ulation could checkpoint and restart. The client was also on the burst buffers in order for the visualiza¬tion to read the checkpoint data. Lustre running on an EMC VNX 7500 served as our PFS. Hardware specifications are listed in Table 1 and a depiction is shown in Figure 2.

The workload consisted of a simulation running for eight timesteps on the compute nodes. After each timestep, the simulation checkpointed using the PLFS driver in the MPI-IO library. Each checkpoint file was then processed by a visualization program which accessed the file via POSIX through a FUSE mount of PLFS.

The simulation was a fluid dynamics model coupled with a wind turbine model [19] used to study effects of terrain and turbine placement on other downstream wind turbines. The software used for the visualiza¬tion was ParaView [1], an open-source, large-scale, parallel visualization and analysis tool. The visual¬ization analyzed the forces on the wind turbine, the air flow, and the vortex cores created by the move¬ment of the wind turbine.

We ran the same workload using two configura-tions. One using the compute nodes and the PFS but without the burst buffers; the second added the burst buffers. For the remainder of the paper, we’ll 

 

refer to the first configuration as NoBB and the sec¬ond as WithBB. Using NoBB the visualization was post-processed: it did not run until the eight com-putational timesteps and checkpoints were complete. By adding the burst buffers however, we were able to co-process the visualizations by running them on the burst buffers and pipelining them with the simu¬lation. Additionally, the checkpoint latency between computational timesteps was reduced with WithBB since the checkpoint bandwidth was increased ap¬proximately 400%. Remember that the checkpoints were still ultimately saved on the PFS; WithBB just allowed that slower data movement to be asyn¬chronous and pipelined with the computation.

The results are shown in Figure 3. Notice

that the simulation, checkpoint, and visualization phases are serialized using NoBB whereas they are pipelined when using WithBB. These graphs show that WithBB provides four distinct benefits. First, the simulation phase finishes about ten percent more quickly. Second, the complete workload finishes about thirty percent more quickly. Third, because the visualization is co-processed instead of post-processed, computational steering is possible much earlier (after about four minutes instead of thirty) as is shown by the elapsed time when the first visual¬ization completes. Finally, since WithBB allows the visualization and asynchronous copying to the PFS to be done on the burst buffers, there is no jitter intro¬duced and therefore the compute phase is not slowed as it might be if the SSDs were placed inside the com¬pute nodes. This can be observed from the duration of each individual simulation phases (i.e. the widths for the simulation bars are the same).

We acknowledge that it is not completely fair to compare NoBB to WithBB since WithBB uses addi¬tional hardware. However, we believe our results are still valid for several reasons. First, the changing eco¬nomics of disks and SSDs seem to dictate the need for SSDs in the exascale era. Our work is an impor¬tant early exploration in this area. Additionally, the thirty percent speedup that we observe in our work 

 

 

Figure 2: Our Prototype Burst Buffer System. This figure shows the prototype exascale storage system we built using eight compute nodes in which the wind-turbine simulation saved checkpoints through PLFS into two burst buffers. Our visualization program then ran on the burst buffers and read the wind-turbine data through PLFS as the data was simultaneously copied onto the parallel file system. After it was no longer needed, the copy on the burst buffers was removed in order to reclaim space.

load possibly under-represents the potential of burst buffers because we evaluated only one small workload that may not be completely reflective of anticipated exascale workloads.

One reason is that the simulation we used check-pointed only five percent of the total RAM in the compute cluster (16 out of 320 GB). Exascale appli¬cations may checkpoint a much larger ratio; there¬fore reductions in checkpoint time will reduce total time to completion more than we have shown here. A second reason is that our simulation was config¬ured to checkpoint at a fixed interval of computation. Figure 3 shows that the simulation checkpointed the same number of times even though it ran for a shorter time. In production settings, applications tend to adjust themselves to checkpoint at a fixed interval of wallclock time. Therefore, a reduction in check-point latency will increase computational efficiency even more than we have shown here.

4 Related Work

Performing visualization and analysis simultaneously with a running simulation is done using one of two different techniques, in situ and co-processing. In situ performs visualization and analysis in the same pro¬cess space as the simulation code, usually through linked library analysis code. This allows the visual¬ization and analysis to have direct access to simula¬tion memory and to perform computations without data movement. Several large-scale in situ analysis libraries have recently emerged [8, 16, 23] as have sev¬eral that are application-specific [2, 21, 24]. Although in situ avoids a data copy, it can not be pipelined with the simulation as we have done. Both in situ

 

(a) Direct to Lustre

Elapsed Time (s)

(b) Using Burst Buffers

Figure 3: Workflow. Both graphs show the elapsed time running the same workload of eight timesteps of the wind-turbine simulation, eight checkpoints, and eight vi-sualizations. The upper graph shows the total time to completion using the default system without the burst buffers and post-processed visualization; the lower with the burst buffers and the co-processing they enable. In both graphs, the height above the x-axis merely indicates whether processing is occuring (the differing heights for the simulation and the visualization are merely to differ-entiate between the two); the height below the x-axis indi-cates bandwidth (lower is better).

and co-processing will be used in exascale.

Other examples of co-processing analyses have been performed with scientific data staging technolo¬gies, including ADIOS [12] and GLEAN [22]. Our contribution is an approach which leverages file sys¬tem semantics so that unmodified applications can use our burst buffer system using standard POSIX and MPI-IO interfaces.

As opposed to placing storage in the middle as we have studied here, other work [7, 14] studied similar workloads in which they place storage on the compute

 

nodes. Although this was not measured in their work or ours, we fear this placement can slow down the foreground computation due to jitter [11, 17].

Another project to use SSDs between the compute nodes and the storage system [18] also uses SSDs as a temporary store before migrating data to the PFS. Unlike our work however, they use SSDs from a sub¬set of the compute nodes that have SSDs attached. Although this places the SSDs on the compute nodes, they can choose to use them as burst buffers in order to avoid jitter. Additionally, our work leverages the IO nodes in existing HPC architectures which allows a reduction in the size of the storage network.

Finally, IOFSL [3] aggregates the IO from multiple compute nodes onto a smaller number of IO nodes. Similar to our work, SCR [13] has also shown large checkpoint improvements by bursting checkpoints to the memory of neighboring nodes and then asyn¬chronously migrating them to the PFS. We believe that both of these, or similar technologies, are likely to be included in the exascale storage stack. These projects are complementary to ours and we are cur¬rently working with both to integrate all three.

5 Conclusion

Economic trends in storage technologies and the stor¬age requirements for exascale computing indicate that SSDs will be incorporated into the HPC storage stack. In this paper, we have provided an important initial exploration of this large design space. Using commodity hardware, modified HPC file system mid-dleware, we evaluated our design with a real HPC workload consisting of simulation and analysis. We demonstrated that placing SSDs in between the com¬pute nodes and the storage array allow jitter-free co-processing of the visualization and reduce total time to completion by up to thirty percent.

We are currently exploring ways to leverage the burst buffers to enable Map-Reduce style analytics as well as incorporating pre-staging of data sets into the scheduler. Additionally, we are developing an analyt¬ical model to help determine the ratio between burst buffers and compute nodes for large systems. Finally, we are continuing to test our design and software on increasingly large system to study its scalability.

References

[1] J. Ahrens, B. Geveci, and C. Law. Paraview: An end user tool for large data visualization. The Visualization Handbook, pages 717–731, 2005.

[2] J. Ahrens, K. Heitmann, M. Petersen, J. Woodring, S. Williams, P. Fasel, C. Ahrens, C. Hsu, and B. Geveci. Verifying scientific sim-ulations via comparative and quantitative visualization. IEEE Computer Graphics and Applications, 30(6):16–28, 2010.

[3] N. Ali, P. Carns, K. Iskra, D. Kimpe, S. Lang, R. Latham, R. Ross, and L. Ward. Scalable i/o forwarding framework for high-performance computing systems. In IEEE International Conference on Cluster Computing, Cluster 2009, New Orleans, LA, Sept. 2009.

 

[4] A. Amer, D. D. E. Long, E. L. Miller, J.-F. Pris, and T. Schwarz. Design issues for a shingled write disk system. In 26th IEEE Symposium on Massive Storage Systems and Technologies, MSST 2010, May 2010.

[5] J. Bent et al. Plfs 2.0. sourceforge.net/pro jects/plfs/files/.

[6] J. Bent, G. Gibson, G. Grider, B. McClelland, P. Nowoczynski, J. Nunez, M. Polte, and M. Wingate. PLFS: a checkpoint filesystem for parallel applications. In Proceedings of the Conference on High Performance Computing Networking, Storage and Analysis, SC ’09, pages 21:1–21:12, New York, NY, USA, 2009. ACM.

[7] D. Camp, H. Childs, A. Chourasia, C. Garth, and K. I. Joy. Evaluating the Benefits of An Extended Memory Hierarchy for Parallel Streamline Algorithms. In Proceedings of the IEEE Symposium on Large-Scale Data Analysis and Visualization, LDAV. IEEE Press, 2011.

[8] N. Fabian, K. Moreland, D. Thompson, A. C. Bauer, P. Marion, B. Geve-cik, M. Rasquin, and K. E. Jansen. The ParaView copro cessing library: A scalable, general purpose in situ visualization library. In 2011 IEEE Symposium on Large Data Analysis and Visualization (LDAV), pages 89–96. IEEE, Oct. 2011.

[9] G. Grider. Speed matching and what economics will allow. HEC-FSIO 2010, Aug. 2010.

[10] G. Grider, H. Chen, J. Nunez, S. Poole, R. Wacha, P. Fields, R. Mar¬tinez, P. Martinez, S. Khalsa, A. Matthews, and G. A. Gibson. Pas¬cal - a new parallel and scalable server io networking infrastructure for supporting global storage/file systems in large-size linux clusters. In IPCCC’06, pages –1–1, 2006.

[11] J. Lofstead, F. Zheng, Q. Liu, S. Klasky, R. Oldfield, T. Kordenbrock, K. Schwan, and M. Wolf. Managing variability in the io performance of p etascale storage systems. In SC ’10: Proceedings of the Conference on High Performance Computing Networking, Storage and Analysis, New York, NY, USA, 2010. ACM.

[12] J. F. Lofstead, S. Klasky, K. Schwan, N. Podhorszki, and C. Jin. Flexible IO and integration for scientific codes through the adaptable IO system (ADIOS). In Proceedings of the 6th international workshop on Challenges of large applications in distributed environments, CLADE ’08, page 1524, New York, NY, USA, 2008. ACM.

[13] A. Moody, G. Bronevetsky, K. Mohror, and B. R. d. Supinski. Design, modeling, and evaluation of a scalable multi-level checkpointing sys¬tem. In Proceedings of the 2010 ACM/IEEE International Conference for High Performance Computing, Networking, Storage and Analysis, SC ’10, pages 1–11, Washington, DC, USA, 2010. IEEE Computer Society.

[14] X. Ouyang, S. Marcarelli, and D. K. Panda. Enhancing checkpoint per¬formance with staging io and ssd. In Proceedings of the 2010 International Workshop on Storage Network Architecture and Parallel I/Os, SNAPI ’10, pages 13–20, Washington, DC, USA, 2010. IEEE Computer Society.

[15] S. V. Patil, G. A. Gibson, S. Lang, and M. Polte. GIGA+: Scalable Directories for Shared File Systems. In Petascale Data Storage Workshop at SC07, Reno, Nevada, Nov. 2007.

[16] T. Peterka, R. Ross, A. Gyulassy, V. Pascucci, W. Kendall, H. Shen, T. Lee, and A. Chaudhuri. Scalable parallel building blocks for custom data analysis. In 2011 IEEE Symposium on Large Data Analysis and Visualiza¬tion (LDAV), pages 105–112. IEEE, Oct. 2011.

[17] F. Petrini, D. J. Kerbyson, and S. Pakin. The caseof the missing su-percomputer performance: Achieving optimal performance on the 8,192 processors of asci q. In Proceedings of the 2003 ACM/IEEE conference on Supercomputing, SC ’03, pages 55–, New York, NY, USA, 2003. ACM.

[18] R. Prabhakar, S. S. Vazhkudai, Y. Kim, A. R. Butt, M. Li, and M. Kan-demir. Provisioning a multi-tiered data staging area for extreme-scale machines. In Proceedings of the 2011 31st International Conference on Distributed Computing Systems, ICDCS ’11, pages 1–12, Washington, DC, USA, 2011. IEEE Computer Society.

[19] E. K. Rodman R. Linn. Determining effects of turbine blades on fluid motion, 05 2011.

[20] G. Shipman, D. Dillow, S. Oral, and F. Wang. The spider center wide file system: From concept to reality. Cray User Group Conference, May 2009.

[21] A. Tikhonova, C. Correa, and K. Ma. Visualization by proxy: A novel framework for deferred interaction with volume data. Visualization and Computer Graphics, IEEE Transactions on, 16(6):1551–1559, 2010.

[22] V. Vishwanath, M. Hereld, and M. E. Papka. Toward simulation-time data analysis and I/O acceleration on leadership-class systems. In 2011 IEEE Symposium on Large Data Analysis and Visualization (LDAV), pages 9–14. IEEE, Oct. 2011.

[23] B. Whitlock, J. Favre, and J. Meredith. Parallel in situ coupling of simulation with a fully featured visualization system. In Eurographics Symposium on Parallel Graphics and Visualization, pages 100–109, 2011.

[24] J. Woodring, J. Ahrens, J. Figg, J. Wendelb erger, S. Habib, and K. Heit-mann. In situ sampling of a Large Scale particle simulation for inter¬active visualization and analysis. Computer Graphics Forum, 30(3):1151– 1160, June 2011.

 

Proceedings of the 43rd Hawaii International Conference on System Sciences - 2010

 

Internet Crime Reporting: Evaluation of a Crime Reporting and Investigative Interview System by Comparison with a Non-interactive Reporting

Alternative1

 

Alicia Iriberri

Instituto Tecnológico Autónomo de México

alicia.iriberri@itam.mx

Abstract

This study reports on the evaluation of a crime reporting and investigative interview system, i-recall . i-recall emulates a police officer conducting a cognitive interview (CI). It incorporates CI techniques to enhance witness memory retrieval and leverages natural language processing technology to support understanding of witness narratives.

In a controlled user study i-recall was compared to a non-interactive textbox computer system. Sophomore college students acted as witnesses to a videotaped staged crime and reported what they saw using one of the two alternative reporting methods.

Results indicate that i-recall outperformed the textbox system significantly in one of two measures, completeness of report. On average i-recall elicited 14 percent of information from witnesses and the textbox system elicited five (5) percent of information, all with 94 percent accuracy. i-recall is a promising Internet reporting alternative. Future work will evaluate i-recall by comparing it to a human expert cognitive interviewer.

1. Introduction

Every year millions of crimes are committed in the US. Research indicates that the principal determinant to solving crimes is the completeness and accuracy of eyewitness reports [30]. However, 50% of crimes go unreported due to fear and privacy concerns [3]. In addition, police resource shortages (e.g., investigators, training, time to respond and transcribe reports) often lead to incomplete or inaccurate information.

1 The authors wish to acknowledge Dr. Gondy Leroy, Director of the Intelligence Systems Lab, Claremont Graduate University for her advice and support for this project and RA Chih-Hao (Justin) Ku for his work in the creation of the non-interactive reporting method and in the addition to i-recall’s knowledge base.

 

Carlos J. Navarrete

California State Polytechnic University, Pomona

cjnavarrete@csupomona.edu

Criminology researchers have urged the creation of alternative reporting methods that serve as a mediating instance between witnesses fear of contacting the police and police goal for preventing and solving crime [22,28]. Instances of Internet use to report crime currently exist. The FBI Tips and Public Leads System and the Claremont University Consortium’s Silent Witness Program are examples. Using the Internet, these systems address the concerns and fears of victims confronted with the decision to file or not file a report. The Internet provides the convenience to reach authorities 24/7 from any location with Internet access while protecting the victims’ identity. The FBI’s and the CUC’s systems allow victims and witnesses to report incidents to police with the option to provide as much information as they choose. The FBI’s system provides a single textbox for victims to file their report. The CUC’s system presents victims with a standard set of questions regardless of the type of crime being reported. Both systems require that the person filing a report remembers all vital information related to the crime without support for event recollection. Police authorities need to have more accurate and complete information related to a crime. To realize the benefits of Internet-based systems and before large scale implementations of these systems take place, it is necessary to test alternatives to find one that optimally addresses victims concerns while similarly facilitating the filing of accurate and complete reports.

We developed a crime reporting and investigative interview system, i-recall, that extracts information from witness crime narratives and emulates an investigative interview [12]. This system is a tool police can use to collect accurate and complete information when conducting face-to-face interviews it is not feasible or possible. In addition, witnesses can maintain their privacy by reporting information anonymously or using secured IDs. Specifically, the system asks witnesses to provide written narratives of a crime they witnessed and extracts relevant facts from the narrative using natural language information

 

Proceedings of the 43rd Hawaii International Conference on System Sciences - 2010

 

extraction techniques. The system then generates questions and an interview strategy to help witnesses recall missing facts about the crime they witnessed conforming to cognitive interview principles. Last, i-recall produces a written standard police report.

In this study, the newly developed system is compared to a non-interactive textbox computer system similar to those currently in use. This comparison sheds light on the extension of improvement achieved by i-recall. The importance of this contribution lies in the potential of using natural language information extraction technology and Internet-based interviewing systems to gather information from the public. Such a system may help alleviate the shortage of police resources while maximizing the quality of information collected from witnesses. Principles used in the design of i-recall informs the design of usable e-Government applications and services.

2. Cognitive Interview

The design of i-recall incorporates theories and principles from memory and eyewitness research. Findings in these fields suggest the use of specific methods and techniques to retrieve information from witness memory effectively and with better results. Geiselman and Fisher integrated these methods in what is known as the Cognitive Interview (CI) [7,8,9,17]. Evidence from CI experimental and field research indicate that when investigative interviewers conduct such cognitive interviews, as opposed to standard, questionnaire-based investigative interviews, the accuracy and richness of the information obtained from witnesses is significantly higher.

The CI uses various methods to help witnesses retrieve information from memory. These methods induce witnesses to reinstate mentally the context of the event to be remembered, to report in free and narrative form everything that they can remember, and to report information starting with mental images that are richest in details and then continuing with lesser rich images until all details about the event to be remember are exhausted. Using the CI skillful interviewers provide the proper guidance to reduce witness vulnerabilities such as memory blocking or inability to discriminate between relevant and irrelevant information to police investigations.

Even though the CI is a very effective investigative interviewing technique, it places high demands on the investigative interviewer in terms of training, time, ability, and availability [6,23,24]. At the time of a criminal incident, investigators may need to interview various witnesses, but due to limited police resources, 

 

they might choose to focus only on interviewing key witnesses, which leads to losing important pieces of information from other witnesses. Findings also indicate that if witnesses are not interviewed shortly after witnessing the incident, their memory of the incident can suffer from cross-contamination and decay [6,10]. Therefore, it is suggested that investigative interviews be conducted soon after the occurrence of the event.

Evaluations of the CI demonstrate its effectiveness and have made the CI a mandate in current police practice [25,26]. Research indicates CI is able to elicit an average of 41 percent (ranging from 15 percent to 147 percent) more correct details compared to standard police interview with 85 percent accuracy [8,17,23,24]. Using the CI, interviewers elicit substantially more information than interviewers using a standard interview technique [23]. However, conducting a CI poses high demands on interviewers personal qualities, intellectual abilities, motivation, and command of the technique.

Given the demands imposed by the CI, it is not surprising that its use in the field by police officers is very limited [9]. Police officers find the CI technique complex and cumbersome and requiring more time than what usually is available [5,6,15].

By leveraging cognitive interview principles we expect i-recall to serve as a mediating alternative that helps alleviate the shortage of police resources while maximizing the quality of information collected from witnesses. At this stage we are evaluating its efficacy in eliciting quality information from witnesses.

3. Criminal Investigation Process

When a crime is reported to the police department, the first officer arriving on the scene is usually a patrol officer. This responding police officer secures the crime scene, takes an initial report from victims, determines if and what type of crime was committed, and decides to or not to forward the case for further investigation. However, given the potential value of the information at the crime scene criminal justice researchers and practitioners are pushing towards expanding the role of patrol officers to conduct more investigative duties at the scene [11,32].

During their preliminary investigation, police officers look for information on the basic characteristics of the perpetrator and the committed offense, such as gender, height, built and whether a weapon was used. They also look for detailed information that would help them identify the perpetrator and the nature of the crime such as clothing, jewelry, distinguishing features, and the exact

 

Proceedings of the 43rd Hawaii International Conference on System Sciences - 2010

 

type of weapon used [16]. Also, the information collected include description of events, behaviors, feelings, thoughts and intentions. For police officers and witnesses alike recording all these details is a very time-consuming and difficult task, since it represents a challenge to their abilities, skills, and resources that they must overcome.

Despite knowing that the information obtained during preliminary investigations is of primary importance, police officers rarely have the time to conduct thorough investigations at the scene [31]. Several concurrent activities need to be accomplished upon arrival at a crime scene. These activities include assessing the situation to determine the type of crime that occurred, identifying victims, witnesses and suspects, arresting perpetrators, documenting the crime scene, and identifying and collecting evidence. These time-critical activities are usually done under pressure to respond to other calls for service [16]. Additionally, these activities require skills and expertise not acquired with the basic training police officers commonly receive [6,11]. As a result the amount and quality of information they obtain from victims and witnesses is negatively affected.

4. Internet Crime Reporting

Internet crime reporting is a convenient alternative and presumably a good alternative to address unreported crime [22]. Using the Internet individuals can file reports any time of day or night and take the time they need to report in their own words the incident they witnessed. Witnesses can use this alternative to report quality of life threats or suspicious activities that might go otherwise unreported.

Evidence to suggest the potential of Internet crime reporting exists [1,14,20,22]. However existing implementations of Internet crime reporting with fill-in-the-blank forms and email address are not capable of enhancing the quality of information witnesses provide in the way a full CI would [27]. Especially if one considers the challenges witnesses may possibly face at the time of reporting (e.g., stressed, unable to remember, unaware of what is relevant to the investigation). An Internet crime reporting that is interactive may be capable of enhancing witnesses reporting experience, and encourage them to report or report more information. i-recall leverages information extraction technology to create such interactive environment.

 

5. Information Extraction

Information extraction (IE) uses a range of natural language processing (NLP) techniques to produce fixed-format data about domain-specific entities found in written narratives (i.e., texts, documents, articles, WebPages). The resulting data then may be used for database queries or further analysis.

In the development of i-recall we used a knowledge engineering, rule-based approach to IE [2]. We analyzed example narratives to identify such text snippets as named entities, grammars, and text patterns. Then, we created rules that model the grammar and text patterns identified in the example narratives. We fed the rules to i-recall’s IE module to use them to process new crime narratives. When the IE module detects instances of the modeled snippets, it automatically extracts and annotates them. To produce annotations, the IE module preprocesses narratives with tokenizing, sentence-splitting, and part-of-speech tagging tools. The IE module uses the output of these tools and processes it further using lexicon lookups to identify named entities. Finally, the IE module uses the grammar and text pattern rules to produce the required output.

For i-recall’s IE module named entities include people, locations, personal physical attributes, weapons, vehicles, acts, and personal property and a text pattern may be for example, “blue eyes.” When the i-recall’s rule-based IE module recognizes this pattern, it annotates as eye:body part and blue:eye color.

The IE rule-based approach is labor-intensive, but it is useful when the number of example narratives available to creating and testing rules is limited and the level of precision required is high. Rule-based IE systems often achieve these levels [2]. After various round of tests and adjustments using different types of witness narratives, the IE performance of i-recall currently shows that it extract 96% of information from witness narratives with 100% precision [12,13,18,19].

6. The System

i-recall incorporates CI techniques to maximize witness memory recall and IE capabilities to extract crime information from interviews. With incremental refinements we expect i-recall to approximate the performance effectiveness of a human cognitive interviewer.

i-recall comprises Internet, database and Java technology, and leverages open-source IE tools from the General Architecture for Text Engineering (GATE) [4]. Using these technologies, i-recall simulates the

 

Proceedings of the 43rd Hawaii International Conference on System Sciences - 2010

 

tasks that a CI interviewer normally performs. First, it asks witnesses to provide general information (e.g. date, time, type of crime) about the crime and then a written narrative of the incident witnessed. Second, it uses IE tools to extract named entities and relevant facts from this narrative. Third, using the output of the initial extraction, i-recall assembles questions and designs an interviewing strategy, just like a CI interviewer does, and presents these questions to help witnesses recall facts that are missing in their report. To build the interviewing strategy, i-recall uses such data structures as a crime report checklist and an inference (network-like) data structure that represents various possible reporting sequences. Last, i-recall stores both complete narratives and annotated text in a database that can later be used to produce a written standard police report. The detailed system design is described and documented in [12].

7. The Study

The efficacy of i-recall was compared to the efficacy of a computer non-interactive reporting system, i-textbox, expecting to find evidence to indicate that i-recall achieves higher efficacy than the alternative method. To determine the efficacy of i-recall as a reporting system two measures were taken, the accuracy and completeness of witness reports. Participants in the study were randomly assigned to one of two reporting methods: i-recall, an Internet interactive reporting system; or i-textbox, a computer non-interactive (e-mail-like) reporting system similar to the system used by the FBI Tips and Public Leads.

The study was conducted in a laboratory setting. In this experiment, participants acted as witnesses to a staged crime depicted on a video clip (two and a half minutes in length) and reported what they saw in the video. To report the crime they witnessed, participants were randomly assigned to one of the reporting systems.

The following hypotheses represent the expected results:

Hypothesis 1 (H1):

Crime reports resulting from the use of i-recall are more complete than crime reports resulting from the use of i-textbox.

Hypothesis 2 (H2):

Crime reports resulting from the use of i-recall are more accurate than crime reports resulting from the use of i-textbox.

The study was a between subjects experimental design with one independent variable and two 

 

dependent variables. The output generated by the reporting method participants used was scored and means per system were compared.

The independent variable in the study design is the reporting method with two levels:

(a) i-textbox: computer crime reporting system with an e-mail-type interface.

(b) i-recall: Internet crime reporting system with an interactive interview interface.

The dependent variables are report completeness and accuracy. These variables are defined as:

1. Completeness: number of correct items in report (quantity of recall).

2. Accuracy: number of correct items in report over the sum of all items in report including correct items, incorrect items and confabulated items (quality of recall).

Completeness and accuracy are two variables customarily used in the study of the efficacy of the cognitive interview technique [8,17]. To be consistent with previous studies this study uses these same metrics.

One additional set of data was collected. This set included the following participants’ demographic characteristics (control variables): participant’s age, gender, ethnicity, and educational level. These variables are known to affect eyewitness testimonies [21,29]. These variables were measured using items on a pre-test questionnaire.

8. Methodology

Data collection was done on an individual basis in two separate offices. These offices were equipped with one computer with Internet connection, a widescreen monitor, and speakers. Individual participants were assigned to one condition and sat at a desk in front of the computer monitor throughout the study. Participants were randomly assigned to each condition to ensure other influencing variables such as age, ethnicity, and educational level would balance out, affecting equally the results for each reporting condition. That is, random assignment of participants to conditions would even out the effect of extraneous variables on the dependent variables.

A video clip depicting a staged crime with duration of two and a half minutes was used as stimulus. This video is an excerpt of the commercial motion picture As Good As It Gets by James L. Brooks. This movie was rated by the Motion Pictures Association of

 

Proceedings of the 43rd Hawaii International Conference on System Sciences - 2010

 

America for audiences of 13 years of age or older (PG-13). In this video, participants observed a crime being committed by two suspects (Suspect 1 and Suspect 2) aided by an accomplice (Model) on a victim (Artist/Painter). The crime is a home robbery invasion and aggravated assault. The suspects take property from the painter’s apartment while the model, who is posing for a portrait, distracts the painter. The painter is completing a full size portrait of the model. When the painter realizes (by his pet dog’s actions) he is being robbed, he confronts the assailants and is assaulted by one of them, who hits him repeatedly with a coat rack. The two suspects and the accomplice flee the scene taking the property items they had put in a sack. The portrait of the accomplice, which is a very accurate representation of the accomplice, is left behind at the scene.

After watching the video clip, participants were asked to report descriptive information of what they saw in the video and to avoid guessing. They were asked to provide descriptions on the suspects, the weapon, the stolen property items, the location, and the actions during the incident they witnessed.

8.1 Data Collection

In preparation for data collection various preliminary activities were performed, which included provisions to protect the study’s internal validity.

8.1.1 Gold Standard Creation. The stimulus video clip was coded and scored to identify relevant facts for a standard police report. One independent rater coded the stimulus. The video rater watched the video clip, extracted descriptive facts, and created a template grouping extracted facts in seven categories: suspect-1, suspect-2, accomplice, weapon, property items, location, and criminal acts. The total number of facts per category was tallied in the template. The template and the total number of facts represent the gold standard against which the resulting crime reports from each reporting method were compared and scored.

8.1.2 Pilot Testing. A pilot test of the study was run to ensure clarity and face validity of materials. This test served also as an opportunity to practice the study procedure and to verify the well functioning of the reporting methods. The pilot test was conducted a week prior to the study with six volunteers (three per reporting method). The crime reports created by these participants were not included in the study final dataset. Materials, procedures, and reporting methods were refined after the pilot test.

 

8.2 Procedure

Data collection activities were conducted at three different phases, presentation, testing, and data coding.

8.2.1 Presentation Phase. In the presentation phase a facilitator (a) introduced participants to the study, (b) had participants watch the video in isolation on an individual bases, (c) surveyed the demographic profile of participants; and (d) allowed participants to break for 20 minutes and asked them not to discuses the content of the video with anyone.

In reality some time passes by before a witness provides an account of the crime incident to police. Thus, to allow for a more naturalistic reporting situation, the 20-minute delay between viewing the stimulus and reporting what they saw was imposed. During the delay, participants conversed with a study confederate to ensure that they would not have the opportunity to replay the video in their minds trying to remember every detail or to discuss its content with anyone.

8.2.2 Test Phase. In the test phase the facilitator (a) assigned participants randomly to each reporting method; (b) instructed participants on the use of the reporting methods using screenshot slides; (c) instructed participants to begin their reports using the assigned condition; and (d) debriefed participants when they completed their reports and asked them not to discussed the study with anyone likely to participate in the study.

8.2.3 Data Coding Phase. During this phase the facilitator (a) formatted all reports uniformly to eliminate the possibility of identifying the reporting method that generated them, (b) submitted resulting reports to be coded and scored, (c) verified the accuracy and completeness scores of each report, and (d) coded demographic questionnaires results.

8.2.4 Data Coding Detailed Procedures. All reports and questionnaires were identified with a participant’s numeric code, only the primary investigator knew the names of participants and the reporting method they used. An independent rater coded and scored the resulting crime reports. The rater scored reports by hand using color-coding (one color per crime video’s category), and tallied totals using the gold standard template created by the video clip coder.

 

Proceedings of the 43rd Hawaii International Conference on System Sciences - 2010

 

8.3 Statistical Analysis

Two types of statistics were calculated to allow for comparison among the two experimental conditions. These statistics were:

1. Descriptive statistics and chi-squares. These were calculated to obtain a profile of participants’ demographic characteristics and verify random assignment.

2. One-way ANOVAs. These were calculated to test the study’s hypotheses for the two dependent variables completeness and accuracy.

9. Results: Completeness and Accuracy

Resulting reports were prescreened to ensure content validity yielding 48 usable reports. Of the 48 usable reports, 23 resulted from i-recall and 25 from i-textbox. Participants in the study were sophomore students enrolled in various sections of the Management Information Systems (MIS) course in the College of Business Administration of the California State Polytechnic University, Pomona (CalPoly). These students majored in accounting; finance, real state and law; international business; e-business; marketing management; computer information systems; management and human resources; and technology and operations management, and all were computer literate. College students in their second year were selected to allow for a controlled sample in terms of educational level and verbal ability. The aim was to have a representative sample in terms of verbal ability of the U.S. adult population, where 52% have achieved some college education (2000 US Census).

The demographic profile for the 48 participants who provided usable reports showed that the sample was representative of the population which is expected to report using the Internet. In the sample, most participants had moderate verbal ability represented by their educational level (some college). In the sample, eighty-six percent had attained some college education 

 

73 percent of the participants were between 18 and 25 years old and 51 percent were males. Statistics also indicated that the sample was ethnically diverse with all major ethnic background (the resulting ethnicity distribution is representative of the student population of the study site).

Chi-squares for demographic variables were calculated to verify that participants were equally distributed among reporting conditions. This equal distribution of influencing factors ensures results could be attributable to the influence of the independent variable and not to other variables.

9.1 Coding and Scoring of Reports

The 48 witness reports each with an average length of four pages were coded. Scores for each report were calculated by tallying the presence of descriptive items that relate to the video clip gold standard. Three totals were obtained, number of correct items (those that match items in the gold standard), number of incorrect items, and number of confabulated items (those not present in the gold standard).

9.2 Completeness

One-way ANOVA (one-tailed) was used to test for completeness (number of correct items in report) among the two reporting methods. Using an alpha level of 0.05, significant differences in mean completeness were observed. Table 1 shows completeness data and ANOVA main effect. Completeness differed significantly across the two methods F(1, 47) = 30.97, η2 = 0.402, p < .001.

The results show that i-recall elicited 51 correct descriptive items from witnesses’ memory and the textbox system elicited 21 correct items. This indicates that i-recall elicits 143 percent more than i-textbox. Thus, results for completeness suggest that i-recall outperformed i-textbox in completeness and the direction of these results was as expected.

 

Table 1. Completeness Data by Reporting Method and ANOVA Main Effect

Reporting Method N Mean SD SE

i-textbox i-recall 25

23 20.8

51.4 14.2

23.3 2.8

4.9

Sum of Squares df Mean Square F p

Between Groups

Within Groups

Total 11271.704

16740.212

28011.917 1

46

47 11271.04

363.918 30.97* p < .001

* Effect is significant at the 0.05 level

 

Proceedings of the 43rd Hawaii International Conference on System Sciences - 2010

 

Table 2. Mean of Items in Report by Reporting Method

Reporting Method

Mean Correct Mean Incorrect Mean Total in Report

i-textbox 20.8 .56 21.3

i-recall 51.4 3.3 54.7

Table 3. Accuracy Data by Reporting Method and ANOVA Main Effect

Reporting Method N Mean SD SE

i-textbox i-recall 25

23 .965

.938 .045

.044 .009

.009

Sum of Squares df Mean Square F p

Between Groups

Within Groups

Total .008

.092

.101 1

46 .008

.002 4.212* p = .046

* Effect is significant at the 0.05 level

 

9.3 Accuracy

One-way ANOVA (one-tailed) was used to test for accuracy (number of correct items over the sum of correct items plus incorrect items plus confabulated items) between the two reporting methods. Some incorrect (incorrectly remembered) and confabulated items (not present in the video) were obtained from participants across the two methods. Incorrect and confabulated items were added together because they represented less than five percent of total items in each report (Table 2). i-recall (M=3.3) produced more incorrect items than i-textbox (M=0.56), but also elicited more information.

Using an alpha level of 0.05, significant differences in mean accuracy of reports were observed. Table 3 shows accuracy data and ANOVA main effect. Accuracy also differed across the two methods F(1, 47) = 4.121, η2 = 0.084, p = .046, but not significantly.

The results show that i-recall elicited descriptive items from witnesses’ memory with 94 percent accuracy and i-textbox elicited items with 96 percent accuracy. This suggests that 94 percent of the information i-recall elicits is correct. Also, the results indicate that i-recall approached significantly the performance of i-textbox in accuracy, but the direction of the results of the comparison with i-textbox was not as expected. It is important to note that accuracy for the two methods seems to reflect a ceiling effect, where both achieve high levels in the range of 94 to 96 percent. Two possible explanations for achieving such high levels of accuracy are, first, that the 20-minute delay between viewing the stimuli 

 

and reporting the incident was too short and, second, that following instructions of the CI, participants were instructed to avoid guessing and to report only about what they were certain.

10. Discussion

The results obtained in this user study support the hypotheses. Differences among the methods in report completeness were significant, i-recall outperformed i-textbox (51 vs. 21 correct items). Overall, the methods achieved high accuracy. i-recall achieved equal accuracy to i-textbox (94% vs. 96.5%). Koehnken et al.’s [17] meta-analysis of 55 evaluations of the CI indicates that accuracy is affected by age of participants, delay between viewing stimuli and reporting the incident, and laboratory where the studies were conducted. In this study, delay is likely the reason for high accuracy levels for the methods. Other studies of the CI have reported anywhere from 76 percent to 100 percent accuracy.

11. Conclusions

The results obtained in this study lead to four conclusions on the efficacy of i-recall in enhancing witness communication and memory retrieval to elicit accurate and complete testimonies.

First, i-recall enhances witness communication and memory retrieval. In terms of completeness, i-recall outperformed i-textbox significantly. In terms of accuracy, i-recall equaled the performance of i-textbox. The direction of these results was as

 

Proceedings of the 43rd Hawaii International Conference on System Sciences - 2010

 

expected except for the results of the comparison in terms of accuracy between i-recall and i-textbox.

Second, i-recall is a promising Internet reporting method considering existing Internet reporting implementations represented by i-textbox, such as the FBI Tips and Public Leads and the Claremont Colleges’ Silent Witness Program websites.

Third, i-recall’s performance as crime reporting and interviewing system is better than existing Internet reporting systems and maintains privacy protection and convenience qualities. i-recall is a crime reporting method with potential to attract reports from witnesses commonly unwilling to report face-to-face.

Last, natural language processing (NLP) technology allowed i-recall to apply successfully the CI memory-enhancing guidelines. Using these investigative interviewing techniques, i-recall is capable of eliciting accurate and complete witness testimonies. Provided cost-benefit analyses are conducted, computer interviews using NLP are promising technologies to offer interactive and usable interfaces in e-government services to the public.

11.1 Limitations

External validity threats are inherent to experimental designs. In this study these threats were anticipated and reduced to the extent possible, but they remain a limitation. These threats included:

Un-naturalistic reporting situation. Students acted as witnesses of a crime depicted in a video clip.

Raters bias. Efforts were made to hire independent video clip and report raters. This was limited by the resource available to conduct the study. The primary investigator participated directly in training, supervising, and at times coding the resulting 48 reports.

Possible ceiling effect for reporting methods’ accuracy. The alternative methods achieved accuracy levels in the range of 94 to 96 percent. Although these results are common in studies on the evaluation of the CI when compared with other alternative interviewing methods (i.e., standard interview), the results for i-recall’s accuracy should be taken with caution. Longer delays between viewing the stimulus and reporting the crime incident may yield more realistic results.

 

11.2 Implications

Criminology studies indicate that millions of crimes are committed annually, but only about one third of these are reported to police. Unwillingness to report, driven by inconvenience or fear, is cited as the primary reason for unreported crime. Singer advocates the creation of alternatives to mediate between police’s goals and victims’ fears [22,28]. Accordingly, law enforcement agencies launched Internet crime reporting as an alternative to face-to-face reporting. With an offer for convenience and privacy, i-recall is a promising example of Internet crime reporting that offers the benefits of producing crime reports that are more complete and as accurate as existing Internet alternatives. This represents a clear advantage over existing Internet reporting implementations. Future studies will compare the efficacy of i-recall to the efficacy of human expert cognitive interviewers and ultimately test whether i-recall is viable as a mediating alternative in natural settings.

12. References

[1] Alarid, L. F., & Novak, K. J. (2008). “Citizens' Views of Using Alternate Reporting Methods in Policing.” Criminal Justice Policy Review, 19(1), 14.

[2] Appelt, D. & Israel, D. (1999). An introduction to information extraction technology. Tutorial prepared

for the IJCAI Conference.

http://www.ai.sri.com/~appelt/ietutorial.

[3] BJS. (2005). “Reporting crime to police, 1992-2000.” Bureau of Justice Statistics Special Report.

[4] Cunningham, H., et al. (2002). “GATE: A framework and graphical development environment for robust NLP Tools and Applications”. In Proceedings of the 40th Anniversary Meeting of the Association for Computational Linguistics (ACL’02), Philadelphia.

[5] Dando, C., Wilcock, R., Milne, R., & Henry, L. (2008). “A Modified Cognitive Interview Procedure for Frontline Police Investigators.” Applied Cognitive Psychology, Published Online.

[6] Fisher, R. P., & Schreiber, N. (2007). “Interviewing Protocols to Improve Eyewitness Memory.” In M. Toglia, J. Reed, D. Ross & R. Lindsay (Eds.), The Handbook of Eyewitness Psychology: Memory for Events (Vol. 1, pp. 53-80). Mahwah, N.J: Erlbaum Associates.

[7] Geiselman, R.E., Fisher, R.P., MacKinnon, D.P. and Holland, H.L. (1985). “Eyewitness memory enhancement in the police interview: Cognitive retrieval mnemonics versus hypnosis.” Journal of Applied Psychology, 70, 401-412.

[8] Geiselman, R. E. and Fisher, R. P. (1997). “Ten years of cognitive interviewing.” In Intersections in Basic

 

Proceedings of the 43rd Hawaii International Conference on System Sciences - 2010

 

and Applied Memory Research, D. G. Payne and F.G. Conrad, Eds. Lawrence Erlbaum Associates.

[9] Geiselman, R. E. (2006). “The Cognitive Interview: Techniques for Facilitating Memory Retrieval.” Unpublished Seminar Presentation. University of California.

[10] Hope, L., Gabbert, F. and Fisher, R. (2006). “Developing a scene of crime recall tool: Investigating the impact of immediate recall on later crime event memory.” American Psychology and Law Conference, St. Petersburg, Florida.

[11] Horvath, F., Meesing, R. T., & Hyeock, Y. H. (2003). “National Survey of Police Policies and Practices Regarding the Criminal Investigations Process: Twenty-Five Years After Rand”. Submitted to the U.S. Department of Justice.

[12] Iriberri, A. (2009). Internet Crime Reporting: Design and Evaluation of a Computer Investigative Interview System. Unpublished Dissertation, Claremont Graduate University.

[13] Iriberri, A., & Leroy, G. (2007). “Natural Language Processing and E-Government: Extracting Reusable Crime Report Information.” Paper presented at the 2007 IEEE International Conference on Information Reuse and Integration, Las Vegas, NV.

[14] Iriberri, A., Leroy, G., & Garret, N. (2006). “Reporting Crime Online: User Intention to Use.” Paper presented at the 39th Hawaii International Conference on Systems (CD-ROM) Science, January 4-7, Kauai, HI.

[15] Kebbell, M. R., & Milne, R. (1998). “Police Officers' Perceptions of Eyewitness Performance in Forensic Investigations.” The Journal of Social Psychology, 133(3), 323-330.

[16] Kebbell, M. R., & Wagstaff, G. F. (1997). “Why Do the Police Interview Eyewitnesses? Interview Objectives and the Evaluation of Eyewitness Performance.” The Journal of Psychology, 13(6), 6.

[17] Kohnken, G., Milne, R., Memon, A., & Bull, R. (1999). “The Cognitive Interview: A Meta-Analysis.” Psychology, Crime and Law, 5, 24.

[18] Ku, C. H., Iriberri, A., & Leroy, G. (2008a, May 12-13, 2008). “Crime Information Extraction from Police and Witness Narrative Reports.” Paper presented at the 2008 IEEE International Conference on Technologies for Homeland Security, Waltham, MA.

[19] Ku, C. H., Iriberri, A., & Leroy, G. (2008b, May 18-21, 2008). “Natural Language Processing and e-Government: Crime Information Extracting from Heterogeneous Data Sources.” Paper presented at the Proceedings of the 9th Annual International Conference on Digital Government Research (dg.o 2008), Montreal, Canada.

[20] Lasley, J. R., & Palombo, B. J. (1995). “When Crime Reporting Goes High-Tech: An Experimental Test of Computerized Citizen Response to Crime.” Journal of Criminal Justice, 23(6), 10.

 

[21] Loftus, E. (1996). Eyewitness Testimony. Cambridge, Massachusetts: Harvard University Press.

[22] McEwen, T., Spence, D., Wolff, R., Wartell, J., & Webster, B. (2003). Call Management and Community Policing: A Guidebook for Law Enforcement: Institute of Law and Justice and Office of Community Oriented Policing Services, U.S. Department of Justice.

[23] Memon, A., & Bull, R. (1991). “The Cognitive Interview: Its Origins, Empirical Support, Evaluation and Practical Implications.” Journal of community and Applied Social Psychology, 1(291-307).

[24] Memon, A., Holley, A., Milne, R., Koehnken, G., & Bull, R. (1994). “Towards Understanding the Effects of Interviewer Training in Evaluating the Cognitive Interview.” Applied Cognitive Psychology, 8, 641¬659.

[25] NIJ. (1999). Eyewitness Evidence: A Guide for Law Enforcement. Rockville, MD: National Institute of Justice: Technical Working Group for Eyewitness Evidence.

[26] NIJ. (2003). Eyewitness Evidence: A Trainer's Manual for Law Enforcement: National Institute of Justice: Technical Working Group for Eyewitness Evidence.

[27] Peiris, R., Gregor, P., & Alm, N. (2000). The Effects of Simulating Human Conversational Style in a Computer-based Interview. Interacting with Computers, 12, 635-650.

[28] Singer, S. (1988). “The Fear of Reprisal and the Failure of Victims to Report a Personal Crime.” Journal of Quantitative Criminology, 4(3), 289-302.

[29] Sporer, S. L., Malpass, R. S., & Koehnken, G. (Eds.). (1996). Psychological Issues in Eyewitness Identification. Mahwah, New Jersey: Lawrence Erlbaum Associates, Publishers.

[30] Rand. (1975). The criminal investigation process 1-3. Rand Corporation Technical Report. Santa Monica.

[31] Willman, M. T., & Snortum, J., R. (1984). “Detective Work: The Criminal Investigation Process in a Medium-Size Police Department.” Criminal Justice Review, 9(1), 7.

[32] Womack, C. L. (2007). Criminal Investigations: The Impact of Patrol Officers on Solving Crime. Unpublished Thesis, University of North Texas.

 

Author Guidelines for the Preparation of Contributions to Springer Proceedings in Business and Economics

Alfred Hofmann1,*, Ralf Gerstner1, Anna Kramer1, and Frank Holzwarth2

1 Springer-Verlag, Computer Science Editorial, Heidelberg, Germany

{alfred.hofmann,ralf.gerstner,anna.kramer}@springer.com

2 Springer-Verlag, Technical Support, Heidelberg, Germany

frank.holzwarth@springer.com

Abstract. The abstract is a mandatory element that should summarize the contents of the paper and should contain at least 70 and at most 150 words. Abstract and keywords are freely available in SpringerLink.

Keywords: We would like to encourage you to list your keywords here. They should be separated by middots.

1 Introduction

You will find here Springer’s guidelines for the preparation of proceedings papers to be published in one of the following series, in printed and electronic form:

• Lecture Notes in Computer Science (LNCS), incl. its subseries Lecture Notes in Artificial Intelligence (LNAI) and Lecture Notes in Bioinformatics (LNBI), and LNCS Transactions;

• Lecture Notes in Business Information Processing (LNBIP);

• Communications in Computer and Information Science (CCIS);

• Lecture Notes of the Institute for Computer Sciences, Social Informatics and Tele- communications Engineering (LNICST);

• IFIP Advances in Information and Communication Technology (IFIP AICT), for- merly known as the IFIP Series;

• Proceedings in Information and Communication Technology (PICT);

• Proceedings in Business and Economics.

Your contribution may be prepared in LaTeX or Microsoft Word. Technical Instructions for working with Springer’s style files and templates are provided in separate documents which can be found in the respective zip packages on our website.

 

2 Preparation of Your Paper

2.1 Structuring Your Paper

Affiliations. The affiliated institutions are to be listed directly below the names of the authors. Multiple affiliations should be marked with superscript Arabic numbers, and they should each start on a new line as shown in this document. In addition to the name of your affiliation, we would ask you to give the town and the country in which it is situated. If you prefer to include the entire postal address, then please feel free to do so. E-mail addresses should start on a new line and should be grouped per affiliation.

Headings. Headings should be capitalized (i.e., nouns, verbs, and all other words except articles, prepositions, and conjunctions should be set with an initial capital) and should, with the exception of the title, be aligned to the left. Only the first two levels of section headings should be numbered, as shown in Table 1. The respective font sizes are also given in Table 1. Kindly refrain from using “0” when numbering your section headings.

Table 1. Font sizes of headings. Table captions should always be positioned above the tables.

Heading level Example Font size and style 

Title (centered) Lecture Notes 14 point, bold

1st-level heading 1 Introduction 12 point, bold

2nd-level heading 2.1 Printing Area 10 point, bold

3rd-level heading Run-in Heading in Bold. Text follows 10 point, bold

4th-level heading Lowest Level Heading. Text follows 10 point, bold

Words joined by a hyphen are subject to a special rule. If the first word can stand alone, the second word should be capitalized.

Here are some examples of headings: “Criteria to Disprove Context-Freeness of Collage Languages”, “On Correcting the Intrusion of Tracing Non-deterministic Programs by Software”, “A User-Friendly and Extendable Data Distribution System”, “Multi-flip Networks: Parallelizing GenSAT”, “Self-determinations of Man”.

Lemmas, Propositions, and Theorems. The numbers accorded to lemmas, propositions, and theorems, etc. should appear in consecutive order, starting with Lemma 1. Please do not include section counters in the numbering like “Theorem 1.1”.

 

2.2 Length of Papers

We only wish to publish papers of significant scientific content. Very short papers (of fewer than four pages) will not be made available for indexing and will not be visible as individual papers on SpringerLink.

2.3 Page Numbering and Running Heads

There is no need to include page numbers or running heads; this will be done at our end. If your paper title is too long to serve as a running head, it will be shortened. Your suggestion as to how to shorten it would be most welcome.

2.4 Figures and Tables

It is essential that all illustrations are clear and legible. Vector graphics (rather than rasterized images) should be used for diagrams and schemas whenever possible. Please check that the lines in line drawings are not interrupted and have a constant width. Grids and details within the figures must be clearly legible and may not be written one on top of the other. Line drawings are to have a resolution of at least 800 dpi (preferably 1200 dpi). The lettering in figures should not use font sizes

 

Fig. 1. Power distribution of channel at 1555 nm along the link of 383 km (Source: LNCS 5412, p. 323)

 

 

Fig. 2. Artifacts empowered by Artificial Intelligence (Source: LNCS 5640, p. 115)

smaller than 6 pt (~ 2 mm character height). Figures are to be numbered and to have a caption which should always be positioned under the figures, in contrast to the caption belonging to a table, which should always appear above the table.

Captions are set in 9-point type. If they are short, they are centered between the margins. Longer captions, covering more than one line, are justified (Fig. 1 and Fig. 2 show examples). Captions that do not constitute a full sentence, do not have a period.

Text fragments of fewer than four lines should not appear at the tops or bottoms of pages, following a table or figure. In such cases, it is better to set the figures right at the top or right at the bottom of the page.

If screenshots are necessary, please make sure that the essential content is clear to the reader.

Remark 1. In the printed volumes, illustrations are generally black and white (half- tones), and only in exceptional cases, and if the author or the conference organization is prepared to cover the extra costs involved, are colored pictures accepted. Colored pictures are welcome in the electronic version free of charge. If you send colored figures that are to be printed in black and white, please make sure that they really are also legible in black and white. Some colors show up very poorly when printed in black and white.

2.5 Formulas

Displayed equations or formulas are centered and set on a separate line (with an extra line or half line space above and below). Displayed expressions should be numbered for reference. The numbers should be consecutive within the contribution, with numbers enclosed in parentheses and set on the right margin. Please do not include section counters in the numbering.

 

x + y = z (1)

Equations should be punctuated in the same way as ordinary text but with a small space before the end punctuation mark.

2.6 Footnotes

The superscript numeral used to refer to a footnote appears in the text either directly after the word to be discussed or – in relation to a phrase or a sentence – following the punctuation mark (comma, semicolon, or period).1

For remarks pertaining to the title or the authors’ names, in the header of a paper, symbols should be used instead of a number (see first page of this document). Please note that no footnotes may be included in the abstract.

2.7 Program Code

Program listings or program commands in the text are normally set in typewriter font:

program Inflation (Output)

{Assuming annual inflation rates of 7%, 8%, and 10%,...

years};

const MaxYears = 10;

var Year: 0..MaxYears;

Factor1, Factor2, Factor3: Real;

begin

Year := 0;

Factor1 := 1.0; Factor2 := 1.0; Factor3 := 1.0;

WriteLn('Year 7% 8% 10%'); WriteLn; repeat

Year := Year + 1;

Factor1 := Factor1 * 1.07; Factor2 := Factor2 * 1.08;

Factor3 := Factor3 * 1.10;

WriteLn(Year:5,Factor1:7:3,Factor2:7:3, Factor3:7:3)

until Year = MaxYears end.

[Example of a computer program from Jensen K., Wirth

N.: Pascal User Manual and Report. Springer, New York

(1991)]

 

2.8 Citations and Bibliography

For citations in the text, please use square brackets and consecutive numbers. We would write [1,2,3,4,5] for consecutive numbers and [1], [3], [5] for non-consecutive numbers. The numbers in the bibliography section are without square brackets.

Please write all references using the Latin alphabet. If the title of the book you are referring to is, e.g., in Russian or Chinese, then please write (in Russian) or (in Chinese) at the end of the transcript or translation of the title.

In order to permit cross referencing within SpringerLink, and eventually between different publishers and their online databases, Springer standardizes the format of the references. This feature aims to increase the visibility of publications and facilitate academic research. Please base your references on the examples given in the references section of these instructions. References that do not adhere to this style will be reformatted at our end.

We would like to draw your attention to the fact that references to LNCS proceed¬ings papers are particularly often reformatted due to missing editor names or incomplete publisher information. This adjustment may result in the final papers as published by Springer having more pages than the original versions as submitted by the authors. Here is an example:

Reference as formatted in author’s original version:

Assemlal, H.E., Tschumperl6, D., Brun, L.: Efficient Computation of PDF-Based Characteristics from Diffusion MR Signal. In: MICCAI. Volume 5242. (2008) 70–78

Reference after reformatting by Springer:

Assemlal, H.E., Tschumperl6, D., Brun, L.: Efficient Computation of PDF-Based Characteristics from Diffusion MR Signal. In: Metaxas, D., Axel, L., Fichtinger, G., Sz6kely, G. (eds.) MICCAI 2008, Part II. LNCS, vol. 5242, pp. 70–78. Springer, Heidelberg (2008)

One more line is needed for this reference, as a result of Springer’s adjustment.

Please make sure that all your sources are correctly listed in the reference section. Do not include references to pieces of work that are not connected with your paper.

The references section at the end of this paper shows a sample reference list with entries for journal articles [1], an LNCS chapter [2], a book [3], proceedings without editors [4] and [5], as well as a URL [6]. Please note that Springer proceedings are cited with their publication acronyms and volume numbers.

 

2.9 Plagiarism

Springer takes plagiarism seriously. If an author has copied from another author or has used parts of another author’s work (text, tables, figures, etc.), without his or her permission and a reference, then the paper on SpringerLink will be given a “retracted” stamp, and an erratum explaining the reasons for the retraction will be included. In addition, the volume editors and the author’s academic supervisors will be informed that plagiarism has been committed. Please note that a retracted paper remains visible, with its “retracted” stamp. It does not simply disappear.

Acknowledgements. This should always be a run-in heading and not a section or subsection heading. It should not be assigned a number. The acknowledgements may include reference to grants or supports received in relation to the work presented in the paper.

3 Additional Information Required from Authors

3.1 Copyright Form

There are different copyright forms in place for the different Springer Computer Science proceedings book series. A prefilled copyright form is usually available from the conference website. Please send your signed copyright form to your conference publication contact, either as a scanned PDF or by fax or by courier. One author may sign on behalf of all of the other authors of a particular paper. In this case, the author signs for and accepts responsibility for releasing this material on behalf of any and all co-authors. Digital signatures are not acceptable.

3.2 Contact Author Information

Kindly assure that, when you submit the final version of your paper, you also provide the name and e-mail address of the contact author for your paper. These details are used to compile a list of contact authors for our typesetting partners SPS in India. The contact author must be available to check the paper roughly seven weeks before the start of the conference, or before the book is due to leave the printing office, in the case of post-conference proceedings.

 

3.3 Correct Representation of Author Names

Authors’ names should be written out in full at the tops of the papers. They are shortened by us to “initials surname” in the running heads and take the form “surname, given name” in the author index. If you or any of your co-authors have more than one family name, it should be made quite clear how your name is to be displayed in the running heads and the author index. Chinese authors should write their given names in front of their surnames at the tops of their papers. If you only have one (main) name, please make sure that this name is written out in full in the running heads, when you check your final PDF. Names and affiliations cannot be changed once a paper has been published.

4 Typesetting of Your Paper at Springer

Please make sure that the paper you submit is final and complete, that any copyright issues have been resolved, that the authors listed at the top of the chapter really are the final authors, and that you have not omitted any references. Following publication, it is not possible to alter or withdraw your paper on SpringerLink.

4.1 What Will Be Done with Your Paper

If the instructions have been followed closely, then only very minor alterations will be made to your paper. The production team at SPS checks the format of the paper, and if, for example, vertical spacing has been inserted or removed, then this is remedied. In addition, running-heads, final page numbers, and a copyright line are inserted, and the capitalization of the headings is checked and corrected if need be. Finally, the reference section is attuned to our specifications (see also Section 2.7). Light technical copyediting may also be performed for post-proceedings.

 

4.2 Proof Reading Stage

Once the files have been worked upon, SPS sends a copy of the final PDF of each paper to its contact author. The contact author is asked to check through the final PDF to make sure that no errors have crept in during the transfer or preparation of the files. This should not be seen as an opportunity to update or copyedit the paper, which is not possible due to time constraints. Only errors introduced during the preparation of the files will be corrected. Particular attention should be paid to the references section.

If SPS does not receive a reply from a particular contact author, within the timeframe given (usually 72 hours), then it is presumed that the author has found no errors in the paper. The tight publication schedule of our proceedings series does not allow SPS to send reminders or search for alternative e-mail addresses on the Internet.

In some cases, it is the contact volume editor or the publication chair who checks all of the PDFs. In such cases, the authors are not involved in the checking phase.

The purpose of the proof is to check for typesetting or conversion errors and the completeness and accuracy of the text, tables, and figures. Substantial changes in content, e.g., new results, corrected values, title and authorship, are not possible and cannot be processed.

5 Online Publication in SpringerLink

All papers are published in our digital library, SpringerLink. Only subscribers to Springer’s eBook packages or to the electronic book series are able to access the full text PDFs of our online publications. Front and back matter, as well as abstracts and references, are freely available for all users.

6 Checklist of Items to Be Sent to Volume Editor

• The final source files, incl. any non-standard fonts.

• A final PDF file corresponding exactly to the final source files.

• A copyright form, signed by one author on behalf of all of the authors of the paper.

• The name and e-mail address of the contact author who will check the proof of the paper.

• A suggestion for an abbreviated running head, if appropriate.

• Information about correct representation of authors’ names, where necessary.

 

References

[1] Smith, T.F., Waterman, M.S.: Identification of Common Molecular Subsequences. J. Mol. Biol. 147, 195–197 (1981)

[2] May, P., Ehrlich, H.C., Steinke, T.: ZIB Structure Prediction Pipeline: Composing a Com¬plex Biological Workflow through Web Services. In: Nagel, W.E., Walter, W.V., Lehner, W. (eds.) Euro-Par 2006. LNCS, vol. 4128, pp. 1148–1158. Springer, Heidelberg (2006)

[3] Foster, I., Kesselman, C.: The Grid: Blueprint for a New Computing Infrastructure. Mor-gan Kaufmann, San Francisco (1999)

[4] Czajkowski, K., Fitzgerald, S., Foster, I., Kesselman, C.: Grid Information Services for Distributed Resource Sharing. In: 10th IEEE International Symposium on High Perfor-mance Distributed Computing, pp. 181–184. IEEE Press, New York (2001)

[5] Foster, I., Kesselman, C., Nick, J., Tuecke, S.: The Physiology of the Grid: an Open Grid Services Architecture for Distributed Systems Integration. Technical report, Global Grid Forum (2002)

[6] National Center for Biotechnology Information, http://www.ncbi.nlm.nih.gov

 

[March 13, 2014]

 

THE UNIVERSITY OF MICHIGAN

PSYCHOLOGY STUDENT ACADEMIC AFFAIRS

POLICIES & PROCEDURES MANUAL FOR

GRADUATE STUDENTS

2019-2020

This manual outlines the rules, traditions, and regulations of graduate training in the Department of Psychology at The University of Michigan. Individual areas within the department may have

requirements that go beyond the matters considered here; hence the answer to a given inquiry

will sometimes require consultation with the Area Chairs.

 

TABLE OF CONTENTS

INTRODUCTION 2

STUDENT ACADEMIC AFFAIRS OFFICE 3

ACADEMIC REQUIREMENTS 4

Academic Standing 4

Course Requirements 4

Grades 5

Student Evaluations 5

Summer Research 5

Research Experience and the 619 Project 6

MENTORING 7

ACADEMIC PROCEDURES 8

Registration 8

Leaves and Re-Entry 8

THE FIVE YEAR PLAN FOR GRADUATE STUDY 10

GRADCARE AND INSURANCE 11

GRADUATE STUDENT AWARDS 12

TEACHING AND GSI ARRANGEMENTS 13

GSI TEACHING DEVELOPMENT AND ENHANCEMENT 14

U-M GRADUATE TEACHING CERTIFICATE 15

DEBIT FUNDS 16

RESEARCH FUNDS 17

ADVANCING TO CANDIDACY 18

EMBEDDED MASTER’S DEGREE 19

DISSERTATION 20

SEXUAL HARASSMENT AWARENESS WORKSHOPS 23

RESEARCH AND ETHICS POLICIES 24

Protection of Human Subjects 24

Subject Pool 24

PEERRS Certification 24

Supervising Undergraduate Research Assistants 24

Human Subject Incentive Payments 25

ULAM Animal Training 25

GRADUATE CERTIFICATE PROGRAMS 26

Graduate Certificate in Lesbian, Gay, Bisexual, Transgender, and Queer Studies 26

LIFE: International Program in Lifespan Development 26

Graduate Certificate in Women’s Studies 27

Cognitive Science Graduate Certificate 27

APPENDIX A—Core Courses by Area 28

APPENDIX B—Psychology Undergraduate Research Listing Template 33

APPENDIX C—Oral Defense Checklist 34

 

INTRODUCTION

The University of Michigan Psychology Graduate Program is one of the largest and best Ph.D. Psychology training programs in the world. We utilize the full resources of the University of Michigan to provide comprehensive and interdisciplinary training experiences in research and teaching. Graduate students work with many faculty in different areas and at various research centers. More than 90% of the students who began their graduate studies in Psychology since 2000 have completed their Ph.Ds. They also achieved candidacy and completed their degrees more quickly than the average student in the social sciences division of Rackham Graduate School.

The Department of Psychology offers Ph.D. training in six Areas

(Biopsychology, Clinical, Cognition & Cognitive Neuroscience, Developmental, Personality & Social Contexts, and Social) and three joint programs (Combined Program in Education &  Psychology, Joint Program in Social Work & Psychology, and the Joint Program in Women’s  Studies & Psychology). Faculty in each Area and Program determine admissions, establish required courses, and evaluate students’ progress.

Psychology Student Academic Affairs (SAA) in the Department of Psychology oversees the administration, funding, and records of Graduate Students according to the Rackham Graduate  School Academic Policies that govern all graduate programs at the University of Michigan. Students are responsible for knowing the policies and procedures in the Rackham guide, as well as the requirements of the Psychology Department. Students should maintain frequent communication with advisors, Area Chairs, and SAA throughout the course of their studies to ensure that all requirements are fulfilled.

 

STUDENT ACADEMIC AFFAIRS OFFICE

1343 East Hall

Phone: 764-2580, Fax: 764-3520

psych.saa@umich.edu 

Brian Wallace, Director

bwallace@umich.edu, 764-9179

• Manages graduate and undergraduate academic programs, including curriculum coordination, advising, and fiscal support.

Saroya Cicero, Advising Coordinator

saroyaj@umich.edu, 764-5724

• Oversees academic advising for undergraduate students; coordinates psychology student organizations; plans major-related events.

Sheri Circele, Graduate Program Coordinator

scircele@umich.edu,, 764-2580

• Supports graduate program and graduate students in areas of admissions, award applications, records maintenance, funding and events planning.

Chloe Davenport, Academic Support Assistant

chloedav@umich.edu, 764-2580

• Provides academic support services to faculty, staff and students and manages SAA front office services.

Kaydee Fry, Honors Program and AMDP Coordinator, Academic Advisor kayfry@umich.edu, 764-2580

• Provides academic advising to undergraduate students; coordinates the AMDP and Honors programs for Psychology and BCN students.

Tina Griffith, Graduate Program Coordinator

tinagrif@umich.edu, 647-3936

• Provides graduate and undergraduate support, with a particular focus on graduate student funding, admissions and events.

Megan Leonard, Time Scheduler

mwolgast@umich.edu, 764-5605

• Maintains the time schedule and course guide for all graduate and undergraduate courses, processes GSI appointments.

Briana Peikert, Academic Advisor

jliddico@umich.edu, 647-6243

• Advises undergraduate students and assists with event planning and website maintenance.

Sarah Wagner, Subject Pool Coordinator

sarahwag@umich.edu, 764-2580

• Coordinates the Subject Pool for introductory Psychology classes and coordinates Project Outreach.

 

ACADEMIC REQUIREMENTS

A) Academic Standing

A student is in good academic standing if he or she:

• Is making satisfactory progress toward the completion of degree requirements and is within the time limits of the degree program

• Is demonstrating an ability to succeed in the degree program; and

• Has a cumulative grade-point average of 3.00 (B) or better.

B) Course Requirements

After consulting with the faculty advisor and receiving the endorsement of the student’s Area Chair, requests for substitutions and exemptions of required courses should be made in writing to Student Academic Affairs.

1) Statistics—Each first-year student must complete a two-semester sequence of statistics (PSYCH 613 and 614) with a grade of B- or better. Clinical students register for this course as Psych 988. Students may choose alternative courses based on their backgrounds and experience with statistics. Please consult with the SAA Chair to discuss options in training.

2) Core Courses—Faculty in each area determine the core courses required in their areas. See Appendix A.

3) Breadth Requirement—Graduate students must take one breadth course outside their own Area but within the Psychology department, prior to advancement to candidacy. Students in joint programs (i.e., Women’s Studies/Psychology, CPEP, Social Work/Psychology) do not need to take a breadth course within Psychology as they already have a broad curriculum. Clinical students are required to take two Breadth courses (see Appendix A).

4) Cognate Requirement—Rackham Graduate School requires the completion of a minimum of four hours of graduate-level course work in a discipline different from the student’s field of study but related to some aspect of this field. Students should consult with their advisors about which cognates will best supplement their work in Psychology. Clinical students meet the cognate requirement by taking Psych 988.

Students must receive a grade of B- or better in order for the course to fulfill the cognate requirement. Courses offered for "S"/"U" credit fulfill a cognate requirement; however, audited ("VI" grades) courses do not. Graduate-level courses include any courses listed in the Rackham Graduate School: Programs of Study.

Students who have done graduate-level cognate work prior to enrollment at Michigan should consult with their Area Chair and faculty advisor about the possibility of using those courses to fulfill this requirement. Prior cognates must be approved by Rackham. A course in the Psychology program that is cross-listed as a course in another program may satisfy the cognate requirement.

 

5) Teaching Academy— Each first-year student must complete a two-semester sequence of the Psychology Teaching Academy (PSYCH 609). During the winter term students will also complete a .25 GSI assignment as a course component.

6) Ethics Course—The Rackham Graduate School requires each graduate student to complete a course on responsible conduct of research and scholarship. Students will fulfill this requirement by taking PSYCH 506 in the winter term of their first year.

7) Grades—Core courses, statistics requirements, and cognate coursework must be completed with a grade of B- or better. However, an overall average of "B" is required of all students in graduate programs.

Individual readings or research courses (i.e., Psych 619) will be graded "S" or "Y" if the work is in progress. If "Y" grades are used, faculty will submit a grade upon completion of the research.

Students may receive a grade of Incomplete (“I”) only if the work remaining to be done for the course by the end of the semester is minimal and the instructor approves an extension for completing the unfinished work. The instructor must agree to this arrangement and determine a deadline for finishing the assigned work before a grade is given. The notation of “I” remains a permanent part of the academic record. When coursework is completed to the satisfaction of the instructor, the grade will appear on the transcript as, for example, “IB+.” The GPA is based only on hours of coursework completed.

Students cannot choose to take a course "Pass/Fail.”

Students may drop a course within the first three weeks of a term without consequence. If a student drops a course after the third week, a grade of “W” will be posted to the student’s transcript.

C) Student Evaluations

Each student's performance will be evaluated every May by the Area or Program faculty. The evaluation will assess progress in coursework, research, and formal and informal teaching. Students are asked to submit a progress report and vita. The funding plan for the coming year will be provided to each student to discuss with their research mentor. In the event that recommendations are made regarding changes in advisor, transfer to another Area, interruption of study, or recommendation to leave the graduate program, a representative of the Area will meet with the student and the Psychology Associate Chair for Teaching and Student Affairs.

D) Summer Research

Students are required to submit a brief (1-2 pages) summer research proposal to the SAA Office in May. The proposal will describe the student’s planned summer research activities. In early September, students will submit a brief summary of what they accomplished. Both documents will be accompanied by approval of the student’s advisor.

 

E) Research Experience and the 619 Project

Students must complete an independent research project as a requirement for candidacy and a master’s degree. Students begin their Psychology 619 research project during the first term and complete it during their second year. This project should demonstrate the student’s independent research competence in methodology, data analyses, and scholarly

interpretation. Students register for 619 credits with their advisor during the first and second years of graduate study, but grades are deferred until the project is completed. The requirement is a finished research paper that is evaluated by two readers, the student’s advisor and a second faculty member; the second reader may be a faculty member in another Area or department.

 

MENTORING

The department recognizes the fundamental importance of mentoring in the careers of graduate students. Therefore, graduate students are matched with a faculty advisor in their area prior to the start of their first year. The purpose of the faculty advisor is to act as a career and research mentor as well as to provide guidance about what courses to take, program requirements, research planning and implementation, and generally assist with students’ academic plans. In order for this to be successful, it is expected that students and advisors will meet on a regular basis. Students should keep their advisor up to date on their research progress and any changes or additions to their academic experience including course selection, funding plans, research and teaching activities and more. Students are encouraged to seek out additional mentors who will broaden their research interests and abilities. As appropriate, mentors can come from outside the department or University. However, they should add to the quality of the student's research, academic and professional development.

Please see Rackham’s resources on Mentorship for further guidance:

Mentoring and Advising 

How to Get the Mentoring You Want: A Guide for Graduate Students

 

ACADEMIC PROCEDURES

All graduate students are required to register each Fall and Winter term under Rackham’s Continuous Enrollment Policy. Students are also required to register during their term of defense, even if the defense takes place during the Spring/Summer term.

A) Registration

Students can register for most Psychology graduate courses without faculty permission. If students receive the message “permission of instructor required” when they attempt to register for a course, they should contact the instructor directly to receive permission. Instructors will need the student’s UM identification number in order to provide permission to enroll.

For Psych 619, Psych 990, and Psych 995, students simply use the Graduate Student  Override Request Tool for permission to enroll. The instructor does not need to contact the SAA Office for an override into these courses. Once the permissions are processed by the Psychology SAA Office, students receive an e-mail stating they can register for the course through Wolverine Access.

When students are enrolled as Candidates, they must elect PSYCH 995 for 8 hours each term and may elect additional courses. As a GSI, the student can elect to register for additional course work without paying additional tuition. As a fellowship student, Candidates may elect either one additional course OR multiple courses for a total of no more than 4 credits each term. A student who does not take any courses during a term may “bank” the free course for the next term. In the following term only, they can either register for courses totaling up to 8 credits or take no more than two courses totaling more than 8 credits. Students using banked courses should consult with the SAA Office before registering. If a Candidate who is not a GSI elects more than the allowed additional courses, that student will pay the cost of the additional tuition.

After the first term, students will receive registration appointments to begin enrolling in courses. Students are expected to register during the early registration period. Students must register for at least one course before the first day of classes. If students do not register before the first day of class, they will be charged a $50 late fee and may incur late payment fees. The Psychology Student Academic Affairs Office does not pay any late fees.

Students may register for a maximum of 18 credits. Students will need to contact the SAA Office if they will exceed this limit.

B) Leaves and Re-entry

Before taking a leave, students are required to discuss their options with their advisor and Area Chair, and also the Student Academic Affairs Office. International students are required to consult with the International Center before taking a leave, as US immigration regulations may restrict eligibility.

Leaves are requested through Rackham’s Leave of Absence system

online: https://secure.rackham.umich.edu/leave/

 

Students on an official leave are not able to make progress towards their degree, and are not eligible for certain University services. Before taking a leave, it is necessary to create a tentative re-entry plan with your advisor and the SAA Office.

There are four types of leaves:

• Medical: A student will be granted a leave of absence for a serious physical or mental health condition that prevents continued participation in the program.

• Family Necessity or Dependent Care: A student will be granted a leave of absence to take care of a serious circumstance that directly affects a family member, such as a death, serious health condition, financial difficulty, or other critical life situation; or to provide care for a dependent incapable of self-care because of age or disability.

• Military Service: A student will be granted a leave of absence for the duration of a military service obligation to the country of citizenship.

• Personal Reasons: After completing at least one term, a student may request a one-term, non-renewable leave of absence for personal reasons. No additional documentation is required for this type of leave.

When a student is ready to return from a leave, the SAA Office and advisor should be notified to update the re-entry plan. A request to return from a leave of absence is also completed through Rackham’s online Leave of Absence System.

More information and checklists regarding leaves can be found at this

website: http://www.rackham.umich.edu/current-students/policies/doctoral/phd 

students/leave-of-absence 

 

THE FIVE-YEAR PLAN FOR GRADUATE STUDY

The Department provides full financial support for ten academic terms over five years for every Graduate Student in good standing.

In general, students serve as Graduate Student Instructors (GSIs) during six terms, and they are supported by various fellowships, grants, and departmental funds during the remaining four terms.

There can be some variation among students regarding the timing of their teaching terms. In addition, there may be differences due to outside fellowships. However, the SAA Office strives to maintain equity in funding and resources. The general model of our five-year plan is shown below.

Year in

Program Fall Winter Spring/Summer

Year 1 Psychology Fellowship

(T, S, GC) GSI (T, S, GC)

Psych Fellowship (S) Psychology Fellowship

(S, GC)

Year 2 GSI

(T, S, GC) GSI

(T, S, GC) Psychology Fellowship

(S, GC)

Year 3 GSI

(T, S, GC) GSI

(T, S, GC) Psychology Fellowship

(S, GC)

Year 4 GSI

(T, S, GC) Psychology Fellowship

(T, S, GC) Psychology Fellowship

(S, GC)

Year 5 Psychology Fellowship

(T, S, GC) Psychology Fellowship

(T, S, GC)

T = Tuition

S = Stipend

GC = GradCare

The SAA Office can provide students with up-to-date information on their financial awards, dates of payment, and assist with changes to the funding plan. Students are encouraged to contact the SAA Office with questions or concerns.

There are additional funds from Rackham to handle student emergencies. Information about the Rackham Graduate Student Emergency Funds can be found here.

 

GRADCARE AND INSURANCE

All Graduate Students are eligible for medical and dental insurance. When students are on fellowship funding, their medical insurance is GradCare. For a brief overview of GradCare, please go to the following website: http://www.uhs.umich.edu/gradcare 

New students will be automatically enrolled in GradCare and Dental Option 1 beginning September 1st of their first year. The Benefits Office will send an email to new enrollees with instructions on how to elect additional benefits. Students are required to complete online benefits selections within 30 days of their first day of eligibility or they will receive the default insurance option. If you are not electing health care coverage through GradCare, please notify the SAA Office of the alternate source of your health care coverage (764-2580 or psych.saa@umich.edu).

When students are in employment status (as a GSI, GSRA or GSSA), they are eligible for any medical coverage through the Benefits Office. The Benefits Office will send an email to newly appointed students with the additional coverage options.

Information about benefit eligibility can be found on the following

website: http://benefits.umich.edu/benefitgroups/grads.html 

For more information regarding benefits, feel free to contact the Benefits Office at 615-2000 locally, or 1-866-647-7657, or benefits.office@umich.edu.

 

GRADUATE STUDENT AWARDS

There are many financial awards administered by the Department. The Student Academic Affairs Office publicizes the awards and deadlines. A complete list may be found on the Graduate Student Funding Opportunities web page. The SAA Curriculum Committee reviews applications and selects nominees or recipients for the various awards.

APA Dissertation Research Awards

APA Student Travel Award

APF/COGDOP Graduate Research Scholarships

Barbara A. Oleshansky Memorial Fund

Barbara Perry Roberson Fellowship

Blunt Family Fund

Clyde Hamilton Coombs Scholarship in Mathematical Psychology

Daniel Katz Dissertation Fellowship in Psychology

Department of Psychology Dissertation/Thesis Grant

Edward S. Bordin Graduate Research Fund

Elizabeth Munsterberg Koppitz Child Psychology Graduate Fellowship

Eric Bermann Research Award

Hough Summer Research Fellowship for Psychology & Ethics

Marquis Award and Rackham/ProQuest Distinguished Dissertation Award

Mary Malcomson Raphael Fellowship

Naomi Lohr Award

Patricia Gurin Award

Pillsbury Graduate Research Award

Rackham Barbour Scholarship for Asian Women

Rackham International Student Fellowship

Rackham Outstanding GSI Nominations

Rackham Predoctoral Fellowship

Roger W. Brown Award

Ruth C. Hamill Award

Susan Lipschutz, Margaret Ayers Host and Anna Olcott Smith Awards for Women Graduate

Students

 

TEACHING AND GSI ARRANGEMENTS

Teaching and learning are intertwined experiences in graduate school, and the Psychology Department strives to maintain high quality experiences in both endeavors. The Psychology Department endorses the University's commitment to training of future teachers. Therefore, being a Graduate Student Instructor (GSI) in various courses and working to improve one's teaching effectiveness throughout graduate school is considered essential to becoming a skilled professional psychologist.

A) The GSI expectation for all graduate Psychology students on the five-year plan is one term of teaching at a .25 level and five terms of teaching at a .50 level. The .25 level appointment will occur in the winter semester of the student’s first year in conjunction with the Teaching Academy course. Two terms of teaching at the .50 level must occur before the end of the second year, and the other 3 by the end of year four. Students have the opportunity to design and teach their own course in the final term of teaching. The teaching responsibilities are usually in the Department of Psychology, but may be outside the department if approved by the student's advisor and the Student Academic Affairs Chair.

B) GSI workload can be reduced if a student is supported by non-Departmental funds that include tuition, stipend, and GradCare. No exemptions from teaching are allowed for lesser alternative funding. If a student has full funding from outside sources, the teaching requirement will be reduced to a minimum of two total terms of teaching. The rationale is that students benefit from teaching experience even if not needed for support.

C) Types of external funding that qualify for a reduction in teaching requirements

• External Fellowships: Awards include student-initiated applications to federal agencies and philanthropic foundations for Graduate Student funding, such as the National Science Foundation, the Ford Foundation, the American Psychological Association, the Department of Defense, and others.

• University Fellowships: Awards include student-initiated applications to departments such as Rackham, the Institute for Social Research or other departments at the University of Michigan.

• Faculty Funded Awards: Funding may be derived from research grants and fellowship programs in which the student is selected on the basis of individual merit. This includes GSRA Appointments within the Psychology Department, or elsewhere in the University.

 

GSI TEACHING DEVELOPMENT AND ENHANCEMENT

Department of Psychology GSI training is in addition to the mandatory CRLT training required of all first time GSI’s. Participation in the fall training session is mandatory for all GSIs who have not attended at least three fall training sessions.

The morning sessions will include a series of workshops. Presenters will include Psychology GSI training directors Brian Malley and Shelly Schreier, other Department of Psychology faculty, and additional University of Michigan professionals (CRLT, CAPS, Office for Students with Disabilities, Psych IT, and library support staff, etc.). GSIs may select which workshops they will attend. Topics will change from year to year, but will include standard workshops such as:

• QPR (behavioral intervention with distressed students)

• Using technology in the classroom

• Working with different learning styles in the classroom setting

• CTools support

• Teaching Certification Preparation

Afternoon sessions will include a keynote topic selected on a rotating basis such as relaxation training, time management, diversity in the classroom, working with diversity in the classroom, and Teaching Certification preparation.

In addition to the mandatory Fall training, GSIs appointed in the Department of Psychology for the first time will also attend a one day workshop at the end of the Winter term or in the spring/early summer as scheduling permits which goes over the "nuts and bolts" of teaching including syllabus preparation, first day of class concerns and lesson planning.

GSIs teaching for the first time in a Winter term will participate in a mandatory, one-day training in December to go over the "nuts and bolts" of teaching in the Psychology Department. In addition to this training, new Winter GSIs also attend a session at the end of January or beginning of February to serve as a "midterm" evaluation as well as the opportunity to provide additional resources and discuss teaching strategies.

Prior to the Fall GSI training, the program directors will solicit suggestions for topics which the GSIs would like to have addressed.

In terms that GSIs are not attending a departmental training session, they are required to complete one CRLT workshop.

 

U-M GRADUATE TEACHER CERTIFICATE

This program offers graduate students at the University of Michigan an opportunity to document professional development as college-level instructors and prepare for the faculty job search. The Certificate is coordinated by the Rackham Graduate School and the Center for Research on Learning and Teaching (CRLT) at the U-M. Participants who complete all program requirements receive a U-M Graduate Teacher Certificate. The Certificate does not appear on official U-M transcripts, but may be included on one’s curriculum vitae. This program is free and open to all U-M graduate students.

The U-M Graduate Teacher Certificate documents one’s professional development in six areas:

• Orientation to college-level teaching and learning

• Exposure to new teaching strategies through seminars and courses

• Experience as a Graduate Student Instructor for at least 2 semesters at the University of Michigan

• Mentorship on teaching from a faculty member

• Consultation on classroom teaching from an instructional consultant

• Preparation of a teaching philosophy statement

For general information about the U-M Graduate Teacher Certificate and its requirements, see: http://crlt.umich.edu/um.gtc.

 

DEBIT FUNDS

Debit accounts are intended to provide equitable and flexible funding during a student's graduate program. Students must be in good standing in their program. The funds are intended mainly to support students' research and travel expenses to professional conferences. Debit accounts are administered through the Student Academic Affairs Office.

Funding Levels

• The annual funding for debit accounts is $400/year over five years, for a maximum of $2,000. In the event of budget fluctuations that affect the department, the annual allowance may be increased or decreased in a uniform manner for all students.

• Students in the Joint Program in Social Work & Psychology, and Psychology & Women’s Studies will be allocated $200/year by the Psychology Department. Students should consult their respective office for information on similar funding from the joint program.

• Students in CPEP will receive funds through their program office.

Allowable Expenses

Debit accounts can be used to purchase consumable research supplies, copying and printing, research services (such as transcription or data analyses), memberships in professional organizations, journal dues, travel to professional conferences, and for professional development. The Student Academic Affairs Office will be the final authority on allowable expenses.

Procedures & Authorization

• Every student must submit requests for debit fund expenditures to the Student Academic Affairs Office. An electronic request form can be accessed on the Graduate Financial  Matters webpage. The request should include an itemized list of expenses as well as the purpose and explanation of the expenditure. Each request needs to be reviewed by the Student Academic Affairs Chair before authorization is granted for the expenditure.

• If receipts are already available, they should be provided with the funds request. If the purchases have not yet been made, receipts must be submitted within two weeks after the expenses have been incurred. Failure to provide the appropriate documentation will result in a freeze on the student's debit account.

• The expected rate of spending is $400 annually, and the balance of funds in a student's debit account will carry forward each year.

• All student debit accounts expire upon completion of the program.

• Students can monitor their debit funds activity. A link to the tool is located on the Graduate Financial Matters webpage.

 

RESEARCH FUNDS

Research funding is available through Rackham Graduate School by application. Each student is eligible for two non-competitive research awards, one of $1,500 as a pre-candidate, and a second of $3000 as a candidate. The website for application information can be found here.

Once a Rackham Research Grant has been received, students may request reimbursement in writing to the SAA Office. An electronic request form can be accessed on the Graduate  Financial Matters webpage. The request should include an itemized list of expenses as well as the purpose and explanation of the expenditure. Each request needs to be reviewed by the Student Academic Affairs Chair before authorization is granted for the expenditure.

 

ADVANCING TO CANDIDACY

Before Candidacy is achieved, a student must complete all course requirements specified by the student’s Area or program, and pass the preliminary examination. Preliminary exams vary across Areas, and they include different demonstrations of scholarship such as take-home exams, writing grant proposals, and preparing portfolios. Students should consult the Area faculty for details and timelines for completing this requirement. It is usually completed before the start of the student’s third year. In order to advance to candidacy, students must be in good standing and have completed the 619 research project.

Please submit the “Approval of Candidacy Status” form and “619 Requirement Completion” form to the Student Academic Affairs Office. With the endorsement of the Area Chair and following a review to confirm that all requirements have been fulfilled, Psychology Student Academic Affairs will recommend to Rackham that the student be advanced to Candidacy. (All required forms can be found on the Student Resources webpage.)

Term in which student will be eligible

for Candidacy (enrollment in course

995) Tentative Deadline for student’s

completion of ALL Candidacy

requirements (including prelims)

Fall 2015 September 10, 2015

Winter 2016 January 8, 2016

During the term that a student is waiting for Candidacy approval, he or she should register for 8 credits of PSYCH 990. When Candidacy is approved, the Registrar's Office will change all 990 enrollments to 995 for the student.

 

EMBEDDED MASTERS’S DEGREE

Students accepted into the Psychology Ph.D. program may elect to apply for a Master’s degree and will be able to receive one upon satisfying the requirements listed below. For those students who leave the graduate program, the Master’s degree signifies the successful completion of a significant amount of graduate work.

A) Requirements

• Two-semester sequence of statistics (613/614) with a grade of B- or better

• Psychology 619 research project

• Two core courses within their area of specialization

• One breadth course (courses inside Psychology but outside the area of specialization)

• Four credits in cognate coursework (courses outside of Psychology).

• An overall B average or better

• 24 credit hours of course credit; four of these hours are comprised of cognate coursework

When the requirements are fulfilled, students may apply for the degree by completing the “Checklist for the Degree of Masters of Arts” form. Joint Program students should also complete the “Change of Program or Dual Degree Application” form and have it signed by both programs. A student does not have to be enrolled to receive the Master’s degree. (All required forms can be found on the Student Resources webpage.)

B) Deadlines

Students who wish to have their names included in the Commencement booklet must apply for graduation through Wolverine Access by the dates shown on the Rackham Graduate  School website. Applications for graduation will be accepted until the last day of classes of the term in which the student wishes to receive their degree/diploma; however, those students’ names will not appear in the commencement program.

 

DISSERTATION

A) Dissertation Chair

The student should work closely with the research advisor to plan a dissertation and see the process through to completion. It is critical to maintain regular contact with the dissertation Chair throughout the process.

As early as possible, but at least six months prior to the planned completion date, the student and Chair should form an official dissertation committee, and hold a prospectus meeting. It is best to include all committee members in the dissertation planning, and to discuss the timeline for the research with them.

B) Committee Formation

In consultation with a faculty advisor, a student selects the members of the dissertation committee. This committee should be formed as soon as possible after the student has achieved Candidacy. The Psychology Student Academic Affairs Office completes dissertation committee paperwork to forward to Rackham when the Approval of Dissertation Prospectus is submitted to the SAA Office.

Please take note of the following regulations when forming your committee:

1. Three of the four required members of the committee (including the chairperson and the cognate members) must be of professorial rank, i.e., assistant, associate, or full professor. To be eligible, assistant professors must have received their Ph.D.

2. The fourth and any additional members can be non-professorial academic faculty or people from outside the university. However, appointments of a non-professorial person to a dissertation committee must be accompanied by the person’s CV and a supporting memo from the student, co-signed by the dissertation committee chair and the Student Academic Affairs Chair. The Nomination for Special Membership form will be completed by the Psychology Student Academic Affairs Office and forwarded to Rackham.

3. At least two members of the committee must have appointments in the Department of Psychology.

4. Any faculty member with an "unmodified" appointment (i.e., not visiting, not adjunct, etc.) may serve as a cognate member provided he or she is of professorial rank, a regular graduate faculty member in a Rackham doctoral program, and has no appointment in or significant affiliation with the Department of Psychology. Students in the joint programs do not need a cognate member on their dissertation committee.

Please see Rackham’s Guidelines for Dissertation Committee Service and their Quick Reference Chart, for more information on forming your committee.

 

C) Prospectus

Students write a dissertation prospectus containing: (1) an abstract of the specific aims of the investigation; (2) the background and significance of the proposed research, including the conceptual framework; (3) the research design and methods of procedure, including measurement techniques to be used, if applicable; (4) analysis strategies to be followed; (5) a tentative timetable.

The dissertation committee then meets as a group to discuss the proposal. A draft document should be given to committee members two weeks before the meeting date. At the meeting, the student provides a 45 minute talk on the planned research, and the faculty ask questions and discuss the plans. The student then leaves the room, and the committee discusses whether the prospectus will be approved.

Once approval of a dissertation prospectus is indicated on the Approval of Dissertation  Prospectus form, signed by the dissertation committee at the prospectus meeting, and countersigned by the Area Chair, it should be forwarded to the Psychology Student Academic Affairs Office. At that time, the Psychology Student Academic Affairs Office will prepare the Dissertation Committee Form and submit it to Rackham. The Dissertation Committee form should be submitted to Rackham at least six months before the defense.

D) Defense Term Registration and Degree Conferral Deadlines Students register for 8 credits of Psych 995 in the term in which they defend.

Degrees are conferred by the Board of Regents three times per year—in April, August, and December. The Rackham Graduate School has two deadlines in each term (an early deadline and a late deadline) by which defense requirements must be met. Students who want to have their degrees conferred in the term in which they defend must abide by the early deadline. For more information about doctoral degree deadlines, please go to  Rackham’s website. 

E) Pre-Defense

Several steps must be followed at this stage of the program. See Appendix C – Oral Defense Checklist. Prior to the dissertation defense, students register online for a Rackham Group Pre-Defense meeting, at which time format guidelines are reviewed along with the requirements for doctoral degree completion. More information about the Pre-Defense Meeting can be found at: http://www.rackham.umich.edu/doctoral students/

The student and dissertation committee chair are responsible for scheduling the oral defense and insuring that the committee evaluations are submitted to Rackham on time. The dissertation committee is expected to approve the dissertation (or recommend changes required before approval may be granted) after the oral defense. The dissertation chair is responsible for collecting committee signatures and filing the necessary forms with Rackham.

It is highly advisable to circulate a draft of the dissertation as early as possible for committee members. It is wise to meet individually with the committee members to see if any concerns

 

need to be addressed before holding the meeting. A final draft must be circulated at least two weeks before the defense occurs.

F) Oral Defense

The final step in the dissertation process is the presentation and defense of the dissertation to the Committee. This is a formal meeting that is open to any member of the university community. However, as it is an examination, students are advised not to invite friends and family members outside of the academic community. No audio or video recording of the defense is permitted.

At the defense, the meeting begins with the committee gathered alone in the room. After discussion among the committee members, the candidate is invited into the room, along with any other attendees. The candidate presents a formal talk for 45 minutes summarizing the results of the research. Then, committee members ask questions and discuss the research. Finally, the candidate is asked to leave the room, along with any other attendees, while the committee deliberates about the outcome of the examination. The candidate is then invited back in to hear the feedback from the committee. Often, changes to the written dissertation are required for final approval.

Celebration of the completion of the defense is best scheduled for a separate occasion.

G) Post-Defense

Once all revisions are submitted to the committee chair for approval, the student can register online for a Rackham Post-Defense Meeting. The meeting must take place before the appropriate Degree deadline. Refer to the Rackham Degree Deadlines website for dates and guidelines.

 

SEXUAL HARASSMENT AWARENESS WORKSHOPS

The University of Michigan values a campus environment that allows for the safe and respectful exchange of ideas. Sexual harassment undermines such an environment. In an effort to encourage awareness of and conversation about sexual harassment, the College of Literature, Science, and the Arts (LS&A) has partnered with the CRLT Players to develop a series of workshops for faculty and graduate students.

Attendance at one of these workshops is mandatory for all newly admitted LS&A graduate students. More information about how to register will be provided at the New Student Orientation.

 

RESEARCH AND ETHICS POLICIES

The University of Michigan complies fully with the federal regulations regarding education of graduate students about ethical concerns in the field. Seminars are completed twice during training as required by the NSF and NIMH for recipients of funding.

Protection of human subjects

All students in the Department of Psychology are required to submit any research proposal (including, but not limited to 619, 719, 819, prelims, dissertation) that uses human subjects, including the Introductory Psychology subject pool, for Internal Review Board (IRB) approval before any collection of data. Complete details for submitting proposals can be found at the IRB website (http://www.irb.umich.edu/). All proposals must be co-signed by a supervising faculty member.

Subject Pool

Allocation of subjects from the Introductory Psychology Subject Pool (PSYCH 111 & 112) is a process separate from human subject approval. In order to use the subject pool, graduate students must request hours and provide the Psychology Student Academic Affairs Office with copies of the IRB Application, IRB Approval letter, and Consent and Debriefing forms. Any questions regarding the subject pool should be directed to the Subject Pool Coordinator in the Psychology Student Academic Affairs Office in 1343 East Hall (e-mail: subject.pool@umich.edu or Phone: 734-764-2580).

PEERRS Certification

All Psychology Graduate Students are also required to obtain PEERRS Certification before they conduct research with human or animal subjects. See the PEERRS website for

details: http://my.research.umich.edu/peerrs/ It is recommended that you obtain this certification sometime during your first term in the PhD program.

Supervising Undergraduate Research Assistants

With approval from a faculty advisor, graduate students may engage undergraduates as research assistants to work on research projects. Typically, the undergraduate will register for one of the Psychology undergraduate independent study courses; however, graduate students are not able to serve as the main sponsor of an undergraduate honors thesis. The various courses for which the undergraduate may register are described on the Independent Study webpage.

There are restrictions on the number of credits students may take in each of the independent courses. Please encourage prospective research assistants to consult the Psychology Student Academic Affairs Office if they have questions about independent study credit. As always, graduate student researchers shouldn’t hesitate to contact the SAA Office with questions. Researchers may recruit undergraduate research assistants by advertising on the Department of Psychology Research Listings. For the template for research listings, see Appendix B. Requests to have research listings posted to the website should be sent to psych.saa@umich.edu.

 

Human Subject Incentive Payments

All payments given to study participants need to be processed through the University’s Human Subject Incentive Payments (HSIP) system through Wolverine Access. Please see this website for more information: http://www.treasury.umich.edu/hsiptrainingresources.htm. To use debit funds, or funds from a Rackham Research Grant to pay study participants, please contact the SAA Office for more information.

ULAM Animal Training

Graduate students who will be using animals in their research are required to abide by the University Committee on Use and Care of Animals (UCUCA) policies and procedures. The Unit for Laboratory Animal Medicine (ULAM) manages the learning plans for each student based on their individual lab’s activities and needs. Learning plans and required trainings can be accessed through the MLearning system. Students should consult with their PI or lab manager for information on required trainings and access to MLearning.

 

GRADUATE CERTIFICATE PROGRAMS

Graduate Certificate in Lesbian, Gay, Bisexual, Transgender, and Queer Studies

Designed for students already enrolled in a terminal degree program at the University of Michigan, the Certificate in LGBTQ Studies consists of graduate course work totaling 15 credit hours. The Certificate, which can be combined with either a masters or a doctoral degree, aims to:

• Provide an interdisciplinary analysis of the function of sexuality, and particularly sexual identity, in the construction of individuals, as a form of minority discourse, as a signifier of cultural representations, and as a site of power.

• Examine the processes by which sexual desires, identities, and practices are produced, represented, regulated, and resisted in the U.S. and globally, both in the past and in the present.

• Address sexuality in a way that consistently demonstrates its interconnections to gender, race, ethnicity, and class.

Coursework for the Certificate in Lesbian, Gay, Bisexual, Transgender, and Queer Studies involves one core course, Introduction to LGBTQ Studies; one additional course in Women’s Studies on sexuality; and two courses, including one outside the discipline, on sexuality or LGBTQ topics. It culminates in an advanced research project designed to incorporate a LGBTQ perspective.

For more information refer to the webpage for Women’s Studies Graduate Certificate Programs.

***

LIFE: International Program in Lifespan Development

As an International Max Planck Research School, LIFE is a multidisciplinary graduate training program that links together the University of Michigan, the University of Virginia, the University of Zurich, and three institutions in Berlin: Humbolt University, Free University, and the Max Planck Institute for Human Development. Our substantive focus is on the social, behavioral, and neuroscience of human development across the life course. LIFE is a forward thinking “virtual institute” that takes an integrative and interdisciplinary approach to connecting evolutionary with ontogenetic, cultural and institutional perspectives on human development.

LIFE fellows are graduate students who are interested in an interdisciplinary life course perspective in human behavior and who are pursuing a doctorate on aspects of human development in one of several disciplines (e.g., psychology, neuroscience, educational science). The program consists of seminars at the home institution, twice yearly teaching and research Academies (alternating between Ann Arbor, Berlin, Charlottesville, and Zurich), participation in workshops, and collaborative mentorship of dissertation and other projects by LIFE faculty of the partner institutions. LIFE provides opportunities for fellows to work at one of the partner institutions abroad for several months. In addition, LIFE promotes international networking and professional development as an integral part of graduate training.

 

Please visit https://www.imprs-life.mpg.de/en/life-program. For more information about this program, contact either of the program Co-Chairs: Patricia Reuter-Lorenz (parl@umich.edu) or Toni Antonucci (tca@umich.edu)

***

Graduate Certificate in Women’s Studies

Designed for students already enrolled in a terminal degree program at the University of

Michigan, the Certificate in Women’s Studies consists of graduate course work totaling 15 credit hours. The Certificate, which can be combined with either a masters or a doctoral degree, aims to provide:

• analyses of contemporary feminist theoretical frameworks and methodologies, and their implications for academic disciplines and professional practices

• an inclusive approach that examines the intersection of gender and other social identities and categories of analysis

• an opportunity to broaden and enrich analytical skills in one or more disciplines while drawing on the interdisciplinary perspectives of Women’s Studies

• a challenge to the traditional separation of academic theory from political and professional practice

Coursework for the Certificate in Women’s Studies (established in 1982) involves core courses in feminist theory and methodology and advanced courses on women and gender in the discipline. It culminates in an advanced research project designed to incorporate a feminist perspective.

Address inquiries to wspgradinquiry@umich.edu or call the Women’s Studies Program at 734¬0615-7619.

***

Cognitive Science Graduate Certificate

Are you interested in being a part of a rapidly-growing area of study? Cognitive Science is the interdisciplinary scientific study of the mind and its' processes, the brain, and behavior. The Graduate Certificate Program includes the intersection of multiple research disciplines, including psychology, artificial intelligence (computer science), philosophy, linguistics, and neuroscience. The program's emphasis on formal methods in theory and data analysis will widen the opportunities for graduate student careers in government and industry, outside the traditional tenure track academic positions where there is increasing demand for such technical skills.

The Graduate Certificate Program coursework will give students cohesive training in essential approaches to the study of cognition. The flexible program will also provide opportunities for students to engage in creative and interdisciplinary research opportunities with cross-departmental faculty.

For more information refer to the webpage for Cognitive Science Graduate Certificate.

 

APPENDIX A

DEPARTMENT OF PSYCHOLOGY CORE COURSES BY AREA

BIOPSYCHOLOGY

PSYCH 731: Advanced Seminar and Practicum in Physiological Psychology

A total of three advanced lecture or seminar courses relevant to biopsychology must be taken, and at least two of these must be at the 600-level or above. Some examples of these are: the Psych 808 seminars, Psych 831, Psych 530, Psych 532, and Psych 533. The faculty advisors will assume the responsibility for assuring that the student’s course selection is adequate preparation for their professional career.

Students are required to take at least one course in Neuroscience and one in Evolutionary Biology. There are a number of courses that meet this requirement; students should consult with their advisor for the appropriate selection.

CLINICAL

PSYCH 670: Research Design and Evaluation in Clinical Psychology (year 1 or 2)

PSYCH 672: Introduction to Intervention and Clinical Ethics (Winter year 1)

PSYCH 771: Topics in Clinical Science and Practice (taken for credit in years 1 & 2)

PSYCH 776: Prosem: Clinical Science in Historical and Cultural Contexts (year 1 or 2)

PSYCH 778: Psychological Assessment I and II (year 1)

PSYCH 874: Theories of Adult Psychotherapy Introduction

PSYCH 875: Introduction to Child Therapy

PSYCH 877: Lifespan Psychopathology: Childhood and Adolescence

PSYCH 878: Lifespan II: Adult Psychopathology

PSYCH 978: Evidence-Based Psychotherapy Laboratory and Practicum 1

PSYCH 978: Evidence-Based Psychotherapy Laboratory and Practicum 2

COGNITION AND COGNITIVE NEUROSCIENCE

PSYCH 634: Human Neuropsychology

PSYCH 643: Theory of Neural Computation

PSYCH 644: Computational Modeling of Cognition

PSYCH 722: Decision Processes

PSYCH 72l: Mathematical Psychology

PSYCH 741: Basic Processes in Cognition & Cognitive Neuroscience

PSYCH 742: Complex Cognition

PSYCH 743: Human Learning and Memory

PSYCH 745: Psychology of Language

PSYCH 958: Learning, Thinking and Problem Solving

 

DEVELOPMENTAL

PSYCH 759: Proseminar in Developmental Psychology

PSYCH 75l: Cognitive Development: Perception, Learning and Memory

PSYCH 756: The Development of Language and Communication Skills

PSYCH 757: Social Development

PSYCH 758: Developmental Neuroscience of Human Behavior

PSYCH 793: Emotional Development

PSYCH 796: Development in Infancy

PSYCH 797: Development in Adolescence

PSYCH 798: The Psychology of Aging

Students will enroll in Psych 759 for a total of four terms; fall and winter of their first and second year. Students will also take three developmental area core courses over their 5 year academic career. The selection of courses should constitute a balance of cognitive and social aspects of development. Both topic areas are important for gaining a sound understanding of human development.

Note: Not all Developmental Core Courses are offered on a regular basis (in particular 756, 758, 793, 796, 797, and 798). It is suggested that students contact the relevant faculty and Area Chair to obtain information about the likelihood particular courses will be offered during their five years in the graduate program.

PERSONALITY & SOCIAL CONTEXTS

PSYCH 653: P&SC Orientation

PSYCH 654: Classical and Modern P&SC Theories

PSYCH 854: Advanced P&SC Research Methods I

PSYCH 855: Advanced P&SC Research Methods II

Two area seminar courses (taught at 600 level or above by P&SC faculty), which cannot be double-counted to also meet the breadth or cognate requirement.

Joint students need only take one semester of P&SC Research Methods and one P&SC area seminar, but they are encouraged to consult with their advisor about taking the second semester of Methods and a second area seminar.

 

SOCIAL

PSYCH 681: Survey of Social Psychology

PSYCH 682: Advanced Social Psychology

PSYCH 685: Social Psychological Theories

PSYCH 782: Cultural Psychology

PSYCH 785: Group Processes

PSYCH 786: Research Methods in Social Psychology

PSYCH 787: Emotions

PSYCH 788: Attitudes and Social Cognition

During the first year students will enroll in Psych 681 in the fall and winter as well as Psych 682 in fall only. In the fall and winter of year two, students will take Psych 685 as preparation for the prelim exam. In either year one or two, students are expected to take a methods course (Psych 786). Social psychology students are expected to gain substantial conceptual and empirical knowledge within social psychology by electing two of the following four core courses designed to cover the field: Psych 782, 785, 787, or 788.

COMBINED PROGRAM IN EDUCATION AND PSYCHOLOGY (CPEP)

EDBEHAVR 800/PSYCH 861: Proseminar in Educ Psych (first semester) EDBEHAVR 801/PSYCH 862: Proseminar in Educ Psych (second semester)* EDUC 708: Cognition and Instruction in the Classroom (taken in first year) EDUC 721: Human Development and Schooling (taken in first year) EDUC 716: Advanced Seminar on Issues in Education and Psychology** EDUC 898: Professional Development Seminar**

Three (3) PSYCH breadth core courses (one in Developmental and two additional from any area)

One (1) Educational Foundations Course (focused on educational or psychological theory

applied to educational topics) from the following list:

EDUC 709: Motivation in the Classroom

EDUC 720: Social & Personality Psychology Perspectives in Education

EDUC 790: Foundations of Schooling

EDUC 791: Foundations of Teaching and Learning

(Other foundational in the School of Education with advisor and program approval)

See CPEP Student Handbook for statistics, methods and other program requirements.

*Not currently offered for an unspecified trial period. The two semesters have been combined

into the first semester to allow students more flexibility in their schedule.

** For third year and above students. Required for all 4th and 5th year students.

 

JOINT PROGRAM IN SOCIAL WORK AND SOCIAL SCIENCE

(1) Each student will take the pro-seminar (SW 800) plus at least five doctoral courses in social work.

(2) The five courses should include at least one course in three of the four curriculum areas (PIP, SSS, Research Methods, and Social Context). With the approval of the doctoral director, students can take one Independent Study class (DOC 900, 971-974, 975-978) as one of the five courses.

DOC 800: Proseminar in Social Work and Social Science

PRACTICE, INTERVENTION AND POLICY (PIP)

DOC 806: Young People Making Community Change

DOC 810: Principles and Processes of Individual Change

DOC 811: Group Intervention for Individual and System Change

DOC 812: Couples, Marital and Family Intervention

DOC 813: Intervention in Human Service Organizations and Social Service Networks

DOC 814: Community Intervention

DOC 815: Policy Analysis, Development and Implementation

DOC 816: Racial, Ethnic, and Gender Factors in Intervention

DOC 817: Preventive Intervention

SOCIAL SERVICE SYSTEMS (SSS)

DOC 820: Historical Analysis of U.S. Social Service Systems

DOC 821: The Future of Social Services in the U.S.

DOC 822: Structure of the Contemporary U.S. Social Service Systems

DOC 823: Comparative Cross-National Analysis of Social Service Systems

DOC 824: Clinical Services Research: Quality, Effectiveness, Outcome

RESEARCH METHODS FOR PRACTICE AND POLICY

DOC 830: Advanced Methods in Clinical Research (PSYCH 811)

DOC 831: Research Methods for Evaluating Social Programs and Human Service Organizations

DOC 832: Research Methods for Social Policy Analysis

DOC 833: Research and Development for Human Service Innovation

SOCIAL CONTEXT FOR PRACTICE AND POLICY

DOC 840: Individual and Family Functioning and Well Being

DOC 841: Social Participation

DOC 842: Social Equality and Equity

JOINT PROGRAM IN WOMEN’S STUDIES AND PSYCHOLOGY

PSYCH 613/614: Statistical Methods I and II

PSYCH 619: Individual Research

WOMENSTD 501: Proseminar in Women’s Studies

WOMENSTD 530: Feminist Theory

WOMENSTD 891: Advanced Research

 

One of the following (3 credits): 

WOMENSTD 601: Approaches to Feminist Scholarship in the Humanities

WOMENSTD 602: Approaches to Feminist Scholarship in the Social Sciences

Additional coursework in WS or crosslisted in WS: 9 credit hours See the Women’s Studies Ph.D. Handbook for additional information.

 

APPENDIX B

PSYCHOLOGY UNDERGRADUATE RESEARCH LISTING TEMPLATE

Please submit the following information to psych.saa@umich.edu in order to post your research

assistant position on the Psychology Department website.

Project Director or Contact Person:

Project Director or Contact Person’s Phone Number:

Project Director or Contact Person’s E-mail:

Title of Project:

Major Area of Psychology in which the project is located:

Area(s) of Psychology to which the program is related:

Description:

Time Commitment Requested:

Dates of Project:

Qualifications of Student:

Credit Offered ___ Money Offered ___Experience only

Work Study Other

Faculty Members Only:

____This research opportunity may lead to an Honors Thesis in the Psychology Honors

Program.

___

Please return completed form to: Student Academic Affairs Office (1343 EH) or email

to psych.saa@umich.edu 

 

APPENDIX C

ORAL DEFENSE CHECKLIST

q Student: Schedule Prospectus Meeting. This should be done at least 6 months prior to defense. Take Approval of Dissertation Prospectus form to meeting.

Date of Meeting

q Student: Submit Approval of Dissertation Prospectus form to SAA Office.

q SAA Office submits Dissertation Committee Form to Rackham. Date: .

q Student and Chair: Approve committee formation in Wolverine Access. Date: .

q Rackham approves Dissertation Committee. Member names can be viewed on transcript once they have been approved.

q Student: Works with advisor to complete research. Confer with committee members that the thesis is complete and ready for the defense to be scheduled.

q Student: Register online for Rackham group Pre-Defense Meeting

at https://secure.rackham.umich.edu/OARD/predef/. Pre-defense meeting must take place at

least 10 days prior to oral defense.

q Rackham notifies SAA Office when student has attended pre-defense meeting. Date of

Rackham Pre-Defense Meeting: .

q SAA Office: Confirm that student is registered in defense term.

q Student: notify SAA Office of oral defense date. Date:

q Student: Distribute Dissertation Evaluation Forms to committee members, along with copy of dissertation and abstract, at least 10 working days before oral defense.

q Student: Three days before oral defense, view committee on Wolverine Access to confirm evaluation forms have been submitted to Rackham.

q Student: Pick up oral defense packet from Rackham, and hand to Chair at defense.

q Student: Do oral defense of dissertation. It is recommended that you reserve a room well in advance of your defense so that you can test AV equipment and become familiar with the layout of the room.

q Student or Chair: Submit Final Oral Examination Report to Rackham within 48 hours of defense.

q Student: Submit any required revisions to chair.

 

q Chair: Confirm that Final Oral Examination Report and Certificate of Dissertation Committee Approval Form have been completed and submitted to Rackham electronically.

q Student: Register for post-defense meeting. Date .

Submit electronically:

o Final Oral Examination Report (if not previously submitted.)

o Complete, final, pdf copy of dissertation and text of your abstract.

o Certificate of Dissertation Committee Approval Form Survey of Earned Doctorates

o Submit dissertation online in OARD Office. Bring dissertation on flash drive.

o Have completed NSF Survey of Earned Doctorates (required)

q Student: Apply for graduation through Wolverine Access. Notify your mentor about whether you will attend a graduation ceremony at the University.

q Student: Check Wolverine Access for financial status. Reconcile any balance due on account.

q Student: Send new address, email, and position information to psych.saa@umich.edu Please stay in touch with us through the years as your address changes!

q SAA Office: Run Degree Conferral Report. Degree conferred on

 

Journal of Theoretical and Applied Information Technology

15th April 2017. Vol.95. No 7 

© 2005 – ongoing JATIT & LLS  

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

 

MEASURING PROCESS INNOVATION ON DOUBLE 

FLANKED CONCEPTUAL MODEL FOR KNOWLEDGE

SHARING ON ONLINE LEARNING ENVIRONMENT

1S.M.F.D SYED MUSTAPHA, 2BIJU THERUVIL SAYED, 3ROSHAYU MOHAMAD 1Computer Science Department, College of Computers and Information Technology

Taif University, Saudi Arabia

2Department of Computer Science, College of Arts and Applied Sciences,

Dhofar University, PO Box 2509, PCode 211, Salalah, Sultanate of Oman

3School of Science and Technology, Asia e University No4, Jalan Sultan Sulaiman, 50000

Kuala Lumpur, Malaysia

E-mail: syed.malek@tu.edu.sa, b_sayed@du.edu.om, roshayu.mohamad@aeu.edu.my

ABSTRACT

There are various innovation models that were discussed in the literature and the adoption is based on the organizational needs for their business contexts, vision and applications. These innovation models require effective innovation process framework to be followed. SECI Model has been chosen as knowledge creation model to facilitate innovation through knowledge sharing and creation. While literature has shown that SECI model has been applied in various fields such as management, manufacturing, education and business, very few has considered it as innovation tool for online learning environment. Knowledge creation requires community who has enculturated with knowledge sharing as part of the practices. For this purpose, Community of Practice (CoP) has been chosen as the essentialities for the prospective innovative community and consequently to make implementation of SECI model a success. Community with CoP values are postulated to provide the right organizational setting for innovation. It is suggested that both SECI Model and CoP are integrated as new a conceptual model being regarded as double flank strategy that synergizes to prepare the right community setting and to facilitate innovation through knowledge creation. Subsequently, this paper proposed the methods and approaches in measuring innovativeness in online learning environment based on the double flank conceptual model called DFCMI.

Keywords: Knowledge Management, SECI Model, Community of Practice, Online Learning, Measuring Innovation

 

1. INTRODUCTION

Innovation models are defined and practically used in various ways by the academics and industrialists. The variants are due to the numerous contexts of what innovations are needed for in order to bring values to the innovators, consumers and organizations. It varies at different business strategies ranging from process innovation, service innovation, product performance innovation, branding innovation, organization structure innovation, design innovation, consumer network innovations, profit model innovation and many more. To achieve the innovation, the fundamental elements should be in place. Some of these involve community engagement, sharing of vision, volunteerism, ad hoc idea generation and mutual interaction. These are essentials to form an innovative community prior to building a practical innovative community.

 

Innovation in the organization must be participated at community level rather than an individual effort. We believe Community of Practice (CoP) is a suitable platform for establishing a special group to instigate innovation and intellectual forum. CoP was [12] introduced by Ettiene Wenger who had defined the CoP characteristics where the common interest group should fulfill. It is not merely a task-based committee that is formed merely to tackle certain issues and to make recommendations or produce solutions. Rather, the group must be formed and to be in practice for certain duration of time before certain level of engagement emerges within the CoP community. CoP needs to be in place as it overcomes the barrier in communication, aligns shared vision, steers volunteer participation and engagement and shares common knowledge resources in which these are the important foundation to stimulate group of innovators.

 

1499

 

Journal of Theoretical and Applied Information Technology

15th April 2017. Vol.95. No 7 

© 2005 – ongoing JATIT & LLS  

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

 

Many theoretical frameworks have been discussed in the literature, on various types of innovation process [3]. Innovation process is categorized by the factor driving the innovation such as technology [4], market [8], integration between market and R & D [1], interactive [7], network [11] and open [6], to name few. Rothwell [9] described the generation of innovation models based on the industrial technology evolution began from linear model in which the innovation is based on invention through R&D (Research and Development). The subsequent generation of innovation model is market driven that had caused innovation products to be developed as less sustainable due to strong influence of market trend. To overcome the problem, the later generation of innovation model balanced both the market demand and R&D as essential role in innovation, so-called coupled model. Interactive model has similar fashion with network model as described by Manley [7] in her illustration of several more innovation models with similar approaches. Open model is in contrary to the earlier version of innovation model where opportunity for participating innovativeness is non-exclusive and open to public as main character in participating the innovation process.

Innovation model that were described in the literature were adopted at the workplace and industrial environment. Our work is interested to look at implementation of the innovation in the online environment were members are not physically connected. The process of innovation involves on the knowledge creation, implementation, reuse and value creation. For this to happen, we adopted SECI model as the innovation model for knowledge creation and to be implemented in the online environment. Even though SECI model was originally developed for the manufacturing environment, the recent work had shown that it has encroached community learning in the web-based environment [2] [5][13][10].

In our work, we postulate that it is required to establish the CoP environment as the pre-requisite to create a community that stimulates innovative environment, demonstrate how knowledge creation can be implemented in the digital environment and finally a proposed computational framework for measuring innovation. In the subsequent subsections, we do reflection on SECI model based on the previous effort of applying it in online environment, illustrate the innovation with respect to CoP and finally the measurement of innovation in the online environment.

 

2. COMMUNITY OF PRACTICE (COP)

Community of Practice can be described as a natural formation of group members who have common interest, free-will membership in a malleable organizational structure, non-tasked based but have shared set of problems to deal with and mutually involved and participated [23]. It was believed that the structure is socially created rather than formally set up like task force or task-oriented committee [24]. Somewhat in a later development, some practitioners have argued that CoP can be cultivated and designed for a specific group which was not initially established as CoP group, to enhance the performance [25][26]. For example, an organized group of specialists who are not acquainted well with each other assigned to solve a complex unprecedented problem could be designed to be in a CoP-based group in order to obtain promising CoP values. Another important development which is related to our work is the emergence of CoP in the virtual environment such as in online learning [28][29].

There are three crucial characteristics of CoP which are domain, community and practice [30]. Domain is referred to the topics and subjects of interest that are shared. For example, a group of students who face similar problems with new regulations introduced by the authority, gather to share similar topic to discuss why the needs of such rule, how to counter propose and what are the implications of non-compliance. The domain may not necessarily require the members to be the technical experts or specialized topics that are not comprehended by others rather the members are composite of various level of expertise for that particular topics. The community involves mutual engagement and participation such that they learn from each other, contribute to each other’s affair, share related resources and knowledge and have frequent interactions. Group of boys playing cricket daily may not be a community of practice unless, they share cricket techniques and helping to develop skills and knowledge. The practice is the practical aspects of community where they share repertoire such as stories, experiences, information and other resources. The sharing must lead towards generating new artefacts for the benefit of the community. For example, a group of marketing and sales executives from various organizations who shared their problems and know-how may not necessary become the community of practice unless those shared items are further developed into useful items such as “manual guide for strategic marketing”.

 

1500

 

Journal of Theoretical and Applied Information Technology

15th April 2017. Vol.95. No 7 

© 2005 – ongoing JATIT & LLS  

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

 

Table 1 CoP Essentials, Values and Supporting

Technolo

CoP Essentials

(described in

[31] as CoP Characteristics) Supporting

Technology CoP values

Community Structure e-forum,

discussion thread, bulletin board Volatile structure, free-will membership

Learning through Participation and Reification CSCW (Computer Support Collaborative Work), Web-based Collaborative Software Generating new artefacts from the group participations and engagement; new ideas and solutions

Negotiation of Learning Project

Management

Software,

Collaborative

Virtual

Environment Traceable evolutionary of idea generation and new knowledge as a result of repetitive discussion

Learning as Temporal Project

Management

Software,

Collaborative

Virtual

Environment Incremental development of knowledge resources

Boundary

Objects Internet forum,

threaded

discussion Multiple membership to different group discussion and resource sharing across different group

Boundary Encounters Internet forum and multiple channel threaded discussion Multiple membership and sharing similar thought across different group

Mutual Engagement Social Network Analysis In-out relationship demonstrate the level of participation and involvement; analysed the relationship as cordial, animus or collaborative

Joint Enterprise Social Network Analysis Frequent exchange of ideas between individual, balanced participation on selected issue and playing leading role.

Share Repertoire Multimedia objects Individual sharing resources created or modified from others

Identity Social Network Analysis Recognizing individual’s character or role in the online group, such as leadership role, knowledge expert

Community of Practice has recently been implemented in the online and virtual environment. This is made possible as the virtual community has been in existence since the internet revolution and it has benefitting in many sectors mainly in 

 

education, manufacturing, financial services and other knowledge-intensive industries. With the modern supporting technology, the essential elements that are required to be in place to fulfil the criteria for CoPs are now feasible if virtual CoP is to be implemented. The essentials of CoP that we believe can be implemented with the supporting technologies to extract some CoP values, as shown in Table 1. We do not deny that there are other CoP values that are not discernible in computational form.

The presence of the CoP in the online environment is to be detected in digital form. Given the list mentioned in Table 1, while the computer technologies are currently available to capture most of the CoP characteristics, but not all CoP values can be sensed in a straightforward manner. For example, in boundary encounters, the membership of individual learner can be detected through its registration, but to monitor who share the knowledge earned from one CoP group across another CoP group can be a challenging task in computational context. Another example is on mutual engagement where relationship between members need to be traced. Relationship of two individuals can be defined objectively through the summation of response value for each interaction that occur, by assigning some values such as +1 for cordial, - 1 for animus and 0 for collaborative. The relationship is labelled based on the highest value of the response. However, the relationship which is based on emotion and subtle relation will be difficult to extract merely from text processing. In the following section, we describe SECI model and its implementation in online learning environment.

3. SECI MODEL

 

Figure 1 Classical View of SECI Model (By Ibmgroup - Using a image editor, CC BY-SA 3.0,

https://commons.wikimedia.org/w/index.php?curid=18653983)

SECI model is a popularly known model for researchers in various disciplines mainly for those looking at implementing knowledge innovation for their organizations. The model describes the

 

1501

 

Journal of Theoretical and Applied Information Technology

15th April 2017. Vol.95. No 7 

© 2005 – ongoing JATIT & LLS  

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

 

knowledge creation to take place through the conversion of two types of knowledge – tacit knowledge and explicit knowledge. There are four types of conversion so-called socialization, combination, externalization and internalization as shown in Figure 1.

In the first quadrant, Socialization is the process where tacit knowledge is transformed within itself where new tacit knowledge is generated. This occurs during conversations, discussions and meetings [19] where one’s tacit knowledge is shared with another person and probably it is received and perceived in the same substance or biased due to “stickiness” [20]. The second quadrant, Externalization is the process where tacit knowledge is transformed into explicit knowledge through actions such as minutes writing, note taking, diagram drawing and reflection writing [17]. Since the source of the knowledge is in a tacit form, some researchers may pose a challenging question on how one would be able to determine whether the explicit knowledge is derived entirely from tacit knowledge or a mixture of another explicit knowledge. In SECI model, tacit knowledge and explicit knowledge is treated as discrete form and hence the shift movement of the knowledge type is assumed to be in a single-type form. The third quadrant is Combination process where new form of explicit knowledge is produced as the result of reorganization, reclassification, compilation or other means of demonstrating of knowledge regeneration explicitly. This may require combination between tacit knowledge and explicit knowledge. For example, in building a prototype requires one’s tacit knowledge for his design skills and explicit knowledge in a form of reference manual for rules and guidelines. These two forms of knowledge thatcould be earned from previous transitions (socialization and externalization). The final quadrant is Internalization where knowledge is converted into one’s tacit knowledge from the explicit knowledge regenerated from the combination process. At this stage, the learner established a kind of ownership to the knowledge earned and the degree of control towards the knowledge determined by his effort towards it ([21], pp 8). In the context of a learner, one achieves a high level of internalization once he has mastered a specific knowledge and skills. SECI model suggests the knowledge conversion moves in a spiral manner to induce knowledge creation and knowledge sharing.

3.1. Applying Seci Model In Online Learning Environment 

 

Application of SECI model in the online learning environment had been investigated to gauge relationship between e-learning and SECI model; and LMS (Learning Management System) and SECI Model [32]. The outcome of the experiment indicated that knowledge creation occurred and the knowledge creation process prescribed in SECI model took place in the e-learning environment. SECI model had also been applied as a framework in virtual learning environment [33]. While the attempt was to implement in virtual environment, Hosseini admitted that not all processes in SECI model can be implemented in the virtual environment. For example, in the socialization, the actual meeting where face to face interactions were done instead of using face to face online meeting. Not least to mention, Chatti [34] was among the earliest to describe the possibilities of implementing SECI model in web-based environment and he highlighted few possible technologies that could support such SECI model activities. However, the paper did not mention specifically how it can be done. The recent work illustrated the effort to build an integrated platform for facilitating some selected SECI model activities [10] which specifically mention which activities for each quadrant of SECI model and which technology that would support them. Our emphasis in this paper is that CoP is the precondition to SECI model as knowledge creation will not take place without the effective knowledge sharing process. We believe that CoP prepares a solid platform for knowledge and integration between CoP and SECI model as a double-flanked framework for facilitating and measuring innovation in the online learning environment. The framework will be discussed in the subsequent section.

4. DOUBLE-FLANKED CONCEPTUAL MODEL FOR MEASURING INNOVATION (DFCMI)

 

1502

 

Journal of Theoretical and Applied Information Technology

15th April 2017. Vol.95. No 7 

© 2005 – ongoing JATIT & LLS  

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

 

MEASUREMENT

TECHNOLOGY

Figure 2 Double-Flank Conceptual Measuring

Innovation (DFCMI)

The problems with knowledge sharing had been discussed elsewhere and the reasons are many. Vajjhala and Hassan [35] reported that in a small and well-structured organization such as small enterprise where employee numbers are rather small, there still exist a resistance for knowledge exchange and the reasons are mainly on cultural and motivational issue. It is not about cultural differences since the workers are locals but rather the intrinsic cultural values that occur within its own culture and this requires changes. Bureš [36] had elaborated well on the factors contributing to cultural issues at both individual and social levels. The factors stated by Bureš in his paper (page 58), were “loss of power”, “fear from revelation”, “uncertainty”, “illusion of reward deprivation”, “single culture elements”, “difference between awareness and knowledge”, “conflict of motives”, “language”, “conflict avoidance”, “bureaucracy and hierarchy”, “incoherent paradigms” and “underestimating of lower levels”. Ahrend et al [37] identified reasons for barriers in knowledge sharing as “trust among colleagues”, “decision structures where lack of autonomy and flexibility”, “poor incentives for sharing” and “improper ICT infrastructure to support knowledge reposition and sharing”. There are many more literature reporting barriers in knowledge sharing [38][39][40]. In our view, the essence of the barriers to knowledge sharing is on the social factors more than technology, while the latter is crucial for facilitating the knowledge sharing process.

Our proposed innovation measuring model (DFCMI) was based on the two important theories which are CoPs and SECI model where technology is the backbone to support the activities and processes in knowledge sharing since it occurs in the online environment as shown in Figure 2.

 

DFCMI has placed CoP on top of SECI model as prerequisite for establishing social cohesiveness. Social cohesiveness encompasses the values that are prescribed under CoP discussed (refer to Table 1). It is shown that CoP and SECI are an attached entity which signifies that the community who are participating in the SECI model activities are CoP compliances and group members uphold CoP values. Social cohesiveness overcome the issues raised in the earlier literature describing the barriers for knowledge sharing. CoP value emphasizes on shared repertoire which includes the communication language, common jokes, problems, vision, strategy, solution and even knowledge artefacts. Depending on the organizational setting on CoP, the “wall” between the management and workers maybe permeable if both are involved in the CoP. In many cases, management fails to communicate their vision and goals effectively as that information are presented in a formal presentation through verbal or strategic plan manual and this syndrome is called “single culture elements”. In addition, this overcomes the problems in “underestimating lower levels” and “bad appraisal of co-workers” since the management team members are also part of the CoP teams. CoP requires mutual engagement where every member participates in one way or the other and establish him/herself in a notable manner and not merely as listener or viewer. The common perception of “knowledge is power” is the cause for fear of “loss power” as members deemed each other as “contender”. Joint enterprise counter the “fear from revelation” and “uncertainty” as members who reveal the knowledge will have some useful feedbacks as they are acquainted to each other such that the feeling of embarrassment for not getting appropriate feedback will not occur. “illusion of reward deprivation” can be addressed by recognizing one’s identity in the community as identity is an essential characteristic in CoP. One’s continuous contribution to the success of the organizational performance will be noticed and reward is redeemed from this recognition. Resistance to take risk and to avoid “rocking the boat” is common fear among workers who are fear for being blame. CoP encourages mutual development of new ideas rather than individual effort, as such, the risk is taken in collective manner. Boundary encounters breaks the “wall” that prevents transparency in the inter-departmental communication. Other aspects that cause the knowledge sharing barriers such as “trust between colleagues or management”, “ill feelings bad emotions among members” and “pseudo innovators” will diminish gradually as CoP values

 

1503

 

Journal of Theoretical and Applied Information Technology

15th April 2017. Vol.95. No 7 

© 2005 – ongoing JATIT & LLS  

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

 

built up within the CoP members. Traditionally, CoP values are monitored and observed through subjective approach by the consulting companies who are presence physically at the premise of the organization. In the online environment, detecting the presence of CoP activities in the online environment would be different from the physical environment and hence, some methods are needed to be considered as alternative to be able to capture activities that are compliances to the CoP characteristics.

At this stage of discussion, based on the literature and the DFCMI illustration, the following assumptions are made:

a. Online-based for CoP activities – it is possible to perform online activities that leave digital traces as evidences as proof to some selected characteristics of CoP [31].

b. Online-based activities on SECI model – it is possible to monitor the online activities that are supporting the evidences of the online activities based on SECI model for an individual.

Based on the above assumptions, the measurement of innovativeness is proposed in the following section.

5. MEASUREMENT OF INNOVATION BASED ON DFCMI

Literature has shown that innovation in online learning subscribes to three categories: innovation in technology, innovation in the pedagogy and academic administration and innovation in the learner’s learning strategy. In the technological perspective, adaptive learning covers aspects on the knowledge representation techniques, strategies for presenting knowledge based on learner’s preferences, evaluation mechanism and automated feedbacks to the learners (refer to [41], page 4 – 5). Another example is using technology to analyse the learner’s learning behaviour and learning pattern, so-called learner’s analytics for the consumption of the students and faculty members [42]. In the second category, Twigg has provided few case studies and reviews on various aspects in the online learning innovation with regard pedagogy and academic administration [43]. Some aspects encompass increasing the accessibility to higher education, managing cost in higher learning, developing new strategies to increase interest in learning, improve success rate in degree completion, understanding learning styles and 

 

improving learning experiences. Learning strategy had been a field of study by itself independent from the technology for decades. Traditionally, learning strategy focus on the classroom learning, teaching practices, teaching professionalism and cognitive psychology, constructivism and learner’s learning behaviour [44]. With the emerging internet technology, collaborative learning software applications and social media, the learning strategy must embrace to these developments.

In our proposal for DFCMI, activities of the learner in the online activities to stimulate innovation is given the primary focus alongside with the technology, learning strategy and pedagogy. Online learning is a broad area and the scope of defining innovation can be indefinite. DFCMI narrow the innovation process to four dimensions: formation of innovative community, evolution of special interest topics, learner’s participation in knowledge development and social recognition. For each of the dimension, there is an element of innovation which should be measurable. Firstly, formation of innovative community is on voluntarily basis such that the members are passionate about the group’s interest such that the interest, vision, objection, mission of the group must be clear, free-flow registration and de-registration on the group to focus on genuinely and potential members, building profile of him/herself with live video (self-introduction) and background information for each member to expedite process of acquaintance and socialization. The profile of the members must be kept active and up-to-date to ensure reliability and validity of the members. The profile may include member’s personal collections (includes articles, images, videos and links) which are relevant to the subjects of group as indication one one’s commitment to the group. Member earns score that contributes to the building the profile of him/herself from each of these activities. We regard this as innovation process which is significant towards attaining than innovation product. The activity is captured using keystroke loggers to capture user’s activity.

The second dimension is special interest topic which is usually a temporal affair for the group member and usually disperse after resolution is achieved or loss of interest. It evolves because of the demand from the current problem that require immediate attention. The discussion in the online discussion platform focusses on issues relevant to the topic. Hence, there are three major measurable parameters which are the life-span of the topic (when it is activated or become inactive) which can be traced from the log file, the continues growth of

 

1504

 

Journal of Theoretical and Applied Information Technology

15th April 2017. Vol.95. No 7 

© 2005 – ongoing JATIT & LLS  

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

the knowledge sources through knowledge sharing (uploading and downloading of multimedia files), participation on crowd-sourcing and collaborative application software (wiki, blogs, forum etc). While these activities need to be translated into some measurable items, some of them pose some challenges to determine the relevancies of the knowledge sources to the special interest topics. This is due to complexity in analysing the content of the multimedia format such as video, audio, images and so forth. For this purpose, collaborative filtering [45] or crowdsourcing would be the appropriate solution towards using group members to assess and evaluate the relevancy of the content posted by other member. The mechanism to collect input from the members and to sum up the total score based on the crowd input need to be established to compute the relevancy of the content. The overall score for the level of process innovation in establishing special interest group will consider the temporal information, knowledge growth [47] and contributing factors of the member based on the relevancy of the posted content.

Thirdly, the learner’s participation in knowledge development is an inducement to the innovation process for knowledge creation. Online communities are built in different model and among those are for information dissemination such as Twitter, LinkedIn; work-related collaboration such as in e-forum, Blackboard or social purposes such as Facebook, FishBrain etc. Two important aspects for the tools for the software is the facilities for knowledge sharing and tools for assessing and analysing each movement and action of the members in the online environment. The computed values for the member activities are used to indicate the process innovation in knowledge development. For example, the number of articles shared with the group, forum participation in terms of quality and involvement across various discussion channels and enriching the knowledge repository for the group are activities that can be traced, translated to numerical values and calculated some values to assign some scoring to them.

Fourthly, the members participating on online activities have rewards in terms of social recognition. Despite in the online environment, one may dispose oneself as the leader among the members from its dominance over the group discussion and possession of knowledge [48]; or as a follower [49]; as well as domain expert [50][51]. The challenges to deal with the online communities in identifying expert would be the dynamic change of the community structure in terms of membership enrolment, topic of discussions, participatory behaviour of the members such that the roles of the members may change over in temporal manner. The mechanism for measuring the participation of the member and assigning scores to rate the level of expertise and leadership is suggested in [52]. Based on the scores, the innovation process of identifying the social status of the members in the online environment and recognizing them are possible.

The four dimensions recommended for measuring innovation process are based on some of the CoP essentials (ref Table 1). Another aspect of DFCMI is on innovation product which should occur as the consequences to the successful innovation process. SECI model is a conceptual framework for knowledge creation which includes forming new ideas, generation of new artefacts, improvement to policies, services and procedures, design and modelling and others which lead towards innovation. SECI model has been considered as tightly linked to innovation [53] while others regard it as knowledge creation model. In the perspective of online learning, the knowledge creation through online activities are the innovation products. For example, summarization of read articles, solutions to complex problems, critical review of the scientific literature, arguments in intellectual forums are generated from the cycles of knowledge creation.

DFCMI stipulates that the innovative community has to experience the four quadrants of the SECI model in a repetitive manner. Some selected SECI model activities that are implementable in online environment that leave traces as evidences of a member who had involved in the four types of knowledge transformation are described in [10]. In order to ensure the group discussion to stay focus on the specialized topic, the relevancy checking and monitoring are performed as background engine [47]. The checking of the relevancy is however, only on the text-based discussion. There is need to explore further on other types of knowledge media such as video, audio, images without going through complex image processing techniques to determine the content and its relevancy to the topic of discussion.

In our view, measuring innovation may receive many criticism as innovation can be defined broadly and it appears in many situations as tangible product which is not presentable in digital format. However, the process innovation is possibly monitored in the online environment albeit with some limitations which will pave way for us to explore more for further research.

1505

 

Journal of Theoretical and Applied Information Technology

15th April 2017. Vol.95. No 7 

© 2005 – ongoing JATIT & LLS  

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

 

6. CONCLUSION AND FUTURE WORK

In this paper, we present the idea of measuring innovation process based on the proposed double-flank model which integrates CoP and SECI Model. The innovation process is a medium platform for a community to embark to innovative community prior to be able to produce various innovative products such as service innovation, technology innovation, marketing innovation and including process innovation (which is not to be mistaken with innovation process). In the context of online learning, measuring the innovation process emphasizes on knowledge sharing and knowledge creation where knowledge is represented in the form of text such as articles and other multimedia format. The innovation process activities performed in online environment are traceable as they are captured in digital format. Hence, measuring the innovation process based to determine the compliance to CoP values and to monitor the knowledge transformation to the four quadrants of SECI model is feasible. The doubly flank model is introduced here to emphasize the importance of establishing the innovative community as many failures to knowledge creation is due to lacking knowledge sharing within community even in a small enterprise due various factors such as culture, communication skills and job related issues.

The future work for this research is on two aspects: technology and empirical experimentation. The previous work has shown where knowledge flow can be traced computationally for SECI Model [10] but on text-based format. There is also a need to investigate how knowledge that is from other format can also be analyzed to determine its relevancy to the topic of discussion. Another aspect of the technology is on determining the COP values in which some of these values are subtle and hardly to captured merely based on text processing. On the empirical experimentation, SECI model itself is lacking empirical evidences and hence DFCMI ought to be fully experimented in real world situation. It requires some proof that CoP is able to be inculcated in the online environment and hence measuring its presence without using the traditional method of in situ observation. In similar fashion, computational platform based SECI model that has been built ought to be experimented to examine the actual output of knowledge creation.

REFERENCES:

[1] Berkhout, A.J., Hartmann, D., van der Duin, P. and Ortt, R. (2006), Innovating the innovation 

 

process, Int. J. Technology Management, Vol. 34, Nos. 3/4, pp.390–404 - couple

[2] Chatti, M.A., Klamma, R. Jarke, M. & Naeve, A. (2007). The Web 2.0 Driven SECI Model Based Learning Process. Seventh IEEE International Conference on Advanced Learning Technologies (ICALT 2007), Niigata, 18 – 20th July, 2007, pp 780 - 782

[3] du Preez, N.D. and Louw, L, (2008). A Framework for Managing Innovation Process.

Portland International Conference on

Management of Engineering & Technology, PICMET 2008.

[4] Godin, Benoît (2006), The Linear Model of Innovation: The Historical Construction of an Analytical Framework, Science, Technology, and Human Values, 31 (6): 639-667. - linear

[5] Hosseini, S.M. (2011). The application of SECI model as a framework of knowledge creation in virtual learning. Asia Pacific Education Review, May 2011, Volume 12, Issue 2, pp 263 – 270.

[6] Krause, W. and Schutte, C.S.L. (2016). Developing Design Propositions for an Open Innovation Approach for SMEs. South African Journal of Industrial Engineering November 2016 Vol 27(3) Special Edition, pp 37-49. - open

[7] Manley, K. (2001). Frameworks for

Understanding Interactive Innovation

Processes, The International Journal of Entrepreneurship and Innovation, 4(1), 25-36.

[8] Mensch, G., Kaash, K., Kleinknecht, A. and Schnapps, R. (1980), Innovation Trends and Switching between Full- and Under 

employment Equilibrium, 1950-1978,

International Institute of Management, Discussion Paper Series, Berlin, January - market

[9] Rothwell, R. (1994), Towards the Fifth-generation Innovation Process, International Marketing Review, Vol. 11 No. 1, 1994, pp. 7¬31.

[10] S.M.F.D Syed Mustapha. (2016). Towards Building Monolithic Computational Platform for SECI Model. International Journal of Intelligence Science, 2016, 6, 29-41.

[11] Shuman, J., & Twombly, J. (2010). Collaborative networks are the organization: an innovation in organization design and

 

1506

 

Journal of Theoretical and Applied Information Technology

15th April 2017. Vol.95. No 7 

© 2005 – ongoing JATIT & LLS  

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

management. Vikalpa: The Journal for Decision Makers, 35(1), 1-13. - network

[12] Wenger, Etienne and Richard McDermott, and William Snyder (2002) Cultivating communities of practice: a guide to managing knowledge. Cambridge, Mass.: Harvard Business School Press.

[13] Wenhua, L and Fushu, C. (2011). Evaluation research based on SECI model in E-Learning. 2011 International Conference on Electrical and Control Engineering (ICECE), Yichang, pp 6571 – 6573.

[14] Kermally, S. (2002). Effective knowledge man agement: A best practice blueprint. New York: John Wiley & Son, 2002.

[15] Chia-Ying Li & Chang-Tseh Hsieh. (2009). The impact of knowledge stickiness on

knowledge transfer implementation,

internalization, and satisfaction for

multinational corporations. International Journal of Information Management, Vol 29, Issue 6, December 2009, pp 425 – 435.

[16] Murray, A. and Hanlon, P. (2010). An Investigation into the Stickiness of Tacit Knowledge Transfer. 13th. Annual Conference of the Irish Academy of Management, Cork Institute of Technology, 1-3 September, 2010. Download: 26th May 2016, from http://arrow.dit.ie/cgi/viewcontent.cgi?article=  1002&context=buschgracon.

[17] Nonaka, I and Konno, N. (1998). The concept of Ba. Building foundation for knowledge creation. California Management Review, 40(3), pp 40 – 54.

[18] Eerola, A & Jorgensen, B.H. (2002). Technology Foresight in the Nordic Countries. Risø National Laboratory, Roskilde,

September 2002. From:

http://citeseerx.ist.psu.edu/viewdoc/download?  doi=10.1.1.195.2797&rep=rep1&type=pdf Download: 27th May 2016.

[19] Lee, H.Y & Lee, H.W. (2016). Comparing Social Network Analysis of Posts with Counting of Posts as a Measurement of Learners’ Participation in Facebook Discussions. TOJET: The Turkish Online Journal of Educational Technology, January 2016, volume 15 issue 1, pp 11 – 19.

[20] Lis, A. (2014). Knowledge Creation and Conversion in Military Organizations: How the SECI Model is Applied Within Armed Forces. Journal of Entrepreneurship, Management and Innovation: Knowledge Management Special Issue: Connecting Theory and Practice, pp 57 - 78

[21] Natek, S and Zwilling, M. (2016). Knowledge Management Systems Support SECI Model of

Knowledge-Creating Process. Joint

International Conference 2016 Technology, Innovation and Industrial Management, 25 – 27 May, 2016, Timisoara, Romania. From: http://www.toknowpress.net/ISBN/978-961-6914-16-1/papers/ML16-251.pdf Download: 29th May 2016.

[22] Cummings, J. (2003). Knowledge Sharing: A Review of Literature. The World Bank Operations Evaluation Department, [from http://ieg.worldbank.org/Data/reports/knowled ge_eval_literature_review.pdf. Download: 26th May 2016].

[23] Lave, J. and Wenger, E. (1991). Situated Learning: Legitimate Peripheral Participation. Cambridge: Cambridge University Press. ISBN 0-521-42374-0.

[24] Hutchins, E. (1995), Cognition in the Wild, Cambridge MA, MIT press

[25] Roberts, J. (2006). Limits to communities of practice. Journal of Management Studies, 43(3), 623-639.

[26] Saint-Onge, H., & Wallace, D. (2003). Leveraging communities of practice for strategic advantage. New York: Butterworth Heinemann.

[27] Newman, Laurel Vaughan. “Keeping Up and Staying Connected: Contributions From a Community of Practice.” Conference Proceedings Paper. Distance Learning Association. 2014.

[28] Murillo, E. (2008). Searching Usenet for virtual Communities of Practice: using mixed methods to identify the constructs of Wenger's theory. Information Research, 13(4) paper 386.

[29] Kapucu, N. (2012). Classrooms as

Communities of Practice: Designing and Facilitating Learning in a Networked Environment. The Journal of Public Affairs Education, JPAE 18(3), pp 585 – 610.

(Available: 24 Jan 2017:

http://www.naspaa.org/JPAEMessenger/Articl  e/VOL18-3/11_Kapucu.pdf)

[30] Wenger, E. McDermott, R. and Snyder, W. Cultivating communities of practice: a guide to

1507

 

Journal of Theoretical and Applied Information Technology

15th April 2017. Vol.95. No 7 

© 2005 – ongoing JATIT & LLS  

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

managing knowledge. Harvard Business School Press, 2002.

[31] S.M.F.D Syed Mustapha. CoP Sensing Framework on Web-based Environment. J.T. Yao (ed), Chapter in Book, Web-based Support Systems, Springer, 2010, Chapter 15, pp 333 - 357.

[32] Kassem, S., Hammami, S. and Alhousaru, T. (2015). Applying SECI Model to Encourage Knowledge Creation in e-learning Environment. International Journal of Economic Research, 12(4), 2015, 1601 – 1611.

[33] Hosseini, S.M. (2011). The Application of SECI Model as a Framework of Knowledge Creation in Virtual Learning. Asia Pacific Education Review. 12: 263 – 270.

[34] Chatti, M.A., Klamma, R., Jarke, M. and Naeve, A. (2007). The Web 2.0 Driven SECI Model Based Learning Process. International Conference of Advanced Learning Technology, ICALT 2007, pp 770 – 782.

[35] Vajhalla, N.R. and Hassan, M.H. (2013). Barriers to Knowledge Sharing in Medium-Sized Enterprises in Transition Economies. International Conference in Management, Knowledge and Learning, 19 – 21 June, Zadar Crotia, pp 809 – 817.

[36] Bureš, V. (2003). Cultural Barriers in Knowledge Sharing, E+M Ekonomics and Management, Liberec, vol.6, special issue, pp.57-62, ISSN 1212-3609

[37] Ahrend, N., Pittke, F. and Leopold, H. (2014). Barriers and Strategies of process knowledge sharing in public sector organizations. Multikonferenz Wirtschaftsinformatik, pp 1 – 12.

[38] Casimir, G., Lee, K., & Loon, M. (2012). Knowledge sharing: Influences of trust, commitment and cost. Journal of Knowledge Management, 16(5), 740-753.

[39] Farris, G. F., Senner, E. E., & Butterfield, D. A. (1973). Trust, culture, and organizational behavior. Industrial Relations, 12(2), 144-157.

[40] Hansen, M. (1999): The search-transfer problem: The role of weak ties in sharing knowledge across organization subunits. Administrative Science Quarterly, 44(1): p. 82-111.

[41] Sonwalker, N. (2005). Adaptive Learning Technologies: From One Size Fits All to Individualization. EDUCAUSE Center for Applied Research: Research Bulletin, Vol 2005, Issue 7, March 29, 2005, pp 1 – 13. (http://net.educause.edu/ir/library/pdf/erb0507.  pdf Available on 7th Feb 2017)

[42] Atkinson, S.P. (2015). Adaptive Learning and Learning Analytics: a new learning design paradigm. BPP Working Paper. January 2015, pp 1- 9. (http://w3.unisa.edu.au/tiu/TEACHING/Learni  ng_analytics/Adaptive%20Learning%20and%  20Learning%20Analytics%20-%20a%20new%20learning%20design%20para digm.pdf Available on 7th Feb 2017)

[43] Twigg, C.A. (2001). Innovations in Online Learning. The Pew Learning and Technology Program 2001, pp 1 – 36. (http://www.thencat.org/Monographs/Mono4.p  df Available on 7th Feb 2017)

[44] Vieluf S., et. al, (2012). Teaching Practices and Pedagogical Innovation: Evidence from

TALIS, OECD Publishing.

(https://www.oecd.org/edu/school/TalisCeri%2  02012%20(tppi)--Ebook.pdf Available on 7th February 2017)

[45] Xiaoyuan, Su. and Taghi, M. K. (2009). A Survey of Collaborative Filtering Techniques. Advances in Artificial Intelligence, Volume 2009, Article ID 421425, pp 1 – 20.

[46] Brabham, D.C. (2013). Crowdsourcing. Massachusetts Institute of Technology.

[47] S.M.F.D Syed Mustapha. (2012). KFTGA: A tool for tracing knowledge flow and knowledge growth in knowledge sharing environment. Journal of Information-Knowledge-Systems Management, Vol 11, Issue 3, pp 205 – 224.

[48] Zhang, X and Dhong, D. (2008) Ways of Identifying Opinion Leaders in Virtual Communities. International Journal of Business and Management, Vol 3, No 7, pp 21 – 27.

[49] Emilyn J. and Keerthana, R. (2015). Community Detection and Identifying Leaders and Followers in Online Social Networks. International Journal of Scientific and Engineering Research, Vol 6, Issue 4, April 2015, pp 245 – 251

[50] Pal, A., and Konstan, J. A. (2010, October). Expert identification in community question answering: exploring question selection bias. In Proceedings of the 19th ACM International Conference on Information and Knowledge Management, pp. 1505-1508.

[51] El-Korany, A. (2013). Integrated Expert Recommendation Model for Online

1508

 

Journal of Theoretical and Applied Information Technology

15th April 2017. Vol.95. No 7 

© 2005 – ongoing JATIT & LLS  

ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195

Communities. International Journal of Web & Semantic Technology (IJWesT) Vol.4, No.4, October 2013, pp 19 – 29.

[52] S.M.F.D Syed Mustapha. (2017). Theoretical Concept on Case-based Reasoning for Identifying Knowledge Leader within Online Community. Technical Report. Taif University. 12 pages.

[53] Easa, N. F.H. (2012). Knowledge Management and The SECI Model: A Study of Innovation in the Egyptian Banking Sector, PhD Thesis.

University Stirling, Feb 2012.

(http://www.academia.edu/11340484/Knowled 

ge_Management_and_Innovation_in_Banking 

Available on 7th Feb 2017)

1509

 

Social Signal Processing: Understanding Social Interactions

through Nonverbal Behavior Analysis

 

A.Vinciarelli and H.Salamin

Idiap Research Institute

CP592 - 1920 Martigny (Switzerland)

EPFL - 1015 Lausanne (Switzerland)

{vincia,hsalamin}@idiap.ch

Abstract

This paper introduces Social Signal Processing (SSP), the domain aimed at automatic understanding of social interactions through analysis of nonverbal behavior. The core idea of SSP is that nonverbal behavior is machine de¬tectable evidence of social signals, the relational attitudes exchanged between interacting individuals. Social signals include (dis-)agreement, empathy, hostility, and any other attitude towards others that is expressed not only by words but by nonverbal behaviors such as facial expression and body posture as well. Thus, nonverbal behavior analysis is used as a key to automatic understanding of social interac¬tions. This paper presents not only a survey of the related literature and the main concepts underlying SSP, but also an illustrative example of how such concepts are applied to the analysis of conflicts in competitive discussions.

1. Introduction

Imagine watching the television in a country of which you do not know the language. While you cannot under¬stand what is being said, you can still catch a good deal of information about social interactions taking place on the screen. You can easily spot the most important guest in a talk-show, understand whether the interaction is tense or relaxed, guess the kind of relationships people have (e.g.,

The work of A. Vinciarelli and H.Salamin has been supported in part by the European Community’s Seventh Framework Programme (FP7/2007-2013) under grant agreement no. 231287 (SSPNet) and in part by the Swiss National Science Foundation through the National Center of Competence in Re- search on Interactive Multimodal Information Man-agement (IM2).

†The work of M. Pantic has been supported in part by the Euro¬pean Community’s Seventh Framework Programme (FP7/2007-2013) un¬der grant agreement no. 231287 (SSPNet), and in part by the European Research Council under the ERC Starting Grant agreement no. ERC-2007-StG-203143 (MAHNOB).

 

M.Pantic†

Imperial College London

108 Queens Gate London

EEMCS - University of Twente

m.pantic@imperial.ac.uk

whether they are couples or members of the same soccer team), etc.

How can we be so effective in interpreting social inter¬actions without the need of understanding what is being said? Psychologists have been studying this phenomenon for decades and they have shown that extracting social in¬formation from nonverbal communication is hard wired in the human brain [33][54]. Any facial expression, vocal out¬burst, gesture or posture triggers often unconscious analy¬sis of socially relevant information [4]. Furthermore, this mechanism seems to be so deeply rooted in our brain, that we cannot escape it, even when we deal with synthetic faces [10] and voices [42] generated by computers.

If nonverbal communication plays such an important role in our life, shouldn’t we enable computers to sense and interpret social meaning of human user’s nonverbal cues? This is exactly the problem addressed by Social Signal Pro¬cessing (SSP), the new, emerging, domain aimed at un¬derstanding social interactions through machine analysis of nonverbal behavior [51][68][69]. The core idea of SSP is that nonverbal behavioral cues can be detected with micro¬phones, cameras, and any other suitable sensors. The cues can then be used as a machine detectable evidence for auto¬matic analysis and understanding of social behavior shown by the human user.

SSP enables Human-Centred computing paradigm [46], effectively dealing with psychological and behavioral re¬sponses natural to humans, in contrast to computing-centred paradigm that requires people to operate computers fol¬lowing technology-driven criteria. This will have a major impact on various domains of computing technology such as Human-Computer Interaction which will become more adept to social interactions with users [46], multimedia con¬tent analysis which will be analyzed according to the way humans perceive the reality around them [22], computer mediated communication (e.g., see [24]) because transmis¬sion will include social cues necessary for establishing a

 

978-1-4244-3993-5/09/$25.00 ©2009 IEEE 42

Authorized licensed use limited to: UNIVERSITEIT TWENTE. Downloaded on January 15, 2010 at 05:52 from IEEE Xplore. Restrictions apply.

 

Figure 1. Social signals. A constellation of nonverbal behavioral cues (posture, interpersonal distance, gestures, etc.) is perceived as a social signal (hostility, aggressiveness, disagreement, etc.).

natural contact with others, and other domains where com¬puters must seamlessly integrate into the life of people.

The paper starts by introducing the most important as¬pects of nonverbal communication (Section 2). It illustrates the main technological components necessary to analyze social behavior (Section 3) and provides an example show¬ing how SSP principles and ideas are applied to the anal¬ysis of conflicts in competitive discussions (Section 4). It also provides a brief survey of the main SSP applications presented so far in the literature. Section 6 concludes the paper.

2. Nonverbal Behavior and Social Signals

Nonverbal communication includes all the messages other than words that people exchange in interactive con¬texts. In some cases, messages are exchanged consciously, and nonverbal behaviors have a precise meaning attached to them (e.g., the thumbs up gesture). More frequently, non¬verbal behavior gives away messages, leaking information about the state of people, e.g. about their emotions, self-confidence, status, etc. [25].

SSP focuses on human nonverbal communication and, in particular, on social signals [3], the relational attitudes displayed by people during social interactions. Consider Figure 1. It is not difficult to guess that the two individu¬als are a couple and that they are fighting, even if the only information at disposition are their silhouettes. The reason is that the picture shows a sufficient number of nonverbal behavioral cues to correctly understand the kind of interac¬tion taking place. Mouths wide open suggest that the two persons are shouting, the tension of gestures shows that the atmosphere is not relaxed, the distance is too close for per¬sons not sharing an intimate relationship, etc.

For the sake of simplicity, psychologists have grouped all possible nonverbal behavioral cues occurring in social inter¬actions into five major classes called codes [30]. The first is physical appearance, including not only somatic char¬acteristics, but also clothes and ornaments that people use 

 

to modify their appearance. While human sciences have extensively investigated the role of appearance in social interactions (e.g., see [18] for the effect of attractiveness, and [12] for the influence of body shape on social percep¬tions), only few works, to the best of our knowledge, have been dedicated to the automatic analysis of the way people look. These are mostly dedicated to the attractiveness of faces (e.g., [27]) and to the recognition of clothes for track¬ing and surveillance purposes (e.g., [ 15]).

The second code relates to gestures and postures, ex¬tensively investigated in human sciences because they are considered the most reliable cue revealing actual attitude of people towards others (see [54] and references therein). Automatic analysis of gestures is a hot topic in technol¬ogy as well, but the goal is mainly to replace keyboards and mouces with hand movements as computer interfaces (see [72] for recent technologies). Gestures and postures have been also analyzed for their affective content (see [ 28] for a survey). However, there are only a few works pre-sented so far addressing the problem of interpreting ges¬tures and postures in terms of social signals (see [68] for a survey).

Face and eye behavior is a crucial code, as face and eyes are our direct and naturally preeminent means of commu¬nicating and understanding somebodys affective state and intentions on the basis of the shown facial expression [32]. Not surprisingly, facial expressions and gaze behavior have been extensively studied in both human sciences and tech¬nology. The first study on facial expressions dates back to Darwin [16], and a comprehensive framework for the de-scription of facial expressions (and messages they convey) has been elaborated in the last decades [21]. Facial expres¬sion analysis is a well established domain (see [76] for the most recent and extensive survey), and gaze has been the subject of significant attention in the last years [64].

Vocal behavior is the code that accounts for how some¬thing is said and includes the following aspects of spoken communication [33][54]: voice quality (prosodic features like pitch, energy and rhythm), linguistic vocalizations (ex¬pressions like “ehm”, “ah”, etc.) and non-linguistic vocal¬izations (laughter, crying, sobbing, etc.), silence (use of pauses), and turn-taking patterns (mechanisms regulating floor exchange) [53][74]. Each one of them relates to social signals that contribute to different aspects of the social per¬ception of a message. Both human sciences and technology have extensively investigated vocal behavior. The former have shown, e.g., that vocal behavior plays a role in ex¬pression of emotions [57], is a personality marker [56], and is used to display status and dominance [59]. The speech analysis community has worked on the detection, e.g., of disfluencies [58], non-linguistic vocalizations (e.g., partic¬ular laughter [52][62]), or rhythm [40], but with the goal of improving the speech recognition performance rather than

 

43

Authorized licensed use limited to: UNIVERSITEIT TWENTE. Downloaded on January 15, 2010 at 05:52 from IEEE Xplore. Restrictions apply.

 

Preprocessing

Figure 2. Machine analysis of social signals and behaviors: a general scheme. The process includes two main stages. Preprocessing takes as input the recordings of social interaction and gives as output multimodal behavioral streams associated with each person. Social interaction analysis maps the multimodal behavioral streams into social signals and social behaviors.

 

analysing social behavior.

The last code relates to space and environment, i.e. the way people share and organize the space they have at disposition. Human sciences have investigated this code, showing in particular that people tend to organize the space around them in concentric zones accounting for different relationships they have with others [29]. For example, Fig¬ure 1 shows an example of individuals sharing the intimate zone, the concentric area closest to each individual. Tech¬nology has started only recently to study the use of space, but only for tracking and surveillance purposes.

3. State-of-the-art

Figure 2 shows the main technological components (and their interrelationships) of a general SSP system. The scheme does not correspond to any approach in particular, but most SSP works presented in the literature follow, at least partially, the processing chain in the picture (see Sec¬tion 5).

The first, and crucial, step is the data capture. The most commonly used capture devices are microphones and cam¬eras (with arrangements that go from a simple laptop we-bcam to a fully equipped smart meeting room [36][70]), but the literature reports the use of wearable devices [20] and pressure captors [41] (for recognizing posture of sitting people) as well.

In most cases, the raw data involve recordings of dif-ferent persons (e.g., the recording of a conversation where different voices can be heard at different moments in time). Thus, a person detection step is necessary to know which part of the data corresponds to which person (e.g., who talks when in the recording of a conversation). This is typ¬ 

 

ically performed with speaker diarization [61], face detec¬tion [73], or any other kind of technique that allows one to identify intervals of time or scene regions corresponding to specific individuals.

Person detection is the step preliminary to behavioral cues extraction, i.e. the detection of nonverbal signals dis¬played by each individual. Some approaches for this stage have been mentioned in Section 2. Extensive overviews are available in [68][69].

The two main challenges in social behavior understand¬ing are the modeling of temporal dynamics and fusing the data extracted from different modalities at different time scales.

Temporal dynamics of social behavioral cues (i.e., their timing, co-occurrence, speed, etc.) are crucial for the inter¬pretation of the observed social behavior [3][21]. However, relatively few approaches explicitly take into account the temporal evolution of behavioral cues to understand social behavior. Some of them aim at the analysis of facial ex¬pressions involving sequences of Action Units (i.e., atomic facial gestures) [60], as well as coordinated movements of head and shoulders [63]. Others model the evolution of col-lective actions in meetings using Dynamic Bayesian Net¬works [17] or Hidden Markov Models [37].

To address the second challenge outlined above (tempo¬ral, multimodal data fusion), a number of model-level fu¬sion methods have been proposed that aim at making use of the correlation between audio and visual data streams, and relax the requirement of synchronization of these streams (see [76] for a survey). However, how to model multimodal fusion on multiple time scales and how to model tempo¬ral correlations within and between different modalities is

 

44

Authorized licensed use limited to: UNIVERSITEIT TWENTE. Downloaded on January 15, 2010 at 05:52 from IEEE Xplore. Restrictions apply.

 

largely unexplored.

Context Understanding is desirable because no correct interpretation of human behavioral cues in social interac¬tions is possible without taking into account the context, namely where the interactions take place, what is the ac¬tivity of the individuals involved in the interactions, when the interactions take place, and who is involved in the inter¬action. Note, however, that while W4 (where, what, when, who) is dealing only with the apparent perceptual aspect of the context in which the observed human behavior is shown, human behavior understanding is about W5+ (where, what, when, who, why, how), where the why and how are directly related to recognizing communicative intention including social behaviors, affective and cognitive states of the ob¬served person [47]. Hence, SSP is about W5+.

However, since the problem of context-sensing is ex¬tremely difficult to solve, especially for a general case (i.e., general-purpose W4 technology does not exist yet [ 47]), answering the why and how questions in a W4-context-sensitive manner when analysing human behavior is virtu¬ally unexplored area of research.

4. An Example: the Analysis of Conflicts

This section aims at providing a concrete example of how principles and ideas outlined in previous sections are applied to a concrete case, i.e. the analysis of conflicts in competitive discussions. Conflicts have been extensively in¬vestigated in human sciences. The reason is that they influ¬ence significantly the outcome of groups expected to reach predefined targets (e.g., deadlines) or to satisfy members needs (e.g., in families) [35].

This section focuses on political debates because these are typically built around the conflict between two fronts (including one or more persons each) that defend opposite views or compete for a reward (e.g., the attribution of an important political position) that cannot be shared by two parties. The corpus used for the experiments includes 45 debates (roughly 30 hours of material) revolving around yes/no questions like “are you favorable to new laws on en¬vironment protection?”. Each debate involves one moder¬ator, two guests supporting the yes answer, and two guests supporting the no answer. The guests state their answer ex¬plicitly at the beginning of the debate and this allows one to label them unambiguously in terms of their position.

The goal of the experiments is 1) to identify the moder¬ator, and 2) to reconstruct correctly the two groups (yes and no) resulting from the structure outlined above. The next sections show how the different steps depicted in Figure 2 are addressed.

 

4.1. Nonverbal Behavior in Conflicts

Human sciences have studied conversations in depth as these represent one of the most common forms of social in¬teraction [53]. Following [74], conversations can be thought of as markets where people compete for the floor (the right of speaking):

[...] the most widely used analytic approach is based on an analogy with the workings of the market economy. In this market there is a scarce commodity called the floor which can be defined as the right to speak. Having control of this scarce commodity is called a turn. In any situation where control is not fixed in advance, anyone can attempt to get control. This is called turn-taking.

This suggests that turn-taking is a key to understand con¬versational dynamics.

In the specific case of conflicts, social psychologists have observed that people tend to react to someone they disagree with rather than to someone they agree with [53][74]. Thus, the social signal conveyed as a direct reaction is likely to be disagreement. Hence, the corresponding nonverbal be¬havioral cue is adjacency in speakers turns. This social psy¬chology finding determines the design of the conflict analy¬sis approach described in the rest of this section.

4.2. Data Capture and Person Detection

The previous section suggests that turn-taking is the key to understand conversational dynamics in conflicts. The data at disposition are television political debates and the turn-taking can be extracted from the audio channel using a speaker diarization approach (see [61] for an extensive sur¬vey on diarization). The diarization approach used in this work is that proposed in [1]. The audio channel of the po¬litical debates is converted into a sequence S:

S = {(s1, t1, Δt1), ... , (sN, tN, ΔtN)}, (1)

where each triple accounts for a turn and includes a speaker label si  A = {a1, ... , aG} identifying the person speak¬ing during the turn, the starting time ti of the turn, and the duration Δti of the turn (see Figure 3). Thus, the se¬quence S contains the entire information about the turn-taking, namely who talks when and how much. The pu¬rity (see [67] for a definition of the purity) of the resulting speaker segmentation is 0.92, meaning that the groundtruth speaker segmentation is mostly preserved.

The diarization can be considered a form of person de¬tection because it identifies the parts of the data that corre¬spond to each person. In the case of this work, this allows for the identification of speaker adjacencies representing the target cue based on which agreement and disagreement be¬tween debate participants will be detected.

 

45

Authorized licensed use limited to: UNIVERSITEIT TWENTE. Downloaded on January 15, 2010 at 05:52 from IEEE Xplore. Restrictions apply.

 

Figure 3. Turn-Taking pattern. The figure shows an example of turn-taking where three persons are assigned to different states.

 

4.3. Social Signal Understanding

The suggestion that people tend to react to someone they disagree with rather than to someone they agree with can be expressed, in mathematical terms, by saying that speaker si is statistically dependent on speaker si1 (see Figure 3). Statistical dependence between sequence el¬ements that follow one another can be modeled using a Markov Chain where the set Q of the states contains three elements, namely T1 (the first group), T2 (the second group) and M (the moderator).

If ϕ : A  Q is a mapping that associates a speaker si  A with a state qj  Q, then the conflict analysis problem can be thought of as finding the mapping ϕ * satisfying the following expression:

ϕ* = arg maxp(ϕ(s1))

ϕQA

where N is the number of turns in the turn-taking, p(ϕ(s 1)) is the probability of starting with state q1 = ϕ(s1), and p(ϕ(sn)|ϕ(sn1)) is the probability of a transition between state qn = ϕ(sn) and state qn1 = ϕ(sn1).

The expression on the left side of Equation (2) has the same value if all the speakers assigned state T1 are switched to state T2 and viceversa. In other words, the model is sym¬metric with respect to an exchange between T1 and T2. The reason is that T1 and T2 are simply meant to distinguish between members of different groups.

The Markov Model is trained using a leave-one-out ap¬proach: all debates at disposition but one are used as train¬ing set, while the left out one is used as the test set. The experiment is reiterated and each time a different debate is used as the test set. The results show that 64.5% of the de¬bates are correctly reconstructed, i.e., the moderator is cor¬rectly identified and the two supporters of the same answer are assigned the same state. This figure goes up to 75% when using the groundtruth speaker segmentation (and not the speaker segmentation automatically extracted from the audio data). The average performance of an algorithm as-signing the states randomly is 6.5% and this means that the 

 

above model, even if rather simple, still performs ten times better than chance.

5. Main SSP Applications

The first extensive surveys of SSP applications have been proposed in [68][69], after the expression Social Signal Processing has been introduced for the first time in [51] to denote several pioneering works published by Alex Pent¬land and his group at MIT.

The earliest SSP works focused on vocal behavior with the goal of predicting (with an accuracy higher than 70%) the outcome of dyadic interactions such as salary ne¬gotiations, hiring interviews, and speed dating conversa¬tions [14]. One of the most important contributions of these works is the definition of a coherent framework for the anal¬ysis of vocal behavior [48][49], where a set of cues accounts for activity (the total amount of energy in the speech sig¬nals), influence (the statistical influence of one person on the speaking patterns of the others), consistency (stability of the speaking patterns of each person), and mimicry (the imitation between people involved in the interactions). Re¬cent approaches for the analysis of dyadic interactions in¬clude the visual analysis of movements for the detection of interactional synchrony [38][39].

Other approaches, developed in the same period as the above works, have aimed at the analysis of small group interactions [35], with particular empha¬sis on meetings and broadcast data (talk-shows, news, etc.). Most of the works have focused on recogni¬tion of collective actions [17][37], dominance detec¬tions [31][55], and role recognition [7][19][23][34][75]. The approaches proposed in these works are often mul-timodal [17][19][31][37][55][75], and the behavioral cues most commonly extracted correspond to speaking energy and amount of movement. In many cases, the approaches are based only on audio, with features that account for turn-taking patterns (when and how much each person talks) [7][34], or for combinations of social networks and lexical features [23].

Social network analysis has been applied as well

 

46

Authorized licensed use limited to: UNIVERSITEIT TWENTE. Downloaded on January 15, 2010 at 05:52 from IEEE Xplore. Restrictions apply.

 

in [65][66][71] to recognize the roles played by people in broadcast data (movies, radio and television programs, etc.), and in an application domain known as reality min¬ing, where large groups of individuals equipped with smart badges or special cellular phones are recorded in terms of proximity and vocal interactions and then represented in a social network [20][50].

The reaction of users to social signals exhibited by com¬puters has been investigated in several works showing that people tend to behave with machines as they behave with other humans. The effectiveness of computers as social ac¬tors, i.e., entities involved in the same kind of interactions as the humans, has been explored in [42][43][44], where computers have been shown to be attributed a personality and to elicit the same reactions as those elicited by persons. Similar effects have been shown in [13][45], where chil¬dren interacting with computers have modified their voice to match the speaking characteristics of the animated personas of the computer interface, showing adaptation patterns typ¬ical of human-human interactions [9]. Further evidence of the same phenomenon is available in [5][6], where the inter¬action between humans and computers is shown to include the Chameleon effect [11], i.e. the mutual imitation of indi¬viduals due to reciprocal appreciation or to the influence of one individual on the other.

6. Conclusion

The long term goal of SSP is to give computers social intelligence [2]. This is one of the multiple facets of human intelligence, maybe the most important because it helps to deal successfully with the complex web of interactions we are constantly immersed within, whether this means to be recognized as a leader on the workplace, to be a good par¬ent, or to be a person friends like to spend time with. The first successes obtained by SSP are impressive and have at¬tracted the praise of both technology [26] and business [8] communities. However, there is still a long way to go before artificial social intelligence and socially-aware computing become a reality.

Several major issues need to be addressed in this di-rection. The first is to establish an effective collabora-tion between human sciences and technology. SSP is in-herently multidisciplinary, no effective analysis of social behavior is possible without taking into account the basic laws of human-human interaction that psychologists have been studying for decades. Thus, technology should take into account findings of human sciences, and these should formulate their knowledge in terms suitable for automatic approaches. The second issue is the development of ap¬proaches dealing with multiple behavioral cues (typically extracted from several modalities), often evolving at dif¬ferent time scales while still forming a coherent social sig¬nal. This is necessary because single cues are intrinsically 

 

ambiguous, sometimes they actually convey social mean¬ing, but sometimes they simply respond to contingent fac¬tors (e.g., postures can communicate a relational attitude, but also be determined by the search for comfort). Finally, an important issue is the use of real-world data in the ex¬periments. This will lead to more realistic assessments of technology effectiveness and will link research to potential application scenarios.

The strategic importance of the domain is confirmed by several large projects funded at both national and international level around the world. In particular, the European Network of Excellence SSPNet (2009-2014) aims not only at addressing the issues outlined above, but also at fostering research in SSP through the diffusion of knowledge, data and automatic tools via its web portal (www.sspnet.eu). In this sense, the portal is expected to be not only a site delivering information, but also an instrument allowing any interested researcher to enter the domain with an initial effort as limited as possible.

References

[1] J. Ajmera, I. McCowan, and H. Bourlard. Speech/music segmentation using entropy and dynamism features in a HMM classification framework. Speech Communication, 40(3):351–363, 2003. 4

[2] K. Albrecht. Social Intelligence: The new science of success. John Wiley & Sons Ltd, 2005. 6

[3] N. Ambady, F. Bernieri, and J. Richeson. Towards a histol¬ogy of social behavior: judgmental accuracy from thin slices of behavior. In M. Zanna, editor, Advances in Experimental Social Psychology, pages 201–272. 2000. 2, 3

[4] M. Argyle. The Psychology of Interpersonal Behaviour. Pen¬guin, 1967. 1

[5] J. Bailenson and N. Yee. Virtual interpersonal touch and dig¬ital chameleons. Journal of Nonverbal Behavior, 31(4):225– 242, 2007. 6

[6] J. Bailenson, N. Yee, K. Patel, and A. Beall. Detecting digital chameleons. Computers in Human Behavior, 24(1):66–87, 2008. 6

[7] S. Banerjee and A. Rudnicky. Using simple speech based features to detect the state of a meeting and the roles of the meeting participants. In Proceedings of International Con¬ference on Spoken Language Processing, pages 2189–2192, 2004. 5

[8] M. Buchanan. The science of subtle signals. Strat-egy+Business, 48:68–77, 2007. 6

[9] J. Burgoon, L. Stern, and L. Dillman. Interpersonal Adap-tation: Dyadic Interaction Patterns. Cambridge University Press, 1995. 6

[10] J. Cassell. Embodied conversational interface agents. Com¬munications of the ACM, 43(4):70–78, 2000. 1

[11] T. Chartrand and J. Bargh. The chameleon effect: the perception-behavior link and social interaction. Journal of Personality and Social Psychology, 76(6):893–910, 1999. 6

 

47

Authorized licensed use limited to: UNIVERSITEIT TWENTE. Downloaded on January 15, 2010 at 05:52 from IEEE Xplore. Restrictions apply.

 

[12] J. Cortes and F. Gatti. Physique and self-description of tem¬perament. Journal of Consulting Psychology, 29(5):432– 439, 1965. 2

[13] R. Coulston, S. Oviatt, and C. Darves. Amplitude conver-gence in children’s conversational speech with animated per¬sonas. In International Conference on Spoken Language Processing, pages 2689–2692, 2002. 6

[14] J. Curhan and A. Pentland. Thin slices of negotiation: pre¬dicting outcomes from conversational dynamics within the first 5 minutes. Journal of Applied Psychology, 92(3):802– 811, 2007. 5

[15] T. Darrell, G. Gordon, M. Harville, and J. Woodfill. Inte-grated person tracking using stereo, color, and pattern detec¬tion. International Journal of Computer Vision, 37(2):175– 185, 2000. 2

[16] C. Darwin. The Expression of the Emotions in Man and An¬imals. J. Murray, 1872. 2

[17] A. Dielmann and S. Renals. Automatic meeting segmenta-tion using dynamic bayesian networks. IEEE Transactions on Multimedia, 9(1):25, 2007. 3, 5

[18] K. Dion, E. Berscheid, and E. Walster. What is beauti¬ful is good. Journal of Personality and Social Psychology, 24(3):285–290, 1972. 2

[19] W. Dong, B. Lepri, A. Cappelletti, A. Pentland, F. Pianesi, and M. Zancanaro. Using the influence model to recognize functional roles in meetings. In Proceedings of the Interna¬tional Conference on Multimodal Interfaces, pages 271–278, 2007. 5

[20] N. Eagle and A. Pentland. Reality mining: sensing complex social signals. Journal of Personal and Ubiquitous Comput¬ing, 10(4):255–268, 2006. 3, 6

[21] P. Ekman and E. Rosenberg. What the Face Reveals: Ba-sic and Applied Studies of Spontaneous Expression Using the Facial Action Coding System (FACS). Oxford University Press, 2005. 2, 3

[22] A. Elgammal. Human-Centered Multimedia: representations and challenges. In Proc. ACM Intl. Workshop on Human-Centered Multimedia, pages 11–18, 2006. 1

[23] N. Garg, S. Favre, H. Salamin, D. Hakkani-T¨ur, and A. Vin-ciarelli. Role recognition for meeting participants: an ap¬proach based on lexical information and social network anal¬ysis. In Proceedings of the ACM International Conference on Multimedia, pages 693–696, 2008. 5

[24] J. Gemmell, K. Toyama, C. Zitnick, T. Kang, and S. Seitz. Gaze awareness for video-conferencing: A software ap-proach. IEEE Multimedia, 7(4):26–35, 2000. 1

[25] E. Goffman. The presentation of self in everyday life. Anchor Books, 1959. 2

[26] K. Greene. 10 emerging technologies 2008. MIT Technology Review, february 2008. 6

[27] H. Gunes and M. Piccardi. Assessing facial beauty through proportion analysis by image processing and supervised learning. International Journal of Human-Computer Stud¬ies, 64(12):1184–1199, 2006. 2

[28] H. Gunes, M. Piccardi, and M. Pantic. From the lab to the real world: Affect recognition using multiple cues and modalities. In J. Or, editor, Affective Computing: Focus on 

 

Emotion Expression, Synthesis, and Recognition, pages 185– 218. 2008. 2

[29] E. Hall. The silent language. Doubleday, 1959. 3

[30] M. Hecht, J. De Vito, and L. Guerrero. Perspectives on non¬verbal communication. codes, functions and contexts. In

L. Guerrero, J. De Vito, and M. Hecht, editors, The nonver¬bal communication reader, pages 201–272. 2000. 2

[31] D. Jayagopi, H. Hung, C. Yeo, and D. Gatica-Perez. Model¬ing dominance in group conversations using non-verbal ac¬tivity cues. IEEE Transactions on Audio, Speech and Lan¬guage: Special Issue on Multimedia, to appear, 2009. 5

[32] D. Keltner and P. Ekman. Facial expression of emotion. In

M. Lewis and J. Haviland-Jones, editors, Handbook of Emo¬tions, pages 236–249. 2000. 2

[33] M. Knapp and J. Hall. Nonverbal Communication in Human Interaction. Harcourt Brace College Publishers, 1972. 1, 2

[34] K. Laskowski, M. Ostendorf, and T. Schultz. Modeling vocal interaction for text-independent participant characterization in multi-party conversation. In Proceedings of the SIGdial Workshop on Discourse and Dialogue, pages 148–155, 2008. 5

[35] J. Levine and R. Moreland. Small groups. In D. Gilbert and G. Lindzey, editors, The handbook of social psychology, volume 2, pages 415–469. Oxford University Press, 1998. 4, 5

[36] I. McCowan, S. Bengio, D. Gatica-Perez, G. Lathoud, F. Monay, D. Moore, P. Wellner, and H. Bourlard. Model¬ing human interaction in meetings. In Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing, pages 748–751, 2003. 3

[37] I. McCowan, D. Gatica-Perez, S. Bengio, G. Lathoud, M. Barnard, and D. Zhang. Automatic analysis of multi-modal group actions in meetings. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(3):305–317, 2005. 3, 5

[38] L. Morency, I. de Kok, and J. Gratch. Context-based recogni¬tion during human interactions: automatic feature selection and encoding dictionary. In Proceedings of the 10th interna¬tional conference on Multimodal interfaces, pages 181–188, 2008. 5

[39] L. Morency, I. de Kok, and J. Gratch. Predicting listener backchannels: A probabilistic multimodal approach. In Lec¬ture Notes in Computer Science, volume 5208, pages 176– 190. Springer, 2008. 5

[40] N. Morgan, E. Fosler, and N. Mirghafori. Speech recognition using on-line estimation of speaking rate. In Proceedings of Eurospeech, pages 2079–2082, 1997. 2

[41] S. Mota and R. Picard. Automated posture analysis for de-tecting learners interest level. In Proceedings of Conference on Computer Vision and Pattern Recognition, pages 49–56, 2003. 3

[42] C. Nass and S. Brave. Wired for speech: How voice activates and advances the Human-Computer relationship. The MIT Press, 2005. 1, 6

[43] C. Nass and K. Lee. Does computer-synthesized speech manifest personality? Experimental tests of recognition, similarity-attraction, and consistency-attraction. Journal of Experimental Psychology: Applied, 7(3):171–181, 2001. 6

 

48

Authorized licensed use limited to: UNIVERSITEIT TWENTE. Downloaded on January 15, 2010 at 05:52 from IEEE Xplore. Restrictions apply.

 

[44] C. Nass and J. Steuer. Computers and social actors. Human Communication Research, 19(4):504–527, 1993. 6

[45] S. Oviatt, C. Darves, and R. Coulston. Toward adaptive con¬versational interfaces: Modeling speech convergence with animated personas. ACM Transactions on Computer-Human Interaction, 11(3):300–328, 2004. 6

[46] M. Pantic, A. Pentland, A. Nijholt, and T. Huang. Human computing and machine understanding of human behavior: A survey. In Lecture Notes in Artificial Intelligence, volume 4451, pages 47–71. Springer Verlag, 2007. 1

[47] M. Pantic, A. Pentland, A. Nijholt, and T. Huang. Human-centred intelligent human-computer interaction (HCI2): How far are we from attaining it? International Jour¬nal of Autonomous and Adaptive Communications Systems, 1(2):168–187, 2008. 4

[48] A. Pentland. Social dynamics: Signals and behavior. In In¬ternational Conference on Developmental Learning, 2004. 5

[49] A. Pentland. Socially aware computation and communica¬tion. IEEE Computer, 38(3):33–40, 2005. 5

[50] A. Pentland. Automatic mapping and modeling of human networks. Physica A, 378:59–67, 2007. 6

[51] A. Pentland. Social Signal Processing. IEEE Signal Process¬ing Magazine, 24(4):108–111, 2007. 1, 5

[52] S. Petridis and M. Pantic. Audiovisual laughter detection based on temporal features. In Proceedings of the 10th inter¬national conference on Multimodal interfaces, pages 37–44. ACM New York, NY, USA, 2008. 2

[53] G. Psathas. Conversation Analysis - The study of talk-in-interaction. Sage Publications, 1995. 2, 4

[54] V. Richmond and J. McCroskey. Nonverbal Behaviors in interpersonal relations. Allyn and Bacon, 1995. 1, 2

[55] R. Rienks, D. Zhang, and D. Gatica-Perez. Detection and application of influence rankings in small group meetings. In Proceedings of the International Conference on Multimodal Interfaces, pages 257–264, 2006. 5

[56] K. Scherer. Personality markers in speech. Cambridge Uni¬versity Press, 1979. 2

[57] K. Scherer. Vocal communication of emotion: A review of research paradigms. Speech Communication, 40(1-2):227– 256, 2003. 2

[58] E. Shriberg. Phonetic consequences of speech disfluency. Proceedings of the International Congress of Phonetic Sci¬ences, 1:619–622, 1999. 2

[59] L. Smith-Lovin and C. Brody. Interruptions in group discus¬sions: the effects of gender and group composition. Ameri¬can Sociological Review, 54(3):424–435, 1989. 2

[60] Y. Tong, W. Liao, and Q. Ji. Facial action unit recognition by exploiting their dynamic and semantic relationships. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(10):1683–1699, 2007. 3

[61] S. Tranter and D. Reynolds. An overview of automatic speaker diarization systems. IEEE Transactions on Audio, Speech, and Language Processing, 14(5):1557–1565, 2006. 3,4

[62] K. Truong and D. Leeuwen. Automatic detection of laugh¬ter. In Proceedings of the European Conference on Speech Communication and Technology, pages 485–488, 2005. 2

 

[63] M. Valstar, H. Gunes, and M. Pantic. How to distinguish posed from spontaneous smiles using geometric features. In Proceedings of the International Conference on Multimodal Interfaces, pages 38–45, 2007. 3

[64] R. Vertegaal, R. Slagter, G. van der Veer, and A. Nijholt. Eye gaze patterns in conversations: there is more to con-versational agents than meets the eyes. In Proceedings of the SIGCHI conference on Human Factors in computing sys¬tems, pages 301–308, 2001. 2

[65] A. Vinciarelli. Sociometry based multiparty audio recording segmentation. In Proceedings of IEEE International Con¬ference on Multimedia and Expo, pages 1801–1804, 2006. 6

[66] A. Vinciarelli. Speakers role recognition in multiparty au-dio recordings using social network analysis and duration distribution modeling. IEEE Transactions on Multimedia, 9(9):1215–1226, 2007. 6

[67] A. Vinciarelli and S. Favre. Broadcast news story segmenta¬tion using social network analysis and hidden markov mod¬els. In Proceedings of the ACM International Conference on Multimedia, pages 261–264, 2007. 4

[68] A. Vinciarelli, M. Pantic, and H. Bourlard. Social Signal Processing: survey of an emerging domain. Image and Vi-sion Computing, to appear, 2009. 1, 2, 3, 5

[69] A. Vinciarelli, M. Pantic, H. Bourlard, and A. Pentland. So¬cial Signal Processing: State-of-the-art and future perspec¬tives of an emerging domain. In Proceedings of the ACM International Conference on Multimedia, pages 1061–1070, 2008. 1, 3, 5

[70] A. Waibel, T. Schultz, M. Bett, M. Denecke, R. Malkin, I. Rogina, and R. Stiefelhagen. SMaRT: the Smart Meet¬ing Room task at ISL. In Proceedings of IEEE International Conference on Acoustics, Speech, and Signal Processing, pages 752–755, 2003. 3

[71] C. Weng, W. Chu, and J. Wu. Rolenet: Movie analysis from the perspective of social networks. IEEE Transactions on Multimedia, 11(2):256–271, 2009. 6

[72] Y. Wu and T. Huang. Vision-based gesture recognition: A re¬view. In A. Braffort, R. Gherbi, S. Gibet, J. Richardson, and D. Teil, editors, Gesture based communication in Human-Computer Interaction, pages 103–114. 1999. 2

[73] M. Yang, D. Kriegman, and N. Ahuja. Detecting faces in images: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(1):34–58, 2002. 3

[74] G. Yule. Pragmatics. Oxford University Press, 1996. 2, 4

[75] M. Zancanaro, B. Lepri, and F. Pianesi. Automatic detec-tion of group functional roles in face to face interactions. In Proceedings of the International Conference on Multimodal Interfaces, pages 28–34, 2006. 5

[76] Z. Zeng, M. Pantic, G. Roisman, and T. Huang. A survey of affect recognition methods: audio, visual and spontaneous expressions. IEEE Transactions on Pattern Analysis and Ma¬chine Intelligence, 31(1):39–58, 2009. 2, 3

 

49

Authorized licensed use limited to: UNIVERSITEIT TWENTE. Downloaded on January 15, 2010 at 05:52 from IEEE Xplore. Restrictions apply.

 

58th Annual IEEE Symposium on Foundations of Computer Science

 

Hardness Results for Structured Linear Systems

 

Rasmus Kyng

Department of Computer Science

Yale University

New Haven, CT, USA

Email: rasmus.kyng@yale.edu

 

Peng Zhang

School of Computer Science

Georgia Institute of Technology

Atlanta, GA, USA

Email: pzhang60@gatech.edu

 

 

Abstract—We show that if the nearly-linear time solvers for Laplacian matrices and their generalizations can be extended to solve just slightly larger families of linear systems, then they can be used to quickly solve all systems of linear equations over the reals. This result can be viewed either positively or negatively: either we will develop nearly-linear time algorithms for solving all systems of linear equations over the reals, or progress on the families we can solve in nearly-linear time will soon halt.

Keywords-Numerical Linear Algebra; Linear System Solvers; Laplacian Solvers; Multi-commodity Flow Problems; Truss Stiffness Matrices; Total Variation Matrices; Complexity The¬ory; Fine-grained Complexity;

I. INTRODUCTION

We establish a dichotomy result for the families of linear equations that can be solved in nearly-linear time. If nearly 

linear time solvers exist for a slight generalization of the families for which they are currently known, then nearly-linear time solvers exist for all linear systems over the reals1.

This type of reduction is related to the successful research program of fine-grained complexity, such as the result [1]

which showed that the existence of a “truly subcubic” time algorithm for All-Pairs Shortest Paths Problem is equivalent to the existence of “truly subcubic” time algorithm for a

wide range of other problems. For any constant a  1, our result establishes for 2-commodity matrices, and several

other classes of graph structured linear systems, that we can

solve a linear system in a matrix of this type with s nonzeros in time O(s) if and only if we can solve linear systems in

all matrices with polynomially bounded integer entries in time O(s).

In the RealRAM model, given a matrix  R and a

vector  R, we can solve the linear system  =  in O(n) time, where ω is the matrix multiplication constant,

for which the best currently known bound is ω < 2.3727 [2], [3]. Such a running time bound is cost prohibitive for the large sparse matrices often encountered in practice. Iterative methods [4], first order methods [5], and matrix sketches [6]

Rasmus Kyng was supported by ONR Award N00014-16-1-2374.

Peng Zhang was partially supported by the NSF under Grant No. 1637566.

1A full version of this paper is available at https://arxiv.org/abs/1705. 02944.

 

can all be viewed as ways of obtaining significantly better performance in cases where the matrices have additional structure.

In contrast, when  is an n  n Laplacian matrix with m non-zeros, and polynomially bounded entries, the linear system  =  can be solved approximately to e-accuracy in O((m + n) log12+(1) n log(1/e)) time [7], [8]. This result spurred a series of major developments in fast graph algorithms, sometimes referred to as “the Laplacian Paradigm” of designing graph algorithms [9]. The asymptotically fastest known algorithms for Maximum Flow in directed unweighted graphs [10], [11], Negative Weight Shortest Paths and Maximum Weight Matchings [12], Min¬imum Cost Flows and Lossy Generalized Flows [13], [14] all rely on fast Laplacian linear system solvers.

The core idea of the Laplacian paradigm can be viewed as showing that the linear systems that arise from interior point algorithms, or second-order optimization methods, have graph structure, and can be preconditioned and solved using graph theoretic techniques. These techniques could poten¬tially be extended to a range of other problems, provided fast solvers can be found for the corresponding linear systems. Here a natural generalization is in terms of the number of labels per vertex: graph Laplacians correspond to graph labeling problems where each vertex has one label, and these labels interact pairwise via edges. Multi-label variants of these exist in Markov random fields [15], image process¬ing [16], Euclidean embedding of graphs [17], data pro¬cessing for cryo-electron microscopy [18], [19], [20], phase retrieval [21], [22], and many image processing problems (e.g. [23], [24]). Furthermore, linear systems with multiple labels per vertex arise when solving multi-commodity flow problems using primal-dual methods. Linear systems related to multi-variate labelings of graphs have been formulated as the quadratically-coupled flow problem [25] and Graph-Structured Block Matrices [26]. They also occur naturally in linear elasticity problems for simulating the effect of forces on truss systems [27].

Due to these connections, a central question in the Lapla¬cian paradigm of designing graph algorithms is whether all Graph-Structured Block Matrices can be solved in (ap¬proximately) nearly-linear time. Even obtaining subquadratic

 

running time would be constitute significant progress. There has been some optimism in this direction due to the existence of faster solvers for special cases: nearly-linear time solvers for Connection Laplacians [28], 1-Laplacians of collapsible 3-D simplicial complexes [29], and algorithms with runtime about n54 for 2D planar truss stiffness matrices [27]. Furthermore, there exists a variety of faster algorithms for approximating multi-commodity flows to (1 + c) accuracy in time that scales as poly(c1) [30], [31], [32], [33], even obtaining nearly-linear running times when the graph is undirected [34], [35], [36].

The subquadratic variants of these routines also in-teract naturally with tools that in turn utilize Laplacian solvers [25], [33]. These existing tight algorithmic connec¬tions, as well as the solver for Connection Laplacians, and the fact that combinatorial preconditioners partly originated from speeding up interior point methods through precondi-tiong Hessians [37], together provide ample reason to hope that one could develop nearly-linear time solvers for linear systems related to multicommodity flows. Any algorithm that solves such systems to high accuracy in m1+ time would in turn imply multicommodity flow algorithms that run in about n12m1+ time [13], while the current best running times are about n25 [38].

Unfortunately, we show that if linear systems in general 2D truss stiffness matrices or 2-commodity Laplacians can be solved approximately in nearly-linear time, then all linear systems in matrices with polynomially bounded integer entries can be solved in nearly-linear time. In fact, we show in a strong sense that any progress made in developing solvers for these classes of matrices will translate directly into similarly fast solvers for all matrices with polynomially bounded integer entries. Thus developing faster algorithms for these systems will be as difficult as solving all linear systems faster.

Since linear system solvers used inside Interior Point Methods play a central role in the Laplacian paradigm for designing high-accuracy algorithms, this may suggest that in the high-accuracy regime the paradigm will not extend to most problems that require multiple labels/variables per edge or vertex. Alternatively, an algorithmic optimist might view our result as a road-map for solving all linear systems via reductions to fast linear system solvers for Graph-Structured Block Matrices.

A. Our Results

Fast linear system solvers for Laplacians, Connection Laplacians, Directed Laplacians, and 2D Planar Truss Stiff¬ness matrices are all based on iterative methods and only produce approximate solutions. The running time for these solvers scales logarithmically with the error parameter c, i.e. as log(1/~). Similarly, the running time for iterative methods usually depends on the condition number of the matrix, but for state-of-the-art solvers for Laplacians, Connection 

 

Laplacians, and Directed Laplacians, the dependence is log¬arithmic. Consequently, a central open question is whether fast approximate solvers exist for other structured linear systems, with running times that depend logarithmically on the condition number and the accuracy parameter.

Integral linear systems are reducible to Graph-Structured Block Matrices. Our reductions show that if fast approximate linear system solvers exist for multi-commodity Laplacians, 2D Truss Stiffness, or Total Variation (TV) Matrices, then fast approximate linear system solvers exist for any matrix, in the very general sense of minimizing min JJ1122. Thus our result also applies to singular matrices, where we solve the pseudo-inverse problem to high accuracy. Theorem I.1 gives an informal statement of our main result. The result is stated formally in Section III as Theorem III.2.

Theorem I.1 (Hardness for Graph-Structured Linear Sys¬tems (Informal)). We consider three types of Graph-Structured Block Matrices: Multi-commodity Laplacians, Truss Stiffness Matrices, and Total Variation Matrices. Sup-pose that for one or more of these classes, the linear system  =  in a matrix  with s non-zeros can be solved in time _O(s), for some constant a  1, with the running time having logarithmic dependence on condition number and accuracy2. Then linear systems in all matrices with polynomially bounded integer entries and condition number can be solved to high accuracy in time O(s), where again s is the number of non-zero entries of the matrix.

Our results can easily be adapted to show that if fast exact linear system solvers exist for multi-commodity Laplacians, then exact solvers exist for all non-singular integer matrices. However, this is of less interest since there is less evidence that would suggest we should expect fast exact solvers to exist.

The notion of approximation used throughout this paper is the same as that used in the Laplacian solver literature (see Section II-A). To further justify the notion of approximate solutions to linear systems, we show that it lets us solve a natural decision problem for linear systems:

We show that deciding if a vector is approximately in the image of a matrix can be reduced to approximately solving linear systems. We show this in Section 10 of the full version of this paper. We also show that the exact image decision problem requires working with exponentially small numbers, even when the input has polynomially bounded integral entries and condition number. This means that in fixed-point arithmetic, we can only hope to solve an approx¬imate version of the problem. The problem of approximately solving general linear systems can be reduced the problem of approximately solving Graph-Structured Block Matrix

2This is the kind of running time guarantee established for Laplacians, Directed Laplacians, Connection Laplacians, and bounded-weight planar 2D Truss Stiffness matrices.

 

685

 

linear systems. Together, these statements imply that we can also reduce the problem of deciding whether a vector is approximately in the image of a general matrix to the problem of approximately solving Graph-Structured Block Matrix linear systems.

We establish surprising separations between many problems known to have fast solutions and problems that are as hard solving general linear systems. Our results trace out several interesting dichotomies: restricted cases of 2D truss stiffness matrices have fast solvers, but fast solvers for all 2D truss stiffness matrices would imply equally fast solvers for all linear systems. TV matrices can be solved quickly in the anisotropic case, but in the isotropic case imply solvers for all linear systems. Fast algorithms exist for multi-commodity problems in the low accuracy regime, but existing approaches for the high accuracy regime seem to require fast solvers for multi-commodity linear systems, which again would imply fast solvers for all linear systems.

Our reductions only require the simplest cases of the classes we consider: 2-Commodity Laplacians are sufficient, as are (non-planar) 2D Truss Stiffness matrices, and Total Variation Matrices with 2-by-2 interactions. Linear systems of these three classes have many applications, and faster solvers for these would be useful in all applications. Trusses have been studied as the canonical multi-variate problem, involving definitions such as Fretsaw extensions [39] and factor widths [40], and fast linear system solvers exist for the planar 2D case with bounded weights [27]. Total Variation Matrices are widely used in image denoising [41]. The anisotropic version can be solved using nearly-linear time linear system solvers [42], while the isotropic version has often been studied using linear systems for which fast solvers are not known [43], [44], [45]. Multi-commodity flow problems have been the subject of extensive study, with significant progress on algorithms with low accuracy [30], [31], [32], [33], [34], [35], [36], while high accuracy ap¬proaches use slow general linear system solvers.

B. Approximately Solving Linear Systems and Normal Equations

The simplest notion of solving a linear system  = , is to seek an  s.t. the equations are exactly satisfied. More generally, if the system is not guaranteed to have a solution, we can ask for an  which minimizes 22. An

 which minimizes this always exists. In general, it may not be unique. Finding an which minimizes 22 is equivalent to solving the linear system T = T, which is referred to as the normal equation for the linear system  =  (see [46]). The problem of solving the normal equations (or equivalently, minimizing 22), is a generalization of the problem of linear system solving, since the approach works when  is non-singular, while also giving meaningful results when  is singular. The normal equation problem can also be understood in terms 

 

of the Moore-Penrose pseudo-inverse of a matrix , which is denoted t as  = (T)tT is a solution to the normal equations. Taking the view of linear system solving as minimizing 2 2 also gives sensible ways to define an approximate solution to a linear system: It is an that ensures 2 2 is close to min22. In Section II, we formally define several notions of approximate solutions to linear systems that we will use throughout the paper.

An important special case of linear systems is when the matrix of coefficients of the system is positive semi-definite. Since T is always positive semi-definite, solving the normal equations for a linear system falls into this case. Linear systems over positive semi-definite matrices can be solved (approximately) by approaches known as iterative methods, which often lead to much faster algoritms than the approaches used for general linear systems. Iterative methods inherently produce approximate solutions3.

C. Graph-Structured Block Matrices

Graph-Structured Block Matrices are a type of linear system that arise in many applications. Laplacian matrices and Connection Laplacians both fall in this category.

Suppose we have a collection of n disjoint sets of variables X1, ... , Xn, with each set having the same size, Xi = d. Let i denote the vector4 of variables in Xi, and consider an equation of the form ij = 0, where

 and  are both r  d matrices. Now we form a linear system  = 0 by stacking m equations of the form given above as the rows of the system. Note that, very importantly, we allow a different choice of  and  for every pair of i and j. This matrix  Rmrxnd we refer to as a Incidence-Structured Block Matrix (ISBM), while we refer to T as a Graph-Structured Block Matrix (GSBM). Note that  is not usually PSD, but T is. The number of non-zeros in T is O(md2). GSBMs come up in many applications, where we typically want to solve a linear system in the normal equations of .

Laplacian matrices are GSBMs where d = 1 and  =

 = w, where w is a real number, and we allow different w for each pair of i and j. The corresponding ISBM for Laplacians is called an edge-vertex incidence matrix. Connection Laplacians are GSBMs where d = O (1) and  = T = w, for some rotation matrix  and a real number w. Again, we allow a different rotation matrix and scaling for every edge. For both Laplacians and Connection Laplacians, there exist linear system solvers that run in time

3A seeming counterexample to this is Conjugate Gradient which is an iterative method that produces exact solutions in the RealRAM model. But it requires extremely high precision calculations to exhibit this behaviour in finite precision arithmetic, and so Conjugate Gradient is also best understood as an approximate method.

4We use superscripts to index a sequence of vectors or matrices, and we use subscripts to denote entries of a vector or matrix, see Section II.

 

686

 

O(m polylog(n, e-1)) and produce c approximate solutions to the corresponding normal equations.

We now introduce several classes of ISBMs and their associated GSBMs. Our Theorem III.2 shows that fast linear system solvers for any of these classes would imply fast linear system solvers for all matrices with polynomially bounded entries and condition number.

Definition I.2 (2-commodity incidence matrix). A 2-commodity incidence matrix is an ISBM where d = 2 and r = 1, and S = T, and we allow three types of S: S=w(1 0),S=w(0 1)and S = w (1 1), where in each case w is a real number which may depend on the pair i and j. We denote the set of all 2-commodity incidence matrices by MC2. The corresponding GSBM is called a 2-commodity Laplacian. The ISBM definition is equivalent to requiring the GSBM to have the form

( 1 0 ) ( 0 0 ) ( 1 1 /

L1 ® +L2 ® +L1+2 ®

0 0 0 1 1 1

where ® is the tensor product and L1, L2, and L1+2 are all Laplacian matrices.

We adopt a convention that the first variable in a set Xis labelled u and the second is labelled v. Using this convention, given a 2-commodity incidence matrix B, the equation Bx = 0 must consist of scalings of the following three types of equations: u  u = 0, v  v = 0, and u  v  (u  v) = 0.

Definition I.3 (Strict 2-commodity incidence matrix). A strict 2-commodity incidence matrix is a 2-commodity inci-dence matrix where the corresponding 2-commodity Lapla¬cian has the property that L1, L2, and L1+2 all have the same non-zero pattern. We denote the set of all strict 2-commodity incidence matrices by MC0

2 . We denote the set of all strict 2-commodity incidence matrices with integer entries by MC0

2Z.

Linear systems in MC0

2 are exactly the systems that one has to solve to when solving 2-commodity problems using Interior Point Methods (IPMs). For readers unfamiliar with 2-commodity problems or IPMs, we provide a brief explanation of why this is the case in Section 9 of the full version of this paper. The MC0

2 is more restrictive than MC2, and MC0

2Z in turn is even more restrictive. One could hope that fast linear system solvers exist for MC0

2 or MC0

2Z, even if they do not exist for MC2. However, our reductions show that even getting a fast approximate solver for MC0

2Z with polynomially bounded entries and condition number will lead to a fast solver for all matrices with polynomially bounded entries and condition number.

The next class we consider is 2D Truss Stiffness Ma-trices. They have been studied extensively in the numer¬ical linear algebra community [39], [40]. For Planar 2D Trusses with some bounds on ranges of edges, Daitch and 

 

Spielman obtained linear system solvers that run in time O(n54 log(1/e)).

Definition I.4 (2D Truss Incidence Matrices). Let G = (V, E) be a graph whose vertices are n points in 2-dimension: s1, ... , s  R2. Consider X1,..., X where d = 2. A 2D Truss Incidence Matrix is an ISBM where d = 2 and r = 1, and for each i and j, we have S = T and S = w(s  s)T, and w is a real number that may depend on the pair i and j, but s depends only on i and vice versa for s. We denote the class of all 2D Truss Incidence Matrices by T2.

Another important class of matrices is Total Variation Ma¬trices (TV matrices). TV matrices come from Interior Point Methods for solving total variation minimization problem in image, see for example [47] and [48]. Not all TV matrices are GSBMs, but many GSBMs can be expressed as TV matrices.

Definition I.5 (TV matrix and 2-TV Incidence Matrices). Let E1  ...  E be a partition of the edge set of a graph. For each 1 < i < s, let B be the edge-vertex incidence matrix of E, W  be a diagonal matrix of edge weights, and r be a vector satisfying W  >,- r(r)T. Given these objects, the associated total variation matrix (TV matrix) is a matrix M defined as

M =E (B)T (W  r(r)T) B. 1<<

A 2-TV Incidence Matrix is defined as any ISBM whose corresponding GSBM is a TV matrix with W  R212 and r  R2. We denote the class of all 2-TV incidence matrices by V2.

D. Our Reduction: Discussion and an Example

In this section we give a brief sketch of the ideas behind our reduction from general linear systems, over matrices in G, to multi-commodity linear systems, over matrices in MC2, and we demonstrate the most important transforma¬tion through an example.

The starting point for our approach is the folklore idea that any linear system can be written as a factor-width 3 system by introducing a small number of extra variables. Using a set of multi-commodity constraints, we are able to express one particular factor-width 3 equation, namely 2x'' = x + x'. After a sequence of preprocessing steps, we are then able to efficiently express arbitrary linear systems over integer matrices using constraints of this form. A number of further issues arise when the initial matrix does not have full column rank, requiring careful weighting of the constraints we introduce.

Given a matrix A with polynomially bounded integer entries and condition number, we reduce the linear system Ax = c to a linear system By = d, where B is a strict multi-commodity edge-vertex incidence matrix with

 

687

 

integer entries (i.e. in MC0

2Z), with polynomially bounded entries and condition number. More precisely, we reduce T = T to T = T. These systems always have a solution. We show that we can find an c-approximate solution to the linear system T = T by a simple mapping on any  that c'-approximately solves the linear system T = T, where c' is only polynomially smaller than c. If  has s non-zero entries and the maximum absolute value of an entry in  is U, then  will have O(s log(sU)) non-zero entries and our algorithm computes the reduction in time O(s log(sU)). Note that T has r = O(s log(sU)) non-zeros, because every row of  has O (1) entries. All together, this means that getting a solver for T = T with running time O(rlog(1/e)) will give a solver for  with O(s log(1/e)) running time.

We achieve this through a chain of reductions. Each reduction produces a new matrix and vector, as well as a new error parameter giving the accuracy required in the new system to achieve the accuracy desired in the original system.

1) We get a new linear system Z2Z2 = Z2 where

Z2 has integer entries, and the entries of each row of Z2 sum to zero, i.e. Z21 = 0, and finally in every row the sum of the positive coefficients is a power of two.

2) Z2Z2 = Z2 is then transformed to  = , where  is a 2-commodity edge-vertex incidence matrix.

3)  =  is then transformed to 0 = 0, where 0 is a strict 2-commodity edge-vertex incidence matrix.

4) 0 = 0 is then transformed to 0Z = 0Z, where0Z is a 2-commodity edge-vertex incidence matrix with integer entries.

We will demonstrate step 2, the main transformation, by example. When the context is clear, we drop the superscripts of matrices for simplicity. The reduction handles each row (i.e. equation) of the linear system independently, so we focus on the reduction for a single row.

Consider a linear system  = , and let us pick a single row (i.e. equation)  =  5. We will repeatedly pick pairs of existing variables of , say x and x', based on their current coefficients in  = , and modify the row by adding C (2x'' (x+x')) to the left hand side where x'' is a new variable and C is a real number we pick. As we will see in a moment, we can use this pair-and-replace operation to simplify the row until it eventually becomes a 2-commodity equation. At the same time as we modify , we also store an auxiliary equation C(x +x'2x'') = 0. Suppose initially that  =  is satisfied. After this modification of  = , if the auxiliary equation is satisfied,  =  is still

5We use i to denote the ith row of , and i to denote the ith entry of , see Section II.

 

satisfied by the same values of x and x'. Crucially, we can express the auxiliary equation C(x + x'  2x'') = 0 by a set of ten 2-commodity equations, i.e. a “2-commodity gadget” for this equation. Our final output matrix will not contain the equation C(x + x'  2x'') = 0 as a row, but will instead contain 10 rows of 2-commodity equations from our gadget construction. Eventually, our pair-and-replace scheme will also transform the row  =  into a 2-commodity equation on just two variables.

Next, we need to understand how the pair-and-replace scheme makes progress. The pairing handles the positive and the negative coefficients of  separately, and eventually ensures that  =  has only a single positive and a single negative coefficient in the modified row  = , in particular it is of the form axax' =  for two variables x and x' that appear in the modified vector of variables , i.e. it is a 2-commodity equation.

To understand the pairing scheme, it is helpful to think about the entries of  written using binary (ignoring the sign of the entry). The pairing scheme proceeds in a sequence of rounds: In the first round we pair variables whose 1st (smallest) bit is 1. There must be an even number of variables with smallest bit 1, as the sum of the positive (and respectively negative) coefficients is a power of 2. We then replace the terms corresponding to the 1st bit of the pair with a new single variable with a coefficient of 2. After the first round, every coefficient has zero in the 1st bit. In the next round, we pair variables whose 2nd bit is 1, and replace the terms corresponding to the the 2nd bit of the pair with a new single variable with a coefficient of 4, and so on. Because the positive coefficients sum to a power of two, we are able to guarantee that pairing is always possible. It is not too hard to show that we do not create a large number of new variables or equations using this scheme.

For example, let us consider an equation

31 + 52 + 43 + 44  165 = 1.

Replace 1 +2 by 26. Add auxiliary equation 1 +2  26 = 0. The equation above becomes

2(1 + 22 + 6 + 23 + 24)  165 = 1.

Replace 2( 1 + 6) by 4 7. Add auxiliary equation 2( 1 + 6  27) = 0. We now have

4(2 + 7 + 3 + 4)  165 = 1.

Replace 4(2 + 7) by 88, and 4(3 + 4) by 89. Add auxiliary equations 4(2 +7 28) = 0, and 4(3 +4  29) = 0. Our equation above becomes

8(8 + 9)  165 = 1.

Replace 8(8 + 9) by 1610. Add auxiliary equation 8(8 + 9  210) = 0. Finally, the equation above has become a 2-commodity equation:

1610  165 = 1.

 

688

 

Now, let us build some intuition for how to replace the equation C(x + x 2x) = 0 by ten 2-commodity equations, i.e. a 2-commodity gadget. Recall that each index i corresponds to a -variable  and a -variable . We think of x, x, x all as -variables. Roughly speaking, the 2-commodity equations of the form  = 0 and  = 0 allow us to set two variables equal, although the effect is more complicated when considering an overconstrained system. The 2-commodity equations of the form  () = 0 are even more important to us: The constraint x + x 2x = 0 can be obtained by adding two equations: x  (x) = 0 and x (x) = 0. The appearance of the term in both equations, though with opposite sign, gives a degree of freedom that ensures that we do not impose additional constraints on x, x, and x. We get the desired constraints listed above by starting with two 2-commodity equations using fresh variables with new indices a, b, c, d:  () =0 and  () =0. We then link together these variables using constraints on pairs of variables of the same type (i.e.  or ). First, we link the  and x variables:  x = 0,  x = 0,  x = 0,  x = 0. Secondly, we constrain the variables to give only the single degree of freedom we need: For technical reasons, we introduce two more new indices f and gand set  = 0,  = 0, and  = 0,  = 0. Now we are left with a system essentially equivalent to the two equations x  (x) = 0 and x (x) = 0. These equations impose exactly the constaint x + x 2x = 0 that we want.

In this way, we process  =  to produce a new set of equations  =  where  is a 2-commodity matrix. If  =  has an exact solution, this solution can be obtained directly from an exact solution to  = . We also show that an approximate solution to  =  leads to an approximate solution for  = , and we show that does not have much larger entries or condition number than .

The situation is more difficult when  =  does not have a solution and we want to obtain an approximate minimizer arg minRn22 from an approximate solution to arg minRn'22. This corresponds to approximately applying the Moore-Penrose pseudo-inverse of  to . We deal with the issues that arise here using a carefully chosen scaling of each auxiliary constraint to ensure a strong relationship between different solutions.

In order to switch from a linear system in a general 2-commodity matrix to a linear system in a strict 2-commodity matrix, we need to reason very carefully about the changes to the null space that this transformation inherently produces. By choosing sufficiently small weights, we are nonetheless able to establish a strong relationship between the normal equation solutions despite the change to the null space.

 

II. PRELIMINARIES

We use subscripts to denote entries of a matrix or a vector: let  denote the ith row of matrix  and denote the (i, j)th entry of ; let  denote the ith entry of vector  and : (i < j) denote the vector of entries , +1, . . ., . We use superscripts to index a sequence of matrices or vectors, e.g., 1, 2, ... , and 1, 2, . .., except when some other meaning is clearly stated.

We use  to denote the Moore-Penrose pseudo-inverse of a matrix . We use im() to denote the image of a matrix . We use •2 to denote the Euclidean norm on vectors and the spectral norm on matrices. When  is an n  n positive

semidefinite matrix, we define a norm on vectors  R

def 

by  = . We let nnz() denote the number

of non-zero entries in a matrix . We define 

E

max, 1 = max

E and  =

max. We let min+() = mins.t. ij =0 . Given a matrix  R and a vector  R for some m, n, we call the tuple (, ) a linear system. Given matrix  R, let A def=A(AA)†A, i.e. the orthogonal projection onto im(A). Note that A = Aand A =  2A.

A. Approximately Solving A Linear System

In this section we formally define the notions of ap-proximate solutions to linear systems that we work with throughout this paper.

Definition II.1 (Linear System Approximation Problem, LSA). Given linear system (, ), where  R, and  R, and given a scalar 0  E  1, we refer to the LSA problem for the triple (, , E) as the problem of finding  R s.t.

2  E 2 ,

and we say that such an  is a solution to the LSA instance (, , E).

This definition of a LSA instance and solution has several advantages: when im() = R, we get A = I, and it reduces to the natural condition Axc2 E c2, which because im(A) = Rm, can be satisfied for any E, and for E = 0 tells us that Ax = c.

When im() does not include all of R, the vector Acis exactly the projection of c onto im(A), and so a solution can still be obtained for any E. Further, as (IA)c is orthogonal to Ac and Ax, it follows that

22 = ()22 + 22 . Thus, when  is a solution to the LSA instance (, , E), then  also gives an E222 additive approximation to

min 2 2 = ()22 .(1)

Rn

 

689

 

Similarly, an  which gives an additive e2 I I Ac22 approx¬imation to Problem (1) is always a solution to the LSA instance (A, c, c). These observations prove the following (well-known) fact:

Fact II.2. Let  E arg minR- II  - 22, then for every ,

11  - 22 <  - 22 +E222 if and only if  is a solution to the LSA instance (, , c).

When the linear system  =  does not have a solution, a natural notion of solution is any minimizer of Problem (1). A simple calculation shows that this is equivalent to requir¬ing that  is a solution to the linear system  = , which always has a solution even when  =  does not. The system  =  is referred to as the normal equation associated with  =  (see [46]).

Fact II.3.  E arg minR„ II  - 2 2, if and only if  = , and this linear system always has a solution.

This leads to a natural question: Suppose we want to approximately solve the linear system  = . Can we choose our notion of approximation to be equivalent to that of a solution to the LSA instance (, , c)?

A second natural question is whether we can choose a notion of distance between a proposed solution  and an optimal solution  E arg minR„ 11  - 22 s.t. this distance being small is equivalent to  being a solution to the LSA instance (, , c)? The answer to both questions is yes, as demonstrated by the following facts:

Fact II.4. Suppose  E arg minR„ 11  - 2

2 then

1)  (T)t = 2 =

 - T.

2) The following statements are each equivalent to being a solution to the LSA instance (, , c):

a) 

(T)t 6  (T)t if

and only if  is a solution to the LSA instance (, , c).

b)  - T 'E T if and only if  is a solution to the LSA instance (, , c).

For completeness, we prove Fact II.4 in Appendix A of the full version of this paper. Fact II.4 explains connection between our Definition II.1, and the usual convention for measuring error in the Laplacian solver literature [7]. In this setting, we consider a Laplacian matrix , which can be written as  =  E R, and a vector  s.t. ΠATAb = b. This condition on b is easy to verify in the case of Laplacians, since for the Laplacian of a connected graph, ΠATA = In111. Additionally, it is also equivalent to the condition that there exists c s.t. b = Ac. For Laplacians 

 

it is possible to compute both  and a vector  s.t.  =  in time linear in nnz(). For Laplacian solvers, the approximation error of an approximate solution  is

measured by the c s.t. 

(T)t e (T)t.

By Fact II.4, we see that this is exactly equivalent to  being a solution to the LSA instance (, , c).

B. Measuring the Difficulty of Solving a Linear System

Running times for iterative linear system solvers generally depend on the number of non-zeros in the input matrix, the condition number of the input matrix, the accuracy, and the bit complexity.

In this section, we formally define several measures of complexity of the linear systems we use. This is crucial, because we want to make sure that our reductions do not rely on mapping into extremely ill-conditioned matrices, and so we use these measures to show that this is in fact not the case.

Definition II.5.

1) Given a matrix  E R, we define the maxi 

mum singular value σmax(A) in the usual way as

/1/T T

σmax(A) = maxxR„x=0 xTx .

2) Given a matrix  E R which is not all zeros, we

define the minimum non-zero singular value σmin(A)

/1/T T

as σmin(A) = minxR„x null(A) xTx .

3) Given a matrix  E R which is not all zeros, we define the non-zero condition number of  as r.() =

max()

min() .

Definition II.6. The sparse parameter complexity of an LSA instance (, , c) where  E Z and nnz() > max(m, n), and c > 0, is

(, , e) def= (nnz(), U(), r.(), e1)

( )

where U() = max max , max ,  1 

min+() , 1 .

min+()

Note in the definition above that when  =0 and  =0 have only integer entries, we trivially have min+() >_ 1 and min+() > 1. However, including min+() , and 1 

C. Matrix Classes and Reductions Between Them

We use the term matrix class to refer to an infinite set of

matrices . In this section, we formally define a notion of

efficient reduction between linear systems in different classes of matrices.

Definition II.7 (Efficient f-reducibility). Suppose we

have two matrix classes 1 and 2, and there ex 

ist two algorithms 12 and 12 s.t. given an LSA

instance ( 1, 1, c), where  1 E 1, the call

 

690

 

12(M 1, c1, c1) returns an LSA instance (M2, c2, c2) s.t. if x2 is a solution to LSA instance (M2, c2, c2) then x1 = 12(M 1, M 2, x2) is a solution to LSA instance

(M 1, c1, c1).

Consider a function of f : R4+ R4+ s.t. every output co¬ordinate is an increasing function of every input coordinate. Suppose that we always have

(M2, c2, c2)  f ((M 1, c1, c1)),

and the running times of 12(M 1, c1, c1) and

12(M 1, M2, x2) are both bounded by O(nnz(M 1)). Then we say that 1 is efficiently f-reducible to 2, which we also write as

1 2.

Lemma II.8. Suppose 1 2 and 2 3. Then 1 ◦3.

Proof: The proof is simply by the trivial composition of the two reductions.

Definition II.9. We let  denote the class of all matrices with integer valued entries s.t. there is at least one non-zero entry in every row and column6.

III. MAIN RESULTS

In this section, we use the notions of sparse parameter complexity and matrix class reductions to prove our main technical result, Theorem III.1, which shows that linear systems in general matrices with integer entries can be efficiently reduced to linear systems in several different classes of Incidence Structured Block Matrices. From this result, we derive as corollary our main result, Theorem III.2, which states that fast high accuracy solvers for several types of ISBMs imply fast high accuracy solvers for all linear systems in general matrices with integer entries.

Theorem III.1. Let f (s, U, K, c) = (O(s log(sU)), poly(UKc1s), poly(UKc1s), poly(UKc1s)), then

1) 0

2Z.

2) 2.

3) 2.

Theorem III.2. Suppose we have an algorithm which solves every Linear System Approximation Problem (A, c, c) with sparse parameter complexity (A, c, c)  (s, U, K, e1) in time O(s polylog(s, U, K, c1)) for some a  1, whenever

I

A for at least one of 0

2Z, 2, 2 . I.e.

we have a “fast” solver7 for one of the matrix classes

0

2Z, 2, or 2. Then every Linear System Approximation

6If there is a row or column with only zeros, then it can always be handled trivially in the context of solving linear systems

7The reduction requires only a single linear system solve, and uses the solution in a black-box way. So the reduction also applies if the solver for the class R only works with high probability or only has running time guarantees in expectation.

 

Problem (A, c, c) where A  with sparse parameter complexity (A, c, c)  (s, U, K, e1) can be solved in time O(s polylog(s, U, K, c1)).

Proof: The theorem is a immediate corollary of Theo¬rem III.1.

Definition III.3. We let z2 denote the class of all matrices with integer valued entries s.t. there is at least one non-zero entry in every row and column, and every row has zero row sum, and for each row, the sum of the positive coefficients is a power of 2.

Lemma III.4. Let f (s, U, K, c) = (O(s), O (c1s92U3) , O (C1s8U3K) , O (s52U2c1)), then

z2.

Lemma III.5. Let f (s, U, K, c) = (O(s log(sU)),

O(s32U log12(sU)), O(Ks4U2log2(sU)), O(sU2c1)),

then

z2 2.

Lemma III.6. Let f (s, U, K, c) = (O(s), O (c1U2K) , O (c1s2U2K) , O(c1)), then

2 0

2 .

Lemma III.7. Let f (s, U, K, c) = (s, c1sU, 2K, O (c1)), then

0

2 0

2Z.

Lemma III.8. Let f (s, U, K, c) be as defined in Lemma III.5

then

z2 2.

Lemma III.9. Let f (s, U, K, c) = (s, U, K, c1), then

2 2.

Proof of Theorem III.1: Follows by appropriate com 

position (Lemma II.8) applied to the the Lemmas above, i.e.

III.4, III.5, III.6, III.7, III.8 and III.9.

 

The full version of this paper, available at https://arxiv.org/ abs/1705.02944, presents proofs of all the lemmas stated in this section. We also give proofs of Lemmas III.8 and III.9 in the following sections.

IV. 2D TRUSSES

In this section, we prove Lemma III.8. We show that the reduction algorithm used in proving z2 2 constructs a 2D Truss Incidence Matrix as per Definition I.4. It follows that for any function f, 2 implies 2. The key is to show that a 2-commodity gadget in the reduction corresponds to a 2D truss subgraph, which we call the 2D-truss gadget.

Without loss of generality, we let u-variables correspond to the horizonal axis and v-variables to the vertical axis of the 2D plane. According to Definition I.2 and I.4:

 

691

 

1) an equation  = 0 in a 2-commodity linear system corresponds to a horizontal edge in the 2D plane;

2) an equation  = 0 in a 2-commodity linear system corresponds to a vertical edge in the 2D plane;

3) an equation () = 0 in a 2-commodity linear system corresponds to a diagonal edge in the 2D plane.

Note that our reduction here heavily relies on the ability to choose arbitrary weights. In particular, the weights on the elements are not related at all with the distances between the corresponding vertices.

Our strategy to pick the coordinates of the vertices of the constructed 2D truss is the following: we first pick the coordinates of the original n vertices randomly, and then determine the coordinates of the new vertices constructed in the reduction to satisfy all the truss equations.

For the n original vertices, we pick their -coordinates arbitrarily and pick their -coordinates randomly. Specifi¬cally, we pick an n-dimensional random vector  uniformly distributed on the n-dimensional sphere centered at the origin and with radius R = n10; we then round each entry of  to have precision δ = 10-10, so that each entry has constant bits. Let y˜ be the vector after rounding. We assign the -coordinate of the ith vertex to be the ith entry of ˜.

We then pick the coordinates of the new vertices in the order they are created. Note that each time we replace two vertices in the current equations, say 1, 2, whose coordinates have already been determined, we create a 2D truss gadget with 7 new vertices, say , +1, .. . , +6 8. According to the construction of this gadget, the new vertices +1, . . . , +6 only appear in this single gadget, whose coordinates do not affect other vertices. Figure 1 is the cor¬responding subgraph which satisfies all the equations in the 2D truss gadget. Note the two triangles (+3, +5, +6) and (+3, +4, +5) need to be isosceles right triangles, which implies  = (1 + 2)/2. We can always place , +1, . . . , +6 to get the desired equations, provided 1 =2, which is guaranteed with high probability by Lemma IV.1.

We prove the following lemma in the full version of this paper.

Lemma IV.1. Let a  R be a fixed vector such that 2 a 2, i  [n] and aT1 = 0. Let ˜ be a vector picked as above. Then,

Pr aT˜ = 0  2δn2

a2 R.

By our construction of the truss, for each vertex, its -coordinate can be written as a fixed convex combination

8See Algorithm 2 MC2GADGET in the full version for a detailed construction.

 

 

+1 +4 1

Figure 1. Geometric realization of the 2D truss gadget

of ˜, say T˜ in which T1 = 1 and  0, i  [n]. Consider a pair of two arbitrary vertices, let 1 and 2 be their coefficient vectors corresponding to the convex combinations. These two vertices have same -coordinate if and only if (12)T˜ = 0. Let a def=1 2. Then, 2  a 2, i  [n], aT1 = 0, and

a2  1 

 2 +  2 

 2  1  1 + 2   1 =2.

By Lemma IV.1,

Pr 1T˜ = 2T˜  δRn2 .

By Lemma III.5, the total number of the vertices in the truss is at most O n2 log n . By a union bound, the probability that there exist two different vertices with same -coordinate is at most

O n2 log n2 = O

Proof of Lemma III.8: Since the linear system for 2D trusses is the same as the linear system for 2-commodity, all complexity parameters of these two linear systems are the same.

V. ISOTROPIC TOTAL VARIATION MINIMIZATION

In this section, we prove Lemma III.9. We show the reduc¬tion algorithm used in proving z2 2 constructs an 2-TV Incidence Matrix defined in Definition I.5. It follows that for any function f, 2 implies 2.

Lemma III.9 follows

from the claim below, which shows  1 1 

that the type 1+2 2-commodity matrices

1 1

that occur in GSBMs for 2-commodity ISBMs can be constructed as Total Variation matrices. The same is true for the other types of entries that occur in 2-commodity GSBMs.

Claim V.1. For each edge (i, j) in the graph corresponding to 1+2, there exists an edge-vertex incidence matrix , a

 

692

 

diagonal matrix W and a vector r such that W >,- r r T and

( 1 1 )

L1+2

  = N T (W  rrT) N 

1 1

Claim V.1 is proven in the full version of this paper. Proof of Lemma III.9: Since the linear system related to the isotropic total variation minimization matrix is the same as the linear system for 2-commodity, all complexity parameters of these two linear systems are the same.

ACKNOWLEDGMENT

We thank Richard Peng and Daniel Spielman for help-ful comments and discussions. We thank our anonymous reviewers for pointing out several typos.

REFERENCES

[1] V. V. Williams and R. Williams, “Subcubic equivalences between path, matrix and triangle problems,” in Proceedings of the 2010 IEEE 51st Annual Symposium on Foundations of Computer Science, ser. FOCS ’10. Washington, DC, USA: IEEE Computer Society, 2010, pp. 645–654. [Online]. Available: http://dx.doi.org/10.1109/FOCS.2010.67

[2] V. Strassen, “Gaussian elimination is not optimal,” Numer. Math., vol. 13, no. 4, pp. 354–356, Aug. 1969. [Online]. Available: http://dx.doi.org/10.1007/BF02165411

[3] V. V. Williams, “Multiplying matrices faster than coppersmith-winograd,” in Proceedings of the Forty-fourth Annual ACM Symposium on Theory of Computing, ser. STOC ’12. New York, NY, USA: ACM, 2012, pp. 887–898. [On¬line]. Available: http://doi.acm.org/10.1145/2213977.2214056

[4] Y. Saad, Iterative Methods for Sparse Linear Systems,

2nd ed. Philadelphia, PA, USA: Society for Industrial

and Applied Mathematics, 2003, available at http://www-users.cs.umn.edu/˜saad/toc.pdf.

[5] S. Boyd and L. Vandenberghe, Convex Optimization. Camebridge University Press, 2004, available at https://web.stanford.edu/˜boyd/cvxbook/bv cvxbook.pdf.

[6] D. P. Woodruff, “Sketching as a tool for numerical linear algebra,” Theoretical Computer Science, vol. 10, no. 1-2, pp. 1–157, 2014, available at http://arxiv.org/abs/1411.4357.

[7] D. Spielman and S. Teng, “Nearly linear time algorithms for preconditioning and solving symmetric, diagonally dominant linear systems,” SIAM Journal on Matrix Analysis and Ap¬plications, vol. 35, no. 3, pp. 835–885, 2014, available at http://arxiv.org/abs/cs/0607105.

[8] M. B. Cohen, R. Kyng, G. L. Miller, J. W. Pachocki, R. Peng, A. B. Rao, and S. C. Xu, “Solving sdd linear systems in nearly mlog1/2n time,” in Proceedings of the Forty-sixth Annual ACM Symposium on Theory of Computing, ser. STOC ’14. New York, NY, USA: ACM, 2014, pp. 343–352. [Online]. Available: http://doi.acm.org/10.1145/2591796.2591833

[9] S.-H. Teng, “The Laplacian Paradigm: Emerging Algorithms for Massive Graphs,” in Theory and Applications of Models of Computation, 2010, pp. 2–14.

 

[10] A. Madry, “Navigating central path with electrical flows: From flows to matchings, and back,” in Proceedings of the 2013 IEEE 54th Annual Symposium on Foundations of Computer Science, ser. FOCS ’13. Washington, DC, USA: IEEE Computer Society, 2013, pp. 253–262. [Online]. Available: http://dx.doi.org/10.1109/FOCS.2013.35

[11] ——, “Computing maximum flow with augmenting electrical flows,” in Proceedings of the 2016 IEEE 57th Annual Sym¬posium on Foundations of Computer Science, ser. FOCS ’16. Washington, DC, USA: IEEE Computer Society, 2016.

[12] M. B. Cohen, A. Madry, P. Sankowski, and A. Vladu, “Negative-weight shortest paths and unit capacity minimum cost flow in ˜O(m10/7log w) time: (extended abstract),” in Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms, ser. SODA ’17. Philadelphia, PA, USA: Society for Industrial and Applied Mathematics, 2017, pp. 752–771. [Online]. Available: http://dl.acm.org/citation.cfm?id=3039686.3039734

[13] Y. T. Lee and A. Sidford, “Path finding methods for linear programming: Solving linear programs in

˜O(



rank) iterations and faster algorithms for maxi¬mum flow,” in Foundations of Computer Science (FOCS),

2014 IEEE 55th Annual Symposium on. IEEE, 2014,

pp. 424–433, available at http://arxiv.org/abs/1312.6677 and http://arxiv.org/abs/1312.6713.

[14] S. I. Daitch and D. A. Spielman, “Faster approximate lossy generalized flow via interior point algorithms,” in Proceedings of the 40th annual ACM symposium on Theory of computing, ser. STOC ’08. New York, NY, USA: ACM, 2008, pp. 451– 460, available at http://arxiv.org/abs/0803.0988. [Online]. Available: http://doi.acm.org/10.1145/1374376.1374441

[15] R. Szeliski, R. Zabih, D. Scharstein, O. Veksler, V. Kol-mogorov, A. Agarwala, M. Tappen, and C. Rother, “A com¬parative study of energy minimization methods for markov random fields with smoothness-based priors,” IEEE Transac¬tions on Pattern Analysis and Machine Intelligence, vol. 30, pp. 1068–1080, 2008.

[16] Y. Wang, J. Yang, W. Yin, and Y. Zhang, “A new alternating minimization algorithm for total variation image reconstruc¬tion,” SIAM Journal on Imaging Sciences, vol. 1, no. 3, pp. 248–272, 2008.

[17] M. B. Cohen, Y. T. Lee, G. Miller, J. Pachocki, and A. Sid-ford, “Geometric median in nearly linear time,” in Pro¬ceedings of the 48th Annual ACM SIGACT Symposium on Theory of Computing. ACM, 2016, pp. 9–21, available at: https://arxiv.org/abs/1606.05225.

[18] A. Singer and Y. Shkolnisky, “Three-dimensional structure determination from common lines in cryo-em by eigenvectors and semidefinite programming,” SIAM journal on imaging sciences, vol. 4, no. 2, pp. 543–572, 2011.

[19] Y. Shkolnisky and A. Singer, “Viewing direction estimation in cryo-em using synchronization,” SIAM journal on imaging sciences, vol. 5, no. 3, pp. 1088–1110, 2012.

[20] Z. Zhao and A. Singer, “Rotationally invariant image rep-resentation for viewing direction classification in cryo-em,” Journal of structural biology, vol. 186, no. 1, pp. 153–166, 2014.

 

693

 

[21] B. Alexeev, A. S. Bandeira, M. Fickus, and D. G. Mixon, “Phase retrieval with polarization,” SIAM Journal on Imaging Sciences, vol. 7, no. 1, pp. 35–66, 2014.

[22] S. Marchesini, Y.-C. Tu, and H.-t. Wu, “Alternating pro-jection, ptychographic imaging and phase synchronization,” arXiv preprint arXiv:1402.0550, 2014.

[23] O. Ozyesil, A. Singer, and R. Basri, “Stable camera motion estimation using convex programming,” SIAM Journal on Imaging Sciences, vol. 8, no. 2, pp. 1220–1262, 2015.

[24] M. Arie-Nachimson, S. Z. Kovalsky, I. Kemelmacher-Shlizerman, A. Singer, and R. Basri, “Global motion esti-mation from point matches,” in 3D Imaging, Modeling, Pro¬cessing, Visualization and Transmission (3DIMPVT), 2012 Second International Conference on. IEEE, 2012, pp. 81–88.

[25] J. A. Kelner, G. L. Miller, and R. Peng, “Faster approximate multicommodity flow using quadratically coupled flows,” in Proceedings of the 44th symposium on Theory of Computing, ser. STOC ’12. New York, NY, USA: ACM, 2012, pp. 1–18, available at http://arxiv.org/abs/1202.3367. [Online]. Available: http://arxiv.org/abs/1202.3367

[26] D. A. Spielman, “Nsf award 1562041: Generalized algebraic graph theory: Algorithms and analysis,” ALGORITHMIC FOUNDATIONS, 2016.

[27] S. I. Daitch and D. A. Spielman, “Support-graph preconditioners for 2-dimensional trusses,” CoRR, vol. abs/cs/0703119, 2007. [Online]. Available: http://arxiv.org/abs/cs/0703119

[28] R. Kyng, Y. T. Lee, R. Peng, S. Sachdeva, and D. A. Spielman, “Sparsified cholesky and multigrid solvers for connection laplacians,” in Proceedings of the Forty-eighth Annual ACM Symposium on Theory of Computing, ser. STOC ’16. New York, NY, USA: ACM, 2016, pp. 842–850. [Online]. Available: http://doi.acm.org/10.1145/ 2897518.2897640

[29] M. B. Cohen, B. T. Fasy, G. L. Miller, A. Nayyeri, R. Peng, and N. Walkington, “Solving 1-laplacians in nearly linear time: Collapsing and expanding a topological ball,” in Pro¬ceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2014, Portland, Oregon, USA, January 5-7, 2014, 2014, pp. 204–216.

[30] N. Garg and J. K¨onemann, “Faster and simpler algorithms for multicommodity flow and other fractional packing prob¬lems,” in In Proceedings of the 39th Annual Symposium on Foundations of Computer Science, 1998, pp. 300–309.

[31] A. Madry, “Faster approximation schemes for fractional mul-ticommodity flow problems via dynamic graph algorithms,” in STOC ’10: Proceedings of the 42nd ACM symposium on Theory of computing. New York, NY, USA: ACM, 2010, pp. 121–130.

[32] L. K. Fleischer, “Approximating fractional multicommodity flow independent of the number of commodities,” SIAM Journal on Discrete Mathematics, vol. 13, pp. 505–520, 2000.

 

[33] T. Leighton, F. Makedon, S. Plotkin, C. Stein, E. Tardos, and S. Tragoudas, “Fast approximation algorithms for multicom-modity flow problems,” in JOURNAL OF COMPUTER AND SYSTEM SCIENCES, 1991, pp. 487–496.

[34] J. Sherman, “Nearly maximum flows in nearly linear time,” in Proceedings of the 2013 IEEE 54th Annual Symposium on Foundations of Computer Science, ser. FOCS ’13. Washington, DC, USA: IEEE Computer Society, 2013, pp. 263–269. [Online]. Available: http: //dx.doi.org/10.1109/FOCS.2013.36

[35] J. A. Kelner, Y. T. Lee, L. Orecchia, and A. Sidford, “An almost-linear-time algorithm for approximate max flow in undirected graphs, and its multicommodity generalizations,” in Proceedings of the Twenty-fifth Annual ACM-SIAM Symposium on Discrete Algorithms, ser. SODA ’14. Philadelphia, PA, USA: Society for Industrial and Applied Mathematics, 2014, pp. 217–226. [Online]. Available: http://dl.acm.org/citation.cfm?id=2634074.2634090

[36] R. Peng, “Approximate undirected maximum flows in o(mpoly log(n)) time,” in Proceedings of the Twenty-seventh Annual ACM-SIAM Symposium on Discrete Algorithms, ser. SODA ’16. Philadelphia, PA, USA: Society for Industrial and Applied Mathematics, 2016, pp. 1862– 1867. [Online]. Available: http://dl.acm.org/citation.cfm?id= 2884435.2884565

[37] P. M. Vaidya, “Speeding-up linear programming using fast matrix multiplication,” in Proceedings of the 30th Annual Symposium on Foundations of Computer Science. Washington, DC, USA: IEEE Computer Society, 1989, pp. 332–337. [Online]. Available: http://portal.acm.org/citation. cfm?id=1398514.1398712

[38] Y. T. Lee and A. Sidford, “Efficient inverse maintenance and faster algorithms for linear programming,” in Founda¬tions of Computer Science (FOCS), 2015 IEEE 56th Annual

Symposium on. IEEE, 2015, pp. 230–249, available at:

https://arxiv.org/abs/1503.01752.

[39] G. Shklarski and S. Toledo, “Rigidity in finite-element matrices: Sufficient conditions for the rigidity of structures and substructures,” SIAM J. Matrix Analysis Applications, vol. 30, no. 1, pp. 7–40, 2008. [Online]. Available: http://dx.doi.org/10.1137/060650295

[40] E. G. Boman, D. Chen, O. Parekh, and S. Toledo, “On factor width and symmetric h-matrices,” Linear algebra and its applications, vol. 405, pp. 239–248, 2005.

[41] T. Chan and J. Shen, Image Processing and Analysis: Varia¬tional, PDE, Wavelet, and Stochastic Methods. Philadelphia, PA, USA: Society for Industrial and Applied Mathematics, 2005.

[42] I. Koutis, G. L. Miller, and D. Tolliver, “Combinatorial pre-conditioners and multilevel solvers for problems in computer vision and image processing,” Computer Vision and Image Understanding, vol. 115, no. 12, pp. 1638–1646, 2011.

[43] D. Goldfarb and W. Yin, “Second-order cone programming methods for total variation-based image restoration,” SIAM J. Sci. Comput, vol. 27, pp. 622–645, 2004.

 

694

 

[44] B. Wohlberg and P. Rodriguez, “An iteratively reweighted norm algorithm for minimization of total variation function-als,” Signal Processing Letters, IEEE, vol. 14, no. 12, pp. 948 –951, dec. 2007.

[45] H. H. Chin, A. Madry, G. L. Miller, and R. Peng, “Runtime guarantees for regression problems,” in Proceedings of the 4th conference on Innovations in Theoretical Computer Science, ser. ITCS ’13. New York, NY, USA: ACM, 2013, pp. 269– 282.

[46] L. N. Trefethen and D. Bau III, Numerical linear algebra. Siam, 1997, vol. 50.

[47] D. Goldfarb and W. Yin, “Second-order cone programming methods for total variation-based image restoration,” SIAM Journal on Scientific Computing, vol. 27, no. 2, pp. 622–645, 2005.

[48] H. H. Chin, A. Madry, G. L. Miller, and R. Peng, “Runtime guarantees for regression problems,” in Proceedings of the 4th conference on Innovations in Theoretical Computer Science. ACM, 2013, pp. 269–282.

 

Functional Partitioning to Optimize End-to-End

Performance on Many-core Architectures

Min Li', Sudharshan S. Vazhkudai2, Ali R. Butt', Fei Meng3, Xiaosong Ma2,3, Youngjae Kim2,

Christian Engelmann2, and Galen Shipman2

'Virginia Tech, 2Oak Ridge National Laboratory, 3North Carolina State University

{limin, butta}@cs.vt.edu, {vazhkudaiss, kimy1, engelmannc, gshipman}@ornl.gov, {fmeng, ma}@cs.ncsu.edu

 

Abstract—

Scaling computations on emerging massive-core super-computers is a daunting task, which coupled with the significantly lagging system I/O capabilities exacerbates applications’ end-to-end performance. The I/O bottleneck often negates potential performance benefits of assigning additional compute cores to an application. In this paper, we address this issue via a novel functional partitioning (FP) runtime environment that allocates cores to specific application tasks — checkpointing, de-duplication, and scientific data format transformation — so that the deluge of cores can be brought to bear on the entire gamut of application activities. The focus is on utilizing the extra cores to support HPC application I/O activities and also leverage solid-state disks in this context. For example, our evaluation shows that dedicating 1 core on an oct-core machine for checkpointing and its assist tasks using FP can improve overall execution time of a FLASH benchmark on 80 and 160 cores by 43.95% and 41.34%, respectively.

I. INTRODUCTION

As growth in processor frequency has stagnated, chip designers have turned to increasing the number of pro¬cessing cores per socket to meet Moore’s law scaling of processor capability. In the near future, each socket may contain 8, 16, or even 80 or more cores, e.g., Intel’s

80-core chip prototype [1]. This push towards increasing peak CPU throughput in High Performance Computing

(HPC) systems is not matched by a similar push towards improving the access bandwidth to other components: sustained I/O bandwidth significantly lags behind pro 

cessor improvements [2]. With many-core processors driving up the per-socket memory and I/O bandwidth

requirements, the “storage wall” problem that has long perplexed designers of parallel computing clusters is now moving to within each compute node.

Consider the current No. 1 machine on the Top500 list [3], the 224,256-core Jaguar petaflop supercomputer.

c 2010 IEEE Personal use of this material is permitted. How¬ever, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

SC10 November 2010, New Orleans, Louisiana, USA 978-1-4244-7558-2/10/$26.00

 

A balanced petaflop system should sustain a bit of I/O operation per instruction, requiring the parallel file system (PFS) to provide 100 TB/s of I/O bandwidth. In reality, the currently used state-of-the-art PFS, Lustre [4], supports a peak I/O bandwidth of 254 GB/s [5] (based on an IOR benchmark), which is two orders of magnitude less than the ideal bandwidth. Furthermore, applications typically only realize a fraction of this peak performance due to software overhead or resource contention.

Simply assigning cores to an application does not scale: End-to-end application performance is not ex-pected to grow linearly with the number of cores [6], mainly due to the bottleneck-prone HPC storage hier-archy and the contention for on-chip resources in the “sea of cores” of modern multicore systems. Adding to this is the overhead from using the needed sophisticated but complex programming techniques both in symmet¬ric [7]–[11] and asymmetric [12], [13] multicores.

To underscore these challenges, we tested two com-mon parallel programs, namely mpiBLAST [14], an I/O-intensive biological sequence alignment application, and FLASH [15], [16], a computation-intensive astrophysics simulation. We executed the applications on a clus¬ter comprising four oct-core machines, using standard scheduling to map processes to cores on the four nodes. We used a 24 GB workload and Sod 3D for mpiBLAST and FLASH, respectively, and studied them with a fixed total problem size, i.e., under strong scaling.

Figure 1 shows the speedup achieved

with increasing

number of cores compared to the case of one core. It is observed that for the selected

workloads, using

increasing number of cores does not provide corresponding improvement in execution time. This behavior can be attributed to: (i) the storage wall as more cores contend for the data

 

from PFS, and (ii) the memory wall as more cores in a node compete for the shared memory.

Functional partitioning of cores: The solution to the compute-I/O imbalance problem must scale with core count. In the post-petascale environment, everything off socket is too far away in terms of “clocks.” Many system resources and tools will need to be present on a compute node and integrated and composed into system-level services at scale. Rather than continuing to assign more of the available cores to computation and intensifying the pressure on the memory and secondary storage systems, some of the cores may serve applications’ overall performance better if they can intercept and absorb part of its data-intensive tasks. Such asymmetric division of labor among cores is not new. There exists I/O libraries that dedicate processors [17], [18] or use separate threads [18], [19] for handling parallel I/O op¬erations. Similarly, BlueGene/L [20] uses distinct cores for compute and networking tasks. However, we propose a functional partitioning (FP) runtime environment as a generalized way to dedicate a subset of cores within a compute node to data processing services to help alleviate the I/O bottleneck. This way, our approach can improve the overall system resource utilization and speed up eventual application turnaround. As a proof-of-concept study, we explore several sample data ser¬vices, such as peer checkpointing, data analytics, and file format conversion. The partitioning techniques we develop can also provide the basis for other activities, such as monitoring, result verification and validation, shadow computation, compression, and encryption.

A suitable platform for demonstrating the usefulness of FP is driven by the observation that future HPC systems are likely to be equipped with non-volatile devices such as solid-state disks (SSDs). SSDs offer excellent read/write throughput (e.g., up to GB/s with PCIe cards) when compared to secondary storage and larger capacity when compared to DRAM (hundreds of GBs). SSDs can be used either as slow memory or a faster cache and posses very desirable properties such as low power consumption and persistence. HPC systems are beginning to explore the use of SSDs in the storage hierarchy (e.g., SDSC’s Gordon [21]) as a means to mitigate the pressure on storage systems. However, there is a lack of a coherent architecture in HPC to use SSDs in a unified fashion and in concert with secondary storage. Further, there is no clear set of guidelines as to where to place SSDs in the supercomputer (e.g., node-local or system nodes). Thus, in this paper, we also investigate the use of SSDs towards addressing the I/O bandwidth bottleneck and illustrate how FP can be used to dedicate cores that are geared towards performing different I/O services on the SSDs. However, the benefits of FP hold irrespective of whether a system employs SSDs or not.

 

A. Contributions

This paper makes the following contributions.

Functional Partitioning (FP) Design Paradigm:We propose a FP runtime environment as a novel generalized approach to partitioning many-core systems. Rather than focusing on raw performance scaling, FP enables the concerted use of the plethora of cores towards an applica-tion’s own diverse set of activities. The runtime lays out an architecture that enables: (a) applications to specify the assist tasks in a job script, (b) an auxiliary application (aux-app) model for the assist tasks so they can co-execute with the applications, and (c) a foundation on which dynamic adaptation of the provisioning of cores to aux-apps can be performed in the future.

Building-blocks for FP: We show, via implemen-tation and experimentation, that dedicating checkpoint and de-duplication cores is in fact a concrete first step towards functionally distinguishing cores.

Aggregate, Intermediate SSD Device: We built an architecture to harness SSD storage, in user space, from individual nodes to create a scalable, aggregate interme¬diate SSD device that can sustain high-speed writes. The device also facilitates diverse data operations potentially offered by our proposed service cores. Our approach is by far the first to propose a concerted use of distributed SSDs in a supercomputer environment.

Evaluation: We undertake a thorough evaluation of the FP approach using a large 160-core testbed, studying the resulting I/O throughput, impact of varying different parameters such as number of dedicated cores, and the overall impact on a real-world application’s performance.

II. RELATED WORK

Resource management in heterogeneous and special¬ized multi-processor systems has gained much research attention recently [22]–[26] with focus on scientific applications. There have also been studies on utiliz¬ing available cores for online execution monitoring or security checking [27], [28]. However, to the best of our knowledge, FP on general-purpose, homogeneous cores has not been studied for mainstream HPC applications. Although there exists I/O libraries that dedicate proces¬sors [17], [18] or use separate threads [18], [19] for handling parallel I/O operations, our FP approach for multicores directly targets the on-chip parallel compu¬tation efficiency problem, and presents a more general and versatile service model for balanced utilization of the increasing number of cores.

Several research efforts have also advocated a pipelined model — that assigns various computational tasks of an application to different cores — for homo¬geneous [29]–[32] and heterogeneous [33]–[36] systems for parallelizing applications. In contrast, FP is a novel runtime environment for core allocation and, in this

 

paper, we highlight its use in servicing I/O and the compute-intensive tasks related to it in order to achieve high overall system throughput.

This paper develops checkpointing and the compute-intensive operations surrounding it as sample services to be provided by dedicated core(s). Our implementation focuses on utilizing aggregated memory/SSD spaces, however, FP is general enough for incorporating other peer-checkpointing schemes, such as [37], [38].

Recently, supercomputers are being equipped with SSDs (e.g., Gordon [21]). Besides illustrating one po-tential mechanism of utilizing node attached SSDs, FP aggressively overlaps SSD accesses with computation, which may help in hiding from the user the SSD per¬formance variance problem revealed by a recent study characterizing scientific I/O workloads on SSDs [39].

III. FP: RATIONALE AND ENVIRONMENT

The advent of multicores implies that data production rates of computer systems are fast surpassing the con¬sumption rates of the associated storage systems, thus creating a fundamental imbalance between the two. As applications try to scale to tens of cores within a single node, the storage-compute performance gap leads to I/O bottleneck. In essence, even though cores are available, they may not yield expected performance benefits.

There are numerous application activities, e.g., check-pointing, file reformatting, etc., in a typical application workflow that can benefit if just a few cores were exclu¬sively allocated for the activities. Prior work in this area has often been relegated to application-specific solutions of running a few support threads on cores. Instead, we create a generic runtime environment for dedicating a portion of the cores allotted to an application towards application activities besides computation.

Partitioning Cores: To handle the aforementioned application activities, we specialize cores by assigning to them specific sets of functionality. The analogy being: just as a large supercomputer has compute, I/O and ser¬vice nodes for different functions, we enable a “system on chip”-like design by partitioning the cores (within a node) based on their functionality, e.g., compute cores, checkpoint cores, shadow-computation cores, verifica¬tion cores, etc. To achieve such a holistic solution, how¬ever, we need a sophisticated runtime environment for many-core systems. We argue that such an outlook brings a novel perspective to current multicore research, the vast majority of which is focused only on raw computational scaling of applications. Instead, we aim to achieve a concerted use of the thousands of cores available to an application on a whole gamut of application activities.

Functionally partitioning cores to conduct assist tasks in-situ, with the application, can be compared against partitioning in time approaches that schedule operations 

 

in a performance-optimal manner. FP offers several advantages in terms of programmability, transparency and simplicity. In contrast, partitioning in time does reduce resource contention, which might be an issue with functionally partitioned cores (e.g., contention for memory resources). However, the storage wall is still a looming issue with the time-partitioning model, which is a critical problem that we strive to alleviate with FP.

A. Runtime Environment for FP

FP is supported through a dynamic runtime component comprising a suite of application support services. We adopt a bottom-up approach and construct several sample data-oriented services, which illustrate the concrete steps in building and enabling application support services. They also serve as proof-of-concept case studies to evaluate the effectiveness of FP.

Realizing the Runtime: The support services are run as “auxiliary applications” (aux-apps) on the dedicated cores (co-located with the main applications). An aux-app monitors its associated main application and trans¬parently performs the support tasks, e.g., an aux-app can be used to create an aggregate distributed storage device (using SSDs) for checkpointing. Moreover, the aux-apps do not have to run on each node; multiple aux-apps can be aggregated on specialized nodes from where they can monitor applications on other nodes.

The first step towards realizing the FP runtime is to provide an interface to the application writers for specifying what aux-apps should run along with the main application. The job submission script is the logical place where such information can be specified. Note that the scripts are used to specify the location of aux-apps as well as to invoke implemented aux-apps, and are not means for implementing the aux-apps.

The runtime utilizes a FUSE [40] enabled driver com¬ponent that exports a mountpoint to allow interactions between the main application and the aux-apps. The driver supports an interface, the aux-app API, through which the aux-apps can be invoked on the data accessed by the application through the mountpoint. The aux-apps are thus implemented as pluggable modules be¬hind the FUSE-based driver component. A number of standard aux-apps, e.g., checkpoint management, etc., are provided by the runtime, and customized aux-apps can be developed by the application programmer using the aux-app API. Consequently, the aux-app operation is transparent to the main application during execution, as it simply accesses data through the mountpoint. This ensures that the design and development of aux-apps is decoupled from that of the main application.

For each aux-app the driver creates a separate thread (or a set of threads), which is then scheduled on the dedicated core(s). The aux-app approach also supports

 

in t aux app write (void * output buffer , in t size){

in t result =SUCCESS ;

//process output in chunks

while( ( chunk = get chunk( & out buffer, size))! = n u l l ){

// compute hash on output buffer chunks

char* hash= sha 1 ( chunk ) ;

//write the new chunk

if (!hash t a b l e get(hash) )

result= da ta write(chunk) ;

// update dedu p hasht ab l e

hash ta ble update( & r e s u l t, chunk, hash ) ;

}

return result;

}

Fig. 2. Example write extension for a checkpointing aux-app.

advanced usage scenarios, e.g., aux-apps from different nodes can work together to provide I/O aggregation across nodes for reducing load on secondary storage.

Consider an aux-app for de-duplicating checkpoint data. The user specifies the aux-app in the job submis¬sion script, which is then executed as a thread by the FUSE-based driver component on application execution. When the application writes to our mountpoint, the de-duplication function (Figure 2) is invoked on the data. Note that although only the write function is shown, all I/O functions supported by the FUSE API are supported by the aux-app API.

Discussion: Our current implementation of the run-time supports a static partitioning of cores, which means that core allocation to aux-apps cannot change during an application run. Such dynamic re-allocation can be useful, as we illustrate later in our evaluation, if an aux-app can benefit from more cores or if an application can make sufficient progress without an additional core. The need for such flexibility is also warranted from the usability aspect of the runtime. For example, how will a user know what is an optimal partitioning? The ideal solution would be to start with a conservative partition and then to let the runtime provision the al¬location based on an agreed upon progress metric. The advantage of the aux-app approach is that it provides the basis upon which such a dynamic provisioning of cores can be built. While the number of cores used by the compute component of an application is typically fixed, the functionally partitioned cores can be used to support multiple services as required to improve the application performance. For example, if two cores are available for running the aux-apps, dynamic provisioning may use one core for de-duplication and data compression and another for checkpointing, instead of using both for checkpointing. Or it may even run an additional format transformation service that shares the available cores with other services. The goal being to improve overall application performance. Such a dynamic approach will allow the aux-apps to adjust to the application demands 

 

while keeping the performance impact to a minimum. In this paper we present and evaluate static FP, and identify the need for future work on dynamic provisioning.

Another aspect that impacts the utility of FP is whether an application is designed to utilize all cores available to it, i.e., no dedicated cores are available for aux-apps. We argue that application writers should examine end-to-end performance, i.e., all tasks and not just compute, and decide what is best for their applica¬tions. Nonetheless, based on our experience with user allocation requests, such a scenario is highly unlikely as users, without exception, over-provision cores, which can then be used by aux-apps.

In summary, the runtime provides for flexibly synthe¬sizing application support services and dynamically us¬ing the allocated cores to improve end-to-end application performance, and not just raw compute performance.

IV. SAMPLE CORE SERVICES

In this section, we present several sample services that can be performed by dedicated cores. Note that a single service may be carried out by multiple cores if necessary, conversely, multiple services can be assigned to a single core. We illustrate dedicating cores to application tasks using a checkpointing service. Checkpointing and its associated tasks such as data draining, de-duplication, and format transformation provide an insightful case study for the application I/O activities that can be expedited using the functional partitioning runtime. We demonstrate the benefits of performing each of these operations in-situ by the aux-apps. This is in contrast to the extant approach of conducting such operations in an offline manner, which exacerbates the storage wall issue due to the constant writing and re-reading of TBs of data from secondary storage.

A. SSD-based Checkpointing

Why Checkpointing? Checkpointing is an important data operation routinely performed by parallel appli¬cations, both for fault tolerance and for user-initiated execution restart. Checkpointing is becoming increas¬ingly expensive relative to computation, especially for large-scale jobs, and is a key consideration in design¬ing supercomputers. For example, the Argonne Intrepid BG/P supercomputer was designed not to meet balanced machine criteria (a bit of I/O per second per instruction per second), but rather to be able to dump the contents of the entire system memory to secondary storage in 30 minutes [41]. This makes checkpointing an appealing candidate for being “outsourced” to spare cores that cannot further help towards improving the end-to-end application execution time.

Rather than simply handing the periodic checkpoint-ing I/O to dedicated cores to store in their associated

 

memory (an approach exploited previously on dedicated processors within an SMP box [18]), in this paper we explore checkpointing using non-volatile memories. One can argue that checkpointing is I/O intensive and, therefore, does not benefit from a dedicated checkpoint core. We counter this hypothesis with the observation that checkpointing can benefit from a host of other operations, such as de-duplication and compression (both of which are compute-intensive), draining and format transformation, which when performed in-situ through FP can alleviate the storage wall problem mentioned earlier. Consequently, the checkpointing service core, in this work, can perform more sophisticated tasks beyond just performing background I/O from the main memory to PFS, which is what most applications are currently faced with and stymied by. Another question that arises in this context is whether the task that would have been performed on the dedicated core can be run on the same core as the application. This is not possible even when the application is performing I/O, unless the application is modified to run a helper thread that is pinned to another spare core. However, Leadership machines, such as Jaguar, are not designed to be time-sharing systems and an application’s allocation of cores is its own for the entirety of the run. For these reasons, the checkpointing service can benefit from the FP runtime.

Why SSDs? Among the several types of non-volatile memories, flash memory-based SSDs are gaining popu¬larity for persistent data storage. SSDs offer a number of benefits over the conventional mechanical disks, such as fast data access time, low power consumption, light weight, higher resilience to external shocks and high temperatures. SSDs become especially helpful as the system memory bandwidth, as a function of computation throughput (byte/FLOP), has been consistently dropping in the Top500 supercomputers over the past decade. In fact, the ratio for a 2018 exascale machine is projected to be 0.01, which has been explicitly identified by DOE as one of the chief exascale problems to be addressed [42]. Furthermore, the non-volatile nature of SSDs can help provide intermediate storage for checkpointing data and reduce disk I/O (and communication) load.

Sustainability of SSD-based checkpointing: To justify the use of SSDs for HPC checkpointing — a write-intensive and write-once workload — one must address two inter-dependent concerns of cost and durability. This is because there is a significant $/GB difference between current SSDs and hard disks, and the number of erase cycles supported on the SSDs is fixed (limited compared to the hard disks). To check whether it is feasible to use SSDs for HPC checkpointing, we examine production run checkpointing characteristics of seven leadership-class DOE applications on the ORNL Jaguar machine (Table I). At the time of the runs, Jaguar had four quad 

 

TABLE I

CHECKPOINT SIZES AND ESTIMATED NODES/SS D1.

Appl.

Data size (MB/Core) Data size (MB/Node) # Nodes/SSD

C1 C2

GTC 180 2880 21 11

XGC1 120 920 31 16

GTS 220 3520 17 9

Chimera 10 160 380 200

S3D 14 224 271 142

GEM 20 320 190 100

M3D-k 14 224 271 142

(1)

• C2 – capacity: To allow lazy draining of data to secondary storage, each SSD should be large enough to hold one complete checkpoint from the nodes sharing the device.

Table I shows the checkpoint characteristics for the applications considered, as well as the estimated number of nodes that can share a single SSD under C1 and C2. There are two key observations that can be made from the table. First, the lifetime is not the limiting constraint for sharing an SSD, rather the limit is determined by the SSD capacity. Second, even with both the constraints, a large number of nodes can share a single SSD. This is promising, as it indicates that an SSD-based checkpoint-ing solution can be economically feasible.

Checkpointing Architecture: In our design, compute nodes contribute one or more cores and their associated SSDs as checkpoint cores to construct an aggregate, distributed checkpoint device. Each checkpoint core runs a benefactor process that contributes available, node-local SSD space (or a partition of it) to a manager process (running on one of the participating nodes) that aggregates such distributed SSD spaces and presents a

1We thank Scott Klasky for providing us with application data sizes.

 

collective intermediate storage device to checkpointing clients. Management tasks (such as benefactor status monitoring, space mapping, and data striping) can be done in a similar way as in existing storage aggregation systems [46]–[48]. For each checkpoint, the manager also maintains a striping map that contains information about where all the different parts (chunks) of the checkpoint are stored.

The aggregate SSD storage is made available to clients via a transparent file system mount point, /AggregateSS-Dstore, using FUSE [40] as discussed in Section III. Here, we leverage our prior work on mounting an aggregate storage of node-local disks [48]. An appli¬cation core that checkpoints data to the mount point will be redirected to the aggregate SSD storage, without requiring any other code modification.

While a single compute node’s local memory is likely to be much smaller than its local SSD, checkpointing to aggregated SSD space from multiple nodes has several advantages. First, it provides fault tolerance in the event of compute node failure, which may render the persistent SSD-resident checkpoint data inaccessible. The globally accessible aggregate storage space also facilitates easy replication, e.g., using a simple copy in the aggregate space, of the individual node’s checkpoint (or chunks of striped checkpoints) across multiple nodes, which would otherwise be complex, visible to the application, and cumbersome if nodes managed their associated SSDs in¬dividually. Second, when the SSDs are distributed across a set of system nodes, aggregation and access through a file system mount point offers an elegant abstraction to transparently access them from the numerous compute nodes, thus decoupling the placement of SSDs from the compute nodes and allowing for sharing of SSDs across multiple nodes. Finally, although we expect the growth in memory sizes to be matched with proportional growth in SSD space on all nodes, even if that is not the case and there is an imbalance, this abstraction allows for data to be striped over to other node-local SSDs. For example, the 512-core DASH system [49] at SDSC (the precursor to 8192-core Gordon system) is equipped with 4 TB of flash storage, compared to its 3 TB of DRAM. Currently, a high-end Fusion I/O PCIe MLC SSD card (io Drive Duo) at 640 GB is priced around $15K. Much like disk storage, SSD storage is increasing in capacity and decreasing in cost. Thus, growth in SSD space is currently outpacing memory increases.

B. De-duplication

Another sample data service function is de-duplication, which is used to identify and store unique data copies. For HPC applications, this service is useful when used in conjunction with checkpointing, to de¬tect the similarity between two successive checkpoint 

 

images and only store their dissimilar parts. Such in-cremental checkpointing techniques have been explored earlier [48], [50]. The challenge in doing de-duplication on dedicated cores simultaneously with the main com¬putation is to avoid significant memory contention.

To this end, we have built a service that computes hashes of the checkpoint data and identifies and removes duplicates. A dedicated de-duplication core (a check¬point core can also double as a de-duplication core) is assigned the task of computing the chunk hashes that are then stored as metadata for that particular dataset at the manager (in checkpoint architecture above). When the checkpoint image for the next timestep, t, is to be written, the chunk hashes from (t  1) are compared against the new, incoming image. A matching chunk hash indicates a duplicate and the chunk is not written, only the checkpoint’s striping map is updated to point to the previously stored chunk. Consequently, depending on the degree of similarity between two successive checkpoints, the size of the checkpoint data written and the time to write it can be significantly reduced.

C. Format Transformation

Another potential data service for using dedicated cores is file format transformation. Large-scale parallel scientific simulations (and subsequent analysis/visualiza-tion tools processing their computation results) do not read/write data in plain binary formats. Instead, they often use high-level I/O libraries to create and access data in special scientific data formats. Intermediate checkpoint snapshot data is also saved in a specific data format so that it can be used by applications as a restart file in case of failure. Well-adopted formats, such as HDF5 [51] and netCDF [52], produce self-explanatory and self-contained files, with support for binary porta¬bility. However, accessing these files, especially through their parallel interfaces, has been substantially slower than reading/writing binary files [18], [53]. Checkpoint-ing, while already cumbersome due to the storage wall, is often further stymied due to the need of being in proper scientific data formats. Recently, researchers have exploited dedicated data service nodes to form staging areas, where output data can be dumped in internal, faster formats, then asynchronously converted to HDF5 files on hard disks, producing a significant I/O performance improvement [54]. With dedicated cores, similar format transformation can be performed, especially with the SSD-based storage layer. As existing intermediate file formats (such as BP [53]) dump data in quite manageable units, the format transformation core can easily perform the conversion in a streaming manner. This reduces the memory requirement and performance perturbation to the computation and other concurrent data tasks running on the other cores. It helps to perform such format

 

conversions while the data is in transit (either in memory or SSD) and has not yet reached secondary storage. Further, performing conversion operations on each core within the compute node makes the data deluge more manageable. If the entire checkpoint or result snapshot is written as binary data to disk and format conversion is performed offline, the entire workflow suffers from con¬stant re-reading of data. Offline format conversion also means that if a failure occurred right after a checkpoint, a valid restart file may not be ready yet, which wastes significant resources and delays job turnaround.

D. Adaptive Checkpoint Data Draining

Although an SSD can store data persistently, and its capacity will typically be manifold compared to node-local memory, large-scale, long-running jobs can gener¬ate overwhelming volumes of data that results in space on the SSD running out. This is especially true when not every node has an SSD attached to it, instead only a select set of system nodes has SSDs (due to budget concerns for example). Fortunately, checkpointing for fault tolerance does not require keeping all checkpoints: typically files are overwritten and saving up to two most recent checkpoints is enough. However, writing checkpoints to secondary storage supported via a parallel file system may still be needed: for some applications, checkpoint data doubles as result data for future analy-sis/visualization, or needs to be saved for elective restart. Even though draining is I/O bound, it cannot be done offline as the aggregate SSD space needs to be vacated for future checkpoint data. Growing memory size and the resulting increasing checkpoint size further stresses the need for in-situ data draining. With a checkpointing core, issues arising from growing memory sizes can be mitigated, and draining from SSDs to the secondary storage can be done in flexible and intelligent ways. As draining is I/O bound, it can be overlapped with other CPU bound checkpoint assist tasks.

The checkpointing core may decide to drain once every k checkpoints, in addition to maintaining the two most recent ones. The parameter k may even be con¬figured and coordinated at runtime, through additional monitoring functions performed by the checkpointing core (such as watching the client checkpoint frequency). When the compute cores are back in the next compu¬tation phase, the checkpoint cores can collectively and lazily drain selected checkpoints to secondary storage. To enable this, the runtime supplies the aux-apps on different nodes with the location of the manager process. The system uses a soft-state protocol, where the aux-apps periodically announce their availability and sharing preferences, e.g., available SSD space, to the manager using keep alive updates. This not only allows the aux-apps to locate and communicate with each other, but also

 

Fig. 3. High level FP system architecture.

provides them with flexibility to change their preferences over time. The manager uses this information to instruct the aux-apps about the secondary storage location to where the checkpoints from the SSDs can be drained.

V. IMPLEMENTATION

As a proof-of-concept, we have used the FP runtime to implement SSD-based checkpointing and various sup¬port services discussed in the previous section, using about 22.1 K lines of C code.

Figure 3 shows the components of our software that is run at the manager, benefactors, and clients. Note that every client node can also be a benefactor if it decides to provide its associated SSD to the aggregate SSD store. An example distribution of cores is also shown, where the white cores are used for computation and the shaded cores run aux-apps for services such as data draining and de-duplication. The manager (running on a separate node) works with the compute nodes to create a virtual aggregated SSD store, which serves as a transparent interface to the distributed SSDs. It also supports lazy draining to the very large but slower secondary disk-based storage. The aux-apps coordinate with each other across nodes using socket communication, and remain transparent to the main application.

The sequence of events when an application check-points is as follows. Upon receiving a request to write the checkpoint data, the FUSE module invokes the client component, which interacts with the manager to deter¬mine the benefactor that will handle the checkpoint for the client. If the client is on a node that has an associated SSD it is given preference and is utilized. The exception to this is if the local SSD is out of storage space, when a remote benefactor is chosen. Regardless of whether the client and benefactor are on the same node, the client directly contacts the benefactor to determine its current availability, divides the checkpoint data into fixed-size chunks, and transfers the chunks to the benefactor. The

 

TABLE II

TESTBED CONFIGURATION.

# of processing nodes 20

Capacity of storage server Network Interconnect 2 TB

Infiniband QD 40 Gbit/s

HDD model

Bandwidth

Capacity WD3200AAJS SATAII

85 MB/s

320 GB

Cores per node

Memory per node

Max. cores available 8

8 GB

160

TABLE III

SSD SPECIFICATIONS OF INTEL X 2 5-E [56].

Model Intel X25-E Extreme

Features

Capacity

Bandwidth

I/O Per Second SATA-II SLC Flash Technology

32GB

Sequential read: 250 MB/s

Sequential write: 175 MB/s

Random 4KB reads: >35K IOPS

Random 4KB writes: >3.3K IOPS

benefactor stores the data on its associated SSD. Asyn¬chronously, as discussed in Section IV the benefactor may drain data from the SSD to the secondary storage system. Once the checkpoint is complete, the benefactors inform the manager. The manager can then also invoke a merge component on the benefactors, which reads the checkpoint chunks from the secondary storage system and rearranges them into a merged checkpoint file, ready to be used by standard restart mechanisms if needed.

Finally, we have also built the checkpoint data ma-nipulation services as discussed in Section IV, such as basic data draining, replication, and de-duplication.

VI. EVALUATION

In this section, we evaluate our implementation of FP and study its impact on application performance.

A. Methodology

Testbed Setup: Table II shows the configuration of our testbed, which uses 20 nodes from the systemG machine at Virginia Tech. All of the participating nodes are identical and run Linux Kernel 2.6.27.10. Each node is also equipped with an emulated SSD that has been validated against a real product (Table III) for sequential I/O throughputs within an error margin of 0.57%. The device uses DRAM for storage and emulates a real SSD by introducing artificial delays [55].

Our setup is not equipped with a PFS, so we used node-local disks for checkpoint data. While typical HPC setups do not employ node-local disks, we use them as a high-throughput alternative to an NFS server.

Workloads: We employ a real-world astrophysics simulation code, FLASH [15], [16], which generates checkpoint files in HDF5. We modified the Sod 3D version of FLASH for our evaluation: all the compute 

 

processes carry out parallel I/O, including checkpointing, using MPI-IO. The problem size remains fixed as the number of compute processes is increased. For more de¬tailed testing, we also use a synthetic benchmark, which is a simple checkpoint application that generates same sized checkpoint data every barrier step. Specifically, we created an MPI program with 160 processes, each writing 0.25 GB of data per checkpoint, thus creating a total checkpoint of 40 GB per barrier step. Finally, we use static functional partitioning for the experiments.

B. Impact of FP

In our first set of experiments, we determine the impact of FP on overall application performance. We use the notation, FP(X, Y), to denote a setup with a total of Y cores per node of which X have been functionally partitioned for support services. For this test, we use FLASH with a checkpoint size of 6.8 GB. No format transformation is performed on the checkpoint data. We consider four cases. (i) Local disk non-FP(0,8): The baseline performance where all 8 cores per node are used for application computation. Checkpoint data from all the 8 cores in every node is written to the local disk on that compute node. (ii) Local disk non-FP(0,7): Repeat (i), but with only 7 cores per node for application computation and the remaining core is left idle. All the 7 cores per node write to the local disk on that compute node. (iii) Aggregate disk FP(1,8): FP where 1 core out of 8 on each node is used as a dedicated checkpoint core. An aux-app is run on these dedicated cores, which assists with checkpointing and its associated tasks. In this case, FP allows us to build sophisticated structures, such as an aggregate distributed store of node-local disks by pooling the aux-app services on each checkpoint core as explained in Sections IV and V. The 7 cores per node checkpoint to this aggregated storage, which stripes the data in parallel to distributed aux-app services. (iv) Aggregate SSD FP(1,8): Similar to (iii), but with checkpointing to aggregate SSD storage.

Figure 4 shows the result for 80 and 160 cores, under strong scaling. First, we observe that removing a core from the computation, local disk non-FP(0,7), does not affect the overall performance significantly; in fact the 2.23% average difference between that and local disk non-FP(0,8) is within the error margin. Note that the small increase in execution time from 80 to 160 cores is due to contention in our testbed. Moreover, it can be observed that dedicating one core to handle the check¬point, aggregate disk FP(1,8), can improve the execution time by 15.42% and 27.05% for 80 and 160 cores, respectively. The benefit is also in part due to the ability to write to an aggregate store of node-local disks, pooled from the aux-apps on the dedicated checkpoint cores. Thus, FP is a viable approach and it also lets us build rich

 

80 160 0 20 40 60 80 100 120 5 10 15 20 25 30 35 40 45

Number of Total Cores Number of Compute Cores Number of Benefactors

Fig. 4. Impact of FP on execution time. Fig. 5. Memory vs. SSD I/O rate. Fig. 6. Impact of varying benefactors.

 

composite services that are much beneficial. Moreover, the use of aggregate SSD FP(1,8) provides an additional gain of 43.95% and 41.34% compared to checkpointing on disks for 80 and 160 cores, respectively. This is promising as SSDs can enable efficient checkpointing in HPC setups that do not have node-local disks.

C. Checkpointing to Memory versus SSD Storage

In the next set of experiments, we compare SSD-based checkpointing with our previous work on in-memory checkpointing that aggregates memory buffers across nodes and employs triple-buffering to provide improved throughput. The workload comprises of a synthetic benchmark with up to 125 clients, each writing a checkpoint of 0.25 GB. For SSD-based checkpointing, each node has an SSD, thus the number of aux-apps (run as benefactors) for this case is 20, “SSD(20)”. The number of benefactors for in-memory triple buffer is set to 35, “Triple Buffer(35)”, which is large enough to avoid draining the data to secondary storage, and gives the best case performance of in-memory checkpointing.

Figure 5 shows the results. As the number of clients increases, the benefactor utilization increases and so does the sustained throughput. We also observe that I/O throughput for checkpointing to an aggregate SSD device is 14.7% lower compared to the in-memory technique, which is obvious. Nonetheless, our SSD-based check-pointing does not require the dedicated cores to give up their memory, which may reduce application perfor¬mance, as can happen for in-memory checkpointing.

Next, we determine the effect of the number of bene¬factors on sustained checkpoint I/O throughput. We fix the client cores to 120 and vary the benefactors from 5 to 40. We perform in-memory checkpointing. Figure 6 shows that the I/O throughput does not increase beyond 25 benefactors and there is even a slight decrease, implying the futility of simply adding more benefactors as it takes cores away from computation.

1) Varying number of application processes: The next experiment limits the number of in-memory benefactors and compares its throughput against checkpointing to SSD. We repeat the experiment of SSD(20) and compare against in-memory triple buffering, but with only 20

 

Fig. 7. Impact of insufficient benefactors.

benefactors, “Triple Buffer(20)”. We use a maximum of 140 compute cores. Figure 7 shows that the initial performance of Triple Buffer(20) is similar to Triple Buffer(35). However, as the client cores are increased beyond 80, the benefactor buffers fill up faster than they can be drained, causing a significant throughput reduction. As the compute cores are further increased, the bandwidth eventually reduces to almost that of a direct checkpoint to disk, as the draining agent has now become I/O bound. Note that the steps pattern on the tail of the graph is an artifact of how clients and benefactors are distributed in our test. In contrast, SSD-based checkpointing achieves better throughput with increasing number of clients simply due to larger available space (32 GB SSD versus 1 GB memory aggregated).

2) Varying number of SSD benefactors: In this experi¬ment, we study the impact of varying the number of SSD benefactors. Figure 8 shows the results. As expected, more SSD benefactors result in better I/O throughput. However, once a sufficient number of benefactors were available, i.e., >= 5, the overall throughput did not change much. A remarkable coincidence is that I/O throughput is limited by the available bandwidth of SSDs, and not the number of SSD benefactors (beyond a certain number, i.e. 5 in this case). Conversely, unless the proper number of SSD benefactors is available, the checkpoint nodes can be a bottleneck for the entire system performance in the worst case.

D. De-duplication of Checkpointing Data

We have seen the benefit of FP on reducing overall execution time. In the following, we observe how of 

 

Fig. 8. I/O rates with ’N’ benefactors.

 

Fig. 9. Checkpoint I/O rates with FP(1,8).

 

Fig. 10. De-dup data size with FP(1,8).

 

 

Fig. 11. Impact of increasing FP cores normalized to FP(1,8).

floading different application activities to dedicated cores can further improve overall application performance.

For this experiment, in particular, we consider the benefits of de-duplicating across consecutive checkpoints in reducing the amount of data that needs to be written. We use our synthetic benchmark program with one core on each node dedicated to de-duplication and check-pointing, and vary the number of compute cores from 20 to 140. We also observe the impact of varying the amount of de-duplication. Figure 9 shows the effective I/O rates for the checkpoint for different de-dup ratios (shown in parenthesis). Observe that although de-duplication reduces the amount of data that needs to be transferred, the I/O rates achieved for writing the checkpoint on average across the considered de-dup ratios are 30% lower than that without de-duplication. This is because of the compute-intensive nature of the de-duplication pro¬cess and indicates that the service is under-provisioned (sharing one core with checkpointing). Figure 10 shows the amount of data transferred for a de-dup ratio of 0.25. As expected, de-duplication reduces the amount of data that needs to be written to the disk by 25% or as much as 8208 MB for 140 cores. Such decrease in the amount of checkpoint data to be written (which is all write I/Os) can help improve the lifetime of SSDs used for intermediate aggregated storage.

In the next experiment, we allocated two service cores, one bound for de-duplication and the other for checkpointing. The core services can be overlapped. Figure 11 shows the benefits of allocating more cores to support services by sacrificing computation resources. 

 

Overall, we see that with functional partitioning, FP(2,8) can provide higher throughput than FP(1,8). For instance, the I/O throughput is improved by about 60% with a de-duplication ratio of 0.25, when we use 120 cores for application computation. We speculate that these benefits are mainly obtained from pipelined effect of two cores executing in parallel. Finally, this experiment illustrates the need for dynamic and autonomic service core allocation and resource provisioning, as discussed earlier. We plan to pursue this as part of our future work.

VII. CONCLUSION

We have discussed FP of cores in large multicore systems to support different application activities, in contrast to the extant approach to allocating all cores to computation. We have applied FP to the critical problem of handling checkpoint I/O in supercomputers, where the large number of cores can result in a significant amount of application execution time spent in checkpointing. We have developed a flexible SSD-based checkpointing system that allows for transparent sharing of SSDs across different nodes, thus providing an economically viable solution. Our evaluation using a real implementation shows that our core allocation model is viable, and can provide significant benefits with minimal impact and even increase overall performance (dedicating 1 core on an oct-core machine for checkpointing can improve overall execution time of a FLASH benchmark on 80 and 160 cores by 43.95% and 41.34%, respectively). In summary, our work demonstrates FP’s usability and in our future work, we will apply such partitioning to support other mission-critical application activities.

ACKNOWLEDGMENT

We are thankful to the anonymous reviewers and our shepherd, Dr. Toni Cortes, for their valuable feedback. We also thank Dr. John Cobb for several useful discus¬sions. This work was sponsored in part by the LDRD program of ORNL, managed by UT-Battelle, LLC for the U.S. DOE (Contract No. DE-AC05-00OR22725), and by NSF grants CCF-0937827, CCF-0746832, CCF-0621470, and CCF-0937690, as well as Xiaosong Ma’s joint appointment between ORNL and NCSU.

 

REFERENCES

[1] Intel. Advancing Multi-Core Technology into the Tera-scale Era, 2009. http://techresearch.intel.com/articles/Tera-Scale/1449.htm.

[2] Henry Monti, Ali R. Butt, and Sudharshan S. Vazhkudai. Timely offloading of result-data in hpc centers. In Proc. ACM ICS, 2008.

[3] Top500 supercomputer sites. http://www.top500.org/.

[4] Philip Schwan. Lustre: Building a File System for 1,000-node Clusters. In Proc. Ottawa Linux Symposium, 2003.

[5] Galen Shipman, Dave Dillow, Sarp Oral, and Feiyi Wang. The spider center wide file system: From concept to reality. In Proc. Cray User Group, 2009.

[6] Ami Marowka. Parallel computing on any desktop. Commun. ACM, 50(9):74–78, 2007.

[7] Herb Sutter. The free lunch is over: A fundamental turn toward concurrency in software. Dr. Dobb’s Journal, 30(3):202–210, 2005.

[8] Nalini Vasudevan and Stephen A. Edwards. Celling shim: com-piling deterministic concurrency to a heterogeneous multicore. In Proc. ACM SAC, 2009.

[9] Jack Dongarra. The impact of multicore on math software and exploiting single precision computing to obtain double precision results. In Proc. ICPP, 2006.

[10] Silas B. Wickizer, Haibo Chen, Rong Chen, Yandong Mao, Frans Kaashoek, Robert Morris, Aleksey Pesterev, Lex Stein, Ming Wu, Yuehua Dai, Yang Zhang, and Zheng Zhang. Corey: An operating system for many cores. In Proc.Usenix OSDI, 2008.

[11] Herb Sutter and James Larus. Software and the concurrency revolution. ACM Queue, 3(7):54–62, 2005.

[12] Duc Vianney, Gad Haber, Andre Heilper, and Marcel Zal-manovici. Performance analysis and visualization tools for cell/b.e. multicore environment. In Proc. ACM IFMT, 2008.

[13] Sain-Zee Ueng, Melvin Lathara, Sara S. Baghsorkhi, and Wen-Mei W. Hwu. Cuda-lite: Reducing gpu programming complexity. In Proc. LCPC, 2008.

[14] Aaron E. Darling, Lucas Carey, and Wu-chun Feng. The design, implementation, and evaluation of mpiblast. In Proc. ClusterWorld, 2003.

[15] Anshu Dubey, Katie Antypas, Murali K. Ganapathy, Lynn B. Reid, Katherine Riley, Daniel J. Sheeler, Andrew Siegel, and Klaus Weide. Extensible component-based architecture for Flash, a massively parallel, multiphysics simulation code. Parallel Computing, 35(10-11):512–522, 2009.

[16] Robert Rosner, Alan Calder, Jonathan Dursi, Bruce Fryxell, Donald Q. Lamb, Jens C. Niemeyer, Kevin Olson, Paul Ricker, Frank X. Timmes, James W. Truran, Henry Tufo, Yuan-Nan Young, Michael Zingale, Ewing Lusk, and Rick Stevens. Flash code: Studying astrophysical thermonuclear flashes. Computing in Science and Engineering (CSE), 2(2):33–41, 2000.

[17] K. E. Seamons, Y. Chen, P. Jones, J. Jozwiak, and M. Winslett. Server-directed collective I/O in Panda. In Proc. SC, 1995.

[18] X. Ma, J. Lee, and M. Winslett. High-level buffering for hiding periodic output cost in scientific simulations. IEEE Transactions on Parallel and Distributed Systems, 17(3):193–204, 2006.

[19] S. More, A. Choudhary, I. Foster, and M. Q. Xu. MTIO: a multi-threaded parallel I/O system. In Proc. International Parallel Processing Symposium, 1997.

[20] A. Gara, M. A. Blumrich, D. Chen, G. L.-T. Chiu, P. Coteus,

M. E. Giampapa, R. A. Haring, P. Heidelberger, D. Hoenicke, G. V. Kopcsay, T. A. Liebsch, M. Ohmacht, B. D. Steinmacher-Burow, T. Takken, and P. Vranas. Overview of the blue gene/l system architecture. IBM Journal of Research and Development, 49(2/3):195–212, 2005.

[21] Supercomputer uses flash to solve data-intensive problems 10 times faster. http://www.sdsc.edu/News%20Items/PR110409 gordon.html, 2009.

[22] Christos D. Antonopoulos, Filip Blagojevic, Andrey N. Chernikov, Nikos P. Chrisochoides, and Dimitrios S. Nikolopou-los. A multigrain delaunay mesh generation method for multicore smt-based architectures. J. Parallel Distrib. Comput., 69(7):589– 600, 2009.

 

[23] Ligang He, Stephen A. Jarvis, Daniel P. Spooner, and Graham R. Nudd. Dynamic scheduling of parallel real-time jobs by mod-elling spare capabilities in heterogeneous clusters. In Proc. IEEE ICCC, 2003.

[24] Jorge Manuel Gomes Barbosa and Belmiro Daniel Rodrigues Moreira. Dynamic job scheduling on heterogeneous clusters. In Proc. IEEE ISPDC, 2009.

[25] M. Mustafa Rafique, Ali R. Butt, and Dimitrios S. Nikolopoulos. Designing accelerator-based distributed systems for high perfor-mance. In Proc. IEEE/ACM CCGrid, 2010.

[26] M. Mustafa Rafique, Benjamin Rose, Ali R. Butt, and Dim-itrios S. Nikolopoulos. Supporting mapreduce on large-scale asymmetric multi-core clusters. SIGOPS Oper. Syst. Rev., 43(2):25–34, 2009.

[27] Shimin Chen, Babak Falsafi, Phillip B. Gibbons, Michael Kozuch, Todd C. Mowry, Radu Teodorescu, Anastassia Ailamaki, Limor Fix, Gregory R. Ganger, Bin Lin, and Steven W. Schlosser. Log-based architectures for general-purpose monitoring of de¬ployed code. In Proc. Architectural and System Support for Improving Software Dependability Workshop, 2006.

[28] Edmund B. Nightingale, Daniel Peek, Peter M. Chen, and Jason Flinn. Parallelizing security checks on commodity hardware. In Proc. ACM ASPLOS, 2008.

[29] Kue-Hwan Sihn, Baik Hyunki, Kim Jong-Tae, Bae Sehyun, and Song Hyo Jung. Novel approaches to parallel h.264 decoder on symmetric multicore systems. In Proc. IEEE ICASSP, 2009.

[30] S. Arash Ostadzadeh, Roel J. Meeuws, Kamana Sigdel, and Koen Bertels. A multipurpose clustering algorithm for task partitioning in multicore reconfigurable systems. In Proc. IEEE CISIS, 2009.

[31] Turgay Altilar and Yakup Paker. Minimum overhead data partitioning algorithms for parallel video processing. In Proc. International Conference on Domain Decomposition Methods, 2001.

[32] Robert Ennals, Sharp, and Mycroft. Task partitioning for multi-core network processors. In Proc. Cluster Computing, 2005.

[33] Marc de Kruijf and Karthikeyan Sankaralingam. MapReduce for the Cell B.E. Architecture. IBM Journal of Research and Development, 53(5):10:1–10:12, 2009.

[34] Bingsheng He, Wenbin Fang, Qiong Luo, Naga K. Govindaraju, and Tuyong Wang. Mars: a mapreduce framework on graphics processors. In Proc. ACM PACT, 2008.

[35] F.T. Hady, T. Bock, M. Cabot, J. Chu, J. Meinecke, K. Oliver, and W. Talarek. Platform level support for high throughput edge applications: the twin cities prototype. IEEE Network, 17(4):22– 27, 2003.

[36] Sanjay Kumar, Gavrilovska, Karsten Schwan, and Srikanth Sun-daragopalan. C-core: Using communication cores for high performance network services. In Proc. IEEE NCA, 2005.

[37] J. S. Plank, K. Li, and M. A. Puening. Diskless Checkpoint-ing. IEEE Transactions on Parallel and Distributed Systems, 9(10):972–986, 1998.

[38] G. Bronevetsky and A. Moody. Scalable I/O Systems via Node-Local Storage: Approaching 1 TB/sec File I/O. LLNL Technical Report LLNL-TR-415791, Lawrence Livermore National Labo-ratory, 2009.

[39] S. Park and K. Shen. A Performance Evaluation of Scientific I/O Workloads on Flash-Based SSDs. In Proc. Workshop IASDS, 2009.

[40] FUSE. File System in Userspace, 2007. http://fuse.sourceforge. net/.

[41] Kamil Iskra, John W. Romein, Kazutomo Yoshii, and Peter H. Beckman. Zoid: I/o-forwarding infrastructure for petascale ar-chitectures. In Proc. ACM SIGPLAN PPoPP, 2008.

[42] International exascale software project roadmap. In Proc. Cross-cutting Technologies for Computing at the Exascale Workshop, 2010.

[43] B. Schroeder and G. A. Gibson. Disk failures in the real world: What does an mttf of 1,000,000 hours mean to you? In Proc. USENIX FAST, 2007.

[44] Feng Chen, David A. Koufaty, and Xiaodong Zhang. Under-standing intrinsic characteristics and system implications of flash memory based solid state drives. In Proc. SIGMETRICS/Perfor-mance, 2009.

 

[45] Numonyx. Wear leveling in nand flash memories. http://www. numonyx.com/Documents/Application%20Notes/AN1822.pdf.

[46] Ali R. Butt, Troy A. Johnson, Yili Zheng, and Y. Charlie Hu. Kosha: A peer-to-peer enhancement for the network file system. Journal of Grid Computing: Special issue on Global and Peer-to-Peer Computing, 4(3):323–341, 2006.

[47] S. Vazhkudai, X. Ma, V. Freeh, J. Strickland, N. Tammineedi, and S. Scott. Freeloader: Scavenging desktop storage resources for bulk, transient data. In Proc. SC, 2005.

[48] Samer Al-Kiswany, Matei Ripeanu, Sudharshan S. Vazhkudai, and Abdullah Gharaibeh. stdchk: A checkpoint storage system for desktop grid computing. In Proc. ICDCS, 2008.

[49] Dash User Guide: Technical Summary, June 2010. http://www. sdsc.edu/us/resources/dash/index.html.

[50] Eliezer Levy, Avi Silberschatz, and Avi Silberschatz. Incremental recovery in main memory database systems. IEEE Transactions 

 

on Knowledge and Data Engineering, 4(6):529–540, 1992.

[51] http://hdf.ncsa.uiuc.edu/HDF5/doc/. HDF5 - A New Generation of HDF.

[52] http://www.unidata.ucar.edu/packages/netcdf/docs.html. NetCDF Documentation.

[53] Jay Lofstead, Fang Zheng, Scott Klasky, and Karsten Schwan. Adaptable, metadata rich io methods for portable high perfor-mance io. In Proc. IPDPS, 2009.

[54] Hasan Abbasi, Jay Lofstead, Fang Zheng, Scott Klasky, Karsten Schwan, and Matthew Wolf. Extending i/o through high perfor-mance data services. In Proc. Cluster Computing, 2009.

[55] Pavan Konanki and Ali R. Butt. An exploration of hybrid hard disk designs using an extensible simulator, 2008. Masters Thesis, Virginia Tech.

[56] Intel. Intel x25-e extreme sata solid-state drive. http://www.intel. com/design/flash/nand/extreme/index.htm.

 

1. Software Process Models

(Sommerville Chapters 4, 17, 19, 12.4)

A software process model is a standardised

format for

• planning

• organising, and

• running

a development project.

 

Hundreds of different models exist and are

used, but many are minor variations on a

small number of basic models.

In this section we:

• survey the important basic models, and

• consider how to choose between them.

 

2

 

1.1. Planning with Models

SE projects usually live with a fixed financial

budget. (An exception is maintainance?)

Additionally, time-to-market places a strong

time constraint.

There will be other project constraints such

as staff.

3

 

 

Project constraints

 

 

Examples of Project Constraints

 

Project planning is the art of scheduling/constraint

solving the project parameters, along various

dimensions:

time, money, staff ...

in order to optimise:

• project risk [low] (see later)

• profit [high]

• customer satisfaction [high]

• worker satisfaction [high]

• long/short-term company goals

 

Questions: 

1. What are project parameters?

2. Are there good patterns of organisation

that we could copy?

 

Project parameters describe the whole project,

but we must at least describe:

• resources needed

(people, money, equipment, etc)

• dependency & timing of work

(flow graph, work packages)

• rate of delivery (reports, code, etc)

It is impossible to measure rate of progress

except with reference to a plan.

7

 

In addition to project members, the following

may need access to parts of the project plan:

• Management

• Customers

• Subcontractors (outsourcing)

• Suppliers (e.g. licenses, strategic partners)

• Investors (long term investment)

• Banks (short term cash)

 

1.2. Project Visibility

Unlike other engineers

(e.g. civil, electronic, chemical ... etc.)

software engineers do not produce anything

physical.

It is inherently difficult to monitor an SE

project due to lack of visibility.

9

 

This means that SE projects must produce

additional deliverables (artifacts) 

which are visible, such as:

• Design documents/ prototypes

• Reports

• Project/status meetings

• Client surveys (e.g. satisfaction level)

 

1.3. What is a Software Process

Model?

Definition.

A (software/system) process model is a

description of the sequence of activities

carried out in an SE project, and the relative

order of these activities.

 

11

 

It provides a fixed generic framework that

can be tailored to a specific project.

Project specific parameters will include:

• Size, (person-years)

• Budget,

• Duration.

project plan =

process model + project parameters

 

12

 

There are hundreds of different process models

to choose from, e.g:

• waterfall,

• code-and-fix

• spiral

• rapid prototyping

• unified process (UP)

• agile methods, extreme programming (XP)

• COTS ...

But most are minor variations on a small

number of basic models.

 

By changing the process model, we can

improve and/or tradeoff:

• Development speed (time to market)

• Product quality

• Project visibility

• Administrative overhead

• Risk exposure

• Customer relations, etc, etc.

 

Normally, a process model covers the entire

lifetime of a product.

From birth of a commercial idea

to final de-installation of last release

i.e. The three main phases:

• design,

• build,

• maintain. (50% of IT activity goes here!)

 

We can sometimes combine process models e.g.

1. waterfall inside evolutionary – onboard shuttle

software

2. Evolutionary inside waterfall – e.g. GUI

prototyping

We can also evolve the process model together with

the product to account for product maturity,

e.g. rapid prototyping  waterfall

 

1.4. The Waterfall Model

• The waterfall model is the classic process

model – it is widely known, understood and used.

• In some respect, waterfall is the ”common

sense” approach.

• R.W. Royce, Managing the Development of Large

Software Systems: Concepts and Techniques,

Proc. IEEE Westcon, IEEE Press, 1970.

17

 

 

Advantages

1. Easy to understand and implement.

2. Widely used and known (in theory!)

3. Fits other engineering process models: civil, mech etc.

4. Reinforces good habits: define-before- design, design 

before-code

8. Identifies deliverables and milestones

9. Document driven: People leave, documents don’t

Published documentation standards: URD, SRD, ... etc. ,

e.g. ESA PSS-05.

10. Works well on large/mature products and weak teams.

19

 

Disadvantages I

1. Doesn’t reflect iterative nature of exploratory

development.

2. Sometimes unrealistic to expect accurate requirements

early in a project

3. Software is delivered late, delays discovery of serious

errors.

4. No inherent risk management

5. Difficult and expensive to change decisions, ”swimming

upstream”.

6. Significant administrative overhead, costly for small

teams and projects. 20

 

1.5. Code-and-Fix

This model starts with an informal general

product idea and just develops code until a

product is ”ready” (or money or time runs

out). Work is in random order.

Corresponds with no plan! (Hacking!)

 

Advantages

1. No administrative overhead

2. Signs of progress (code) early.

3. Low expertise, anyone can use it!

4. Useful for small “proof of concept”

projects, e.g. as part of risk reduction.

 

Disadvantages

1. Dangerous!

1. No visibility/control

2. No resource planning

3. No deadlines

4. Mistakes hard to detect/correct

2. Impossible for large projects,

communication breakdown, chaos.

 

1.6. Evolutionary Development

Types

Type 1: Exploratory Development: customer

assisted development that evolves a product  from

ignorance to insight, starting from core, well

understood components (e.g. GUI?)

Type 2: Throwaway Prototyping: customer

assisted development that evolves requirements

from ignorance to insight by means of

lightweight disposable prototypes.

24

 

1.7. Type 1: Spiral Model

Extends waterfall model by adding iteration to explore 

/manage risk

Project risk is a moving target. Natural to progress a project

cyclically in four step phases

1. Consider alternative scenarios, constraints

2. Identify and resolve risks

3. Execute the phase

4. Plan next phase: e.g. user req, software req, architecture

... then goto 1

25

 

In 1988 Boehm developed the spiral model as

an iterative model which includes risk

analysis and risk management.

Key idea: on each iteration identify and solve

the sub-problems with the highest risk.

 

26

 

 

Advantages

1. Realism: the model accurately reflects the

iterative nature of software development

on projects with unclear requirements

2. Flexible: incoporates the advantages of the

waterfall and evolutionary methods

3. Comprehensive model decreases risk

4. Good project visibility.

28

 

Disadvantages

1. Needs technical expertise in risk analysis

and risk management to work well.

2. Model is poorly understood by non 

technical management, hence not so

widely used

3. Complicated model, needs competent

professional management. High

administrative overhead.

 

1.8. Type 2: Rapid Prototyping

Key idea: Customers are non-technical and

usually don’t know what they want.

Rapid prototyping emphasises requirements

analysis and validation, also called:

• customer oriented development,

• evolutionary prototyping

30

 

The Rapid

Prototype Workflow

 

Advantages

1. Reduces risk of incorrect user requirements

2. Good where requirements are changing/

uncommitted

3. Regular visible progress aids management

4. Supports early product marketing

32

 

Disadvantages I

1. An unstable/badly implemented prototype

often becomes the final product.

(Migration to a type 1 process!)

2. Requires extensive customer collaboration

– Costs customers time/money

– Needs committed customers

– Difficult to finish if customer withdraws

– May be too customer specific, no broad

market 33

 

Disadvantages II

 

3. Difficult to know how long project will

last

4. Easy to fall back into code-and-fix without

proper requirements analysis, design,

customer evaluation and feedback.

 

34

 

1.9. Type 1 :

Agile Software Processes

Need for an adaptive process model suited to changes in:

• User requirements

• Customer business models

• Technology

• In-house environment

De-emphasise documentation, esp. URD!

Emphasise change management e.g. reverse engineering

design!

Examples include XP, Scrum, Agile modeling etc

35

 

1.9.1. Agile Principles

(C.f Sommerville Fig 17.3)

• Incremental delivery of software

• Continuous collaboration with customer

• Embrace change

• Value participants and their interaction

• Simplicity in code,

 

1.9.2. XP Release Cycle

For a sample story see Sommerville Figure 17.6

Same as use-case? 37

 

1.9.3. XP Practices (Summary)

1. Incremental planning

2. Small releases

3. Simple design

4. Programming in pairs (egoless programming, see 7.)

5. Test-driven development

6. Software refactoring (needs UML?)

7. Collective ownership: metaphors, standards, code

8. Continuous integration

9. Sustainable pace (No overtime!)

10. On-site customer!

38

 

Advantages

1. Lightweight methods suit small-medium

size projects

2. Produces good team cohesion

3. Emphasises final product

4. Iterative

5. Test-based approach to requirements and

quality assurance

 

Disadvantages

1. Difficult to scale up to large projects where

documentation is essential

2. Needs experience and skill if not to

degenerate into code-and-fix

3. Programming pairs is costly (but see

productivity literature)

4. Test case construction is a difficult and

specialised skill.

40

 

1.10. Rational Unified Process

(RUP)

• Hybrid model inspired by UML and Unified

Software Development Process.

• A generic component-based process?

• Three views on the process

– Dynamic view: RUP phases

– Static view: RUP activities

– Practise view: RUP best-practise

 

41

 

Details

• Lifetime of a software product in cycles:

• Birth, childhood, adulthood, old-age,

death.

• Identify product maturity stages

• Each project iteration cycle is a phase,

culiminating in a new release (c.f. Spiral

model)

 

UP process – RUP phase workflow

(drawn as a UML Statechart!)

 

RUP Phases

Phases are goal directed  and yield deliverables:

• Inception – Establish the business case. Identify

external entities (actors, systems). Estimate ROI.

• Elaboration – Understand problem domain.

Establish architecture, and consider design

tradeoffs. Identify project risks. Estimate and

schedule project. Decide on build vs. buy.

• Construction – Design, program and test.

Components are bought and integrated.

• Transition – release a mature version and deploy

in real world.

 

RUP Workflows

• RUP separates what and when into two orthogonal 

process views.

• When modeled by phases 

• What modeled by workflows (c.f. Sommerville

Figure 4.13)

• Any workflows can be active in any phases.

• Anything that instantiates the following diagram is

an instance of RUP!!

• An agile instantiation exists (Carman 2002)

45

 

 

RUP Practise View

1. Develop software iteratively

2. Manage requirements

3. Use component-based architectures

4. Visually model software

5. Verify software quality

6. Control software changes

 

Use Case Model UML class diagram! 

4. Visually model software =

UML Model-based Development Test Model

48

 

Advantages/ Disadvantages

• Difficult to judge without knowing the actual

chosen instantiation of the RUP

• Unique use of the UML philosophy of SE.

• Can be as good/better than any other process

• Also as bad/worse than ...

 

1.11. COTS

• COTS =

Commercial Off-The-Shelf software

• Integrate a solution from existing

commercial software components using

minimal software plumbing

• All projects seek some software re-use –

Holy Grail of SE

• See also Sommerville Chapter 19.

 

Possible Approaches

1. Third-party vendors: component libraries, Java beans,

2. Integration solutions: CORBA, MS COM+, Enterprise

Java Beans, software frameworks ...

3. Good software engineering: application generators

(Yacc, Lex ... Sommerville 18.3), generic programming

(types, polymorphism etc)

4. Commercial finished packages: databases, spread

sheets, word processors, web browsers, etc.

5. Design to open source interfaces: e.g. XML, ISO

standards, etc.

 

Distributed COTS :

Web Service Providers

• Integration of web-services: third party service

providers using (XML-based?) service models:

• JINI – extension of Java for service discovery

• SOAP (Simple Object Access Protocol)

• WSDL (Web Services Description Language)

• UDDI (Universal Description, Discovery and

Integration)

52

 

Advantages

1. Fast, cheap solution

2. Explore solutions with existing products

(c.f. your project work!)

3. May give all the basic functionality

4. Well defined integration project, easy to

run.

5. Open to outsourcing (c.f. Point 4)

6. Build strategic supplier partnerships

 

Disadvantages

1. Limited functionality/ compromise/ requirements

drift.

2. Component identification can be tricky –

mistakes occur!

3. Licensing problems: freeware, shareware, etc.

4. Customer lock-in: license fees, maintainance

fees, upgrades ...

5. Compatibility issues/loss of control:

1. Between components (doesn’t start!)

2. During maintenance (suddenly stops!)

54

 

PROSIDING KOMMIT 2012

(KOMPUTER DAN SISTEM INTELIJEN)

Volume 7 – 2012

TEKNOLOGI INFORMASI DAN KOMUNIKASI

(TIK) UNTUK KETAHANAN NASIONAL

ISSN: 2302-3740

PENERBIT

Lembaga Penelitian Universitas Gunadarma

 

Alamat Editor:

Lembaga Penelitian Universitas Gunadarma

Jl. Margonda Raya 100 Pondok Cina

Depok, 16424

Telp. +62-21-78881112 ext. 455

Fax. +62-21-7872829

e-Mail: kommit@gunadarma.ac.id

Laman: http://penelitian.gunadarma.ac.id/kommit

 

Prosiding KOMMIT, Volume 7 - 2012

Editor:

Tety Elida, Moh. Okki Hardian, Wahyu Rahardjo, Fitrianingsih, Tri Wahyu Retno Ningsih

Disain sampul: Wira Catur

Penerbit: Lembaga Penelitian Universitas Gunadarma

Hak cipta © 2012 oleh Universitas Gunadarma. Dilarang memperbanyak atau memindahkan sebagian atau seluruh isi prosiding ini dalam bentuk apapun, baik secara eletronis maupun mekanis, termasuk memfotocopy, merekam atau dengan sistem penyimpanan lainnya tanpa izin tertulis dari penerbit.

ISSN: 2302-3740

 

DEWAN REDAKSI

Penanggung Jawab:

Dr. Ir. Hotniar Siringoringo, MSc.

Ketua Dewan Editor:

Dr. Ir. Tety Elida Siregar, MM.

Editor Pelaksana:

Moh. Okki Hardian, ST., MT.

Wahyu Rahardjo, SPsi., MSi.

Fitrianingsih , SKom., MMSi.

Tri Wahyu Retno Ningsih, SSas., MM.

Reviewer:

Prof. Dr. I Wayan Simri Wicaksana, S.Si, M.Eng.

Prof. Dr.rer.nat. Achmad Benny Mutiara, SSi, SKom.

Prof. Dr. Busono Soerowirdjo

Prof. Dr. Sarifuddin Madenda

Prof. Dr. dr. Johan Harlan

Prof. Dr. Ir. Eriyatno MSAE.

Dr. Tb. Maulana Kusuma, SKom., MEngSc.

Dr.-Ing. Adang Suhendra, SSi,SKom,MSc.

Prof. Dr. Ir. Kudang Boro Seminar, MSc.

Drs. Agus Harjoko MSc., PhD.

Dr. Ir. Joko Lianto Buliali

PENERBIT

Lembaga Penelitian Universitas Gunadarma

Jl. Margonda Raya 100 Pondok Cina

Depok, 16424

Telp. +62-21-78881112 ext. 455

Fax. +62-21-7872829

e-Mail: kommit@gunadarma.ac.id

Laman: http://penelitian.gunadarma.ac.id/kommit

iii

 

PANITIA PELAKSANA SEMINAR

 

Penasehat:

Prof. Dr. E.S. Margianti, S.E., MM.

Prof. Suryadi Harmanto, SSi., M.MS.I.

Agus Sumin, S.Si., MM.

Penanggung Jawab:

Prof. Dr. Yuhara Sukra, MSc.

Prof. Dr. Didin Mukodim, MM.

Ketua Pelaksana:

Dr. Ir. Hotniar Siringoringo, MSc.

Wakil Ketua Pelaksana:

Dr. Bertalya

Sekretariat:

Ida Ayu Ari Angreni, ST., MMT.

Dr. Jacobus Belida Blikololong

MS. Harlina, S.Kom., MM.

Sarana Prasarana:

Drs. Hardjanto Sutedjo, MM.

Rino Rinaldo, SE., MM

Riyanto, ST.

 

iv

 

KATA PENGANTAR

Pertukaran informasi merupakan kebutuhan masyarakat modern, sehingga Teknologi Informasi dan Komunikasi (TIK) menjadi hal yang sangat penting. Secara kasat mata, setiap orang dapat menyaksikan perkembangan TIK yang sangat pesat. Perkembangan TIK sampai saat ini masih didominasi oleh negara-negara maju. Kondisi ini harus direposisi.

Indonesia memiliki sumber daya manusia yang handal dan banyak, di antaranya berada di perguruan tinggi. Sumber daya manusia ini terkesan bekerja masih sendiri-sendiri. Penelitian di lingkungan perguruan tinggi maupun litbang sering disalahartikan sebagai pemuas akademis, sementara di kalangan industri lebih tertarik pada penyelesaian ekonomis jangka pendek. Permasalahan ini dapat diatasi dengan memulai kolaborasi antara dunia pendidikan, litbang, industri dan pemerintah.

KOMMIT merupakan seminar nasional di bidang komputer dan teknik yang mendukung pengembangan teknologi komputer maupun aplikasi komputer dalam berbagai bidang. Seminar ini bertujuan menyediakan wadah bagi peneliti, akademisi dan praktisi untuk saling bertukar informasi, berdiskusi dan berkolaborasi sehingga dapat menghasilkan produk siap pakai di dalam bidang sistem informasi.

Topik yang menjadi pembahasan pada KOMMIT ke 7 ini adalah: sistem informasi manajemen, sistem informasi geografis, sistem informasi medis, enterprise resource planning, information retrieval, matematika aplikasi, sistem keamanan, aplikasi multimedia, pengolahan sinyal dan citra, computer vision, open source & open content, e-government, e-business, e-education, data semantik, information system interoperability, distributed, parallel, grid, P2Pp, mobile information management, mobile tecnology, green computing, telekomunikasi dan jaringan komputer, sistem kontrol, instrumentasi dan diagnosis, mekanika dan elektronika, energi terbarukan, cognitive science, soft computing, perceptual science, bioinformatika dan geoinformatika, collaborative network, dan electron devices.

Artikel yang disajikan pada seminar ini setelah melalui proses peer review, berjumlah seratus satu, yang berasal dari 15 Perguruan Tinggi di Indonesia. Beberapa artikel yang terpilih akan di publikasikan pada Jurnal Ilmiah yang diterbitkan oleh Universitas Gunadarma.

Semoga seminar ini dapat memberikan masukan bagi pengembangan teknologi informasi dan komunikasi di negara kita. Kami ucapkan terima kasih kepada para reviewer yang telah bersedia melakukan review, juga kepada pembicara tamu dan nara sumber yang telah berkontribusi pada acara ini, serta kepada semua pihak yang telah membantu proses produksi prosiding ini.

Ketua Pelaksana

Dr. Ir. Hotniar Siringoringo, MSc.

V

 

 

DAFTAR ISI

DEWAN REDAKSI iii

PANITIA PELAKSANA SEMINAR iv

KATA PENGANTAR v

DAFTAR ISI vii

DAFTAR ARTIKEL:

1. Sistem Informasi Manajemen Penanggulangan Kemiskinan (Studi Kasus Kabupaten Ogan Komering Ilir Provinsi Sumatera Selatan)

Ahmad Haidar Mirza 1

2. Optimasi Pencarian dengan Knowledge Graph

Abidin Ali, Dina Rifdalita, Juliana Putri Lestari, Lintang Yuniar Banowosari 11

3. Analisis Teknik Reduksi Data dan Minimalisasi Ukuran File APK pada Mobile Application Pengenalan Budaya Indonesia Berbasis Android Serta

Pengembangannya

Adhika Novandya, Debyo Saptono 18

4. Aplikasi Manajemen File Berbasis Web untuk Monitoring Status Kegiatan

Akhmad Fauzi, Tri Sulistyorini 27

5. Penerapan Metode Dijkstra dalam Pencarian Jalur Terpendek pada Perusahaan Distribusi Film

Albert Kurnia, Friska Angelina, Windy Dwiparaswati 36

6. Penyembunyian Informasi (Steganography) Audio Menggunakan Metode LSB (Least Significant Bit) Menggunakan Matlab

Ari Santoso, Irfan, Nazori AZ 42

7. Standardisasi Sistem Informasi Kesehatan Berjenjang Open E-Health Gunadarma Information System, Mewujudkan Layanan Kesehatan Prima

Aries Muslim, AB Mutiara, Teddy Oswari, Riyandari Auror, Irdiah Amsawati 51

8. Pengembangan Web sebagai Upaya Penunjang Optimalisasi Produk Asuransi

Armaini Akhirson 59

9. Protokol Autentikasi Berbasis One Time Password untuk Banyak Entitas

Avinanta Tarigan, D.L. Crispina Pardede 67

10. Peningkatan Keamanan Kartu Kredit Menggunakan Sistem Verifikasi Sidik Jari di Indonesia

Bima Shakti Ramadhan Utomo, Denny Satria, Lulu Mawaddah Wisudawati 72

11. Rancangan Aplikasi Pencarian Barang Pada Metro Pacific Place dengan Menggunakan Macromedia Dreamweaver 8

Triyanto, Bramantyo Sukarno, Miftah Andriansyah 78

vii

 

12. Sistem Pengambilan Keputusan Bela Negara Non-Fisik untuk Daerah Depok

dengan Metode AHP (Analytic Hierarchy Process)

Damai Subimawanto, Surya Thiono Wijaya, Yusuf Triyuswoyo, I Wayan Simri

Wicaksana, Detty Purnamasari 85

13. Penerapan Teknologi Informasi dan Komunikasi (TIK) pada UMKM dengan Menggunakan Technology Acceptance Model (TAM) (Studi Kasus di Depok dan Qingdao)

Deboner Hillery, Dharma Tintri, Pandam R Wulandari 94

14. Faktor Kunci Sukses dalam Pelaksanaan Sistem Enterprise Resource Planning

Delvita Dita Putri Anggrayni, Dewi Agushinta R. 101

15. Model Penentuan Posisi Siaga Lift sebagai Pemanfaatan Penghematan Energi pada Sistem Kerja Lift

Denmas Muhammad Ridwan, Donny Ejie Baskoro, Faisal Yafi, Lily Wulandari 110

16. Pemanfaatan Jaringan Akses Telepon sebagai Jaringan Broadband Layanan Internet dengan Teknologi Asymmetric Subscriber Line

Djasiodi Djasri 116

17. Evaluasi Website JobsDBTM Mobile dengan Metode Usability Heuristic Esty Purnamasari, Helen Wijayanti, Yosfik Alqadri, Dewi Agushinta Rahayu,

Fani Yayuk Supomo 123

18. Perancangan dan Implementasi Sistem Informasi Peralatan dengan Penerapan Konsep Three Tier (Studi Kasus: Gardu Induk Prabumulih UPT Palembang)

Evi Yulianingsih, Marlindawati 131

19. Faktor-Faktor yang Mempengaruhi Minat Nasabah Menggunakan Internet Banking dengan Menggunakan Anjungan Tunai Mandiri (Studi Kasus pada Bank BCA, BRI dan Bank Syariah Mandiri)

Faramita Dwitama, Mohammad Abdul Mukhyi 139

20. Enkripsi Informasi untuk Pengamanan Pesan Singkat pada Telepon Seluler Berbasis Java MIDP

Farid Thalib, Melba Mauludina Novalestari 148

21. Desain Database e-Supermuseum Batik Indonesia

Fikri Budiman, Slamet Sudaryanto Nurhendratno 157

22. Analisis Perbandingan Kinerja Search Engine Menggunakan Penelusuran Precision dan Recall untuk Informasi Ilmiah Bidang Ilmu Kedokteran

Sukesi, Fitrianingsih 164

23. Membandingkan Web Pengunduhan Perangkat Lunak

Fuji Ihsani, Istiana Idha Aulia, Melisa Chatrine Kamu, Anacostia Kowanda, Trini

Saptariani 172

24. Analisis dan Verifikasi Formal Protokol Non-Repudiasi Zhang-Shi dengan Logika SVO-CP

Hanum Putri Permatasari, Avinanta Tarigan, D. Lucia Crispina Pardede 178

25. Implementasi Kebijakan E-Government pada Pemerintah Kota Palembang

Hardiyansyah 185

 

viii

 

26. Aplikasi Pengingat Jadwal Imunisasi Berbasis Android

Hauliza Rindhayanti, Lintang Yuniar Banowosari 193

27. Model Berbasis Ekstraksi untuk Analisis Gaya Berjalan

Hustinawaty, Miftahul Jannah, Rd. Fazlur Rahman 201

28. Metoda Penumbuhan Kreativitas Berbasis Web: Studi Pengembangan Produk Kerajinan Tenun Ikat dalam Upaya Melestarikan dan Meningkatkan Nilai Tambah

Iman Murtono Soenhadji, Priyo Purwanto, Ida Astuti, Faisal Reza 209

29. Simulasi dan Optimasi Antrian Pelayanan Agen JNE Buaran Isram Rasal, Hardimen Wahyudi, Nadia Rahmah Al Mukarromah, Yuhilza

Nahum 218

30. Aplikasi Data Mining dengan Teknik Decision Tree untuk Mengklasifikasikan Data Pasien Rawat Inap

Julius Santony, Sumijan 226

31. Integrasi Sumber Data Heterogen Menggunakan Ontologi, Studi Kasus: Data Kependudukan Indonesia

Kemal Ade Sekarwati, I Wayan Simri Wicaksana 235

32. Pengenal Ucapan untuk Belajar Bahasa Menggunakan Perangkat Mobile

Kezia Velda Roberta, Raden Supriyanto 241

33. Sistem Pakar Pendeteksi Prediksi Kemungkinan Penyakit Stroke

Linda Atika 247

34. Analisis Sektor Unggulan dalam Perekonomian DKI Jakarta

Lita Praditha, Mohammad Abdul Mukhyi 254

35. Kapabilitas Proses Konstruksi Perangkat Lunak pada Perusahaan Pengembang Perangkat Lunak di Bali Menggunakan Kerangka Kerja ISO/IEC 15504

Luh Gede Surya Kartika, Kridanto Surendro 262

36. Sistem New Media pada Aplikasi Internet Radio Berbasis Android

Lulu Mawaddah Wisudawati, Avinanta Tarigan 269

37. Kajian Awal Hibridisasi Toyota Soluna dengan Konfigurasi Parallel HEV

Mohamad Yamin, Agung Dwi Sapto 276

38. Pemodelan dan Analisis Rem Cakram dan Rem Tromol dengan Software CATIA V5

Mohamad Yamin, Darmawan Sebayang 283

39. Deteksi Sonority Peak untuk Penderita Speech Delay Menggunakan Speech Filing System

Muhammad Subali, Tri Wahyu Retno Ningsih, M. Kholiq 289

40. Penerapan Periklanan di Internet dan Pemasaran Melalui E-Mail untuk Meningkatkan Pemasaran Produk UMKM di Wilayah Depok

Mujiyana, Lana Sularto, M. Abdul Mukhyi 296

41. Monitoring Sistem Pengendalian Suhu dan Saluran Irigasi Hydroponik pada Greenhouse Berbasis Web

Nia Maharani Raharja, Iswanto 303

 

ix

 

42. Disain Rangkaian Detektor Mini Doppler

Nur Sultan Salahuddin, Paulus Jambormias, Erma Triawati 311

43. Prototipe Sistem Pemrosesan Limbah Medis

Nur Sultan Salahuddin, Adi Hermansyah, RR Sri Poenomo Sari 317

44. Audit TIK pada Sistem Penerbitan Surat Perjalanan Republik Indonesia (SPRI) di Kantor Imigrasi Bogor

Nurul Adhayanti, Karmilasari 323

45. Aplikasi Pencarian Lokasi Sekolah Menggunakan Telepon Selular Berbasis Android

Nuryuliani, Selvi Isni Hadisaputri, Miftah Andriansyah 331

46. Faktor Penentu Efektifitas IT Governance: Studi Kasus pada Perusahaan di DKI Jakarta

Pandam Rukmi Wulandari, Samuel David Lee, Renny Nur'ainy 340

47. Aplikasi Mobile Panduan Diet Berdasarkan Golongan Darah Berbasis Android

Parno, Swesti Mahardini 345

48. Studi Terhadap Konstruksi Model Pengklasifikasi Regresi Logistik

Retno Maharesi 352

49. Karakteristik dan Model Matematika Aliran Lumpur pada Pipa Spiral

Ridwan 360

50. Implementasi Mikrokontroler untuk Deteksi Drop Tegangan pada Instalasi Sederhana

Rif'an Tsaqif As Sadad, Iswanto 368

51. Analisis Pendeteksian Nodul Citra Sinar-X Paru

Rodiah, Sarifuddin Madenda, Dewi Agushinta Rahayu 377

52. Composite Range List Partitioning pada Very Large Database

Rosni Gonydjaja, Yuli Karyanti 384

53. Analisis Perbandingan Waktu untuk Layanan Email dan SMS pada Jaringan Interkoneksi untuk Kajian Efektivitas Dukungan Media Komunikasi Dosen-Mahasiswa

S N M P Simamora, Karina Datty Putri, Robbi Hendriyanto 389

54. Desain Prototipe Aplikasi Sistem Keamanan pada Rumah Berbasis Pengenalan Wajah dengan Algoritma Jaringan Saraf Tiruan dan Fitur Fft

Shinta Puspasari, Hendra 398

55. Analisis Implementasi Algoritma Propagasi Balik pada Aplikasi Identifikasi Wajah Secara Waktu Nyata

Shinta Puspasari, Alfan Sucipta 405

56. Sistem Pemantau Ruangan dengan Penangkapan Gambar Otomatis Menggunakan Sensor Infra Merah Pasif

Singgih Jatmiko, R. Supriyanto, R.N. Nasution 412

 

x

 

57. Sistem Pengenalan Ekspresi Wajah Berdasarkan Citra Wajah Menggunakan Metode Eigenface dan Nearest Feature Line

Sulistyo Puspitodjati, Tyas Arie Wirana 418

58. Ekstraksi Data pada Halaman Web Database Mining Akademik Menggunakan Simple Tree Matching (STM)

Sumijan, Julius Santony 426

59. Perancangan dan Implementasi Software Penyelesaian Persamaan Non Linier dengan Metode Fixed Point Iteration

Vivi Sahfitri 447

60. Perhitungan Panjang Janin pada Citra Ultrasonografi untuk Memprediksi Usia Kehamilan

Wahyu Supriyatin, Bertalya 456

61. Model Translator Notasi Algoritmik ke Bahasa C

Wijanarto, Achmad Wahid Kurniawan 464

62. Simulasi Dinamika Molekular Sistem Molekul Argon dan Graphene dengan Menggunakan Perangkat Lunak Dl_Poly

Ahmad Rifqi Muchtar, Wisnu Hendradjit, Agus Samsi 473

63. Pengidentifikasian Otomatis Bentuk Kista Ovarium Menggunakan Deteksi Circle dan Deteksi Tepi Laplacian dan Prewitt.

Yenniwarti Rafsyam, Jonifan 482

64. Pengaruh Karakteristik, Sikap dan Pelatihan terhadap Penggunaan Teknologi Informasi dan Kinerja Pegawai untuk Penerapan Pemerintah Elektronik di Pedesaan

Yuventus Tyas Catur Pramudi, Karis Widyatmoko 489

65. Perancangan Sistem Informasi Alur Kerja (Work Flow) Dokumen Pengajuan Proposal Skripsi

Zulfiandri, Sarip Hidayatullah, Wahyudianto 500

66. Aplikasi Pengenalan Budaya dari 33 Provinsi di Indonesia Berbasis Android

Adhika Novandya, Ajeng Kartika, Ari Wibowo, Yudhi Libriadiany 508

67. Sistem Informasi Geografis Bengkel Resmi Mercedes-Benz dan BMW di Kota Jakarta Menggunakan Quantum GIS

Agustini Dwi Setia Rahayu, Ana Rizki, Ria Awalliya 514

68. Studi Kasus Konflik PT.XXX dengan Pelanggan Kereta Kelas Ekonomi Berdasar Ilmu Teori Organsisasi Umum

Albert Kurnia Himawan, Juliana Putri Lestari, Aris Budi Setiawan 517

69. Aplikasi Pengenalan Dasar-Dasar Bahasa Inggris untuk Anak Usia Dini Menggunakan Adobe Flash CS 3 Professional

Alfa Marlin, Siti Andini, Sri Wahyuni 519

70. Eksploitasi Celah Keamanan Piranti Lunak Web Server Vertrigoserv pada Sistem Operasi Windows Melalui Jaringan Lokal

Andrias Suryo Widodo, Maria Magdalena Merry, Stefanus Dwi Putra Medisa 524

 

xi

 

71. Sistem Pengambilan Keputusan Kelayakan Sekolah Mendapatkan Status RSBI Studi Kasus SMA RSBI Di DKI Jakarta

Ardhani Reswai Yudistari, Odheta, Tryono Taqwa 529

72. Penerapan Algoritma Kruskal dan Pengimplementasiannya dalam Kasus Pendistribusian Majalah "UG News" Antar Universitas Gunadarma

Ardisa Pramudhita, Mahisa Ajy Kusuma, Nur Fisabilillah 535

73. Implementasi Algoritma Dijkstra untuk Menentukan Rute Terpendek Antar Museum di Yogyakarta Berbasis Web

Ardo Rama, Citra Ika Wibawati, Rizka Fajriah 538

74. Pembuatan Aplikasi Permainan Labirin 2D untuk Handphone

Aries Afriliansyah 542

75. Konfigurasi Trixbox Server Untuk VoIP pada Jaringan Peer to Peer

Arif Liberto Jacob, Muhammad Muhijar, Ferry Wisnuargo 547

76. Sistem Penunjang Keputusan Memilih Kriteria Lagu Pop Indonesia yang Baik

Ario Halik, Virgiawan Ananda Pratama 550

77. Evaluasi Algoritma Prim dan Kruskal Terhadap Pemasangan Kabel Telepon di DKI Jakarta

Atikah Luthfiyyah, Voni, Wahyu Pratama 553

78. Aplikasi Pemetaan Pusat Perbelanjaan Kota Bekasi Menggunakan Android Awal Arifianto, Muhammad Yunus, Andrika Siman, Agung Rahmat Dwiardi,

Deny Nugroho 556

79. Penerapan Algoritma Greedy pada Studi Kasus Pencarian Rumah Sakit Terdekat di Jakarta Selatan

Bagus Fitroh Alamsyah, Maulana Malik Ibrahim, Prakasita Wigati 559

80. Implementasi Algoritma Dijkstra Guna Optimasi Jalur Pendistribusian Produk Seluler

Banu Adi Witono, Dhita Angreny, Randy Aprianggi 561

81. Face Recognition Menggunakan Metode Linear Discriminant Analysis (LDA)

Bayu Adi Yudha Prasetya 563

82. Pembuatan Game Arasen untuk Latihan Soal Tes Potensi Akademik Menggunakan RPG Studio

Daisy Patria, Hayu Wasna Sari, Riyandari Asrita 570

83. Pemodelan Spasial Tingkat Kerawanan Kecelakaan Lalu Lintas di Kota Depok

Eriza Siti Mulyani, Muhammad Arsah Novel Simatupang 576

84. Sistem Log Monitoring Jaringan (LAN) Menggunakan Bahasa Pemrograman Pascal

Fendy Christian, Stefanus Goutama, Afrilia Nita Anjani 582

85. Website Surat Pembaca Sebagai Media Komunikasi dalam Penyampaian Aspirasi Masyarakat

Hamisati Muftia, Nabiyurrahmah 584

 

xii

 

86. Aplikasi Pendidikan Bagi Anak di Bawah Umur 7 Tahun

Helmi, Muhammad Subentra, Randy Aditiya Yusuf 586

87. Sistem Pencarian Fasilitas Umum Terdekat Menggunakan Augmented Reality dengan Minimum Spanning Tree

Hifshan Riesvicky, Prita Dessica, Tatang Fanji Permana 592

88. Aplikasi Multimedia Audio Video Player dengan Menggunakan Visual Basic .Net 2008

Inggrit Parnandes, Rias Astria, Meilisa Ndaru Hermiyanti 595

89. Aplikasi Energy Usage Calculator untuk Menghitung Penggunaan dan Biaya Energi Listrik Berbasis Python Versi 3.2.3

M Haidar Hanif, Herio Susanto 599

90. Implementasi Algoritma Kruskal untuk Optimasi Pengangkutan Sampah

Meilidyaningtyas Cantika Ryadiani, Nurul Ardianingsih, Robby Matheus 602

91. Pemilihan Aplikasi Permainan untuk Perkembangan Motorik dan Simbolik Anak Usia 1 - 7 Tahun

Michael Satrio Prakoso, Detty Purnamasari 605

92. Sistem Informasi Geografis SMA di Bogor

Muhamad Ramadani Silatama, Narendra Paskarona, Ary Wahyudi 608

93. Pembuatan Website World Watch Shop Menggunakan Magento Commerce

Rahma Eka Putri, Septiana Dewi Saputri, Sheila Rizka 614

94. Pembuatan Aplikasi Pemetaan Tempat Usaha di Sekitar Kampus Depok Gunadarma Menggunakan Android 2.1

Rangga Adhitya Pradiptha, Titik Rahayu Mariani, Winda Utari 616

95. Aplikasi Penjualan Makanan Khas Garut pada Toko Aneka Sari dengan Menggunakan Visual Basic .Net

Rangga Septian Putra, Rion Saputra, Ryan Oktario 619

96. Pengembangan E-Government pada Layanan Informasi Publik Pemerintahan Daerah Sulawesi Barat Menuju Good Governance

Rizka Fajriah, Windy Dwiparaswati, Aris Budi Setyawan 625

97. Perlunya Penerapan Teknologi Web Semantik pada Situs Pencarian Lowongan Pekerjaan di DKI Jakarta

Robby Matheus Gultom, Tatang Fanji Permana, Aris Budi Setyawan 628

98. Program Aplikasi Enkripsi dan Dekripsi SMS pada Ponsel Berbasis Android dengan Algoritma DES

Rudy Hendrayanto, A. Ramadona Nilawati 631

99. Penentuan Keputusan untuk Membantu Program Genre Bagi Pasangan Muda

Sandi Agung Harseno, Moh. Ropiyudin, Dessy Wulandari 634

100. Pembuatan Aplikasi Pembelajaran Bahasa Jerman Berbasis Mobile Android

Satrio Wibisono, Lisda 638

101. Aplikasi Foodcourt Menggunakan Microsoft Visual Studio 2008

Tri Hardiyanti, Shelly Gustika Septiani 644

 

xiii

 

Prosiding Seminar Ilmiah Nasional Komputer dan Sistem Intelijen (KOMMIT 2012) ol. 7 September 2012

Universitas unadarma Depok 18 1 September 2012 ISSN: 2302-3740

RANCANGAN APLIKASI PENCARIAN BARANG PADA METRO

PACIFIC PLACE DENGAN MENGGUNAKAN MACROMEDIA

DREAMWEAVER 8

Triyanto1

Bramantyo Sukarno2

Miftah Andriansyah3

123Prodi Teknik Informatika, Sekolah Tinggi Teknik Cendekia-Tangerang, Banten

1tri3yanto@yahoo.co.id

2,3{bram.sukarno, miftah.andriansyah}@gmail.com

Abstrak

Seiring dengan perkembangan zaman dan teknologi yang semakin pesat, sebagian besar masyarakat telah menggunakan komputer di berbagai bidang pekerjaannya. Namun masalah yang dipantau pada sebuah Pusat Perbelanjaan adalah proses pencarian barang yang masih manual dan belum memiliki basis data, sehingga sulit dalam meninjau perubahan data yang telah terjadi. Untuk itulah penelitian dilakukan dan dibuatlah Aplikasi Pencarian Barang ini, yang kemudian digunakan untuk mengefisiensikan waktu dan memudahkan pencarian barang yang diinginkan oleh Pelanggan, pada gudang penyimpanan barang yang dilakukan secara komputerisasi. Piranti lunak yang digunakan untuk membuat Aplikasi Pencarian Barang ini adalah Macromedia Dreamweaver 8 dan PHP. dengan metode penelitian yang digunakan terdiri dari metode pengumpulan data dan metode pengembangan sistem, yang terdiri dari metode perancangan basis data konseptual, logical, fisikal, dan perancangan aplikasi. Di akhir tahapan, dilakukan implementasi pada Pusat Perbelanjaan tersebut dan dipantau tingkat keberhasilannya. Diharapkan aplikasi ini dapat menjawab kebutuhan dan memecahkan permasalahan yang ada.

Kata Kunci : Aplikasi Pencarian Barang, Pusat Perbelanjaan Metro, Macromedia Dreamweaver 8

 

PENDAHULUAN

Latar Belakang Penelitian

Teknologi informasi saat ini ber-pengaruh besar terhadap kinerja perusa-haan dalam menjalankan proses bisnis-nya, untuk menempatkan perusahaan pada posisi terdepan dalam persaingan bisnis saat ini, tidak semua bidang usaha dan terutama Pusat Perbelanjaan terkenal memanfaatkan komputer sebagai perang-kat cari untuk memudahkan pencarian barang. Dalam penelitian awal studi pus-taka, dalam sebuah jurnal Internasional menunjukkan bahwa pelanggan membu-tuhkan layanan cepat pencarian barang sebagai suatu standar kepuasan pelang- 

 

gan, mendukung terciptanya EcoMarket. [Karipidis, 2010] Namun bentuk layanan cepat tersebut tidak semua diterjemahkan dengan pemanfaatan komputer sebagai alat bantunya. Seperti Pusat Perbelanjaan Isetan di Singapura, masih menerapkan pencarian barang dengan mengandalkan Store Assistant di Wilayah Barang yang menguasai tempat barang yang dimaksud.

Seiring dengan perkembangan era globalisasi yang semakin maju, peranan komputer sangat diperlukan dalam proses pencarian barang untuk memudahkan bagi penjual demi terciptanya kepuasan pelanggan, salah satunya di Jakarta yaitu Pusat Perbelanjaan METRO PACIFIC PLACE, yang pada beberapa cabangnya

 

78 Triyanto dkk, Rancangan Aplikasi Pencarian...

 

Prosiding Seminar Ilmiah Nasional Komputer dan Sistem Intelijen (KOMMIT 2012) ol. 7 September 2012

Universitas unadarma Depok 18 1 September 2012 ISSN: 2302-3740

 

masih melakukan pencarian barang secara manual, tercatat pada selembar kertas yang kemudian ditempel pada rak yang tersedia. Tentunya hal ini menghambat proses pencarian barang tersebut karena harus melihat satu persatu catatan yang ada pada beberapa rak.

Pemecahan masalah pencarian barang pada Pusat Perbelanjaan Metro dengan membuat aplikasi pencarian ba-rang yang digunakan untuk membantu serta memudahkan Penjual dalam mem-berikan layanan yang memuaskan bagi pelanggan-pelanggannya. Dengan ada-nya aplikasi ini, Penjual dapat mengukur sampai di mana tingkat kecepatan dan ketepatan proses pencarian barang ter-sebut.

METODE PENELITIAN

Dalam penelitian ini, tahapan yang dilakukan adalah:

1. Observasi; Dalam tahapan ini, data didapatkan langsung pada obyek pene-litian yang berhubungan dengan masa-lah pokok pada penulisan ini. Teknik atau cara yang dilakukan adalah mewawancarai beberapa orang.

 

2. Analisis; Pada tahap ini dilakukan analisis terhadap proses yang sedang berjalan saat ini untuk kemudian menentukan apa yang akan menjadi kebutuhan dalam mendukung aplikasi yang akan dibuat. Analisis dilakukan dengan menganalisa hasil observasi, kebutuhan informasi, dan persyaratan sistem.[Bunafit, et all; 2004]

3. Pembuatan Aplikasi; Pada tahapan ini dilakukan perancangan aplikasi yang berbentuk basis data, baik konseptual, logikal dan fisikal dengan User Inter-face berbasis PHP. Selanjutnya imple-mentasi atas rancangan yang telah dibuat, baik berbentuk basis data mau-pun yang berbantuk interface dengan menggunakan piranti lunak Macrome-dia Dreamweaver versi 8.0. Namun pembangunan aplikasi ini tetap mem-perhatikan format PHP dan MySQL yang sangat erat hubungannya dengan basis data. [Indrajit,et all; 2002]

Gambar 1 merupakan ERD konseptual dari rancangan dan Gambar 2 merupakan ERD logika aplikasi yang dibuat.

 

 

Gambar 1. Rancangan ERD Konseptual Gambar 2. Rancangan ERD Logikal

Triyanto dkk, Rancangan Aplikasi Pencarian... 7

 

Prosiding Seminar Ilmiah Nasional Komputer dan Sistem Intelijen (KOMMIT 2012) ol. 7 September 2012

Universitas unadarma Depok 18 1 September 2012 ISSN: 2302-3740

 

HASIL DAN PEMBAHASAN

Hasil dari penelitian yang dilakukan, yaitu berupa aplikasi, dengan mengacu ERD yang dibuat:

a. Perancangan Diagram Fungsi

Gambar 3 diagram yang menggam-barkan tentang siapa saja yang dapat menjalankan aplikasi dan otoritas yang dapat dilakukan, yaitu admin dan user. Yang selanjutnya dapat di-gambarkan diagram aktifitasnya. Na-mun tidak ditampilkan dalam tulisan ini.

 

b. Perancangan Basis Data

Basis Data yang dibuat menerapkan Siklus Hidup Pengembangan Data-base yang digagas oleh Connolly [Connolly; 2005], dengan beberapa langkah. Gambar 4 adalah gambar rancangan basis data konseptual, Gambar 5 merupakan perancangan basis data logika dan Gambar 6 merupakan perancangan basis data fisikal. Rancangan struktur menu admin dapat dilihat pada Gambar 7.

 

 

Gambar 3. Diagram Fungsi Sistem

 

Gambar 4. Rancangan Basis Data Konseptual

80 Triyanto dkk, Rancangan Aplikasi Pencarian...

 

Prosiding Seminar Ilmiah Nasional Komputer dan Sistem Intelijen (KOMMIT 2012) ol. 7 September 2012

Universitas unadarma Depok 18 1 September 2012 ISSN: 2302-3740

Gambar 5. Rancangan Basis Data Logikal

Gambar 6. Rancangan Basis Data Fisikal

Gambar 7. Rancangan Stuktur Menu Admin

Hasil dari rancangan tersebut dibangun aplikasi dengan piranti lunak Macromedia Dreamweaver 8, PHP dan MySQL.

• Form Login

Pada Form ini User memasukkan username dan password yang akan diperiksa apakah memiliki otoritas untuk mengakses aplikasi ini atau tidak. Fasilitas Login dapat juga disatukan dengan fasilitas standar aplikasi utama Sistem Database Pengelolaan Barang yang selama ini telah ada (Gambar 8). Proses dari Form Login akan masuk ke Form Master Barang yang merupakan fasilitas utama dari Sistem Database Pengelolaan Barang yang telah ada dan tidak dibahas di forum ini.

Triyanto dkk, Rancangan Aplikasi Pencarian... 81

 

Prosiding Seminar Ilmiah Nasional Komputer dan Sistem Intelijen (KOMMIT 2012) ol. 7 September 2012

Universitas unadarma Depok 18 1 September 2012 ISSN: 2302-3740

 

Gambar 8. Form Login

 

• Form Cari Barang

Pada Form Cari Barang inilah Pencarian Barang mulai dilakukan. Aplikasi ini hanya mencari keberadaan barang, baik melalui Form Barang Masuk ataupun Form Barang Keluar yang merupakan fasilitas utama yang berada di Sistem Utama Pengelolaan Barang (Gambar 9).

Pencarian Barang dilakukan ber-dasarkan indeks Satuan Kualitas Unit (SKU) yang akan diproses dan langsung ditunjukkan Jumlah Barang, 

 

Harga, Lokasi Gudang, dll yang dibu-tuhkan. Pendefinisian Jenis dan Satuan Barang dilakukan di Sistem Utama Database Pengelolaan Barang, melalui Master Barang.

• Form Master Barang

Dari Master Barang ini kemu-dian didefinisikan apakah barang tersebut merupakan barang masuk ataukah barang keluar dari gudang (Gambar 10, Gambar 11 dan Gambar 12).

 

 

Gambar 9. Form Cari Barang

82 Triyanto dkk, Rancangan Aplikasi Pencarian...

 

Prosiding Seminar Ilmiah Nasional Komputer dan Sistem Intelijen (KOMMIT 2012) ol. 7 September 2012

Universitas unadarma Depok 18 1 September 2012 ISSN: 2302-3740

 

Gambar 12. Form Barang Keluar

 

Secara proses, informasi kebera-daan barang langsung ditampilkan di Form Pencarian Barang, yang juga berfungsi sebagai Tampilan Output, 

 

untuk mendapatkan hasil segera se-hingga memudahkan Sales Assistant dalam operasionalnya.

 

Triyanto dkk, Rancangan Aplikasi Pencarian... 83

 

Prosiding Seminar Ilmiah Nasional Komputer dan Sistem Intelijen (KOMMIT 2012) ol. 7 September 2012

Universitas unadarma Depok 18 1 September 2012 ISSN: 2302-3740

 

SIMPULAN DAN SARAN

Simpulan

Aplikasi yang dibuat ini hanya apli 

kasi tambahan yang dimasukkan sebagai add-on ke dalam Sistem Database Utama Pengelolaan Barang, yang sebelumnya tidak ada.

1. Dari penelitian yang dilakukan meng 

hasilkan sebuah solusi yang mem-bantu, serta mempermudah pekerjaan Sales Assistant dalam melakukan pro-ses pencarian barang lebih cepat dari sebelumnya sehingga diharapkan dapat mendukung proses operasional perusa-haan dalam mengontrol barang yang ada di masing-masing gudang penyim-panan stock barang.

2. Aplikasi pencarian barang memudah-kan pihak Pusat Perbelanjaan Metro Indonesia dalam pembuatan laporan, walapun dalam hasil dari pencetakan laporan masih bersifat umum.

Saran

Dari penelitian yang dilakukan,

aplikasi yang dihasilkan masih terdapat 

 

ketidak sempurnaan dalam hubungan antara persediaan barang keluar dengan barang masuk karena aplikasi ini bukan memantau keadaan terkini dari stok barang.

DAFTAR PUSTAKA

Bunafit, N. 2004 PHP & MySql dengan Editor Dreamweaver MX Penerbit Andi Yogakarta.

Connolly, Thomas, Begg, Carolyn. 2005 Database Systems: A Practical Approach to Design, Implementation, and Management Edisi ke-4 Addison-Wesley USA.

Indrajit, R. E., Prastowo, B.N., Syukri, M. 2002 Aplikasi Web Database Menggunakan PHP & MySQL Elex Media Komputindo Jakarta.

Karipidis, P., Tsakiridou, E. 2010

“Consumers' purchasing and store switching intentions in respect of eco 

marked products” International

Journal of Economics and Business Research 6 (2) 511-524.

 

84 Triyanto dkk, Rancangan Aplikasi Pencarian...

 

'Jf5i7tV~G:JU_,)"<75iAZ t1)'/'Jf

Clustering based on Graph Cuts*

I~ ~ fi YA--(~,7 +~

Jan Poland and Thomas Zeugmann

wS'f'fbÅ1ÑfvÑ

Graduate School of Information Science and Technology, Hokkaido University

Abstract: Clustering algorithms based on a matrix of pairwise similarities (kernel matrix) for the data are widely known and used, a particularly popular class being spectral clustering algorithms. Here we propose a clustering algorithm based on the SDP relaxation of the max-k-cut of the graph of pairwise distances. We compare the algorithm with a spectral relaxation of a norm-k-cut. Moreover, we propose a simple heuristic for dealing with missing data.

 

1 Introduction

Consider a set of n objects or data points, x1, ... , xn. We might not know anything about the objects, but assume that their pairwise distances dij = d(xi, xj) are known. Here, d: M x M  is a distance measure over a set M, i.e., d(x, y) > 0, d(x, y) = d(y, x) for all x, y E M, and d(x, y) = 0 iff x = y. Then we can cluster the data, i.e., as¬sign the xi to k distinct groups such that the dis¬tances within groups are small and the distances between the groups are large. This is done as fol¬lows. Construct a graph from the pairwise distances and choose an algorithm from the large class of re¬cently published methods based on graph-theoretic cut criteria. As the cuts are usually -hard to optimize, appropriate relaxations have been sub¬ject to intensive research. Two types of relaxations are particularly important:

(1) Spectral methods, where the top eigenvectors of the graph’s adjacency matrix are used to project the data into a lower dimensional space. This gives rise to new theoretical investigations of the popular spectral clustering algorithms.

(2) Semi-definite programming (SDP), where the discrete constraints of the cut criterion are replaced by continuous counterparts. Then convex solvers can be used for the optimization.

Surprisingly, all of the clustering approaches sug 

Support by JSPS 21st century COE program C01 and the MEXT Grand-in-Aid for Scientific Research on Priority Areas under Grant No. 18049001 is gratefully acknowledged. A Yc: T060-0814~Ligi~4LK4L14~'t-,R9TFI

E-mail: jan,thomas@ist.hokudai.ac.jp

 

gested so far work on a graph of similarities rather than distances. This means that, given the dis¬tances, we need one additional step to obtain simi¬larities from distances, e.g., by applying a Gaussian kernel. This also involves tuning the kernel width, a quantity which the clustering algorithm is quite sensitive to. Hence, it is natural to avoid this step by using a cut criterion that directly works with the distance graph, e.g., max-cut. We follow Frieze and Jerrum [4] and solve the max-cut problem via an SDP relaxation. We compare this method with a representative of spectral clustering algorithms, namely the spectral relaxation of the normalized cut criterion [12].

As a second contribution of this paper, we pro¬pose a simple heuristic for dealing with missing data, i.e., the case where some of the pairwise dis¬tances dij are unknown. Then, our aim is to substi¬tute the missing dij by a value which is most likely to leave the values of the cuts intact. This turns out to be the mean of the observed dij.

One motivation for considering missing data is given by the application we shall use to test the algorithms: Clustering of natural language terms using the Google distance. The Google distance [2] is a means of computing the pairwise distance of any searchable terms by just using the rela¬tive frequency count resulting from a web search. The Google API provides a convenient way for au¬tomating this process, however with a single key (which is obtained by prior registration) the maxi¬mum amount of daily queries is currently limited to 1000. Hence, by querying an incomplete sparse dis 

 

tance matrix rather than a full one, one can speed up considerably the overall process, as we will see.

The paper is structured as follows. In Section 2, we introduce the algorithm based on the max-k-cut relaxation, and recall some theory. We then briefly review the algorithm based on norm-k-cut in Sec¬tion 3. In Section 4 we address the missing data problem. Section 5 confronts the two algorithms, looking on the exact cut criteria rather than the relaxations, and compares the computational re¬sources required. In Section 6 we present experi¬mental results with the Google distance. Relation to other work is discussed and conclusions are given in Section 7.

2 Max-k-cut

Given a fully connected, weighted graph G= (V, D) with vertices V = x1, ... , xn and edge weights

D = dij  0  1  i, j  n which express pairwise distances, a k-way-cut is a partition of V into k disjoint subsets S1, ... , Sk. Here k is assumed to be given. We define the predicate A(i, j) = 0 if xi and xj happen to be in the same subset, i.e., if t[1 B  k, 1  i, j  n and i, j  SP], and A(i, j) = 1, otherwise.The weight of the cut (S1, ... , Sk) is defined as

n

dijA(i, j) .

i,j =1

The max-k-cut problem is the task of finding the

partition that maximizes the weight of the cut. It

can be stated as follows: Let a1, ... , ak k2 be

the vertices of a regular simplex, where

d = x d+1x2 = 1

is the d-dimensional unit sphere. Then the inner

product ai • aj =  1 

k1 whenever i = j. Hence,

finding the max-k-cut is equivalent to solving the

following integer program:

IP : maximize k1 

k dij(1 yi • yj)

i<j

subject to yj a1, ... , ak for all 1  j  n.

Frieze and Jerrum [4] propose the following semidef 

inite program (SDP) in order to relax the integer

program:

SDP: maximize k1 

k dij(1  vi •vj)

i<j

subject to vj n1 for all 1  j  n and

vi • vj  1 

k1 for all i = j

(necessary if k  3).

 

The constraints vi • vj  1 

k1 are necessary for k  3 because otherwise the SDP would prefer so-lutions where vi •vj = 1, resulting in a larger value of the objective. We shall see in the experimental part that this indeed would result in invalid approx-imations. The SDP finally can be reformulated as a convex program:

~CP : minimize dijYij (1a)

i<j

subject to Yjj = 1 for all 1  j  n, (1b)

Yij  1 

k1 for all i = j (if k  3) and (1c)

Y = (Yij)1<i,j<n satisfies Y  0. (1d)

Here, for the matrix Y nn the last condition Y  0 means that Y is positive semidefinite. Effi¬cient solvers are available for this kind of optimiza¬tion problems, such as CSDP [1] or SeDuMi [10]. In order to implement the constraints Yij  1  k1 with these solvers, positive slack variables Zij have to be introduced together with the equality con-straints Yij Zij =  1 

k1 .

Finally, for obtaining the partitioning from the vectors vj or the matrix Y, Frieze and Jerrum [4] propose to sample k points z1, ... , zk randomly on n1, representing the groups, and assign each vj to the closest group, i.e., the closest zj. They show ap-proximation guarantees generalizing those of Goe-mans and Williamson [5]. In practice however, the approximation guarantee does not necessarily im¬ply a good clustering, and applying the k-means algorithm for clustering the vj gives better results here. We use the kernel k-means (probably intro¬duced for the first time by [9]) which directly works on the scalar products Yij = vi • vj, without need of recovering the vj. We recapitulate the complete algorithm:

Algorithm. Clustering as an SDP relaxation of max-k-cut

Input: Distance matrix D = (dij) .

1. Solve the SDP via the CP (1a) through (1d).

2. Cluster the resulting matrix Y using kernel k-means.

3 Normalized k-cut

The normalized cut criterion has emerged as one of the most widely accepted cut criteria for clustering. It is defined on a graph G = (V, W) of pairwise sim¬ilarities rather than distances: W = wij  wij [0, 1], 1  i, j  n. Here, we identify the edges of G with their weights given by the similarities. For a k-way-cut, i.e., a partition of V into k dis¬joint subsets S1, ... , Sk, the norm-k-cut criterion is

 

defined as (cf. Yu and Shi [12])

~

~iS,e,j /S,e wij iS,e ,jV wij

=: 1  knassoc(S1, ... , Sk), (2)

where knassoc(S1, ... , Sk) is called the k-way nor-malized associations criterion. Therefore, minimiz¬ing the norm-k-cut value is equivalent to maximiz¬ing the norm-knassoc value.

Optimizing (2) can be restated as solving an in-teger problem, which is shown to be -hard. Yu and Shi [12] therefore derive a spectral relaxation, which we shall use in the following (see also [8]). We briefly point to the fact that the EM proce¬dure used in this spectral relaxation algorithm is different from, but nevertheless closely related to the k-means algorithm used in the algorithm based on max-k-cut and many other spectral clustering algorithms. For details, please see [12].

4 Missing data

Assume that either the distance matrix D or the similarity matrix W is not fully specified, but a portion of the off-diagonal entries is missing. One motivation for considering this case could be the desire to save resources by computing only part of the entries (e.g., for the Google distance dis¬cussed below, normal user registration permits only a limited amount of queries a day). Suppose that M = (Mij)1<i,j<n 0, 1n×n is a matrix such that Mii = 0 for all i = 1, ..., n and Mij = 1 if and only if dij (or wij, respectively) is not missing. As¬sume that the diagonal of D is zero and that of W is one, and denote the ith column of a matrix X by X [i]. Define the mean of the observed values,

D ¯ =  1

Ei,j Mij

W¯ =  1

Ei,j Mij

Then, replacing the missing entries in D with the value ¯D, the resulting distance matrix D is an un-biased estimate for the original full matrix, if the positions of the missing values are sampled from a uniform distribution. Hence, the resulting max-k-cut criterion for each partition is an unbiased es¬timate for the criterion respective to the original matrix, and this is the best we can do to achieve our goal that the optimal k-way-cuts of the original and the completed matrix are the same.

 

Also in the case of a similarity matrix W, the missing values should be replaced by the mean of the observed values. Asymptotically for n , this also yields an unbiased estimate for the norm-k-cut criterion. However, the reasoning is more difficult here, since the norm-k-cut criterion is a sum of quotients, and for two random variables X and Y, we have E[X/Y] = E[X]/E[Y ]. Still, the actual values of numerator and denominator are close to their expectations, as one can verify us¬ing concentration inequalities, e.g., Hoeffding’s in-equality. Then, for large n, with high probability the quotient is close to the corresponding quantity for the original (full) similarity matrix.

5 Max-k-cut versus norm-k-cut

In this section, we compare the max-cut and norm-cut criteria on distance and similarity matrices that are small enough to allow for a brute-force compu-tation of the exact criteria. We start from 10  10 matrices D0 and W0 consisting of two blocks of each size 5,

0 ••• 0 1 •••



 .

.

. . . . .

.

. .

.

. ..

.

D0 =



 0

1

..

. •••

•••

. . . 0

1

.

.

. 1

0

.

.

. •••

•••

. . .

1 ••• 1 0 •••

1 ••• 1 0 •••



 ... ... ... ... ...

W0 =

 1

0

... •••

•••

... 1

0

... 0

1

... •••

•••

...

 0 ••• 0 1 •••

From these matrices, distance matrices D and similarity matrices W are obtained by (1) perturb¬ing the value by Gaussian noise of varying am¬plitude, (2) making the matrices symmetric and rescaling them to the interval [0, 1], (3) removing a fraction of the off-diagonal values and replacing them by the mean of the remaining values. Another matrix we use for the norm-cut criterion is a kernel matrix obtained from the distance matrix using a Gaussian kernel, WD = exp( 1 

2σ D2) (all opera-tions are meant in the pointwise sense here). Since the values of the distance matrix are normalized to [0, 1], we use a fixed σ = 31. The missing values of WD are replaced by the mean of the observed values in W D.

All values displayed in Figures 1 through 3 below are means of 1500 independent samples. Figure 1 shows that, when using the max-cut criterion, the

 

Figure 1: Average fraction of incorrect cluster-ings by max-k-cut on a noisy distance matrix with missing data.

Figure 2: Difference of the average fraction of incorrect clusterings by norm-k-cut relative to max-k-cut, where the similarity matrix WD was obtained from the distance matrix D as WD = exp(1

2σ D2).

relative number of experiments that result in a dif-ferent clustering than the originally intended one, grows if either the noise amplitude or the fraction of missing values increases. Of course this was ex-pected. The max-cut criterion even yields always the correct clustering if both noise amplitude and missing data fraction are sufficiently low.

The same holds in principle for the norm-cut cri-terion, both for the directly generated similarity matrices W and for those matrices WD derived from the distance matrix by means of the Gaus¬sian kernel. However, in Figure 2, where the av¬erage difference of the error rates of the norm-cut clustering of WD to the max-cut clustering of D is displayed, we can see: The norm-cut clustering 

 

Figure 3: Average fraction of incorrect clus-terings by norm-k-cut: Difference of a WD = exp( 1 

2σ D2) matrix to a directly generated similarity matrix W.

always produces a higher error rate. The error rate is even more significantly higher for large fractions of missing values.

Did we introduce this increased error artificially by the additional transformation with the Gaussian kernel? Figure 3 indicates that this is not the case, as it shows nowhere a significantly positive value. Precisely, Figure 3 displays the difference of the er-ror rates of the norm-cut clusterings of WD relative to the directly generated matrices W.

Next, we turn to the computational resources re-quired by the algorithms. Both max-cut and norm-cut are -hard to optimize, so let us look at the relaxations. The spectral decomposition of a n  n matrix can be done in time O(n3), and if only the top k eigenvectors are desired, the effort can be even reduced to O(kn2) by an appropriate Lanc-zos method. Therefore the norm-cut/spectral al¬gorithm has quadratic or (depending on the imple¬mentation) at most cubic complexity.

On the other hand, solving the SDP in order to approximate max-cut is more expensive. The re-spective complexity is O(n3+ m3) (see [1]), where m is the number of constraints. If k = 2, then m = n and the overall complexity is cubic. However, for k  3, we need m = O(n2) constraints, resulting in an overall computational complexity of O(n6).

Finally, note that the analysis of the eigenvalues of the similarity matrix can yield a quite useful cri-terion to automatically determine the number k of clusters, in case that k is not known. We do not know of a corresponding method based on the dis-tance matrix. We shall not further discuss this issue here and assume in the following that k is known.

 

6 Experimental results with the Google distance

We evaluate both clustering algorithms from Sec-tions 2 and 3 on a set of natural language terms clustering tasks [2, 8]. We used the following datasets, which are all available at

http://www-alg.ist.hokudai.ac.jp/datasets.html.

The dataset people2 contains the names of 25 famous classical composers and 25 artists (i.e., two intended clusters), people3 contains all names from people2 plus 25 bestseller authors, people4 is ex¬tended by 25 mathematicians, and people5 ad¬ditionally contains 25 classical composers. The dataset alt-ds contains not terms in natural lan¬guage, but rather titles and authors’ last names from (almost all of) the papers from the ALT 2004 and DS 2004 conferences. Furthermore we use the datasets math-med-fin containing 20 terms each from the mathematical, medical, and finan¬cial terminology, finance-cs-j contains 20 finan¬cial and 10 computer science terms in Japanese, phil-avi-d has 98 terms from philately and 100 terms from aviation in German, and math-cuisine has 254 mathematical and 346 cuisine-related terms (in English). The distance matrices of the last two data sets are not fully given: in phil-avi only 50% of the entries are known, in math-cuisine it is only 30%.

For the norm-cut based algorithm, we need to convert the distance matrix to a similarity ma¬trix. We do this by using a Gaussian kernel WD = exp( 1 

2σ D2) and set the width parame¬ter σ = ¯D/2, which gives good results in prac¬tice. Another almost equally good choice is σ = 31, which can be justified by the fact that the Google distance is scale invariant and mostly in [0, 1].

Table 1 shows the number of clustering errors, i.e., the number of data points that are sorted to a different group than the intended one, respectively, on the data sets just described. One can see that both algorithms perform well in principle, in fact many of the “errors” displayed are in reality am-biguities of the data, e.g., the only misclustering in the math-med-fin data set concerns the term “average” which was intended to belong to the mathematical terms but ended up in the financial group.

We remark that the constraints (1c) and the re-sulting huge SDP size were really necessary in or¬der to get reasonable results: Without these con¬straints, e.g., clustering the people5 data set with the SDP algorithm, the resulting average number of errors is 36.

 

missing data

Figure 4: Comparison of the max-cut/SDP (dark bars) and the norm-cut/spectral algo-rithm (light bars) with variable fraction of missing data and variable number of clus¬ters and data set size, on the data sets people2-people5.

Looking on the computation times in Table 1 (measured on a 3 Ghz Pentium IV), the spectral method is clearly much faster than the SDP, in particular for k = 3 or more clusters. Here the quadratic number of constraints in the SDP and the resulting 6th order computation time are re¬ally expensive. Actually, the available SDP soft¬ware (CSDP, SeDuMi) do not even work at all with much larger problems if k  3.

Next we consider a situation with varying frac¬tion of missing data, shown in Figure 4 for the data sets people2-people5. Here the max-cut/SDP algorithm consistently outperforms the norm-cut/spectral algorithm, in particular if the number of clusters or the fraction of missing data grows. The same can be observed for other data sets. Both algorithms work quite well until about 70% missing data, after that the error increases sharply. The figure is based on 20 independent sam¬ples of missing data each, where the missing data locations were sampled in a balanced way such that each row and column of the distance matrix has the same fraction of missing values.

7 Relations to other work and Conclusions

There are many papers on clustering based on sim-ilarity matrices, in particular spectral clustering. It seems that norm-(k-)cut is quite established as an ideal criterion here, but there are different, such

 

data set information clustering errors and comp. time

name size #clusters missing data max-cut/SDP norm-cut/spectral

people2 50 2 0% 0 (4 sec) 0 (0.07 sec)

people3 75 3 0% 0 (90 sec) 2 (0.13 sec)

people4 100 4 0% 1 (876 sec) 5 (0.2 sec)

people5 125 5 0% 4 (2544 sec) 8 (0.35 sec)

alt-ds 64 2 0% 1 (3 sec) 1 (0.1 sec)

math-med-fin 60 3 0% 1 (36 sec) 1 (0.1 sec)

finance-cs-j 30 2 0% 4 (1.8 sec) 1 (0.05 sec)

phil-avi-d 198 2 50% 5 (12 sec) 6 (2 sec)

math-cuisine 600 2 70% 23 (137 sec) 22 (16.6 sec)

Table 1: Empirical comparison of the algorithms on the basic data sets (without removing additional data).

 

as min-max cut [3]. But also SDP has been used in connection with spectral clustering and kernels: [11] propose a SDP relaxation for norm-k-cut clus¬tering based on a similarity matrix, while [6] and [7] use SDP for completion and learning of kernel matrices, respectively.

To our knowledge, this is the first time that a dis-tance matrix and a max-(k-)cut criterion for simi¬lar clustering tasks have been used, which is natu¬ral in many applications where distances are given instead of similarities. We have seen that a SDP relaxation works quite well and yields results which tend to be superior to the spectral clustering re¬sults, in particular if the fraction of missing values grows. However, the SDP relaxation is expensive for k = 3 or more clusters. Thus we conclude with the open question of how to obtain a more efficient relaxation of max-k-cut, for instance a spectral one.

References

[1] B. Borchers and J. G. Young. Implementation of a primal-dual method for sdp on a shared memory parallel architecture. March 27, 2006.

[2] R. Cilibrasi and P. M. B. Vit´anyi. Automatic meaning discovery using Google. Manuscript, CWI, Amsterdam, 2006.

[3] C. H. Q. Ding, X. He, H. Zha, M. Gu, and H. D. Simon. A min-max cut algorithm for graph parti¬tioning and data clustering. In ICDM ’01: Proceed¬ings of the 2001 IEEE International Conference on Data Mining, pages 107–114. IEEE Computer So¬ciety, 2001.

[4] A. Frieze and M. Jerrum. Improved algorithms for MAX k-CUT and MAX BISECTION. Algorith-mica, 18(1):67–81, 1997.

[5] M. X. Goemans and D. P. Williamson. .879-approximation algorithms for MAX CUT and MAX 2SAT. In STOC ’94: Proceedings of the 

 

Twenty-Sixth Annual ACM Symposium on Theory of Computing, pages 422–431. ACM Press, 1994.

[6] T. Graepel. Kernel matrix completion by semidef-inite programming. In ICANN ’02: Proceedings of the International Conference on Artificial Neural Networks, pages 694–699. Springer-Verlag, 2002.

[7] G. R. G. Lanckriet, N. Cristianini, P. Bartlett, L. El Ghaoui, and M. I. Jordan. Learning the ker¬nel matrix with semidefinite programming. JMLR, 5:27–72, 2004.

[8] J. Poland and T. Zeugmann. Clustering pairwise distances with missing data: Maximum cuts versus normalized cuts. Accepted to the 9th International Conference on Discovery Science (DS), 2006.

[9] B. Sch¨olkopf, A. Smola, and K.-R. M¨uller. Non¬linear component analysis as a kernel eigenvalue problem. Neural Computation, 10(5):1299–1319, 1998.

[10] J.F. Sturm. Using SeDuMi 1.02, a MATLAB tool-box for optimization over symmetric cones. Op-timization Methods and Software, 11(12):625–653, 1999.

[11] E. P. Xing and M. I. Jordan. On semidefinite re-laxation for normalized k-cut and connections to spectral clustering. Technical Report UCB/CSD-03-1265, EECS Department, University of Califor¬nia, Berkeley, 2003.

[12] S. X. Yu and J. Shi. Multiclass spectral clustering. In ICCV ’03: Proceedings of the Ninth IEEE In¬ternational Conference on Computer Vision, pages 313–319. IEEE Computer Society, 2003.

 

Sensor management: Past, Present, and Future

Alfred O. Hero III, Fellow, IEEE and Douglas Cochran, Senior Member, IEEE

 

Abstract—Sensor systems typically operate under re¬source constraints that prevent the simultaneous use of all resources all of the time. Sensor management becomes relevant when the sensing system has the capability of actively managing these resources; i.e., changing its op¬erating configuration during deployment in reaction to previous measurements. Examples of systems in which sensor management is currently used or is likely to be used in the near future include autonomous robots, surveillance and reconnaissance networks, and waveform-agile radars. This paper provides an overview of the theory, algorithms, and applications of sensor management as it has developed over the past decades and as it stands today.

Index Terms—Active adaptive sensors, Plan-ahead sens¬ing, Sequential decision processes, Stochastic control, Multi-armed bandits, Reinforcement learning, Optimal decision policies, Multi-stage planning, Myopic planning, Information-optimized planning, Policy approximation, Radar waveform scheduling

I. INTRODUCTION

Advances in sensor technologies in the last quarter of the 20th century led to the emergence of large numbers of controllable degrees of freedom in sens¬ing devices. Large numbers of traditionally hard-wired characteristics, such as center frequency, bandwidth, beamform, sampling rate, and many other aspects of sensors’ operating modes started to be addressable via software command. The same period brought remarkable advances in networked systems as well as deployable autonomous and semi-autonomous vehicles instrumented with wide ranges of sensors and interconnected by networks, leading to configurable networked sensing systems. These trends, which affect a broad range of sensor types, modalities, and application regimes, have continued to the present day and appear unlikely to abate: new sensing concepts are increasingly manifested with device technologies and system architectures that are well suited to providing agility in their operation.

A. O. Hero is with the Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI 48109-2122, USA. D. Cochran is with the School of Mathematical and Statistical Sciences and the School of Electrical, Computer, and Energy Engineering, Arizona State University, Tempe, AZ 85287-5706, USA.

 

The term “sensor management,” as used in this pa¬per, refers to control of the degrees of freedom in an agile sensor system to satisfy operational constraints and achieve operational objectives. To accomplish this, one typically seeks a policy for determining the optimal sensor configuration at each time, within constraints, as a function of information available from prior mea-surements and possibly other sources. With this per-spective, the paper casts sensor management in terms of formulation and approximation of optimal planning policies. This point of view has led to a rich vein of research activity that extends and blends ideas from control, information theory, statistics, signal processing, and other areas of mathematical, statistical, and com-putational sciences and engineering. Our viewpoint is also slanted toward sensor management in large-scale surveillance and tracking systems for civilian and de¬fense applications. The approaches discussed have much broader utility, but the specific objectives, constraints, sensing modalities, and dynamical models considered in most of the work summarized here have been drawn from this application arena.

Within its scope of attention, the intention of this paper is to provide a high-level overview; references are given to guide the reader to derivations of mathematical results, detailed descriptions of algorithms, and specifications of application scenarios and systems. The list of references, while extensive, is not exhaustive; rather it is represen¬tative of key contributions that have shaped the field and led to its current state. Moreover, there are several areas relevant or related to sensor management that are not within the scope of this survey. These include purely heuristic approaches to sensor management and schedul¬ing as well as adaptive search methods, clinical treatment planning, human-in-the-loop systems such as relevance feedback learning, robotic vision and autonomous navi¬gation (path planning), compressive and distilled sensing, and robust sensing based on non-adaptive approaches.

The most comprehensive recent survey on sensor management of which the authors are aware is the 2008 book [1]. This volume consists of chapters written collaboratively by numerous current contributors to the field specifically to form a perspicuous overview of the main methods and some noteworthy applications. The 1998 survey paper by A. Cassandra [2], while not

 

2

 

devoted to sensor management, describes a few appli-cations of partially observed Markov decision process (POMDP) methods in the general area of sensor man-agement and scheduling, thereby illustrating conceptual connections between sensor management and the many other POMDP applications summarized in the paper. The earlier 1982 survey paper by G. E. Monahan [3] does not consider sensor management applications, but gives an excellent overview of the base of theory and algorithms for POMDPs as they were understood a few years before sensor management was becoming established as an appreciable area of research. A 2000 paper by G. W. Ng and K. H. Ng [4] provides an overview of sensor management from the perspective of sensor fusion as it stood at that time. This point of view, although not emphasized in this paper or in [1], continues to be of interest in the research literature. Another brief survey from this period is given by X.-X. Liu et al. in [5], and a short survey of emerging sensor concepts amenable to active sensor management is given in [6].

Several doctoral dissertations on the topic of sensor management have been written in the past fifteen years. Most of these include summaries of the state of the art and relevant literature at the time they were composed. Among these are the dissertations of G. A. McIntyre (1998) [7], D. Sinno (2000) [8], C. M. Kreucher (2005) [9], R. Rangarajan (2006) [10], D. Blatt (2007) [11],

J. L. Williams (2007) [12], M. Huber (2009) [13], and

K. L. Jenkins (2010) [14].

The remainder of this paper is organized as follows. Section II describes the basic goals and defines the main components of a sensor management system. In Section III, the emergence of sensor management is recounted within a historical context that includes both the advancement of statistical methods for sequential definition, collection, and analysis of samples and the rise of sensor technologies and sensing applications enabling and calling for sensor management. Section IV gives an overview of some of the current state of the art and trends in sensor management and Section V describes some of the future challenges and opportunities faced by researchers in the field.

II. DESCRIPTION OF SENSOR MANAGEMENT

The defining function of sensor management is dy-namic selection of a sensor, from among a set of available sensors, to use at each time during a mea-surement period in order to optimize some metric of performance. Time is usually partitioned into a sequence of epochs and one sensor is to be chosen in each epoch, thereby creating a discrete-time problem. The 

 

term “sensor management” most often refers to closed-loop solutions to problems of this nature; i.e, the next sensor to employ is chosen while the sensor system is in operation and in view of the results obtained from prior sensor measurements. The term “sensor scheduling” is sometimes used to refer to feed-forward schemes for sensor selection, though this usage is not standardized and the two expressions are used interchangeably in some literature. In current applications of sensor man-agement, and especially in envisioned future applica-tions, the sensors available for selection in each time epoch are actually virtual sensors, each representing one choice of configuration parameters affecting the physical configurations and operating modes of a collection of sensors, sensor suites, sensor platforms, and the way data are processed and communicated among interconnected subsystems. With this perspective, selecting a sensor really means determining the values to which the avail-able controllable degrees of freedom in a sensor system should be set.

Figure 1 illustrates the basic elements and operation of a closed-loop sensor management system. Once a sensor is selected and a measurement is made, information relevant to the sensing objective is distilled from the raw sensor data. This generally entails fusion of data representing disparate sensing modalities (e.g., optical and acoustic) and other properties, and further combining it with information gleaned from past measurements and possibly also side information from sources extrinsic to the sensor system. The fusion and signal processing com-ponents of the loop may produce ancillary information, such as target tracks or decisions about matters external to the sensor manager (e.g., direct an aircraft to take eva-sive action to avoid collision). For the purposes of sensor management, they must yield a state of information on the basis of which the merit of each possible sensor selection in the next time epoch may be quantified. Such quantification takes many forms in current approaches, from statistical (e.g., mean risk or information gain) to purely heuristic. From this point, the sensor manager must optimize its decision as to which sensor to select for the next measurement.

The notion of state is worthy of a few additional words. Heuristically, the state of information should rep-resent all that is known about the scenario being sensed, or at least all that is relevant to the objective. Often this includes information about the physical state of the sensor system itself (e.g., the position and orientation of the air vehicle carrying one of the video sensors), which may constrain what actions are possible in the next step and thus the set of virtual sensors available to select in the upcoming epoch. Knowledge of the physical state

 

Fig. 1. Conceptual block diagram of a sensor management system. The sensor selector selects among sensor actions S1, S2, and S3 based on the output of the optimizer. The optimizer attempts to optimize a system performance metric, such as information gain or mean risk associated with decisions or estimates produced by signal processing algorithms that operate on fused sensor data.

 

frequently has utility extrinsic to the sensor manager, so some literature distinguishes physical and information states and their coupled dynamical models as depicted in Figure 2. This diagram evinces the similarity of sensor management and feedback control in many important respects, and indeed control theory is an important ingredient in current perspectives on sensor management. But sensor management entails certain aspects that give it a distinctive character. Chief among these is in the role of sensing. In traditional feedback control, sensors are used to ascertain information about the state of a dynamical plant. This information informs the control action through a control law or policy which in turn affects the state. In sensor management, the state of information is directly affected by the control action; i.e., rather than helping to decide what control action to invoke, the act of sensing is itself the control action.

Sensor management is motivated and enabled by a small number of essential elements. The following para 

Control State

Sensor

Management

Fig. 2. A control-theoretic view of sensor management casts the problem as that of optimally controlling a state, sometimes regarded as consisting of separate information and physical components, through the selection of measurement actions.

 

graphs describe these and explain the roles they play in the current state of the subject. First, a summary of waveform-agile radar is given to provide the context of a current application for the more general descriptions that follow.

A. Sensor management application – Waveform-agile radar

Among the most well developed focus applications of sensor management is real-time closed-loop scheduling of radar resources. The primary feature of radar systems that makes them well suited for sensor management is that they offer several controllable degrees of freedom. Most modern radars employ antenna arrays for both the transmitter and receiver, which often share the same antenna. This allows the illumination pattern on transmit as well as the beam pattern on receive to be adjusted sim-ply by changing parameters in a combining algorithm. This ability has been capitalized upon, for example, by adaptive signal processing techniques such as adaptive beamforming on both transmit and receive and more recently by space-time adaptive processing (STAP). The ability for the transmitter to change waveforms in a limited way, such as switching between a few pre-defined waveforms in a library, has existed in a few radar systems for decades. Current radar concepts allow transmission of essentially arbitrary waveforms, with constraints coming principally from hardware limitations such as bandwidth and amplifier power. They also remove traditional re¬strictions that force the set of transmit antenna elements to be treated as a phased array (i.e., all emitting the same waveform except for phase factors that steer the beam pattern), thereby engendering the possibility of the

 

4

 

transmit antennas simultaneously emitting completely different waveforms. This forms the basis of one form of so-called multi-input multi-output (MIMO) radar.

Two more aspects of the radar application stand out in making it a good candidate for sensor management. One is that pulse-Doppler radars have discrete time epochs intrinsically defined by their pulse repetition intervals and often also by their revisit intervals [15]. Also, in radar target tracking applications there are usually well defined performance metrics and well developed dynam-ical models for the evolution of the targets’ positions, velocities, and other state variables. These metrics and models directly enhance the sensor manager’s ability to quantitatively predict the value of candidate measure-ments before they are taken.

In view of these appealing features, it is no surprise that radar applications have received a large amount of attention as sensor management has developed. The idea of changing the transmitted waveform in a radar system in an automated fashion in consideration of the echo returns from previously transmitted waveforms dates to at least the 1960s, though most evidence of this is anecdotal rather than being documented in the research literature. The current generation of literature on closed-loop waveform management as a sensor management application began with papers of D. J. Kershaw and R. J. Evans [16], [17] and S. M. Sowelam and A. H. Tew-fik [18], [19] in the mid-1990s, roughly corresponding to the ascension of sensor management literature in broader contexts. Among the early sensor management papers that focused on closed-loop beam pattern management were those of V. Krishnamurthy and Evans in the early 2000s [20], [21]. Several contributions by numerous authors on these and related radar sensor management applications have appeared in the past decade. Among the topics addressed in this recent literature are radar waveform scheduling for target identification [22], target tracking [23], clutter and interference mitigation [24], [25], and simultaneously estimating and tracking pa-rameters associated with multiple extended targets [26]. There has also been recent interest in drawing insights for active radar and sonar sensor management from biological echolocation systems [27] and in designing optimal libraries of waveforms for use with radar systems that support closed-loop waveform scheduling [28].

B. Controllable Degrees of Freedom

Degrees of freedom in a sensor system over which control can be exercised with the system in operation provide the mechanism through which sensors can be managed. In envisioned applications, they include di¬ 

 

verse sets of parameters, including physical configu¬ration of the sensor suite, signal transmission charac¬teristics such as waveform or modulation type, signal reception descriptors ranging from simple on/off state to sophisticated properties like beamform. They also include algorithmic parameters that affect local versus centralized processing trade-offs, data sharing protocols and communication schemes, and typically numerous signal processing choices.

Many characteristics of current and anticipated sensor systems that are controllable during real-time opera¬tion were traditionally associated with subsystems that were designed independently. Until relatively recently, transduction of physical phenomena into electrical sig-nals, analog processing, conversion to digital format, and digital processing at various levels of information abstraction were optimized according to performance criteria that were often only loosely connected with the performance of the integrated system in its intended function. Further integrated operation of such subsystems generally consisted of passing data downstream from one to the next in a feed-forward fashion. Integrated real-time authority over controllable degrees of freedom spanning all of this functionality not only allows joint optimization of systemic performance metrics but also accommodates adaptation to changing objectives.

In the radar sensor management example, the ease and immediacy of access (i.e., via software command) to crucial operating parameters such as antenna patterns and waveforms provides the means by which a well conceived algorithm can manage the radar in each time epoch.

C. Constraints

The utility of sensor management emerges when it is not possible to process, or even collect, all the data all the time. Operating configurations of individual sensors or entire sensor systems may be intrinsically mutually exclusive; e.g., the transmitter platform can be in position A or in position B at the time the next waveform is emitted, but not both. One point of view on configurable sensors, discussed in [29], imagines an immense suite of virtual sensor systems, each defined by a particular operating configuration of the set of physical sensors that comprises the suite. Limitations preventing an individual sensor from being in multiple configurations at the same time are seen as constraints to be respected in optimizing the configuration of the virtual sensor suite. This is exactly the case in the waveform-agile radar example, where only one waveform can be transmitted on each antenna element at any given time.

 

5

 

Restrictions on communications and processing re-sources almost always constrain what signal processing is possible in networked sensor applications. Collecting all raw data at a single fusion center is seldom possible due to bandwidth limitations, and often to constraints imposed by the life and current production of batteries as well. So it is desirable to compress raw data before transmission. But reducing the data at the nodes requires on-board processing, which is typically also a limited resource.

D. Objective Quantification

When controllable degrees of freedom and constraints are present, sensor management is possible and war-ranted. In such a situation, one would hope to treat the selection of which sensing action to invoke as an optimization problem. But doing so requires the merit of each possible selection to be represented in such a way that comparison is possible; e.g., by the value of a cost or objective functional.

The value of a specified set of data collection and processing choices generally depends on what is to be achieved. For example, one set of measurements by a configurable chemical sensor suite may be of great value in determining whether or not an analyte is an explosive, but the best data to collect to determine the species of a specimen already known to be an explosive may be quite different. Moreover, the objective may vary with time or state of knowledge: once a substance is determined to be an explosive, the goal shifts to determining what kind of explosive it is, then how much is present, then precisely where it is located, etc. Consequently, predic¬tively quantifying the value of the information that will be obtained by the selection or a particular sensing action is usually difficult and, at least in principle, requires a separate metric for each sensing objective that the system may be used to address. The use of surrogate metrics, such as information gain discussed in Section IV, has proven effective in some applications. With this approach, the role of a metric designed specifically for a particular sensing objective is undertaken by a proxy, usually based on information theoretic measures, that is suited to a broader class of objectives. This approach sacrifices specificity in exchange for relative simplicity and robustness, especially to model mismatch.

Management of radar beamforms and waveforms for target tracking, though not trivial, is one of the most tractable settings for objective quantification. The pa-rameters can be chosen to optimize some function of the track error covariance, such as its expected trace or determinant, at one or more future times; e.g., after 

 

the next measurement, after five measurement epochs, or averaged over the next ten epochs. Computation or ap-proximation of such functions is assisted by the tracker’s underlying model for the dynamical evolution of the target states. The use and effectiveness of waveform management in such applications is discussed in [1, Ch. 10], which also cites numerous references.

III. HISTORICAL ROOTS OF SENSOR MANAGEMENT

It has long been recognized that appropriate collec¬tion of data is essential in the design of experiments to test hypotheses and estimate quantities of interest. R. A. Fisher’s classical work [30], which encapsulated most of the ideas on statistical design of experiments developed through the first part of the 20th century, primarily addressed the situation in which the compo¬sition of the sample to be collected is to be determined in advance of the experiment. In the early 1950s, the idea of using closed-loop strategies in experiment de¬sign emerged in connection with sequential design of experiments. In his 1951 address to the Meeting of the American Mathematical Society [31], H. Robbins observed:

A major advance now appears to be in the making with the creation of a theory of the sequential design of experiments, in which the size and composition of the samples are not fixed in advance but are functions of the observations themselves.

Robbins attributes the first application of this idea to Dodge and Romig in 1929 [32] in the context of indus-trial quality control. They proposed a double sampling scheme in which an initial sample is collected and analyzed, then a determination about whether to collect a second sample is based on analysis of the first sample. This insight was an early precursor to the development of sequential analysis by Wald and others during the 1940s [33], and ultimately to modern methods in statistical signal processing such as sequential detection [34]. In the interim, H. Chernoff made substantial advances in the statistical study of optimal design of sequences of experiments, particularly for hypothesis testing and parameter estimation [35], [36]. Many results in this vein are included in his 1972 book [37]. Also in 1972, V. V. Fedorov’s book [38] presented an overview of key results, many from his own research, in optimal experimental design up to that time. The relevance of a portion of Fedorov’s work to the current state of sensor management is noted in Section IV.

One view of the raison d’être for sensors, particularly among practitioners of sensor signal processing, is to

 

6

 

collect samples to which statistical tests and estimators may be applied. From this perspective, the advancement of sensor signal processing over the latter half of the 20th century paralleled that of experimental design. By the early 1990s, a rich literature on detection, estimation, classification, target tracking and related problems had been compiled. Nearly all of this work was predicated on the assumption that the data were given and the goal was to process it in ways that are optimally informative in the context of a given application. There were a few notable cases in which it was assumed the process of data collection could be affected in a closed-loop fashion based on data already collected. In sequential detection theory, for example, the data collection is continued or terminated at a given time instant (i.e., binary feedback) depending on whether a desired level of confidence about the fidelity of the detection decision is supported by data already collected. An early example of closed-loop data collection involving a dynamic state was the “measurement adaptive problem” treated by L. Meier et al. in 1967 [39]. This work sought to simultaneously optimize control of a dynamic plant and the process of collecting measurements for use in feedback. Another is given in a 1972 paper of M. Athans [40] that considers optimal closed-loop selection of the linear measurement map in a Kalman filtering problem.

One of the first contexts in which the term “sensor management” was used in the sense of this discussion1 was in automating control of the sensor systems in military aircraft (see, e.g., [42]). In this application, the constrained resource is the attention of the pilot, particularly during hostile engagement with multiple adversaries, and the objective of sensor management is to control sensor resources in such a way that the most important information (e.g., the most urgent threats) are emphasized in presentation to the pilot. Applications associated with situational awareness for military aircraft continue to be of interest, and this early vein of applica-tion impetus expanded throughout the 1990s to include scheduling and management of aircraft-based sensor assets for surveillance and reconnaissance missions (see, e.g., [43], [44] and [1, ch. 11]).

Also beginning in the 1980s, sensor management was actively pursued under the label of “active vision” for applications in robotics [45]. This work sought to exercise feedback control over camera direction and sometimes other basic parameters (e.g., zoom or focal distance) to improve the ability of robotic vision systems

1The phrase comes up in various literature in ways that are related to varying degrees to our use in this paper. To maintain focus, we have omitted loosely related uses of the term, such as in clinical patient screening applications [41].

 

to contribute to navigation, manipulation, and other tasks entailed in the robot’s intended functionality.

The rapid growth of interest in sensor management be-ginning in the 1990s can be attributed in large part to de-velopments in sensor and communications technologies. New generations of sensors, encompassing numerous sensing modalities, are increasingly agile. Key operating parameters, once hard-wired, can be almost instantly changed by software command. Further, transducers can be packaged with A/D converters and microprocessors in energy efficient configurations, in some cases on a single chip, creating sensors that permit on-board adaptive processing involving dynamic orchestration of all these components. At the same time, the growth of networks of sensors and mobile sensor platforms is contributing even more controllable degrees of freedom that can be managed across entire sensor systems. From a purely mathematical point of view, it is almost always advantageous to collect all available data in one location (i.e., a “fusion center”) for signal processing. In to-day’s sensor systems, this is seldom possible because of constraints on computational resources, communication bandwidth, energy, deployment pattern, platform motion, and many other aspects of the system configuration. Even highly agile sensor devices are constrained to choose only one configuration from among a large collection of possibilities at any given time.

These years spawned sensor management approaches based on the modeling sensor management as a deci¬sion process, a perspective that underpins most current methods as noted in Section IV. Viewing sensor man-agement in this way enabled tapping into a corpus of knowledge on control of decision processes, Markov decision processes in particular, that was already well established at the time [46]. Initial treatments of sensor management problems via POMDPs, beginning with D. Castañón’s 1997 paper [47], were followed shortly by other POMDP-based ideas such as the work of J. S. Evans and Krishnamurthy published in 2001–2002 [48], [49]. These were the early constituents of a steady stream of contributions to the state of the art summarized in Section IV-B. A formidable obstacle to the practicality of the POMDP approach is the computational complexity entailed in its implementation, particularly for methods that look more than one step ahead. Consequently, the need for approximation schemes and the potential merit of heuristics to provide computational tractability was recognized from the earliest work in this vein.

The multi-armed bandit (MAB) problem is an impor-tant exemplar of a class of multi-stage decision problems where actions yielding large immediate rewards must be balanced with others whose immediate rewards are

 

7

 

smaller, but which hold the potential for greater long¬term payoff. While two-armed and MAB problems had been studied in previous literature, the origin of index policy solutions to MAB problems dates to J. C. Gittins in 1979 [50]. As discussed in Section IV-C, under certain assumptions, an index solution assigns a numerical index to each possible action at the current stage of an infinitely long sequence of plays of a MAB. The indices can be computed by solving a set of simpler one-armed bandit problems and their availability reduces the decision at each stage to choosing the action with the largest index. The optimality of Gittins’ index scheme was addressed by P. Whittle in 1980 [51].

As with POMDPs, the MAB perspective on sensor management started receiving considerable research at-tention around 2000. Early applications of MAB method-ology to sensor management include the work of Kr-ishnamurthy and R. J. Evans [20], [21] who considered a multi-armed bandit model with Markov dynamics for radar beam scheduling. The 2002 work of R. Washburn et al. [52], although written in the context of more general dynamic resource management problems, was influential in the early develop of MAB approaches to sensor management.

A theory of information based on entropy concepts was introduced by C. E. Shannon in his classic 1948 paper [53] and was subsequently extended and applied by many others, mostly in connection with communi¬cation engineering. Although Shannon’s theory is quite different than that of Fisher, sensor management has leveraged both in various developments of information-optimized methods. These were introduced specifically to sensor management in the early 1990s by J. Manyika and H. Durrant-Whyte [54] and by W. W. Schmaedeke [55]. As remarked in Section IV, information-based ideas were applied to particular problems related to sensor management even earlier. Fisher’s information theory was instrumental in the development of the theory of optimal design of experiments, and numerous examples of applications of this methodology have appeared since 2000; e.g., [56], [57]. Measures of information led to sensor management schemes based on information gain, which developed into one of the central thrusts of sensor management research over the past decade. Some of this work is summarized in Section IV-D, and a more complete overview of these methods in provided in [1, Ch. 3].

From foundations drawing on several more classical fields of study, sensor management has developed into a well-defined area of research that stands today at the crossroads of the disciplines upon which it has been built. Key approaches that are generally known to researchers 

 

in the area are discussed in the following section of this paper. But sensor management is an active discipline, with new work and new ideas appearing regularly in the literature. Some noteworthy recent developments include work by V. Gupta et al. which introduces random scheduling algorithms that seek optimal mean steady state performance in the presence of probabilistically modeled effects [58], [59], [60]. K. L. Jenkins et al. very recently proposed the use of random set ideas, similar to those applied in some approaches to multi-target track-ing, in sensor management [61], [62]. These preliminary investigations have resulted in highly efficient algorithms for certain object classification problems. Also very recently, D. Hitchings et al. introduced new stochastic control approximation schemes to obtain tractable algo-rithms for sensor management based on receding horizon control formulations [63]. They also proposed a stochas-tic control approach for sensor management problems with large, continuous-valued state and decision spaces [64].

Despite ongoing progress, sensor management still holds many unresolved challenges. Some of these are discussed in Section V.

IV. STATE OF THE ART IN SENSOR MANAGEMENT

The theory of decision processes provides a unifying perspective for the state of the art in sensor management research today. A decision process, described in more detail below, is a time sequence of measurements and control actions in which each action in the sequence is followed by a measurement acquired as a result of the previous action. With this perspective, the design of a sensor manager is formulated as the specification of a decision rule, often called a policy, that generates realizations of the decision process. An optimal policy will generate decision processes that, on the average, will maximize an expected reward; e.g., the negative mean-squared tracking error or the probability of detection. A sound approach to sensor management will either approximate an optimal policy in some way or else at-tempt to analyze the performance of a proposed heuristic policy. In this section we will describe some current approaches to design of sensor management policies. The starting point is a formal definition of a decision process.

A. Sensor management as a decision process

Assume that a sensor collects a data sample yt+1 at time t after taking a sensing action at. It is typically assumed that the possible actions are se-lected from a finite action space , that may change

 

8

 

over time. The selected action ak depends only on past samples yk, yk1.... , y1 and past actions ak1, ak2...., a0, and the initial action a0 is de¬termined offline. The function that maps previous data samples and actions to current actions is called a policy. That is, at any time t, a policy specifies a mapping -yt and, for a specific set of samples, an action at = -yt(akk<t, ykk<t). A decision process is a sequence ak, yk+1k0 = a0, y1, a1, y2, a2, y3 ..., , which is typically random and can be viewed as a realization from some generative model specified by the policy and the sensor measurement statistics.

A well designed sensor manager will formulate the policy with the objective of maximizing an av¬erage reward. The reward at time t is a function Rt(akk<t, skk<t) of the action sequence akk>0 and a state sequence skk>0, describing the environ¬ment or a target in the environment. The state sk might be continuous (e.g., the position of a moving target) or discrete (e.g., sk = 1 when the target is moving and sk = 0 when it is not moving). It is customary to model the state as random and the data sample yk as having been generated by the state sk in some random manner. In this case, there exists a conditional distribution of the state sequence given the data sequence and the average reward at time t can be defined through the statistical expectation E[Rt(akk<t, skk<t)].

An optimal action policy will maximize the average award at each time t during the sensor deployment time period. The associated optimization must be performed over the set of mappings -yt defined on the cartesian product spaces tk=1 k1  and mapping to t for t = 0, 1, .... Due to the high dimensionality of the cartesian product spaces, no tractable methods exist for determining optimal action policies under this degree of generality. Additional assumptions on the statistical distributions of the decision process and state process are needed to reduce the dimensionality of the optimization spaces.

When the unknown state sk is not recoverable from yk then the decision process is called a partially observ¬able decision process. The partially observable case is common in actual sensing systems where the measure-ments yk are typically contaminated by noise or clutter. However, policy optimization generally presents more mathematical difficulties in the partially observable case than in the perfectly observable case.

B. Markov decision processes

A natural way to simplify the task of policy optimiza-tion is to assume that the general decision process de¬ 

 

scribed in Section IV-A satisfies some additional Marko-vian properties. To make the general decision process Markovian one imposes the assumption that the state sequence is dependent only on the most recent state and action given the entire past. Specifically, we assume that P(st+1sk, akk<t) = P(st+1st, at), the conditional state transition probability, and P(ytsk, akk<t) = P(ytst, at), the measurement likelihood function given action at.

We additionally restrict the reward to be additive over time and only consider policies that depend on the most recent measurement, i.e., Rt(akk<t, skk<t) = ~tk=0 Rt(ak, sk) and the associated mapping -yt is re¬stricted to be from t1 to t. When the state can be recovered from the measurements the resultant process is called a Markov decision process (MDP). When the state is not recoverable from the measurements the resultant process is called a partially observable Markov decision process (POMDP).

For MDP or POMDP models the optimal restricted policy can be determined by backwards induction over time. In particular, there is a compact recursive formula, known as Bellman’s equation, for determining the map-ping -yt1 from the mapping -yt. In special cases where the state and the measurements obey standard dynamical stochastic state models (e.g., the linear-Gaussian model assumed in the Kalman filter), this optimal restricted policy is in fact the overall optimal policy. That is, the overall optimal policy only depends on the most recent measurements. Furthermore, as shown by E. J. Sondik [65], the optimal policy can be found by linear pro-gramming. For more details on MDPs, POMDPs, and Bellman’s equation and solutions, the reader is referred to [1, Ch. 2].

As noted in Section III, the application of MDP and POMDP methods to sensor management problems can be traced back to the mid 1990’s. In their 1994 overview of the field of sensor management [43], S. Musick and R. Malhotra suggested that a comprehensive mathe¬matical framework was needed to assess and optimize scheduling over sensor and inter-sensor actions. Antici¬pating the future application of POMDP and reinforce¬ment learning approaches, went on to suggest adaptive control, state space representations, and mathematical programming as the components of a promising frame¬work. However, to be applied to practical large scale sensor management problems approximate solutions to the POMDP would be necessary. Castañón’s 1997 policy rollout approximation [47] was the earliest successful application of the POMDP to sensor management.

Several types of approximations to the optimal POMDP sensor management solution are discussed in

 

9

 

[1, Ch. 2] under the heading of approximate dynamic programming. These include: offline learning, rollout, and problem approximation techniques. Offline learning techniques use offline simulation to explore the space of policies and include the large class of reinforcement learning methods [44], [66], [67]. Rollout uses real-time simulation to approximate the rewards of a sub¬optimal policy [47], [68]. Problem approximation uses a simpler approximate model or reward function for the POMDP as a proxy for the original problem and includes bandit and information gain approaches, discussed in the following subsections.

POMDP approaches have been applied to many dif-ferent sensing systems. One of the most active areas of application has been distributed multiple target tracking, see for example [69] and references therein. When target dynamics are non-linear and environments are dynami-cally changing, the states of targets can be tracked by a particle filter [70]. This filter produces an estimate of the posterior density of the target tracks that is used by the scheduler to predict the value of different sensing actions [67]. Managers for many other sensing systems have been implemented using POMDPs and reinforcement learning, for example, multifunction radar [71], underwater sensing applications [72], passive radar [73], and air traffic management [74].

C. Multi-armed bandit decision processes

A multi-armed bandit (MAB) is a model for sequential resource allocation in which multiple resources (the arms of the bandit) are allocated to multiple tasks by a controller (also called a processor). When a particular arm at of the bandit is pulled (a control action called a “play”) at time t the MAB transitions to a random state xt and pays out a reward depending on the state. As in a MDP, successive MAB control actions produce a sequence of actions and states. When the MAB action-state sequence is Markovian it is a special case of a MDP or POMDP process.

In some cases, the optimal policy for a k-arm MAB problem can be shown to reduce to a so-called index policy. An index policy is a simpler mapping that assigns a score (or index) to each arm of the MAB and pulls only the arm having maximum score at a given time. The key to the simplification is that these scores, the Gittins indices mentioned in Section III, can be determined by solving a much simpler set of k different single-armed bandit problems. Gittins index policies exist when the actions are not irrevocable; meaning that any available actions not taken at the present time can be deferred to the future, producing the same sequence of future 

 

rewards, except for a discount factor. The significance of Gittins index policies is that they are frequently much simpler to compute than backwards induction solutions to optimal policies for MDPs and POMDPs. Thus they are sometimes used to approximate these optimal poli¬cies; e.g., using rollout with MAB index-rules as the base policy [75]. See [1, Ch. 6] for further discussion of index policies and their variants.

As a simple example, consider the aforementioned wide area search problem for the case of a single non-moving target that could be located in one of k locations with equal probability. Assume that in each time epoch a sensor can look at a single location with specified probabilities of correct detection and false alarm. Further assume that the reward is decreasing in the amount of time required by the sensor to correctly find the target. Identify each sensing action (location) as an arm of the MAB and the un-normalized posterior probability of the true target location as the state of the MAB. Under these assumptions, the optimal MAB policy for selecting arms is an index policy and specifies the optimal wide area search scheduler. For further details on this application of MAB to sensor management see [1, Ch. 7].

Bandit models were proposed for search problems like the above several decades ago [76], but their application to sensor management is relatively recent. Early applications of the multi-armed bandit model to sensor management were Krishnamurthy’s treatment of the radar beam scheduling for multiple target tracking problem [20], [21] and Washburn et al.’s application to general problems of sensor resource management [52]. As another example, arm acquiring bandits have been proposed by Washburn [1, Ch. 7] for tracking targets that can appear or disappear from the scene. Also discussed in [1, Ch. 7] are restless bandits, multi-armed bandits in which the states of the arms not played can evolve in time. Sensor management application of restless bandits include radar sensor management for multi-target tracking (see, e.g., [77]).

D. Information-optimized decision processes

The MDP/POMDP and MAB approaches to sensor management involve searching over multi-stage look-ahead policies. Designing a multi-stage policy requires evaluating each available action in terms of its impact on the potential rewards for all future actions. Myopic sensor management policies have been investigated as low complexity alternatives to multi-stage policies. My-opic policies only look ahead to the next stage; i.e., they compute the expected reward in the immediate future to determine the best current action. Such greedy policies

 

10

 

benefit from computational simplicity, but at the expense of performance loss compared to multi-stage optimal policies. Often this loss is significant. However, there are cases where the myopic loss approach gives acceptable performance, and indeed is almost optimal in special cases.

The most obvious way to obtain myopic sensor scheduling policies is to only consider the effect of the control action on the immediate reward; i.e., to truncate the future reward sequence in the multi-stage POMDP scheduling problem. This approach is called the optimal one-step look-ahead policy. However, it has often been observed that a myopic policy can achieve better overall performance by maximizing a surrogate reward, such as the mutual information between the data and the target. The information gain, discussed in more detail below, has the advantage that it is a more fundamental quantity than a task-specific reward function. For example, unlike many reward functions associated with estimation or detection algorithms, the mutual information is invariant to invertible transformations of the data. This and other properties lead to myopic policies that are more robust to factors such as model mismatch and dynamically chang-ing system objectives (e.g., detection versus tracking), while ensuring a minimal level of system performance. For further motivation and properties of information theoretic measures for sensor management the reader may wish to consult [1, Ch. 3].

Information theoretic measures have a long history in sensor management. Optimization of Fisher information was applied to the related problem of optimal design of experiments (DOE) by Fisher, discussed in Section III, in the early part of the twentieth century [30]. Various functions of the Fisher information matrix, including its determinant and trace, have been used as reward functions for optimal DOE [38]. More recently, sensor management applications of optimal DOE have been proposed; e.g., in managed sensor fusion [78], in sensor managed unexploded ordnance (UXO) detection [56], in multi-sensor scheduling [79], and in sensor manage¬ment for robotic vision and navigation [57]. However, Fisher information approaches to sensor management have several drawbacks. Notable among these are that the Fisher information requires specification of a parametric model for the observations. It is also a local measure of information that does not apply to discrete targets or mixtures of discrete and continuous valued targets. Model mismatch and/or discrete valued quantities fre-quently arise in sensor management applications. For ex-ample, discrete values arise when there is categorical side information about the target or clutter, or a target that transitions between two states like stopping and moving. 

 

These are principal reasons that non-local information measures such as entropy and mutual information have become more common in sensor management.

In his 1998 PhD thesis [7], McIntyre cites the work of Barker [80] and Hintz and McVey [81] as the first to apply entropy to sensor management problems in 1977 and 1991, respectively. However, while the problems they treated are special cases of sensor management, they did not treat the general sensor management problem nor did they use the term in their papers. The first papers we know of that applied entropy measures explicitly to sen-sor management were Manyika and Durrant-Whyte [78] in 1992 and Schmaedeke [55] in 1993. The information measure used in these papers was the expected update in posterior entropy, called the information gain, that is associated with a given candidate sensor action.

These early information theoretic sensor management papers assumed Gaussian observations and linear dynam-ics, in which case the entropy and information gain have closed form mathematical expressions. Subsequently, the linear-Gaussian assumptions have been relaxed by using non-parametric estimation of entropy and information gain. Other information gain measures have also been in-troduced, such as the Kullback-Leibler (KL) divergence, the KL discrimination, and the Rényi entropy. The reader can consult the book [1] and, in particular, early papers by Schmaedeke and K. Kastella [82], [83], R. Mahler [84], Hintz and McIntyre [85], and Kreucher et al. [86], [87].

At first information gain sensor management methods were focused on single modality tracking of simple passive targets. In recent years, information gain has been applied to increasingly general models and sen¬sor management tasks. For example, information driven methods have been applied to dynamic collaborative sensing with communication costs [88], multi-sensor information fusion [89], target tracking with uncertain sensor responses [90], multi-target tracking in large dy-namic sensor networks [91], multi-modality multi-target tracking with time varying attenuation and obscuration [92], [67], robot path planning [93], and active camera control for object recognition and tracking using mutual information [94].

A striking mathematical result on the capabilities of information driven sensor management was obtained by J. L. Williams et al. [95] in connection with the general problem of information gathering in the context of graphical models under the assumption of condi¬tionally independent measurements. In 2005 Guestrin et al. [96] showed that the conditional mutual information is submodular in the context of general machine learning problems. The significance of this result for sensor

 

11

 

management is that the maximizer of a submodular ob-jective function can be well approximated using greedy optimization algorithms. Using this insight, Williams et al. established in [95] that greedy sequential methods for measurement planning are guaranteed to perform within a factor of 1/2 of the optimal multi-stage selection method. Furthermore, this bound is independent of the length of the planning horizon and is sharp. The remark¬able results of [95] are significant in that they provide theoretical justification for the computationally simpler myopic strategy and provide the designer with a tool to gauge the expected loss with respect to the optimal, but intractable, multi-stage policy. The bound was used to design resource constrained, information driven sensor management algorithms that exploit the submodularity property. The algorithm monotonically reduces an upper bound on the optimal solution that permits the system designer to terminate computation early with a near-optimal solution. These results are further elaborated in [12], [97], [98], [95].

V. OPPORTUNITIES ON THE HORIZON

Despite intensive research activity over the past fifteen years, and particularly in the past decade, formidable challenges remain to be addressed in order for sensor management to be genuinely viable in large-scale sens¬ing systems. A central issue is computational feasibility of even approximate methods when scaled to problems that involve large numbers of controllable parameters, pose acute time constraints, or can only be adequately addressed by methods that look multiple steps ahead.

One arena of current investigation seeking to address the complexity issue involves sparse convex optimization approaches. The selection of an action sequence among a large number of possible sequences is similar to variable selection in sparse (lasso) regression [99] and compres-sive sensing [100], among other areas. This insight led R. Rangarajan et al. [101] to apply convex relaxation to optimal waveform design. A similar approach was later applied by S. Joshi and S. Boyd [102] to sensor selection. The use of such convex relaxation principles to develop tractable approximations to more complex sen¬sor management combinatorial optimization problems, such as multi-stage planning, may lead to computational breakthroughs.

Another circle of current research offering some promise with regard to mitigating complexity involves the use of statistical machine learning tools. Often dif-ficult problems in one domain can be reduced to equiv-alent problems in another domain for which different and effective solution tools have been developed. For example, the celebrated boosting method of Y. Freund 

 

and R. Schapire for learning optimal classifiers [103] was directly motivated by optimal multi-armed bandit strategies. Conversely, by casting offline learning of optimal POMDP policies as an equivalent problem of learning optimal classifiers [104], [105], D. Blatt et al. [106] developed a boosting approach to learning optimal sensor management policies for UXO and radar sensing applications. It is likely that other advances in statistical machine learning can have positive impact on sensor management.

The authors are aware of ongoing research involving new approximation schemes that adaptively partition the information state space in an MDP problem in a way that allows controllable tradeoff of computational efficiency and approximation fidelity. This work, as yet unpublished, casts fidelity in terms of preserving the ranking of possible actions in terms of expected loss rather than preserving the actual values of expected loss.

The area of adversarial sensor management, which deals with situations where an adversary can control some aspects of the scenario to deliberately confound the sensor manager’s objectives, presents opportunities for new sensor management research directions involv¬ing game theory and other methods. Recent work on POMDP for smart targets, i.e., targets that can react when they sense that they are being probed, is a step in this direction [107],[108]. Adversarial multi-armed bandits [109] and game theoretic solutions to adversarial multimobile sensing have also been proposed [110]. However, there are presently very few fundamental re-sults on performance in adversarial environments; e.g., generalizations of the non-adversarial bounds of Cas-tañón [111] and Williams [12] for POMDPs or those of K. D. Glazebrook and R. Minty for multi-armed bandits [112].

VI. CONCLUDING REMARKS

In this overview article on sensor management we have described the primary models and methods around which recent research in the field has been centered. We have also attempted to expose the historical roots in classical work spanning sequential analysis, optimal design of experiments, information theory, and optimal control. In our discussion of current trends and future research opportunities, we point out formidable chal-lenges to achieving the performance gains in real-world systems that we believe are potentially possible. The computational viability of scaling the methods described in this paper to large-scale problems involving sensing systems with many controllable parameters, applications with fast operating tempos, and scenarios calling for non-myopic optimization depends upon substantial advances

 

in efficient and certifiable approximation in all the main components depicted in Figure 1.

Nevertheless, there is much room for optimism. The past two decades have seen intense research activity that has legitimized sensor management as a field of study and established its mathematical foundations. These have drawn on, adapted, and blended ideas from several established areas, including Markov decision processes, multi-armed bandit scheduling, and information gain myopic planning. The applications and technological advances that spurred the profound growth of interest in sensor management during this period continue to provide more and more opportunities for sensor man-agement, and in some cases demand it. In our own application regime of surveillance and reconnaissance for security and defense applications, future operational concepts envision increasingly versatile networked col-lections of sensor assets, a large fraction of which will be mounted on autonomous or semi-autonomous platforms, providing situational awareness at levels of abstraction considerably higher than target tracks and emission source localizations. We remain hopeful that a combination of significant incremental advances and bona fide breakthroughs will enable sensor management to rise to meet such visions.

In closing, we wish to acknowledge the role of numer-ous sponsored research programs that have enabled and shaped the development of sensor management over the past decade. Some such activities of which we are aware include DARPA’s Integrated Sensing and Processing and Waveforms for Active Sensing programs, which ran from 2001 through 2006. The U.S. Department of Defense has also invested in academic research in sensor management through several Multidisciplinary University Research Initiatives (MURIs) since the early 2000s. These have been managed by DARPA, the U.S. Air Force Office of Scientific Research (AFOSR), and the U.S. Army Research Office (ARO). We are also aware of sponsored work through the U.S. Air Force Research Laboratory, the Australian Defence Science and Technology Organi-sation, a few other government agencies, and several in-dustrial sources. This list is by no means comprehensive, but it illustrates the recognition of sensor management as a valuable emerging area of study by major research or-ganizations. Further, this trend is ongoing. For example, two new MURI projects related to sensor management have recently been initiated, one by AFOSR in 2010 entitled Control of Information Collection and Fusion and the most recent by ARO in 2011 entitled Value of Information for Distributed Data Fusion.

 

12

REFERENCES

[1] A. Hero, D. A. Castañón, D. Cochran, and K. Kastella, Foundations and Applications of Sensor Management. New York: Springer, 2008.

[2] A. R. Cassandra, “A survey of POMDP applications,” in Working Notes of AAAI 1998 Fall Symposium on Planning with Partially Observable Markov Decision Processes. Citeseer, 1998, pp. 17–24.

[3] G. E. Monahan, “A survey of partially observable Markov de¬cision processes: theory, models and algorithms,” Management Science, vol. 28, no. 1, pp. 1–16, January 1982.

[4] G. W. Ng and K. H. Ng, “Sensor management — What, why and how,” Information Fusion, vol. 1, no. 2, pp. 67–75, 2000.

[5] X.-X. Liu, S.-L. Shen, and Q. Pan, “A survey of sensor management and methods,” Acta Electronica Sinica, vol. 30, no. 3, pp. 394–398, 2002.

[6] D. Cochran, “Waveform-agile sensing: Opportunities and chal¬lenges,” in Proc. IEEE Int. Conf. Acoust., Speech, and Sig. Proc., vol. V, March 1995, pp. 877–880.

[7] G. A. McIntyre, “A comprehensive approach to sensor man-agement and scheduling,” Ph.D. dissertation, George Mason University, 1998.

[8] D. Sinno, “Attentive management of configurable sensor sys¬tems,” Ph.D. dissertation, Arizona State University, May 2000.

[9] C. Kreucher, “An information-based approach to sensor re-source allocation,” Ph.D. dissertation, EECS Department, Uni¬versity of Michigan, 2005.

[10] R. Rangarajan, “Resource constrained adaptive sensing,” Ph.D. dissertation, EECS Department, University of Michigan, 2006.

[11] D. Blatt, “Performance evaluation and optimization for infer¬ence systems: Model uncertainty, distributed implementation, and active sensing,” Ph.D. dissertation, EECS Department, University of Michigan, 2007.

[12] J. L. Williams, “Information theoretic sensor management,” Ph.D. dissertation, Massachusetts Institute of Technology, 2007.

[13] M. Huber, Probabilistic Framework for Sensor Management. KIT Scientific Publishing, 2009.

[14] K. L. Jenkins, “Fast adaptive sensor management for feature-based classification,” Ph.D. dissertation, Boston University, 2010.

[15] M. A. Richards, Fundamentals of Radar Signal Processing. McGraw-Hill, 2005.

[16] D. J. Kershaw and R. J. Evans, “Optimal waveform selection for tracking systems,” IEEE Trans. on Inform. Theory, vol. 40, no. 5, pp. 1536–1550, September 1994.

[17] ——,“Waveform selective probabilistic data association,” IEEE Trans. on Aerosp. and Electron. Systems, vol. 33, no. 4, pp. 1180–1188, October 1997.

[18] S. Sowelam and A. H. Tewfik, “Optimal waveform selection in range-doppler imaging,” in IEEE Int. Conf. on Image Processing, November 1994, pp. 441–445.

[19] S. M. Sowelam and A. H. Tewfik, “Optimal waveform se-lection for radar target classification,” in IEEE Int. Conf. on Image Processing, October 1997, pp. 476–479.

[20] V. Krishnamurthy and R. J. Evans, “Hidden Markov model multiarm bandits: a methodology for beam scheduling in mul-titarget tracking,” IEEE Trans. on Signal Processing, vol. 49, no. 12, pp. 2893–2908, 2001.

[21] ——, “Correction to ‘Hidden Markov model multiarm bandits: a methodology for beam scheduling in multitarget tracking’,” IEEE Trans. on Signal Processing, vol. 51, no. 6, pp. 1662– 1663, 2003.

 

13

 

[22] N. A. Goodman, P. R. Venkata, and M. A. Neifeld, “Adaptive waveform design and sequential hypothesis testing for target recognition with active sensors,” IEEE Journ. Selected Topics in Signal Processing, vol. 1, no. 1, pp. 105–113, June 2007.

[23] S. P. Sira, Y. Li, A. Papandreou-Suppappola, D. Morrell,

D. Cochran, and M. Rangaswamy, “Waveform-agile sensing for tracking,” IEEE Signal Processing Magazine, vol. 26, no. 1, pp. 53–64, 2009.

[24] M. Greco, F. Gini, P. Stinco, A. Farina, and L. Verrazzani, “Adaptive waveform diversity for cross-channel interference mitigation,” in Proc. IEEE Radar Conf., May 2008, pp. 1–6.

[25] S. P. Sira, D. Cochran, A. Papandreou-Suppappola, D. Morrell, W. Moran, S. D. Howard, and R. Calderbank, “Adaptive waveform design for improved detection of low-RCS targets in heavy sea clutter,” IEEE Journ. Selected Topics in Signal Processing, vol. 1, no. 1, pp. 56–66, June 2007.

[26] A. Leshem, O. Naparstek, and A. Nehorai, “Information the-oretic adaptive radar waveform design for multiple extended targets,” IEEE Journ. Selected Topics in Signal Processing, vol. 1, no. 1, pp. 42–55, 2007.

[27] M. Vespe, G. Jones, and C. J. Baker, “Lessons for radar,” IEEE Signal Processing Magazine, vol. 26, no. 1, pp. 65–75, 2009.

[28] D. Cochran, S. Suvorova, S. D. Howard, and B. Moran, “Waveform libraries,” IEEE Signal Processing Magazine, vol. 26, no. 1, pp. 12–21, 2009.

[29] D. Sinno and D. Cochran, “Estimation with configurable and constrained sensor systems,” in Defence Applications of Signal Processing, D. Cochran, W. Moran, and L. B. White, Eds. Elsevier, 1998.

[30] R. A. Fisher, The Design of Experiments. Oliver and Boyd, Edinburgh, 1935.

[31] H. Robbins, “Some aspects of the sequential design of experi¬ments,” Bulletin of the American Mathematical Society, 1952.

[32] H. F. Dodge and H. G. Romig, “A method of sampling inspection,” Bell System Technical Journal, vol. 8, pp. 613– 631, 1929.

[33] A. Wald, Sequential Analysis. New York: Wiley, 1947.

[34] C. W. Helstrom, Elements of Signal Detection and Estimation. Prentice Hall, 1994.

[35] H. Chernoff, “Locally optimal designs for estimating parame¬ters,” Ann. Math. Stat., vol. 24, no. 4, pp. 586–602, December 1953.

[36] ——, “Sequential design of experiments,” Ann. Math. Stat., vol. 30, no. 3, pp. 755–770, September 1959.

[37] ——, Sequential analysis and optimal design. SIAM, 1972.

[38] V. V. Fedorov, Theory of Optimal Experiments. Academic Press, 1972.

[39] L. Meier III, J. Perschon, and R. M. Dressier, “Optimal control of measurement subsystems,” IEEE Trans. Automatic Control, vol. 12, no. 5, pp. 528–536, October 1967.

[40] M. Athans, “On the determination of optimal costly mea-surement strategies for linear stochastic systems,” Automatica, vol. 8, no. 4, pp. 397–412, July 1972.

[41] J. R. Landoll and C. A. Caceres, “Automation of data acqui-sition in patient testing,” Proceedings of the IEEE, vol. 57, no. 11, pp. 1941–1953, 1969.

[42] S. G. Bier, P. L. Rothman, and R. A. Manske, “Intelligent sen¬sor management for beyond visual range air-to-air combat,” in Proc. IEEE National Aerosp. and Electron. Conf. (NAECON), vol. 1, May 1988, pp. 264 – 269.

[43] S. Musick and R. Malhotra, “Chasing the elusive sensor man¬ager,” in Proceedings of the IEEE 1994 National Aerospace and Electronics Conference (NAECON). IEEE, 1994, pp. 606–613.

[44] R. Malhotra, E. P. Blasch, and J. D. Johnson, “Learning 

 

sensor-detection policies,” in Proc. IEEE National Aerosp. and Electron. Conf. (NAECON), vol. 2. IEEE, 1997, pp. 769–776.

[45] J. Aloimonos, I. Weiss, and A. Bandyopadhyay, “Active vi-sion,” International Journal of Computer Vision, vol. 1, no. 4, pp. 333–356, 1988.

[46] A. Segall, “Optimal control of measurement subsystems,” IEEE Trans. Automatic Control, vol. 22, no. 2, pp. 179–186, April 1977.

[47] D. A. Castañón, “Approximate dynamic programming for sensor management,” in Proc. of IEEE Conf. Decision and Control, vol. 2, 1997, pp. 1202–1207.

[48] J. S. Evans and V. Krishnamurthy, “Optimal sensor scheduling for hidden Markov model state estimation,” Int. J. Contr., vol. 74, no. 18, pp. 1737–1742, December 2001.

[49] V. Krishnamurthy, “Algorithms for optimal scheduling and management of hidden Markov model sensors,” IEEE Trans. on Signal Processing, vol. 50, no. 6, pp. 1382–1397, 2002.

[50] J. C. Gittins, “Bandit processes and dynamic allocation in-dices,” Journal of the Royal Statistical Society: Series B (Methodological), vol. 41, no. 2, pp. 148–177, 1979.

[51] P. Whittle, “Multi-armed bandits and the Gittins index,” Journ. Royal Statist. Soc. B, vol. 42, no. 2, pp. 143–149, 1980.

[52] R. B. Washburn, M. K. Schneider, and J. J. Fox, “Stochastic dynamic programming based approaches to sensor resource management,” in Proc. Int. Conf. on Inf. Fusion, vol. 1, 2002, pp. 608–615.

[53] C. E. Shannon, “A mathematical theory of communication,” Bell System Technical Journal, vol. 27, pp. 379–423, July 1948.

[54] J. Manyika and H. Durrant-Whyte, Data fusion and sensor management: a decentralized information-theoretic approach. Prentice Hall PTR Upper Saddle River, NJ, USA, 1995.

[55] W. W. Schmaedeke, “Information-based sensor management,” in Proc. Meeting of Intl. Soc. for Optical Engin. (SPIE), vol. 156, 1993.

[56] Y. Zhang, X. Liao, and L. Carin, “Detection of buried targets via active selection of labeled data: Application to sensing subsurface UXO,” IEEE Transactions on Geoscience and Remote Sensing, vol. 42, no. 11, pp. 2535–2543, 2004.

[57] K. Tian and G. Zhu, “Sensor management based on Fisher information gain,” Journal of Systems Engineering and Elec¬tronics, vol. 17, no. 3, pp. 531–534, 2006.

[58] V. Gupta, T. H. Chung, B. Hassibi, and R. M. Murray, “On a stochastic sensor selection algorithm with applications in sensor scheduling and sensor coverage,” Automatica, vol. 42, no. 2, pp. 251–260, February 2006.

[59] ——, “Sensor scheduling algorithms requiring limited com-putation,” in Proc. IEEE Int. Conf. Acoust., Speech, and Sig. Proc., vol. 3, May 2004, pp. 825–828.

[60] T. H. Chung, V. Gupta, B. Hassibi, J. Burdick, and R. M. Murray, “Scheduling for distributed sensor networks with single sensor measurement per time step,” in Proc. IEEE Int. Conf. on Robotics and Automation, vol. 1, April 2004, pp. 187–192.

[61] K. L. Jenkins and D. A. Castañón, “Information-based adaptive sensor management for sensor networks,” in Proc. American Control Conf., June 2010, pp. 4934–4940.

[62] ——, “Adaptive sensor management for feature-based clas-sification,” in Proc. of IEEE Conf. Decision and Control, December 2010, pp. 522–527.

[63] D. Hitchings and D. A. Castañón, “Receding horizon stochas¬tic control algorithms for sensor management,” in Proc. Amer¬ican Control Conf., June 2010, pp. 6809–6815.

[64] ——,“Adaptive sensing for search with continuous actions and observations,” in Proc. of IEEE Conf. Decision and Control, December 2010, pp. 7443–7448.

 

14

 

[65] E. J. Sondik, “The optimal control of partially observable Markov processes,” Ph.D. dissertation, Stanford University, 1971.

[66] L. P. Kaelbling, M. L. Littman, and A. R. Cassandra, “Plan-ning and acting in partially observable stochastic domains,” Artificial Intelligence, vol. 101, pp. 99–134, 1998.

[67] E. K. P. Chong, C. Kreucher, and A. Hero, “Partially ob-servable Markov decision process approximations for adaptive sensing,” Discrete Event Systems, vol. 19, no. 3, pp. 377–422, September 2009.

[68] Y. Li, L. W. Krakow, E. K. P. Chong, and K. N. Groom, “Approximate stochastic dynamic programming for sensor scheduling to track multiple targets,” Digital Signal Process-ing, vol. 19, no. 6, pp. 978–989, 2009.

[69] J. Liu, M. Chu, and J. E. Reich, “Multitarget tracking in dis-tributed sensor networks,” IEEE Signal Processing Magazine, vol. 24, no. 3, pp. 36–46, 2007.

[70] C. Kreucher, K. Kastella, and A. Hero, “Multitarget tracking using a particle filter representation of the joint multitarget probability density,” IEEE Trans. on Aerosp. and Electron. Systems, vol. AES-39, no. 4, pp. 1396–1414, 2005.

[71] V. Krishnamurthy and D. V. Djonin, “Optimal threshold poli¬cies for multivariate POMDPs in radar resource management,” IEEE Trans. on Signal Processing, vol. 57, no. 10, pp. 3954– 3969, 2009.

[72] S. Ji, R. Parr, and L. Carin, “Nonmyopic multiaspect sensing with partially observable Markov decision processes,” IEEE Trans. on Signal Processing, vol. 55, no. 6, pp. 2720–2730, 2007.

[73] T. Hanselmann, M. Morelande, B. Moran, and P. Sarunic, “Sensor scheduling for multiple target tracking and detection using passive measurements,” in Proc. Int. Conf. on Informa¬tion Fusion, 2008, pp. 1–8.

[74] M. J. Kochenderfer, L. P. Espindle, J. K. Kuchar, and J. D. Griffith, “A comprehensive aircraft encounter model of the national airspace system,” Lincoln Laboratory Journal, vol. 17, no. 2, pp. 41–53, 2008.

[75] D. P. Bertsekas and D. A. Castañón, “Rollout algorithms for stochastic scheduling problems,” Journal of Heuristics, vol. 5, no. 1, pp. 89–108, 1999.

[76] S. J. Benkoski, M. G. Monticino, and J. R. Weisinger, “A sur¬vey of the search theory literature,” Naval Research Logistics, vol. 38, no. 4, pp. 469–494, 1991.

[77] B. F. La Scala and B. Moran, “Optimal target tracking with restless bandits,” Digital Signal Processing, vol. 16, no. 5, pp. 479–487, 2006.

[78] J. M. Manyika and H. F. Durrant-Whyte, “Information-theoretic approach to management in decentralized data fu-sion,” in Proc. Meeting of Intl. Soc. for Optical Engin. (SPIE), vol. 202, 1992.

[79] M. L. Hernandez, T. Kirubarajan, and Y. Bar-Shalom, “Mul-tisensor resource deployment using posterior Cramér-Rao bounds,” IEEE Trans. on Aerosp. and Electron. Systems, vol. 40, no. 2, pp. 399–416, 2004.

[80] W. H. Barker, “Information theory and the optimal detection search,” Operations Research, vol. 25, no. 2, pp. 304–314, 1977.

[81] K. J. Hintz and E. S. McVey, “Multi-process constrained estimation,” IEEE Trans. on Systems, Man, and Cybernetics, vol. 21, no. 1, pp. 434–442, January/February 1991.

[82] W. Schmaedeke and K. Kastella, “Event-averaged maximum likelihood estimation and information-based sensor manage¬ment,” in Proc. Meeting of Intl. Soc. for Optical Engin. (SPIE), vol. 2232, Orlando, June 1994, pp. 91–96.

[83] K. Kastella, “Discrimination gain to optimize detection and classification,” IEEE Trans. on Systems, Man and Cybernetics, 

 

Part A: Systems and Humans, vol. 27, no. 1, pp. 112–116, 1997.

[84] R. Mahler, “Global optimal sensor allocation,” in Proceedings of the Ninth National Symposium on Sensor Fusion, vol. 1, 1996, pp. 167–172.

[85] K. J. Hintz and G. A. McIntyre, “Information instantiation in sensor management,” in Proceedings of the SPIE Interna-tional Symposium on Aerospace/Defense Sensing & Control, Orlando, FL, vol. 3374, 1998, pp. 38–47.

[86] C. Kreucher, K. Kastella, and A. O. Hero, “Information based sensor management for multitarget tracking,” in Proc. Meeting of Intl. Soc. for Optical Engin. (SPIE), San Diego, 2003.

[87] ——, “Multi-target sensor management using alpha-divergence measures,” in 3rd Workshop on Information Processing for Sensor Networks, Palo Alto, CA, 2003.

[88] F. Zhao, J. Shin, and J. Reich, “Information-driven dynamic sensor collaboration for tracking applications,” IEEE Signal Processing Magazine, vol. 19, no. 2, pp. 61–72, 2002.

[89] N. Xiong and P. Svensson, “Multi-sensor management for in¬formation fusion: Issues and approaches,” Information Fusion, vol. 3, no. 2, pp. 163–186, 2002.

[90] M. P. Kolba and L. M. Collins, “Information-based sensor management in the presence of uncertainty,” IEEE Trans. on Signal Processing, vol. 55, no. 6, pp. 2731–2735, 2007.

[91] C. M. Kreucher, A. O. Hero, K. D. Kastella, and M. R. More-lande, “An information based approach to sensor management in large dynamic networks,” Proceedings of the IEEE, vol. 95, no. 5, pp. 978–999, May 2007.

[92] C. Kreucher, K. Kastella, and A. O. Hero, “Sensor manage-ment using an active sensing approach,” Signal Processing, vol. 85, no. 3, pp. 607–624, 2005.

[93] G. Zhang, S. Ferrari, and M. Qian, “An information roadmap method for robotic sensor path planning,” Journal of Intelligent and Robotic Systems, vol. 56, no. 1, pp. 69–98, 2009.

[94] J. Denzler and C. M. Brown, “Information theoretic sensor data selection for active object recognition and state esti-mation,” IEEE Trans. on Pattern Anal. and Machine Intell., vol. 24, no. 2, pp. 145–157, 2002.

[95] J. L. Williams, J. W. Fisher III, and A. S. Willsky, “Perfor-mance guarantees for information theoretic active inference,” in Proc. Int. Conf. on Artificial Intelligence and Statistics, M. Meila and X. Shen, Eds., March 2007, pp. 616–623.

[96] A. Krause and C. Guestrin, “Near-optimal nonmyopic value of information in graphical models,” in Uncertainty in Artificial Intelligence, July 2005.

[97] J. L. Williams, J. W. Fisher III, and A. S. Willsky, “Approx-imate dynamic programming for communication-constrained sensor network management,” IEEE Transactions on Signal Processing, vol. 55, no. 8, pp. 3995–4003, August 2007.

[98] ——,“A constraint generation integer programming approach to information theoretic sensor resource management,” in Proc. IEEE Workshop on Statistical Signal Processing, August 2007.

[99] T. Hastie, R. Tibshirani, and J. H. Friedman, The Elements of Statistical Learning: Data Mining, Inference, and Prediction. New York: Springer, 2001.

[100] E. Candés, J. Romberg, and T. Tao, “Stable signal recovery from incomplete and inaccurate measurements,” Comm. Pure Appl. Math., vol. 59, pp. 1207–1223, 2005.

[101] R. Rangarajan, R. Raich, and A. O. Hero, “Single-stage waveform selection for adaptive resource constrained state estimation,” in Proc. IEEE Int. Conf. Acoust., Speech, and Sig. Proc., May 2006.

[102] S. Joshi and S. Boyd, “Sensor selection via convex optimiza¬tion,” IEEE Trans. on Signal Processing, vol. 57, no. 2, pp. 451–462, 2009.

 

[103] Y. Freund and R. Schapire, “A decision theoretic generalization of online learning and an application to boosting,” Journ. of Computer and System Sciences, vol. 55, pp. 119–139, 1997.

[104] J. Langford and B. Zadrozny, “Reducing T-step reinforcement learning to classification,” in Proceedings of the Machine Learning Reductions Workshop, 2003.

[105] D. Blatt and A. Hero, “From weighted classification to policy search,” in Advances in Neural Information Processing Sys¬tems 18, Y. Weiss, B. Schölkopf, and J. Platt, Eds. Cambridge, MA: MIT Press, 2006, pp. 139–146.

[106] D. Blatt and A. O. Hero, “Optimal sensor scheduling via classification reduction of policy search (CROPS),” in Inter-national Conference on Automated Planning and Scheduling, 2006.

[107] C. Kreucher, A. O. Hero, D. Blatt, and K. Kastella, “Adaptive multi-modality sensor scheduling for detection and tracking of smart targets,” in Workshop on Defense Applications of Signal Processing, 2004.

[108] B. Liu, C. Ji, Y. Zhang, and C. Hao, “Blending sensor scheduling strategy with particle filter to track a smart target,” Wireless Sensor Networks, no. 4, November 2009.

[109] T. Uchiya, A. Nakamura, and M. Kudo, “Algorithms for ad-versarial bandit problems with multiple plays,” in Algorithmic Learning Theory. Springer, 2010, pp. 375–389.

[110] M. Wei, G. Chen, E. Blasch, H. Chen, and J. B. Cruz, “Game theoretic multiple mobile sensor management under adversarial environments,” in Proc. Int. Conf. on Inf. Fusion, 2008, pp. 1–8.

[111] D. Castañón, “Stochastic control bounds on sensor network performance,” in Proc. of IEEE Conf. Decision and Control and European Control Conf. IEEE, 2005, pp. 4939–4944.

[112] K. D. Glazebrook and R. Minty, “A generalized Gittins in-dex for a class of multiarmed bandits with general resource requirements,” Mathematics of Operations Research, vol. 34, no. 1, pp. 26–44, 2009.

 

No. 09-35

September 2009

WORKING

PAPER