Jasa Konsultasi Disertasi Keuangan (1) dan Perbankan S3, hubungi WA 0821 2230 7021: Penelitian doktoral bidang knowledge management: Retrieval and learning of design knowledge

Retrieval and learning of design knowledge

CAYMAN ISLANDS

Supplement No. 1 published with Gazette No. 08 dated 22 April, 2013.

PRACTICE DIRECTION No. 1 /2013 (GCR O.1, r.12)

CONSENT ORDERS IN ANCILLARY RELIEF PROCEEDINGS

PRACTICE DIRECTION No. 1 /2013

(GCR O.1, r.12)

CONSENT ORDERS IN ANCILLARY RELIEF PROCEEDINGS

Practitioners are reminded that when the Court is asked to make an order by consent, embodying the terms of an agreed settlement of ancillary relief proceedings, the Court is still obliged to comply with the statutory duty imposed by Section 19 of the Matrimonial Causes Law (2005 Revision).

1. In this regard, the Court will now require the following before exercising its powers under

section 19:

(i) Confirmation signed by both parties that they have made and received sufficient disclosure to enable them to make an informed decision to ask the Court to approve the terms contained within the proposed consent order.

(ii) In the case of unrepresented parties, signed confirmation that they have been advised to seek legal advice in regard to the proposed settlement, and have either done so or have declined to do so.

(iii) The lodging of a fully completed Statement of information for a consent order. (Where all of the parties attend the hearing of an application for financial relief or where the court is satisfied by the content of evidence already filed, the court may dispense with the lodging of a Statement of information and give directions for the information which would otherwise be required to be given in such a statement to be given in such manner as it sees fit).

This Practice Direction shall come into force on the 1st day of May 2013.

DATED this 9th day of April 2013

The Hon. Anthony Smellie

Chief Justice

Grand Court Form

STATEMENT OF INFORMATION FOR A CONSENT ORDER

The full names of the parties

Petitioner:

Respondent:

The age of the parties and duration of the marriage To be completed by the court

Date filed:

Case number:

Petitioner:

Respondent:

Duration of the Marriage:

Important Note. This form is to be lodged when the Court is asked to make an order by consent embodying the terms of an agreed settlement of ancillary relief proceedings.

1 Summary of Means

Give as the date of the statement the approximate amount or value of capital resources and net income of the petitioner and respondent and, where relevant, any minor children of the family. State also the net equity in any property concerned and the effect of its proposed distribution State:

1. Capital Resources less any unpaid mortgage or charges

(i) Petitioner CI/US$

(ii) Respondent CI/US$

(iii) Children CI/US

2. Net Income

(i) Petitioner CI$

(ii) Respondent CI$

(iii) Children CI$

2 Where the parties are to live

Give details of what arrangements are intended for the accommodation of each of the parties and any children of the family

3 Marital Plans

Petitioner Respondent

Yes / No Yes / No has no present intention to marry or cohabit

Yes / No Yes / No has remarried

Yes / No Yes / No intends to remarry

Yes / No Yes /No intends to cohabit with another person

4 Notice to Chargee

To be answered by the applicant where the terms of the order provide for a transfer of property

Has any and every chargee of the property been served with a notice of the application

Yes / No

Has any objection to such a transfer been made by any charge within14 days from the date of service?

Yes / No

5. Give details of any especially significant matters including but not limited to how the attached proposed consent order was reached e.g. discussions between parties, negotiations through attorneys, mediation, or other out of court dispute resolution

Petitioner’s Statement of Truth

I believe that the facts stated in this Statement of information for a consent order are true.

I confirm that I have read the consent order and the completed Statement of information for a consent order from the Respondent and that I have made and received sufficient disclosure to enable me to make an informed decision to ask the Court to approve the terms contained within the consent order attached.

[I confirm that I am unrepresented and I have had the opportunity to seek legal advice in regard to the proposed settlement, and have done so / have declined to do so]. (delete if represented)

I confirm that I have not been pressured in any way to endorse the terms of the consent order attached.

Print full name:

Name of Petitioner’s attorney (complete if represented) :

Signed by Petitioner:

Dated:

Proceedings for contempt of court may be brought against a person who makes or causes to be made, a false statement in a document verified by a statement of truth.

Respondent’s Statement of Truth

I believe that the facts stated in this Statement of information for a consent order are true.

I confirm that I have read the consent order and the completed Statement of information for a consent order from the Petitioner and that I have made and received sufficient disclosure to enable me to make an informed decision to ask the Court to approve the terms contained within the consent order attached.

[I confirm that I am unrepresented and I have had the opportunity to seek legal advice in regard to the proposed settlement, and have done so / have declined to do so]. (delete if represented)

I confirm that I have not been pressured in any way to endorse the terms of the consent order attached.

Print full name:

Name of Respondent’s attorney (complete if represented) :

Signed by Respondent:

Dated:

Proceedings for contempt of court may be brought against a person who makes or causes to be made, a false statement in a document verified by a statement of truth.

www.elsevier.com/locate/aei

Soft computing in engineering design – A review

K.M. Saridakis, A.J. Dentsoras *

Machine Design Laboratory, Department of Mechanical Engineering & Aeronautics, University of Patras, Patras 26500, Greece

Received 6 June 2006; received in revised form 4 October 2007; accepted 5 October 2007

Available online 3 December 2007

Abstract

The present paper surveys the application of soft computing (SC) techniques in engineering design. Within this context, fuzzy logic (FL), genetic algorithms (GA) and artificial neural networks (ANN), as well as their fusion are reviewed in order to examine the capa-bility of soft computing methods and techniques to effectively address various hard-to-solve design tasks and issues. Both these tasks and issues are studied in the first part of the paper accompanied by references to some results extracted from a survey performed for in some industrial enterprises. The second part of the paper makes an extensive review of the literature regarding the application of soft comput¬ing (SC) techniques in engineering design. Although this review cannot be collectively exhaustive, it may be considered as a valuable guide for researchers who are interested in the domain of engineering design and wish to explore the opportunities offered by fuzzy logic, artificial neural networks and genetic algorithms for further improvement of both the design outcome and the design process itself. An arithmetic method is used in order to evaluate the review results, to locate the research areas where SC has already given considerable results and to reveal new research opportunities.

Keywords: Engineering design; Soft computing; Fuzzy logic; Genetic algorithm; Neural networks

1. Engineering design: an overview

Design has been a human activity for thousands of years. In both its creative and routine forms, the scientific community has extensively studied design during the last decades for the establishment of general purpose and domain-independent scientific rules and methodologies. Finger et al. [1], in a review paper, surveyed the issues of design theory, design methodologies and design models. Another survey on the same topic was conducted by Evbuomwan et al. [2] a few years later. Besides summariz¬ing and reviewing the developed design models and methodologies, the aforementioned articles investigated the nature and the characteristics of the design process, classified the design models into categories and located possible research opportunities.

*Corresponding author. Tel./fax: +30 2610997217.

E-mail addresses: saridak@mech.upatras.gr (K.M. Saridakis), dentsora

@mech.upatras.gr (A.J. Dentsoras).

Three general design models are identified in the litera¬ture: the descriptive, the prescriptive and the computer-based [2] model. The descriptive models for design tend to capture the processes, strategies and methods that design¬ers use in order to address certain design problems. These models usually initialize the design process through the generation of a single solution concept, which is then sub¬mitted to further analysis, refinement, development and evaluation. Although descriptive models incorporate past experience, they still meet difficulties in capturing the collaborative activities among designers, especially in vague conceptual phases. On the other hand, the prescrip¬tive models are well addressed in preliminary design stages and are extensively used by the designers since they provide an intuitive sense to the design process. The underlying prescription may be deployed either towards providing guidelines for the design process or towards determining the design attributes of the designed artifact. Additionally, the prescriptive models are mainly algorithmic-based and this fact encourages several analytical activities before the

K.M. Saridakis, A.J. Dentsoras / Advanced Engineering Informatics 22 (2008) 202–221 203

generation of concept solutions. The prescriptive models however, present the following two weaknesses: (a) They are inefficient in mapping the design requirements to the attributes of the artifact, and (b) they present incompatibil¬ity with computer-based design methods – if integration is considered. Finally, the computer-based models enhance features from the two aforementioned categories and are capable of performing many different design activities. They may address decision-making processes or analyze the design knowledge in order to provide a better under¬standing of the considered design problem. Unfortunately, the computer-based models perform optimally only in spe¬cific, well-defined classes of problems, mostly of a quantita¬tive nature. This implies that there is a need for new unified computer-based design models that would include and could manipulate both quantitative and qualitative knowledge.

On the basis of these three general design models, sev-eral design methods have been proposed with the charac-terization ‘design-for-X’, each of them viewing a different aspect of design, e.g. design for assembly, design for man¬ufacture, design for maintenance, design for cost, design for reliability, design for service ability, etc. The applicabil¬ity of a design model or a design methodology highly depends on the type of the problem under consideration. In the literature, three basic categories of design problems are identified: (a) routine design that characterizes the prob¬lems with predefined, explicit and well-structured design knowledge shared among identical or similar problems, (b) redesign, which involves the adaptation of an existing problem representation and/or solution in order to meet some newly defined requirements, and (c) non-routine or original design further classified in innovative and creative design. In innovative design, although new variables, parameters or features may be introduced, the resemblance with an existing design problem is preserved and the new solution is delivered through a synthesis process. On the other hand, in creative design, the newly introduced design entities differentiate the current design problem. The extracted solutions are not delivered via a predefined plan and they may be characterized as unique.

In 1996, the Japan society for the promotion of sciences initiated a new funding scheme, called “Research of Future”program, encouraging fundamental research in engineering and science. In the context of this program, the design problems have been identified as synthesis prob¬lems and have been classified based on the difficulties that are encountered during the solving activities. Ueda [3] pro¬vided an overview of the methodology of emergent synthe¬sis, while discussing analysis, synthesis and emergence in the realm of artifacts. In this theory, the difficulties in syn¬thesis are categorized into three classes. Class I is charac¬terized by systems with a complete description of the design specification and the environment; Class II corre¬sponds to complete design specifications but to incomplete information about the environment; finally, Class III denotes systems with both incomplete specification and

an incomplete environment description. Since a great num¬ber of the design problems belong to Class II and Class III, the existing principles such as analysis and determinism seem to serve inadequately. Therefore, new methodologies based on emergence are considered promising to realize an artifactual system that is capable of addressing these prob¬lems even in unpredictable conditions. The latter ascertain¬ment is the basis of synthesis that plays a key role in almost all stages of the artifacts’ lifecycle, while the related research area has strong relationships with the fields of arti¬ficial life, artificial intelligence, evolutionary and emergent computation, soft computing, complex adaptive systems, reinforcement learning, self-organization, etc.

Through the review of the existing literature, the follow¬ing issues regarding the design process become distinguish¬able: (a) the design knowledge representation (modeling), (b) the search for optimal solutions, (c) the retrieval of pre-exist¬ing design knowledge and the learning of new knowledge. Recent research activity in engineering design is deployed by taking into account one or more of these issues, which, in the context of the present paper, are used as a basis in order to evaluate the proposed approaches as they arise from the relevant literature.

1.1. Design knowledge representation (modeling)

Most design tasks and activities involve decision-making concerning issues such as function, structure, configura¬tion, material and geometry of the designed artifact. The quantity and the quality of the available information and data for decision-making vary along the different design phases. At the preliminary/conceptual design phase, very little information is available. As design advances, more information is produced and the design knowledge is con¬tinuously enriched.

During the very first phases, design may be enhanced by the collaboration of multiple designers or design teams on the basis of a communication and common decision-making framework. In a relevant paper, Wang et al. [4] are interested in collaborative conceptual design and they review the literature for relative approaches and applica¬tions. From this survey, it becomes evident that there is a necessity for handling different types of uncertainty and vagueness – inherently characterizing the design process – through systematic methods and techniques.

The strict formalization of the design process by means of rules, scientific criteria and design axioms [5] has raised several arguments and philosophical issues [6]. By taking into account the nature and the underlying features of the design process, it is probably more reasonable and scientifically more consistent to view design through the synthesis of issues raised when considering many different viewpoints [7] and alternative design knowledge representa¬tion formalisms. Regardless of the design model chosen for each occasion, it is important to capture the design inten¬tion during the design process in order to fully understand the design project and in order to be able to reuse the

204 K.M. Saridakis, A.J. Dentsoras /Advanced Engineering Informatics 22 (2008) 202–221

process in order to accomplish similar projects in the future. The description of rationality in the Cambridge dic¬tionary of philosophy [8] is summarized as follows: “Theo¬retical rationality applies to beliefs (e.g. beliefs that are self evident or derived from self evident beliefs by a reliable procedure) while practical rationality applies to actions. Acting rationally simply means acting in a way that is max¬imally efficient in achieving one’s goals on the basis of uni-versalizable principles, so that what is a reason for one person must be a reason for everyone”. Regardless of the rational or irrational solutions achieved for a given design problem, the designer follows a specific rationale. Regli et al. [9] investigate this design rationale, by analyzing the reviewed literature in terms of knowledge representation schemes, design rationale capture approaches and design rationale retrieval methodologies. In their conclusions they remark the importance of design rationale as a mediator between loosely defined prescriptive models and strict com¬puter-based models. Kryssanov et al. study a computa¬tional theory to ensure creativity for engineering design [10]. In their research, the human activities and creativity are surveyed in the context of design, in terms of which basic notions of algebraic semantics are defined and dis-cussed. The proposed model provides a formal description and investigation of the essential properties of the design process, namely its dynamics and non-determinism.

1.2. Search for optimal solutions

Design is characterized by multiple decision-making actions throughout the design process, starting during its early stages and ending by its completion. It is always expected that these decision-making activities – that are always subjected to significant uncertainty, vagueness and imprecision – would eventually lead to an optimal design solution. Under these circumstances, a systematic design methodology must be based on principles that formalize the decision-making framework [11] and take into account its incomplete nature.

Design principles may be used to establish realistic and robust criteria for the optimality and the efficiency of the derived design solutions. Until now, it has been proved difficult to turn qualitative, quantitative – and usually com¬peting – design objectives into explicit criteria. Further¬more, in the case of multiple designers/design teams, it is necessary for the expression and the aggregation of the design criteria to be formal and systematic in order for the collaboration to be feasible and successful. A recent study in the area of decision support systems (DSS) [12] concludes that the core research interests have been point¬ing to group decision support systems, model management, design and its implementation, while the study of founda¬tions and individual differences has faded away. Moreover, recent advances in computational intelligence are increas-ingly integrated in DSS systems, while cognitive science is identified as a contributing field in the context of principles and disciplines. The enhanced proposed models succeed in

delivering efficient design solutions even with vague design knowledge, despite the fact that their applicability is restricted in specific design domains. Hazelrigg [13] discuss the properties that a design decision methodology should have and use these properties in order to to validate several existing decision and selection methods.

A necessary step in the design process is the evaluation of the alternative solutions in order to obtain the optimal one according to the posed criteria. This search may be per¬formed in solution spaces that may contain discrete and non-continuous subspaces, through qualitative and quanti¬tative design parameters. The selection of the optimization technique, which may be either heuristic (if domain-specific knowledge is available) or exhaustive (if the entire solution space must be searched), depends on the design problem under consideration. Aspects of the optimization methods such as trapping in solution spaces with local optimal points, near-optimal but not absolute optimal solutions, the computational cost, the convergence time needed for reaching optimal solution, etc. must be carefully taken into account.

1.3. Retrieval and learning of design knowledge

A common phenomenon in design problem solving is the extraction of sub-optimal solutions, because of the par¬tial availability (or inexistence) of design knowledge with underlying design experience. The two initiatives of the design process, namely the representation of the design knowledge and the acquisition of the optimal solution, may be facilitated if past design knowledge is available and reusable. A characteristic example of taking advantage of past solutions is the case-based reasoning paradigm, which was firstly introduced in design by Kolodner et al. [14]. Design problem solving by retrieving solutions to sim¬ilar past problems and adapting them to the new design requirements coincides with the natural human analogical thinking. Case-based reasoning was further discussed and analyzed in the context of engineering design in the book authored by Maher et al. [15].

The reuse of past knowledge and experience in new design problems may be implemented in ways other than storing and retrieving design cases. The activity, during which a computer-based system records and learns from notable events and attributes that have occurred during the development of past designs, is characterized as machine-learning (ML). Grecu et al. [16] suggest the follow¬ing five dimensions for this activity: (a) triggers of learning, (b) the elements supporting learning, (c) what gets learned, (d) availability of knowledge to be learned, and (e) methods of learning. The aforementioned research work suggests that it is necessary to analyze the machine learning devel¬opments and opportunities in a systematic manner and has been given further extensions by Sim et al. [17]. Sim et al attempt to provide a foundation of machine learning in design in terms of a different set of key elements: (a) input knowledge, (b) knowledge transformers, (c) output

K.M. Saridakis, A.J. Dentsoras / Advanced Engineering Informatics 22 (2008) 202–221 205

knowledge, (d) goals/reason for learning, and (e) triggers of learning. Considering these elements, a number of machine-learning systems in design have also been reviewed. In many cases, the machine-learning activity relies on the generalization of the available design knowledge, which may lead to imprecise and trivial results. However, the majority of designers still express their convenience about getting every possible support during the design process, even when it comes to managing gener¬alized or approximated design models. In the latter cases, the human factor must have the main role during detailed and/or refinement design phases, despite the efficiency of the machine-learning process.

2. Engineering design in real world situations – open design issues

The outcome of the research activity in the field on engineering design should always be tested in real world problems in order to prove their validity. Additionally, every-day design practice in industry provides the research community with a wide spectrum of problems that form the core material to be further studied and analyzed. In the previous section of the paper, three main tasks of the engineering design were identified. With the initiative of investigating the difficulties, which are observed during the design process, in the present section some major issues regarding the design process in industry are addressed. The main purpose is to identify opportunities for further research and improvement and to emphasize on the necessity of applying new techniques and methodologies – including those based on soft computing – to open problems in the field of design. Some results from a short study done for five enterprises with production activities in Greece are occasion¬ally referred to as indicative examples.

2.1. Uncertainty management and conceptual design

The diversity of types and the degrees of uncertainty for the underlying design knowledge constitute the discrete lim¬its between loosely- and well-defined design problems. In an industrial environment, product development is based on multiple assumptions and, most of the times, requires the solution of multiple combinatorial problems with vague, imprecise partially available underlying design knowledge. Moreover, the design case may consist of many different variants, each of which requires a different way of represen¬tation formalism (visualization, analytical expression, etc.). Under these circumstances most of the undertaken work is executed manually, autonomously and the consolidation is performed during meetings and discussions.

2.2. Collaboration, communication and coordination

The demand for top-quality products, as a result of the complex multi-disciplinary design discourses, increasingly requires the participation and the availability of more than

one designer. The experiences and the personal abilities of each individual designer must be integrated through suit¬able representation and reasoning frameworks, imple-mented via web technologies and Internet services and thus supporting the collaborative activities and the com-munication of the design knowledge. As an example, the surveyed enterprises, all have intranet, internal email com¬munication and communication with other affiliates, three of them utilize a database for new product specifications, all of them keep records of past solutions, two of them use e-room asynchronous communication for major pro¬jects but none utilizes a tool for enhancing collaboration in design in ways other than sharing information.

2.3. Solution extraction under lacking knowledge/uncertainty

In many industrial projects, it is crucial to proceed to design solutions even if the available design knowledge is not adequate in terms of specifications, of technical know-how or of underlying risk. The increasing number of innovative products and the ‘fit’ manufacturing strategy are usually fulfilled in an environment of missing informa¬tion where designers and project engineers make specula¬tions, estimations and assessments. The latter tasks may be performed on the basis of standard practices and tem¬plates, while records and examples from past products and similar projects may be available in corresponding archives. The final outcome is strongly dependent on the human reasoning power and the available expertise to handle these problems. In the surveyed enterprises, no supportive tool or automated process was utilized for the extraction of new solutions with partially available infor-mation/knowledge or on top of past solutions.

2.4. Management and reuse of existing design knowledge

In the previous paragraph the potential of using past solutions in an automated (or semi-automated) way in order to solve loose-defined design problems was remarked as inefficacy of the surveyed enterprises to take advantage of this potential. Most of the times, the main reason for this inefficacy is the inconsistent and/or poor documentation produced during product development. This implies that in case that a similar product has to be developed in the future, the relevant past knowledge and experience may not be reproducible and exploitable. This ascertainment reveals the importance of keeping records and documenta¬tion with proper formalism in order to use this recorded knowledge in the future. As the activity of documenting and recording is time-consuming, the companies should perform it with efficient, rapid and less knowledge-demand¬ing ways.

2.5. Generality and domain-independence

Numerous methodologies and tools are used in engi-neering design, with most of them focusing on specific

206 K.M. Saridakis, A.J. Dentsoras /Advanced Engineering Informatics 22 (2008) 202–221

issues of the design process. For example, a design tool may perform sufficiently in in representing the design problem without providing sufficient support to other tasks such as optimization and design solution extraction dis¬course. Therefore, attention should be paid to preserve a balance among the underlying interacting modules. This implies that each one of the modules participating in the design process should provide generality and sufficient domain-independence in order to address a variety of design problems.

2.6. Extensibility and connectivity to other tools

The applicability of a design methodology to a broad range of design problems as well as its robustness and effi¬ciency depend on the capability of its modules to extend and interface with other well-established tools. A key fac¬tor that may contribute towards this direction is the maxi¬mization of the functional independence of each interacting module. Moreover, it is possible for an enterprise to utilize a specific tool either for design or for other supportive tasks. Then a new design tool should be fully compatible with the existing ones by easily accepting and providing data, information and knowledge. The surveyed enterprises gave three examples of big investments made on specific software that very soon became obsolete without returning any added value. The first was a Computerized Mainte¬nance Management System (CMMS) that could not be integrated with SAP, the second was a design CAD tool that could not use existing libraries of past product solu-tions and the third was a global web-based spare parts management system that could not be extended to cover all spare parts data available in local level.

2.7. Balance between automation and human activities

In computer-assisted engineering design, it is very important to ensure that some design tasks are automated and/or design tools are available in order to support the designers to manually execute some other tasks. The bal-ancing between full automation and supportive implemen¬tation is a very important issue and should be taken into account during the planning phase of the design process. Human and computer-based processes should always be implemented in parallel or in hybrid sessions so that the advantages of each aspect be captured and at the same time, depending on the problem under consideration, the analogy between them should always be a priori estimated. Although both parts are enhanced and better solutions may be extracted through this human–machine interaction, none of the surveyed enterprises presented a systematic way of approaching this goal. Additionally, it became obvi¬ous through the discussion with their representatives that the human insight and creativity cannot be replaced by any computerized system and the efficiency of the com¬puter-based system in specific discrete tasks can be reached by any individual stakeholder.

2.8. Simplicity

In any engineering domain there are conventional estab¬lished techniques used by the majority of the engineers. The reason for that is that these techniques have been repeat¬edly tested and have provided successful results. However, in many cases these techniques proved to be inadequate and inefficient and thus different and more advanced tech¬niques should be deployed. It is a fact that these advanced techniques require more specialized and expert knowledge. As a result of their use, it is highly possible that the final outcome is sub-optimal, as the optimality of the solution strongly depends on adjustments of some operating param¬eters of the mechanism of this advanced technique. For example, if a fuzzy logic based technique is utilized in solv¬ing a problem, then the designer should know exactly the impact of using the alternative methods for the aggregation and defuzzification of the fuzzy components. The sustain-ability and the success of an advanced technique should always be based on simplicity and this simplicity can be originated by the combination of characteristics of the three design models: prescriptive, descriptive and com¬puter-based. The survey in the reviewed enterprises revealed that no advanced tool was sustainable in long terms. Even supportive modules in existing, well-estab¬lished systems in other functions (e.g. maintenance routines in SAP) were abandoned, as they required the dedication of individuals with specific knowledge and expertise in order to follow-up.

3. Soft computing and engineering design

Developments in the computer hardware during the last two decades have made it easier for the artificial intelli¬gence techniques to grow into more efficient frameworks. Moreover, it has been proven that several artificial intelli¬gence techniques may be used as tools in problems where conventional approaches fail or perform poorly. An excel¬lent example to demonstrate this potential is the field of engineering design due to its specific characteristics and requirements. A survey of the existing literature could eas¬ily confirm this fact but it could also reveal the growing interest of the research community on the relatively new field of soft computing from the engineering point of view. Considering this interest, the authors suggest that a new term, namely SCAD (Soft Computing-Aided Design) is adapted in order toto describe the research domain where engineering design meets fuzzy logic, artificial neural net¬works and genetic algorithms.

The relevant research activity has been directed towards the development of architectures and ontologies to apply artificial intelligence in design, in order to assist the design activity and to apply automation to the more complex, conceptual and decision-making tasks [18]. It is also a fact that a parallelism between engineering design and artificial intelligence is noticed in both historical and current trends [19]. However, AI is still not widely used in CAD/CAM

K.M. Saridakis, A.J. Dentsoras / Advanced Engineering Informatics 22 (2008) 202–221 207

systems, which, nevertheless, incorporate developments from the domain of applied mathematics and information technology [20]. The reasons for that could be: (a) the dif¬ficulty of deploying several AI techniques in a concurrent or collaborative framework, (b) the increased demands in computational power, (c) the need to address the issues of conceptualization, visualization and detailed representation.

In the early 1990s, Dr. Zadeh introduced the concept of Soft Computing (SC). Soft Computing is an evolving col¬lection of artificial intelligence methodologies aiming to exploit the tolerance for imprecision and uncertainty that is inherent in human thinking and in real life problems, to deliver robust, efficient and optimal solutions and to fur¬ther explore and capture the available design knowledge. Fuzzy Logic (FL), Artificial Neural Networks (ANN), and Genetic Algorithms (GAs) are the core methodologies of soft computing. SC yields rich knowledge representation (symbol and pattern), flexible knowledge acquisition (machine learning), and flexible knowledge processing (inference by interfacing symbolic and pattern knowledge). Additionally, SC techniques can either be deployed as sep¬arate tools or be integrated in unified and hybrid architec¬tures. The fusion of SC techniques causes a paradigm shift (breakthrough) in engineering and science fields – including engineering design – by solving problems, which could not be solved with the conventional and/or stand-alone com¬putational tools.

Research has been deployed in the direction of applying SC to engineering design in the context of replacing exist¬ing analytical models with approximated models or meta-models. Simpson et al. [21] investigated the potential of soft computing techniques by comparing them to the statistical techniques in meta-modeling and they provided some rec¬ommendations about their appropriate use. Besides meta-modeling, SC techniques may be combined with expert and knowledge-based systems. The common path of expert systems (ES) and SC techniques was surveyed by Liao [22] in a decade (1995–2004) review.

In another literature review, Rao et al. [23] change their focus to the new product development (NPD) activities. In their article, they categorize the applications in five discrete areas, in NPD stages and NPD core elements. The pene¬tration of SC in activities that are strongly related with engineering design is also evident in the research work of Hsu et al. [24]. NPD is also the main research direction in this article, but this time in the context of deploying two SC techniques, namely ANN and FL, in an integrated framework. A general neuro-fuzzy model is suggested, while different formulations of neuro-fuzzy networking are discussed.

The acceptance of SC techniques by the scientific com¬munity is also based on the potential of fusing SC with con¬ventional hard computing techniques. Kamiya et al. [25] investigate this fusion and remark the possibilities of its application in large-scale plants. The advances of success¬fully applying SC in demanding domains like engineering

design, has also promoted the use of SC techniques to a wide spectrum of industrial applications. Dote et al. [26] review a significant number of such SC-based applications.

The consideration that the art of designing still remains a special human activity, which is in most times an enter-tainment for the designer, is remarked by Cross [27]. There¬fore, future advances in engineering design should perhaps be limited to the cognition of this human natural intelli¬gence. Having confirmed the efficiency of SC either in terms of stand-alone techniques or as supportive tools to other conventional design methodologies, many researchers have been investigating the possibility of enhancing the existing design methodologies by simulating the design process through SC techniques. Shakeri et al. [28] simulate the design process using a multi-agent system, with some of the agents being based on SC. These agents mimic the behaviour of a design team and a set of design methodolo¬gies is constructed by using learning techniques.

Numerous SC-based methodologies and applications have been reported in the literature in a variety of scientific domains. In the context of the present paper, representative approaches with high relation to the design process are pre¬sented. The rest of the paper is organized in sections, each of which combining engineering design with (a) fuzzy logic (FL), (b) genetic algorithms (GA), (c) artificial neural net¬works (ANN), (d) combinations of these three soft com¬puting techniques, and (e) applications of soft computing techniques in case-based design. Each of the aforemen¬tioned sections presents references depending on the underlying general theoretic background, the relevant methodologies and applications. Most of these references are related to books or journal articles and are being referred to in chronological order within the last decade, whereas each research field retains space in the current paper proportional to the number of the relevant reviewed approaches. For each section, an overall evaluation of the reviewed approaches is performed versus the open design issues described in Section 2. This evaluation is performed in order to extract the relevant efficiency of the proposed approaches and in order to create a roadmap for the improvement of the design process by integrating it with SC techniques.

3.1. Fuzzy logic and engineering design

Zadeh initially introduced fuzzy logic in the mid 1960s [29]. The transition from the Aristotelian logic (between two competing states one and only one is true) to the fuzzy logic (multiple competing states may be true at the same time, each one at a different degree of truth) was accepted by the scientific community with hesitation. The ability, however, of modeling the uncertainty through fuzzy logic attracted many researchers that contributed to the founda¬tion of various fuzzy logic concepts relative to engineering design [30]. Towards the direction of establishing fur¬ther fuzzy logic theory, Dubois et al. [31] reviewed the

208 K.M. Saridakis, A.J. Dentsoras /Advanced Engineering Informatics 22 (2008) 202–221

possibility of utilizing fuzzy rules systems, while Roy-chowdhury et al. [32] surveyed defuzzification strategies.

In complex domains with a large set of parameters and in a changing environment, the hierarchical organization of the knowledge is considered as common practice. Torra

[33] reviewed the construction of hierarchical fuzzy systems and focused on aspects related to the construction of this hierarchy. Moreover, the advances in object-oriented pro¬gramming influenced a significant number of researchers

[34] in suggesting object models with relations expressed via fuzzy sets. Considering these fuzzy relations, Cross

[35] examined three kinds of relationships inherent to object models: (a) an instance-of, (b) a-kind-of, and (c) a category on the basis of a generalized object model incor-porating the perspectives of semantic data modeling, artifi¬cial intelligence and database systems.

During the last decades, researchers have been using fuzzy logic as a representation framework in design prob-lems characterized by inherent uncertainty during deci-sion-making. Fuzzy logic has also been applied to design activities where there are needs other than knowledge rep¬resentation, e.g. for cognitive support in reverse engineer¬ing [36] and for decision making in demanding domains, such as conceptual design [37]. Design structure matrices (DSM) have also been used as representation and analysis tools that manage the design process under diverse perspec¬tives. Saridakis et al. [38] address the DSM as a communi¬cating design tool among multiple designers by introducing a fuzzy-logical inference mechanism that permits the col¬laboration among designers on the qualitative definition of the interrelations among the design problem’s entities or tasks. Quality function deployment (QFD) has also been used to translate customer needs into technical design requirements, aiming at the increase of customer satisfac-tion. The co-founder of QFD, Dr. Yoji Akao, has recently published a book that describes QFD, through case studies from big Japanese companies and industries, ranging from manufactured and assembled products, to construction, chemical process, service, and software [39]. The QFD methodologies are based on the house of quality (HOQ), which is a matrix providing a conceptual map for the design process. Temponi et al. [40] implement a fuzzy logic-based methodology to business decision making within the context of the HOQ. In this approach, the com¬munication between different team members is facilitated through fuzzy representation of requirements and identifi¬cation of conflicting and cooperative interrelationships. Finally, a fuzzy inference scheme is used for the reasoning of the implicit relationships between requirements. Towards the direction of effectively capturing the genuine requirements of customers (voice of customers-VoC), Yan et al. [41] use repeated single-criterion sorting com-bined with fuzzy evaluation.

The definitions of relative rankings of design require-ments, customer needs and design concepts have also been a topic for research. Buyukozkan et al. [42] use a general form of the analytic hierarchy process (AHP) and more

specifically of the analytic network process (ANP), in order to prioritize the design requirements by taking into account the degree of their interdependence with the customer needs. In this approach, triangular fuzzy numbers are used to improve the quality of the responsiveness to customer needs and design requirements. Jiao et al. [43] introduce a fuzzy ranking methodology for concept evaluation within the framework of configuration design for mass customiza-tion. This methodology uses fuzzy numbers and linguistic terms to model combinations of tangible and intangible cri¬teria. Wang [44] utilizes a fuzzy outranking model to deter¬mine the non-dominating design concepts. The concepts are compared in pairs and the evaluation is performed by using three types of indices.

A significant part of the research activity focuses on fuzzy decision-making in design. In their research work, Sii et al. [45] address multiple criteria for decision-making. The proposed framework expresses the uncertainty of these criteria through fuzzy numbers and solves the design/pro-curement proposal selection problems through underlying conflicting objectives. Hsiao [46] proposes a fuzzy deci¬sion-making methodology on the basis of hierarchical structures of evaluation objectives. In this approach, fuzzy membership functions are used to model the relative contri¬bution of each objective to the overall value of the solution and the degree of approximation of a solution with respect to a given objective. Furthermore, a generalized weighted-mean method is utilized in order to calculate the fuzzy probability ascending level by level. Chan et al. [47] pro¬pose an enhanced weighted fuzzy reasoning algorithm that is capable of evaluating a decision and is based on the com¬bined effect of individual factors in the antecedent of a rule. The importance of individual factors – or the weights assigned to the factors – can also be tailored according to individual user preference. However, the fuzzy analysis may result in severe computational cost in complex design problems. Under this consideration, Jensen et al. [48] intro¬duce a methodology based on approximation concepts and fuzzy calculus. In the context of this approach, explicit approximate responses reduce the number of system anal¬yses that are required in the fuzzy weighted algorithm (FWA) in order to estimate the membership functions for system outputs.

Some other researchers incorporate functions or fea-tures in fuzzy design models. For instance, Fenga et al. [49] propose a framework for mapping the component requirements into feature-related functions through a fuzzy relation system. This mapping is performed in the context of two perspectives: (a) the importance of requirements/ functions and (b) their measurement. In other approaches, functions and functional requirements are modeled through other design entities. In his thesis, Law [50] describes the method of imprecision (MOI), a method that models the design problem via design parameters and per¬formance variables. Both design parameters and perfor¬mance variables may be expressed with fuzzy sets that represent various levels of preferences on specific intervals

K.M. Saridakis, A.J. Dentsoras / Advanced Engineering Informatics 22 (2008) 202–221 209

of values. The proposed reasoning scheme uses mathemat¬ics of fuzzy sets, suitably adapted to the domain of engi¬neering design, in order to provide solutions that meet the highest overall preference.

Scott [51] used this method of imprecision in his thesis as a basis for the development of a framework capable of accommodating the preferences of multiple designers and for the production of formally optimal solutions according to the criterion of maximum overall preference. Thus, collaboration among multiple designers could be accom¬plished through different proposed aggregation strategies. Jones et al. [52] augmented the method of imprecision with a fuzzy knowledge base that includes a library of compo¬nents, parameters of components expressed with fuzzy sets and stored membership functions. The utilization of the aforementioned knowledge base contributed to the reduc-tion of the time required for an exhaustive search of the design space.

The research activity on implementing fuzzy logic in design is also reported through a large number of applica-tions in various domains such as the design of comfort sys¬tems (heating, cooling, ventilation, shading) [53], the design of machining activities [54], the preliminary design of re¬entry aeroshells [55], the preliminary design of vehicle structure [56], the conceptual design of robot grippers [57], etc.

3.2. Genetic algorithms and engineering design

The genetic algorithms are members of a collection of methodologies known as evolutionary computation (EC). These techniques are based on the selection and evolution processes that are met in nature and imitate these principles in many scientific domains. One of the researchers that worked for the establishment of the genetic algorithms’ theory was Holland [58]. A few years later Goldberg, in his book [59], studied several aspects of the implementation of genetic algorithms and examined their potential in the context of optimization and learning for large-scale com¬plex systems. By changing their focus to engineering design, Bullock et al. [60] investigated the advances of the genetic algorithms, while Rosenman [61] performed an even more specific survey on evolutionary models that are applicable to non-routine design problems. Another important review paper about approaches that apply both evolutionary and adaptive search in engineering design problems was authored by Parmee [62]. Genetic algorithms have also been utilized as creative design tools [63] or as support tools to computer-based systems applied to detailed design [64].

The efficiency of the different architectures of evolution¬ary algorithms in comparison to other heuristic techniques has been tested in both generic [65–67] and engineering design [68] problems. Through these tests, the genetic algo¬rithms are identified as robust heuristic tools capable of delivering efficient and robust solutions to diverse design problems. An interesting research topic for the implemen¬

tation of genetic algorithms is the area of assembly design. Lazzerini et al. [69] deploy genetic algorithms to generate optimal assembly plans. The genetic algorithm produces near-optimal assembly plans starting from a randomly ini¬tialised population of assembly sequences in the context of minimizing both the orientation changes of the product and the gripper replacements, while grouping technologi¬cally similar assembly operations. The quality of the gen¬erated assembly sequences is assessed by a space-state search algorithm that adopts a best-first search algorithm and seeks the path that corresponds to a feasible sequence with the lowest total cost. Another approach for the assembly line planning problem is proposed by Chen et al. [70]. In this study, a hybrid genetic algorithm addresses assembly planning with various objectives, including minimizing cycle time, maximizing workload smoothness, minimizing the frequency of tool change, minimizing the number of tools and machines used and minimizing the complexity of assembly sequences. More¬over, a self-tuning method was developed to enhance the effective schemata of chromosomes during the deployment of the proposed genetic operators. Taura et al. [71] studied the implementation of an adaptive-growth-type 3D repre-sentation based on evolutionary algorithms in the configu¬ration design. The evolution of the shapes expressed in the process takes place through the interaction with a defined outside environment and their interaction among the shapes themselves. The underlying framework allows the determination of 3D shapes and their layouts in the same framework and generates a diversity of shapes and config¬urations that help the designer develop his/her insights and ideas.

Other researchers focus on the product’s underlying architecture and develop evolutionary methods to manipu¬late the design knowledge. Souza et al. [72] examine the trade-off between commonality and individual product per¬formance within a product family and introduce a genetic-algorithm-based method in order to find an acceptable balance between commonality in the product family and desired performance of the individual products in the fam¬ily. In this approach, design of experiments is used to help the screening of unimportant factors and identify factors of interest to the product family, while a multi-objective genetic algorithm optimizes the performance of the products in the resulting family by using a non-dominated sorting strategy. Fan et al. [73] propose an approach for synthesizing system-level designs for multi-domain dynamic systems in an automated manner. This approach uses genetic programming as a competent search method for designs, bond graphs as a representation scheme for dynamic systems, while a hierarchical competition model is adopted for preventing the premature convergence often encountered in evolutionary computation. The result is a design environment that accomplishes open-ended topo-logical search for system-level models of various classes of engineering systems, while providing the designer the option to continue with the next step of embodiment of

210 K.M. Saridakis, A.J. Dentsoras /Advanced Engineering Informatics 22 (2008) 202–221

the conceptual designs or, instead, to gain design insight by analyzing the design candidates.

Although genetic algorithms are well suited for design problems where explicit design knowledge is available, they can also be excellent tools for manipulating vague design knowledge. Zechman et al. [74] suggest that the optimal solution in a real optimization problem, in which some decision-making objectives cannot be modeled or quanti¬fied, must be tagged with a set of good alternative solu¬tions. In the latter paper, an evolutionary algorithm is proposed that systematically searches regions of the objec¬tive space, close to the optimal solution in order to locate alternative solutions that are maximally different in the decision space. If sets of diverge solutions are introduced to the designer and/or to the design team, an interactive/ collaborative decision-making discourse may take place that preserves the flexibility of the design process. In the context of supporting the designer during the early phases of the design process, Taura et al. [75] introduce a model that utilizes representation schemes of free-form shape fea¬tures and aims to the generation of a system capable of holding and manipulating the shape features after synthe¬sis. The proposed shape feature generating process model (SFGP model) takes advantage of developmental biology to devise a computational model of the representation (cell division model) by using a classifier system (CS). Although this model has a limited capability of representing shapes, it can hold specific features while showing a variety of alter¬native shapes after combining already existing shapes.

Several researchers have concluded that genetic algo-rithms may be properly modified in order to be used in engineering design. Rasheed et al. [76] presented a method to improve the efficiency of a genetic algorithm to perform optimization in engineering design by utilizing a sequence of previously explored points to guide further exploration. The proposed methodology recognizes both good and bad solution spaces and is suitable for continuous spaces with expensive evaluation functions, such as those that appear in engineering design. Gero et al. [77] deployed a genetic algorithm that operates in conjunction with the problem-solving process in order to automate the knowledge acqui¬sition and reuse in design. The described algorithm relies on the fact that structural features of the genotype influ¬ence the fitness or behaviour of the resulting designs and, if these features could be isolated, they could be used for the benefit of the design process. These genetic features can also be subjected to genetic engineering operations, such as gene surgery and gene therapy. Since the genetic features are in the form of genes, they can be replaced by a single “evolved”gene and thus extend the range of sym¬bols used in the genetic representation. This could allow the achievement of a more focused solution search by using knowledge, which was not previously available.

Considering the presence of multiple conflicting objec-tives in many engineering design problems, Andersson et al. [78] introduce a method that combines the struggle-genetic-crowding algorithm with Pareto-based population

ranking in order to locate both discontinuous Pareto fron-tiers and multiple Pareto frontiers in multi-modal search spaces. In the proposed Pareto multiple-objective strug-gle-genetic algorithm (MOSGA), the ranking of each indi¬vidual is based on how many members of the population are preferred in a Pareto-optimal sense, while parents are selected uniformly from the population before crossover and mutation operators are initiated. Each produced child is then inserted into the population and replaces the most similar individual to itself only if it has a better ranking. The similarity between two individuals is measured through a function which considers the differences in both parameter and attribute space. Parmee [79] describes a dual agent strategy, namely GAANT, which concurrently com¬bines ant colony search with evolutionary search to manip¬ulate both discrete and continuous variable parameter sets. The communication between the two search-agents results to a more efficient search across the system design hierar¬chy – when compared to that achieved by a structured genetic algorithm – and also to a simplification of the chro-mosomal representation. This simplification allows further development of the preliminary design hierarchy.

Other researchers suggest that the underlying knowledge of the considered design problem may be involved in the genetic operations. Ryoo et al. [80] adopt a genetic algo¬rithm to search a globally compatible solution in a decom¬position-based design environment. In this approach, the design problem is decomposed into sub-problems whose solutions are obtained through co-evolution. Mechanisms based on modification of genetic makeup through experi¬ential inheritance (exposure to another species) and through interspecies’ migration are deployed in order to exchange design information among the temporarily decoupled sub-problems. Moreover, different forms of cou¬pling among sub-problems are investigated, ranging from simple coupling through constraints to coupled objective and constraint functions. In their paper, Wallace et al. [81] suggest a specification-based design evaluation method that emulates the way that specifications are used by prod¬uct designers in a concurrent design environment. The specifications are mapped to acceptability functions, which are capable of capturing the subjective probability and the different performance levels of the design attributes of each solution. The proposed methodology is augmented by a genetic algorithm that uses a penalty-based fitness function on the basis of a logarithmic acceptability metric.

Genetic programming may also contribute to the extrac¬tion of knowledge about the design problem under consid¬eration. Ishino et al. [82] have developed a methodology for estimation of design intent (MEDI) on the basis of a staged design evaluation model that uses two basic algorithms. The first algorithm includes multiple genetic programming, while the second involves both principal component analy¬sis and multivariate regression. Therefore, MEDI can pro¬vide an approximate evaluation of how preferable a specific product model is, while both the structure of target perfor¬mance functions and the approximate values of their

K.M. Saridakis, A.J. Dentsoras / Advanced Engineering Informatics 22 (2008) 202–221 211

weights can be estimated through multivariate genetic programming.

A significant common characteristic of several approaches is the evolution of the design problem under consideration through genetic operations. Maher [83] iden¬tifies design as a co-evolutionary process, with parallel search for both design requirements and design solutions. In this approach, the interaction between requirements and solution redefine the current fitness function that is not necessarily related with convergence in computational evolution. This is explained by the fact that fitness is used to determine which individuals survive and convergence occurs when new ideas cannot be found, thus the termina¬tion conditions do not rely on the fitness of individuals. Parmee et al. [84,85] describe an environment that inte¬grates evolutionary search methodologies and exploration technologies as well as other computational intelligence techniques. The latter research work has been deployed in the perspective of capturing (a) knowledge generated from human-based reasoning and (b) activity by providing interaction between the designer and an iterative evolution¬ary search procedure. The proposed methodology may be used in the conceptual phase of the design process, during which the design knowledge processing together with the discussion among the members of the design team in an iterative designer/evolutionary search procedure may result to a problem reformulation. This reformulation will pro¬vide a more explicit definition and greater reliability in the machine-based representation. Cvetkovic et al. [86] introduce an evolutionary conceptual design system. This system is developed on the basis of a simple architecture based on several software agent classes that perform differ¬ent tasks, supporting the designer to solve various design issues. In the context of this research work, the proposed agents are designed for a conceptual design environment where design goals and constraints are still rather vague. The role of the agents in this environment is to provide help to the designer. Liu et al. [87] suggest a framework to dynamically manage cooperative agents in a distributed environment on the basis of a tree-structure-based genetic algorithm and interactive selection. The proposed frame¬work serves as a continuous novelty generator that may stimulate the imagination of designers through the explor¬ative evolution and not as an optimizing tool. Kampis et al. [88] study systems with both incomplete specification and an incomplete environment description and they introduce an agent-based simulation informed from biological evolu¬tion. In their approach, they deal with the problem of per¬sistent species evolution in an artificial evolutionary system and argue that a species evolution process can help addressing design problems. The incomplete knowledge is handled without human intervention, through an iterated process with changing function space and changing attri¬bute space.

The developments in the research field of genetic pro-gramming have been utilized in various application domains, such as the optimization of linear elastic struc¬

tures [89], the conceptual design of supersonic transport aircraft [90], the conceptual design of a micro-air vehicle [91], the domain of fashion design [92], telecommunication network design [93], wind force analysis and analysis, design and optimisation of steel structures [94], the design of elastic flywheels [95], the design rolling element bearing [96], the design of four bar mechanism [97], the design of combinational logic circuits [98], the problem of manufac¬turing cells formation [99] and many others.

3.3. Artificial neural networks and engineering design

The artificial neural networks were first introduced by McCulloch et al. [100], who suggested that the biological function of the human brain could be emulated by a simpli¬fied computational model. The theory of artificial neural networks (ANNs) remained out of interest for a long time. However, during the last three decades a huge growth in this research field has been reported and several ANN architectures have already proved their efficiency in diverse aspects. A good book for reviewing the fundamentals of ANNs through graphs, algorithms and real-world applica¬tions has been authored by Bose et al. [101]. In another book authored by Gallant [102], neural network learning algorithms are discussed from a computational point of view – together with an extensive exploration of neural net¬work expert systems – showing how learning via neural net¬works could automatically generate expert systems. Potter et al. [103] investigate the applicability of inductive machine learning to the design synthesis in the conceptual design tasks and discusses several issues that must be con¬sidered in order to achieve efficient design solutions and qualified underlying design knowledge. Although some researchers still believe that there are other techniques that perform better than ANNs in specific problems [104], the increasing approaches and applications that implement ANNs as core or supportive elements dominate and char¬acterize the current research trend.

Ivezic et al. [105] introduced a simulation-based decision support system (SB-DSS) approach in order to support the early stages of collaborative design, where collaborative, distributed design teams refine selected conceptual design solutions under uncertain design specifications. The pro¬posed system is based on four components: (a) a behav¬iour-evaluation (BE) model, which is used to structure individual, domain-specific decision models and to orga¬nize these models into a collaborative decision model, (b) a probabilistic framework that enables the management of the uncertainty within a constraint satisfaction environ¬ment by using simulation-based knowledge, (c) a statistical neural network, which captures the simulation-based knowledge and builds the probabilistic behaviour models based on this knowledge, and (d) a Monte Carlo simulation mechanism, which samples the trained neural networks and approximates the likelihoods of design variable values.

Hsu et al. [106] proposed a sequential approximation method that utilizes a back-propagation neural network

212 K.M. Saridakis, A.J. Dentsoras /Advanced Engineering Informatics 22 (2008) 202–221

in order to simulate a rough map of the feasible domain formed by the constraints that are modeled as pass– fail functions. The approximate model follows an iterative process in searching for the optimal point in new approxi¬mated feasible domains. This approach can be imple¬mented in design optimization problems characterized by the existence of implicit or binary constraints and discrete variables.

Chen et al. [107] proposed an approach that is based on the laddering technique and the radial-basis function (RBF) neural network. The proposed system facilitates the new product development by eliciting the customer require¬ments through a three level hierarchical laddering technique and by overcoming the qualitative nature of the imposed requirements through the RBF neural network. This proto¬type system makes it possible for the similarities and differ¬ences between among several respondent groups to be studied both psychologically and computationally.

Ding et al. [108] present a feature recognition method that integrates design by features, ANN techniques and a heuristic algorithm in order to handle feature interactions. A unified data structure models the underlying feature clas¬ses. An input representation that is based on F-adjacent and V-adjacent matrices is implemented in the neural network, which is capable of recognizing the features. Moreover, the proposed heuristic algorithm recognizes the interactions of internal features and classifies them in four types: parent– child, connection, non-connection and overlap hiding.

Yasuda et al. [109] introduce a design methodology of a fault-tolerant autonomous multi-robot system (MRS), which integrates the design of an on-line autonomous behaviour acquisition mechanism. This mechanism devel¬ops cooperative roles and assigns them to a robot appropri¬ately in a noisy embedded environment, by applying (a) reinforcement learning that adopts the Bayesian discrimi¬nation method for segmenting a continuous state space and a continuous action space simultaneously, and (b) a neural network to predict the average of the other robots’ postures at the next time step in order to stabilize the rein¬forcement-learning environment.

Barai et al. [110] propose a heuristic approach called the SG (k-NN) ensemble that is utilized for the systematic gen¬eration of good-quality and diverse models. In this approach, several neural network models are used in com¬binations and the results show that even the worst combi¬nation performs better that any single model thus proving that careful generation of ensembles can improve good-quality models created from good-quality data. However, the effectiveness of the ensemble modeling depends on the accuracy and the diversity of the individual networks. Granitto et al. [111] investigate how to tune the ensemble members in order to have an optimal compromise and per¬form an extensive evaluation of several algorithms for ensemble construction, including new proposals and com¬paring them to standard methods used in the literature.

In some approaches the computational neural network model is augmented with a visualization scheme that is

compliant with the designers’ perception. Towards this direction, Hsiao et al. [112] introduce an approach that uti¬lizes back-propagation neural networks to establish the relationships between product–form parameters and adjec¬tive image words. The connections among the design ele¬ments, product images and shape generation rules are stored in a database that is used by a CAD system that helps designers generate 3D models with different images by providing basic design elements and shape generation rules. Wang et al. [113] developed Creative Stimulator (CreaStim), which is an intelligent interface that enhances creativity in pattern design by helping designers explore innovative pattern designs. CreaStim relies on the catastro¬phe theory, which implies that sudden realization in the thinking process of design may lead to creativity and it is based on a neural network-based imagining engine, a data repository, and its learning strategies. The ANN learns the psychological factors and generalizes new patterns with dif¬ferent psychological requirements.

3.4. Fusion of soft computing techniques and engineering design

Fuzzy logic, genetic algorithms and artificial neural net¬works are not competing to each other, but instead, they may be combined on the basis of integrated frameworks to outperform conventional design approaches. Four books that extensively describe architectures and models of the computational intelligence and their possible fusion have been authored by Kosko [114], Kasabov [115], Koza [116,1] and Cordon et al. [117]. Keeping in mind the engi¬neering design process, several approaches are described in the following paragraphs, which implement combinations of soft computing techniques.

Vico et al. [118] consider design synthesis as an optimiza¬tion problem and under this perspective a neural network is utilized to implement a fitness function for a genetic algo¬rithm that searches for the optimal solution. In this frame¬work, the designer supervises the system and each time a new good result is extracted, he/she checks the validity before feeding the neural network, which adapts and fits better the designer’s criteria. Mitra et al. [119] integrate arti¬ficial neural networks, fuzzy sets, genetic algorithms and rough sets in a knowledge-based network for pattern classi¬fication and rule generation. The proposed method applies rough set theory to directly extract dependency rules from a real-valued attribute table consisting of fuzzy membership values. An algorithm involves synthesis of several fuzzy modules, each encoding the rough set rules for a particular class and a genetic algorithm is deployed to refine these knowledge-based modules. The final rules represent refined knowledge extracted from the trained network. Sasaki et al. [120] propose a method to solve fuzzy multiple-objective optimal system design problems with hybridized genetic algorithms (HGA). This approach enables the design of flexible optimal system by applying fuzzy goals and fuzzy constraints. Moreover, generalized upper bounding

K.M. Saridakis, A.J. Dentsoras / Advanced Engineering Informatics 22 (2008) 202–221 213

(GUB) is applied in order to structure the representation of the chromosomes in the genetic algorithm. Wang et al. [121] suggest an interactive evolutionary approach to synthesize component-based preliminary design engineering problems by combining agent-based hierarchical design representa¬tion, set-based design generation with fuzzy design trade¬off strategy and evolutionary synthesis. The proposed framework facilitates the human–computer interaction to define the fitness function of solutions incorporating both multi-criteria evaluation and constraint satisfaction. Lim et al. [122] automate the formulation of fuzzy rules by means of a genetic algorithm. They suggest that, instead of using the entire solution space, the latter must be system-atically partitioned into smaller subspaces, in which the GA could focus for optimal solutions. This process continues iteratively for various subspaces, until finally a compact set of fuzzy rules is derived.

Xiong et al. [123] present a synthetic mixed-discrete fuzzy nonlinear programming (MDFNLP) optimization method that combines the fuzzy formulation with a genetic algo¬rithm and a traditional, gradient-based optimization strat¬egy. This method can find a globally compromise solution for fuzzy optimization problems containing mixed-discrete design variables, even when the objective functions are non-convex and/or non-differentiable. Delgado et al. [124] utilize Takagi–Sugeno fuzzy models on the basis of a co-evolutionary hierarchical collaborative design approach. The proposed framework induces collaboration among individuals of genetically different populations through fit¬ness sharing among individuals of different species. Critical model parameters such as antecedent aggregation opera¬tors, number of fuzzy rules, type, location and shape of membership functions emerge from the underlying co-evo¬lution. Yang et al. [125] suggest three approaches for imple¬menting and constructing fuzzy neural networks: fuzzy neuron based on fuzzy logic operations, fuzzy neural networks based on fuzzy logic blocks and fuzzy neural networks based on fuzzy reasoning computation. Additionally, a genetic algorithm is proposed as neural learning algorithm. Marcelin [126] describes the use of back-propagation neural networks in creating function approximations of computationally intensive finite element calculations in combination with optimization based on genetic algorithms. Su et al. [127] propose a hybrid approach on the basis of integration of a knowledge base, neural networks, genetic algorithm and CAD/CAE/CAM in a single environment, which can be implemented in var¬ious stages of the design process. The genetic algorithm is used to conduct optimization tasks in the context of achiev¬ing the optimal combination of design parameters, as well as the optimal architecture of the artificial neural networks used in this hybrid system. Yeun et al. [128] have developed a hybrid system of neural networks and genetic program¬ming (GP) trees for problem domains where the complete input space can be decomposed into several different sub¬regions, which are represented in the form of an oblique decision tree. The architecture of this system, called feder¬

ated agents, consists of a facilitator, neural networks, used as local agents that are expert in different sub-regions, and genetic programming (GP) trees that serve as boundary agents. A boundary agent is specialized at the borders of sub-regions where discontinuities or different patterns coex¬ist, while the facilitator is responsible for choosing the boundary agent that is suitable for given input data using the information obtained from oblique decision tree.

Tsai et al. [129] suggest that designers may create a new design in shorter time by modifying previous designs and with this perspective they propose an intelligent design retrieval system that utilizes soft computing techniques. Fuzzy relation and fuzzy composition are used for features associations, while a fuzzy neural network is responsible for the composition of object association functions allow¬ing designers to control the similarity of retrieved designs. Zha [130] studies the assemblability and the assembly sequence evaluation in the engineering design through a neuro-fuzzy approach. According to this approach, the fuzziness is a property of the degree of difficulty assigned to the operation which can be represented by a fuzzy num¬ber between 0 and 1. The assembly operations have been evaluated based on various criteria, such as time and equip¬ment required, although the analysis focuses on the difficulty of operation. Moreover, a neural network automatically tunes the membership functions of assembla-bility factors, so as to adjust the assembly difficulty score. Using the neuro-fuzzy approach, the relationships between product definition data, assembly factor, and assemblabil-ity can be formulated followed by sensitivity analysis that could predict how a design parameter change will affect the assemblability. Saridakis et al. [131] represent the design problem in terms of qualitative and quantitative design parameters and their associative relationships of dif¬ferent formalisms, with a genetic algorithm to be deployed to find the optimum solution according to a specific optimi¬zation criterion. During genetic optimization, the best solu¬tions are recorded and are submitted to a neuro-fuzzy process that limits the number of inputs and outputs and resolves problem’s complexity by substituting existing asso¬ciative relations with a fuzzy rule system.

The fusion of soft computing techniques has also been reported in a significant number of design applications, such as civil engineering (pre-stressed concrete pile diagnosis, concrete mix design, design of industrial roofs) [132], design of adaptive car-following indicator [133], the purchasing decisions at crude oil market [134], the explosive cutting process of plates by shaped charges [135], the maintenance of road pavements [136], the optimization of clamping forces in a machining operation [137], the human–machine workstation design and simulation [138], etc.

3.5. Utilization of soft computing techniques in case-based design

The efficiency of soft computing techniques in activi-ties related to machine learning has been noticed in the

214 K.M. Saridakis, A.J. Dentsoras /Advanced Engineering Informatics 22 (2008) 202–221

literature [139]. As mentioned in a previous section of the present paper (Section 1.3), an important reasoning meth¬odology that resembles the natural human perception is case-based reasoning (CBR), which in the context of engi¬neering design may be referred to as case-based design (CBD). The inclusion of soft computing techniques gains in popularity, as the proposed hybrid systems seem to out¬perform the conventional CBR systems in demanding sci¬entific domains such as engineering design. Dubois et al. [140] study the opportunities of implementing fuzzy logic to perform different activities, which are incorporated in CBR cycle. Pal et al. [141] authored an important book about the synergy of CBR with soft computing and demon¬strates several different architectures and applications on this topic. In a more focused research, Mitra et al. [142] investigated alternative methods for the adaptation proce-dure of the CBR cycle, and they have classified the reviewed methods according to three aspects: (a) domain knowledge requirement, (b) adaptive capabilities of the case adaptation methods, and (c) type of adaptation knowledge required.

In general, case-based reasoning algorithms rely on domain knowledge and heuristics in order to adapt past designs to new problems. The large amount of domain-spe¬cific knowledge and domain- and task-specific heuristics required for such knowledge-based adaptation is surpassed in GENCAD [143] with a general-purpose knowledge lean method based on genetic algorithms for the subtask of case adaptation. GENCAD is used in the domain of structural design of tall buildings. The CREATOR and its successor CREATOR II [144] are case-based creative design systems in the domain of digital circuit design. CREATOR com¬prises four different modules: reasoning, knowledge case, evaluation and meta-control. The design solution is pre¬sented in the form of a hierarchy of the device structures. Adaptation is performed through thematic abstraction (generation of new solution through a single case) and composition (splits or merges case pieces generating new solutions). CIGAR [145] addresses design as an optimiza¬tion problem whose solution is extracted by a GA aug¬mented by a case-based memory that periodically injects similar design solutions to the GA population. Rosenman [146] utilizes an evolutionary algorithm to perform adapta¬tion tasks in a CBR system and evaluates his method in 2D spatial design of houses.

In the context of acquiring knowledge that is used in retrieval and adaptation tasks in CBD systems, researchers have been investigating the impact of specific-to-general and general-to-specific learning on adaptation knowledge [147]. Towards the aforementioned research directions, Craw [148] proposes a GA method that optimizes the decision tree index and applies a “leave-one-out”k-NN method for retrieval. In Soft-CBR [149], in a component-based architecture, fuzzy logic concepts are integrated with an evolutionary algorithm that facilitates the optimization and the maintenance of the system. Tsai et al. [150] com¬bine the fuzzy logic theory with the adaptive resonance the¬

ory (ART) to introduce model, which helps designers searching for similar design cases on both geometric and technological basis through vague and incomplete two-stage queries. The beneficial effect of case-based reasoning, if used as a pre-processor to the genetic optimization, is investigated by Louis et al. [151]. In the proposed frame¬work, solutions from similar past problems are retrieved and then the genetic algorithm’s population is being injected with these retrieved solutions. Kraslawski et al. [152] use a fuzzy neural network to model the problem of selection heat-exchange equipment in mixing tanks and to generate possible solution cases that are incorporated by a case-based reasoning system. Saridakis et al. [153] have developed Case-DeSC, a case-based design system that uses: (a) fuzzy preferences to model the design objectives, (b) hybrid genetic algorithms to execute optimization tasks, and (c) a competitive neural network to retrieve similar past design cases. Several different modules exist in the pro¬posed system, providing the designer with the capability to extract design solutions for ill-structured/creative design problems through case-based retrieval [154] or to support the optimization procedure with the case-based retrieved solutions [155]. Moreover, they have developed a frame¬work, named CopDeSC, which addresses parametric design in the context of collaborative development of fuzzy objectives on design parameters [156].

4. Soft computing and engineering design: evaluation and research opportunities

In the previous sections of the paper, a significant num¬ber of references have been discussed and commented on. The last part of the paper attempts to identify strong and weak points of the proposed SCAD techniques and to per¬form a gap analysis by road-mapping future opportunities of SCAD implementation in the industry. For the evalua¬tion of the reviewed approaches a simple arithmetic meth¬odology is utilized. The methodology may not capture the qualitative nature that a gap analysis should have, but it can be used as a good starting point for a discussion about potential areas of improvement in SCAD. First, a table is created (see Table 1) that contains the strengths of the dependencies between the three identified design tasks (see Section 1) and the eight design issues commented in Section 2. The dependency strengths may have arithmetic values from zero (0) to three (3), corresponding to four qualitative grades (from none to strong). The authors of the present paper have determined these strengths. For example, in Table 1 it is stated that the simplicity of a design approach strongly depends on the design knowledge representation.

Next, an evaluation is performed for each area of soft computing technique in the context of identifying how effi¬ciently each approach of the specific area addresses the design tasks. For example, all referenced approaches of the research area of applications of fuzzy logic on engineer¬ing design are evaluated depending on how well they

K.M. Saridakis, A.J. Dentsoras / Advanced Engineering Informatics 22 (2008) 202–221 215

Table 1

Dependency strength between design tasks and design issues

Design knowledge representation Search for optimal solutions Retrieval and learning of design knowledge

Uncertainty management and conceptual design 3 2 2

Collaboration, communication and coordination 3 2 1

Solution extraction under lacking knowledge/ uncertainty 2 3 3

Management and reuse of existing design knowledge 2 1 3

Generality and domain-independence 2 1 2

Extensibility and connectivity to other tools 2 1 1

Balance between automation and human activities 2 1 2

Simplicity 3 2 2

19 13 16

3 = strong, 2 = medium, 1 = light, 0 = none.

address the three design tasks. Then, an average evaluation metric for all approaches in the specific field is extracted ranging from one to three. This evaluation metric is then multiplied by the strength of their dependency with the design issues, thus providing an evaluation denoted with a percentage about how well the design issues are addressed by the referenced approaches (see relation (1) below). The results of this arithmetic evaluation process are depicted in the diagrams shown in Table 2.

As an example, consider again the section of fuzzy logic approaches. These approaches are evaluated with three (3) for the knowledge representation, with two (2) for the solu¬tion extraction and with one (1) for the retrieval/learning of design knowledge. The three design tasks have respectively strong (=three), medium (=two) and medium (=two) dependencies with “Uncertainty management and concep¬tual design”issue (Table 1-first row). Then the evaluation score (ES) of the fuzzy logic approaches for the specific issue is calculated using the following method:

(ES) = (3 x 3 + 2 x 2 + 1 x 2)

/(maximum possible score) (%)

=(3 x 3 + 2 x 2 + 1 x 2)

/(3 x 3 + 3 x 3 + 3 x 3) = 56(%) (1)

The results of the described arithmetic evaluation are summarized in Table 3. From this table some qualitative conclusions may be drawn. Very low percentages reveal a large efficiency gap and significant opportunity for further research. It is also evident that the fusion of SC techniques provides more robust design frameworks, which address the set of the identified design issues more efficiently. Finally, the combination of SC techniques with case-based design approaches provides even more efficient results regarding these issues, resembling the way that human designers provide solutions (through analogical reasoning).

Design is a process that may integrate diverse activities (i.e. exploratory, creative, rational, interactive, decision-making, iterative, etc.) during the three main design steps of design knowledge representation, optimal solution

search and design knowledge learning. The review in the literature shows that each of these steps may be supported by soft computing techniques, thus providing more efficient results than those provided by the conventional methodol¬ogies. For example, the fuzzy logic-based approaches pro¬vide the capability of modeling uncertainty in design and managing this uncertainty under individual or collabora¬tive frameworks on the basis of both qualitative and quan¬titative design entities. The optimal solution could be searched through a set-based optimization method such as the genetic algorithm, which provides the possibility of avoiding local extremes, resulting in both optimal and robust ‘elite’ solutions, which can be further used in vari¬ous ways. Furthermore, the neural networks may act as powerful learning or approximation tools and be imple-mented at specific design phases. An important research opportunity is to use the SC techniques with the same aforementioned success to learn and develop descriptive processes, related with designers’ best practices and strate¬gies. Furthermore, it could be possible to use SC techniques to imitate/simulate/emulate processes and systems from other scientific or social domains, which could be applied in engineering design under specific aspects.

A significant research activity has been reported towards addressing the collaboration, communication and coordi¬nation in engineering design, but according to the authors’ beliefs there are a lot of opportunities in the context of inte¬grating the three aforementioned issues with the support of SC techniques. Although collaborative discourses may enhance the design approaches in detailed parametric design by communicating the knowledge and reasoning about the individual objectives, the creative and conceptual design problems require a different perspective. This per¬spective should preserve the rationality and consistency of the design knowledge representation, whereas the designers’ intuition and experience should be utilized in a systematic way. Identifying this difference of perspectives on developing approaches for design problems, different development paths should be followed, according to the degree of the encapsulated design knowledge. Although,

216 K.M. Saridakis, A.J. Dentsoras /Advanced Engineering Informatics 22 (2008) 202–221

Table 2

Evaluation diagrams per research area versus identified design tasks and design issues

Fuzzy logic & Engineering design

Genetic Algorithms & Engineering design

Neural Networks & Engineering design

Fusion of SC techniques & Engineering design

Case-Based Design & SC

Design Tasks

AA: Design knowledge representation, BB: Search for optimal solutions, CC: Retrieval and learning of design knowledge.

Design Issues

A: Uncertainty management and conceptual design, B: Collaboration, communication and coordination, C: Solution extraction under lacking knowledge/

uncertainty, D: Management and reuse of existing design knowledge, E: Generality and domain-independence, F: Extensibility and connectivity to other

tools, G: Balance between automation and human activities, H: Simplicity.

K.M. Saridakis, A.J. Dentsoras / Advanced Engineering Informatics 22 (2008) 202–221

Table 3

Summary of the evaluation of the reviewed SC approaches 217

FL (%) GA (%) ANN (%) SC FUSION (%) CBD and SC (%) AVG (%)

Uncertainty management and conceptual design 56 35 41 59 67 51

Collaboration, communication and coordination 52 31 30 52 56 44

Solution extraction under lacking knowledge/uncertainty 56 48 52 70 81 61

Management and reuse of existing design knowledge 41 26 44 48 59 44

Generality and domain-independence 37 22 33 41 48 36

Extensibility and connectivity to other tools 33 19 22 33 37 29

Balance between automation and human activities 37 22 33 41 48 36

Simplicity 56 35 41 59 67 51

Percentages represent the degree of how well the specific issue is addressed by the SC domain.

SC techniques are applied to both problem types, there is no unified SC-based approach that addresses both concep¬tual and detailed designs. This remark is also enforced by the fact that there are still difficulties in modeling and solv¬ing geometry-based and visualized/conceptualized design problems. Furthermore, from the survey performed on the enterprises (see Section 2), it becomes evident that the SCAD approaches are not capable of addressing specific design issues that commonly underlie in industries design activities even in the case that the SC techniques are inte¬grated and combined in different ways of hybridism. In order to have these design issues addressed, there is a need for holistic approaches using information and AI technol-ogies as integrated frameworks capable of accommodating the SCAD modules as autonomous and independent agents. These integrated frameworks should be deployed through web-based platforms enhancing collaborative and argumentative discourses among designers. Although the ideal SCAD system should provide the maximum level of automation and/or support to the designer, its underly¬ing architecture should preserve a human-centric character that takes advantage of human intuition, creativity and experience.

5. Conclusions

The present article has been authored with the intention of focusing on the latest advances and to highlight the opportunities provided by the implementation of soft com¬puting techniques in engineering design. During the last decades the application of artificial intelligence techniques and methods has proved to be beneficial for the expansion of our knowledge on design process and the achievement of better products and artifacts. The combination of these fields leads to more complex architectures able to address multiple – and more complex – design issues such as repre¬sentation, solution search, learning, etc. Additionally, the evaluation of the reviewed approaches, which is attempted in the context of the present article, may lead to some valu¬able conclusions.

As the scientific domain of SC is related with increasing research activity and newly introduced SCAD approaches seem to be capable of outperforming the existing design

approaches, every design framework that utilizes SC tech¬niques ought to provide extendable architecture. Conse¬quently, a design method (either SC-supported or not) should axiomatically be characterized by the principle of continuous improving design, which may be achieved under different perspectives (improved strategies, increas¬ing quantity and quality of existing design knowledge, etc.). Provided that the design method is characterized by flexibility and extensibility, then it is possible for the design product to adapt to specific customer requirements, pro¬viding a tailor-oriented design process, which has recently become a new trend in the market reality.

However, it seems that -until now- the soft computing techniques are mainly utilized to perform specific design tasks (e.g. representation, optimization, etc.) and the approaches that deploy soft computing methods in an inte¬grated manner are rare (if not existent at all). Additionally, although many of the existing SCAD approaches perform efficiently in domain-specific design problems, they miss applicability or robustness in case of problems with multi¬ple compensating characteristics (e.g. co-existence of quan¬titative and qualitative design parameters, common models for the conceptual and detailed design phases). Therefore, the expansion of SCAD systems in industrial activities ren¬der the deployment of supportive technologies such as agent-based, web-based and AI systems imperative in order to address the identified obstructive design issues. The abovementioned remarks may be considered as future opportunities for the SCAD research community, with the perspective of developing integrated multi-tasking and multi-purpose SCAD methodologies/systems in applied design centers and in the industry.

Acknowledgements

The present research work has been done within the framework of the project Pythagoras II (EPEAEK). Uni-versity of Patras is a member of the EU-funded I*PROMS Network of Excellence.

References

[1] S. Finger, J.R. Dixon, A review of research in mechanical engineer¬ing design. Part I: descriptive, prescriptive, and computer-based

218 K.M. Saridakis, A.J. Dentsoras /Advanced Engineering Informatics 22 (2008) 202–221

models of design processes. Part II: representations, analysis and design for the life cycle, Research in Engineering Design (1) (1989) 51–67, see pp. 121–137.

[2] N.F.O. Evbuowman, S. Sivaloganathan, A. Jebb, A survey of design philosophies, models, methods and systems, Proc Insts Mech Engrs Part B: Journal of Engineering Manufacture 210 (1996) 301–320.

[3] K. Ueda, Synthesis and emergence: research overview, Advanced Engineering Informatics 15 (2001) 321–327.

[4] L. Wang, W. Shen, H. Xie, J. Neelamkavil, A. Pardasani, Collab-orative conceptual design-state of the art and future trends, Computer-Aided Design (34) (2002) 981–996.

[5] N. Suh, Axiomatic Design: Advances and Applications, Oxford University Press, 2001.

[6] A.D. Dimarogonas, Philosophical issues in engineering design, Design Transactions of SDPS 1 (1) (1997) 54–75.

[7] C.L. Dym, Engineering Design: A Synthesis of View, Cambridge University Press, New York, 1994.

[8] R. Audi, The Cambridge dictionary of philosophy, Cambridge University Press, second ed., 1999, ISBN: 0521637228.

[9] W.C. Regli, X. Hu, M. Atwood, W. Sun, A survey of design rationale systems: approaches, representation, capture and retrieval, Engineering with Computers (16) (2000) 209–235.

[10] V.V. Kryssanov, H. Tamaki, S. Kitamura, Understanding design fundamentals: how design synthesis and analysis drive creativity, resulting in emergence, Advanced Engineering Informatics 15 (2001) 329–342.

[11] P. Papalambros, D. Wilde, Principles of Optimal Design, Springer, New York, 1988.

[12] S.B. Eom, Decision support systems research: current state and trends, Industrial Management & Data Systems 99 (5) (1999) 213– 220.

[13] G.A. Hazelrigg, Validation of engineering design alternative selection methods, Engineering Optimization 35 (2) (2003) 103– 120.

[14] J.L. Kolodner, Extending problem solving capabilities through case-based inferences, in: Proceedings of Fourth International Workshop on Machine Learning, 1987, pp. 167–178.

[15] L.M. Maher, B.M. Balachandran, M.D. Zhang, Case-based Reason-ing in Design, Lawrence Erlbaum Associates, Publishers, Mahwah, New Jersey, 1995.

[16] D.L. Grecu, D.C. Brown, Dimensions of machine learning in design, Artificial Intelligence for Engineering Design, Analysis and Manu¬facturing (12) (1998) 117–121.

[17] S.K. Sim, A.H.B. Duffy, A foundation for machine learning in design, Artificial Intelligence for Engineering Design, Analysis and Manufacturing (12) (1998) 193–209.

[18] J.R. Dixon, M.K. Simons, An architecture for application of artificial intelligence to design, in: 21st Design Automation Confer-ence, IEEE, Paper 40.2, 1984, pp. 634–640.

[19] W.P.S. Dias, Reflective practice, artificial intelligence, and engineer¬ing design: common trends and interrelationships, Artificial Intelli¬gence for Engineering Design, Analysis and Manufacturing (16) (2002) 261–271.

[20] A. Papavasileiou, E. Gegov, K. Gavros, Application of the artificial intelligence methods in CAD/CAM/CIM systems, Annual of the University of Mining and Geology “St. Ivan Rilski”, Mechaniza¬tion, electrification and automation in mines, vol. 44–45, part III, Sofia, 2002, pp. 75–79.

[21] T.W. Simpson, J.D. Peplinski, P.N. Koch, J.K. Allen, Metamodels for computer-based engineering design: survey and recommenda-tions, Engineering with Computers (17) (2001) 129–150.

[22] S.H. Liao, Expert system methodologies and applications—a decade review from 1995 to 2004, Expert Systems with Applications (28) (2005) 93–103.

[23] S.S. Rao, A. Nahm, Z. Shi, X. Deng, A. Syamil, Artificial intelligence and expert systems applications in new product devel-opment – a survey, Journal of Intelligent Manufacturing (10) (1999) 231–244.

[24] C.H. Hsu, B.C. Jiang, E.S. Lee, Fuzzy neural network modeling for product development, Mathematical and Computer Modeling (29) (1999) 71–81.

[25] A. Kamiya, S.J. Ovaskab, R. Roy, S. Kobayashi, Fusion of soft computing and hard computing for large-scale plants: a general model, Applied Soft Computing (5) (2005) 265–279.

[26] Y. Dote, S.J. Ovaska, Industrial applications of soft computing: a review, in: Proceedings of the IEEE, 2001, pp. 1243–1265.

[27] N. Cross, Natural intelligence in design, Design Studies (20) (1999) 25–39.

[28] C. Shakeri, D.C. Brown, Constructing design methodologies using multiagent systems, Artificial Intelligence for Engineering Design, Analysis and Manufacturing (18) (2004) 115–134.

[29] L.A. Zadeh, Fuzzy sets, Information Control (8) (1965) 338–353.

[30] I. Gazdik, Zadeh extension principle in design reliability, Fuzzy Sets and Systems (83) (1996) 169–178.

[31] D. Dubois, H. Prade, What are the fuzzy rules and how to use them, Fuzzy Sets and Systems (84) (1996) 169–185.

[32] S. Roychowdhury, W. Pedrycz, A survey of defuzzification strate-gies, International Journal of Intelligent Systems (16) (2001) 679– 695.

[33] V. Torra, A review of the construction of hierarchical fuzzy systems, International Journal of Intelligent Systems (17) (2002) 531–543.

[34] J. Lee, J.Y. Kuo, N.L. Xue1, A note on current approaches to extending fuzzy logic to object-oriented modeling, International Journal of Intelligent Systems (16) (2001) 807–820.

[35] V.V. Cross, Fuzzy extensions for relationships in a generalized object model, International Journal of Intelligent Systems (16) (2001) 843–861.

[36] J.H. Jahnke, Cognitive support in software reengineering based on generic fuzzy reasoning nets, Fuzzy Sets and Systems (145) (2004) 3– 27.

[37] S.H. Yeo, M.W. Mak, S.A.P. Balon, Analysis of decision-making methodologies for desirability score of conceptual design, Journal of Engineering Design 15 (2) (2004) 195–208.

[38] K.M. Saridakis, A.J. Dentsoras, A fuzzy rule-based approach for the collaborative formation of design structure matrices, in: 25th SGAI International Conference on Innovative Techniques and Applica¬tions of Artificial Intelligence, Cambridge, 2005.

[39] Y. Akao, Quality Function Deployment: Integrating Customer Requirements into Product Design, Productivity Press, 2004, ISBN 1563273136.

[40] C. Temponi, J. Yen, W.A. Tiao, House of quality: a fuzzy logic based requirements analysis, European Journal of Operational Research (117) (1999) 340–354.

[41] W. Yan, C.H. Chen, L.P. Khoo, An integrated approach to the elicitation of customer requirements for engineering design using picture sorts and fuzzy evaluation, Artificial Intelligence for Engi-neering Design, Analysis and Manufacturing (16) (2002) 59–71.

[42] G. Buyukozkan, T. Ertay, C. Kahraman, D. Ruan, Determining the importance weights for the design requirements in the house of quality using the fuzzy analytic network approach, International Journal of Intelligent Systems (19) (2004) 443–461.

[43] J. Jiao, M.M. Tseng, Fuzzy ranking for concept evaluation in configuration design for mass customization, Concurrent Engineer-ing: Research and Application 6 (3) (1998) 189–206.

[44] J.R. Wang, Ranking engineering design concepts using a fuzzy outranking preference model, Fuzzy Sets and Systems (119) (2001) 161–170.

[45] H.S. Sii, J. Wang, I. Jenkinson, J.B. Yang, L. Liu, A design-decision support framework for evaluation of design options/proposals using a fuzzy-logic-based composite structure methodology, Journal of Engineering Design 15 (5) (2004) 493–514.

[46] S.W. Hsiao, Fuzzy logic based decision model for product design, International Journal of Industrial Ergonomics (21) (1998) 103–116.

[47] C.W. Chan, P. Lau, Representing user preference in engineering design domains using an enhanced weighted fuzzy reasoning algorithm, Artificial Intelligence in Engineering (13) (1999) 1–10.

K.M. Saridakis, A.J. Dentsoras / Advanced Engineering Informatics 22 (2008) 202–221 219

[48] H.A. Jensen, A.E. Sepulveda, Use of approximation concepts in fuzzy design problems, Advances in Engineering Software (31) (2000) 263–273.

[49] C.X. Fenga, P.G. Lib, M. Liang, Fuzzy mapping of requirements onto functions in detail design, Computer-Aided Design (33) (2001) 425–437.

[50] S. Law, Evaluating imprecision in engineering design, PhD thesis, California Institute of Technology, Pasadena, CA, 1996.

[51] M.J. Scott, Formalizing negotiation in engineering design, PhD thesis, California Institute of Technology, Pasadena, CA, 1999.

[52] J.D. Jones, Y. Hua, A fuzzy knowledge base to support routine engineering design, Fuzzy Sets and Systems (98) (1998) 267–278.

[53] A.I. Dounis, M.J. Santamouris, C.C. Lefas, A. Argiriou, Design of a fuzzy set environment comfort system, Energy and Buildings (22) (1995) 1–87.

[54] Y. Chen, A. Hui, R. Du, Fuzzy expert system machining operations, Int. Journal Mach. Tools Manufact. 35 (12) (1995) 1605–1621.

[55] M.J. Scott, R.W. Kaiser, M. Dilligan, R.J. Claser, Erik K. Antonsson, Managing uncertainty in aeroshell design analysis, in: Proceedings of DETC’97 ASME Design Engineering Technical Conferences, September 14–17, Sacramento, California, 1997.

[56] M.J. Scott, E.K. Antonsson, Preliminary vehicle structure design: an industrial application of imprecision in engineering design, in: Proceedings of DETC’98 ASME Design Engineering Technical Conferences, September 13–16, Atlanta, Georgia, 1998.

[57] V.C. Moulianitis, N.A. Aspragathos, A.J. Dentsoras, A model for concept evaluation in design – an application to mechatronics design of robot grippers, Mechatronics (14) (2004) 599–622.

[58] J.H. Holland, Adaptation in Natural and Artificial Systems, The University of Michigan Press, Ann Arbor, 1975.

[59] D.E. Goldberg, Genetic Algorithms in Search: Optimization and Learning, Addison-Wesley Publishing Co., Reading, 1989.

[60] G.N. Bullock, M.J. Denham, I.C. Parmee, G. Wade, Developments in the use of the genetic algorithm in engineering design, Journal of Design Studies (16) (1995) 507–524.

[61] M.A. Rosenman, An exploration into evolutionary models for non-routine design, Artificial Intelligence in Engineering (11) (1997) 287–293.

[62] I.C. Parmee, A review of evolutionary and adaptive search in engineering design, Evolutionary Optimization: An International Journal 1 (1) (1999) 13–39.

[63] D.E. Goldberg, The Design of Innovation, Kluwer Academic Publishers, Massachusetts, 2002.

[64] G. Renner, A. Ekart, Genetic algorithms in computer-aided design, Computer-Aided Design (35) (2003) 709–726.

[65] S. Areibi, M. Moussa, H. Abdullah, A comparison of genetic/ memetic algorithms and other heuristic search techniques, in: ICAI 2001, Las Vegas, Nevada, 2001.

[66] H. Youssef, S.M. Sait, H. Adiche, Evolutionary algorithms, simu-lated annealing and tabu search: a comparative study, Engineering Applications of Artificial Intelligence (14) (2001) 167–181.

[67] E. Elbeltagi, T. Hegazy, D. Grierson, Comparison among five evolutionary-based optimization algorithms, Advanced Engineering Informatics (19) (2005) 43–53.

[68] L. Giraud-Moreau, P. Lafon, A comparison of evolutionary algorithms for mechanical design components, Engineering Optimi-zation (34) (2002) 307–322.

[69] B. Lazzerini, F. Marcelloni, A genetic algorithm for generating optimal assembly plans, Artificial Intelligence in Engineering (14) (2000) 319–329.

[70] R.S. Chen, K.Y. Lu, S.C. Yu, A hybrid genetic algorithm approach on multi-objective of assembly planning problem, Engineering Applications of Artificial Intelligence (15) (2002) 447–457.

[71] T. Taura, I. Nagasaka, Adaptive-growth-type 3D representation for configuration design, Artificial Intelligence for Engineering Design, Analysis and Manufacturing (13) (1999) 171–184.

[72] B. Souza, T.W. Simpson, A genetic algorithm based method for product family design optimization, Engineering Optimization 35 (1) (2003) 1–18.

[73] Z. Fan, K. Seo, J. Hu, E.D. Goodman, R.C. Rosenberg, A novel evolutionary engineering design approach for mixed-domain sys-tems, Engineering Optimization 36 (2) (2004) 127–147.

[74] E.M. Zechman, S.R. Ranjithan, An evolutionary algorithm to generate alternatives (EAGA) for engineering optimization prob-lems, Engineering Optimization 36 (5) (2004) 539–553.

[75] T. Taura, I. Nagasaka, A. Yamagishi, Application of evolutionary programming to shape design, Computer-Aided Design 30 (1) (1998) 29–35.

[76] K. Rasheed, H. Hirsh, Learning to be selective in genetic-algorithm-based design optimization, Artificial Intelligence for Engineering Design, Analysis and Manufacturing (13) (1999) 157–169.

[77] J.S. Gero, V. Kazakov, Machine learning in design using genetic engineering-based genetic algorithms, in: Workshop on Machine Learning in Design Research, Artificial Intelligence in Design, Worcester, 2000, pp. 1–6.

[78] J. Andersson, D. Wallace, Pareto optimization using the struggle genetic crowding algorithm, Engineering Optimization (34) (2002) 623–643.

[79] I.C. Parmee, Evolutionary and adaptive strategies for efficient search across whole system engineering design hierarchies, Artificial Intel¬ligence for Engineering Design, Analysis and Manufacturing (12) (1998) 431–445.

[80] J. Ryoo, P. Hajela, Decomposition-based design optimization method using genetic co-evolution, Engineering Optimization 36 (3) (2004) 361–378.

[81] D.R. Wallace, M.J. Jakiela, W.C. Flowers, Design search under probabilistic specifications using genetic algorithms, Computer-Aided Design 28 (5) (1996) 405–421.

[82] Y. Ishino, Y. Jin, Estimate design intent: a multiple genetic programming and multivariate analysis based approach, Advanced Engineering Informatics (16) (2002) 107–125.

[83] M.L. Maher, A model of co-evolutionary design, Engineering with Computers (16) (2000) 195–208.

[84] I.C. Parmee, C.R. Bonham, Towards the support of innovative conceptual design through interactive designer-evolutionary com-puting strategies, Artificial Intelligence for Engineering Design, Analysis and Manufacturing (14) (2000) 3–16.

[85] I.C. Parmee, Improving problem definition through interactive evolutionary computation, Artificial Intelligence for Engineer¬ing Design, Analysis and Manufacturing (16) (2002) 185– 202.

[86] D. Cvetkovic, I. Parmee, Agent-based support within an interactive evolutionary design system, Artificial Intelligence for Engineering Design, Analysis and Manufacturing (16) (2002) 331–342.

[87] H. Liu, M. Tang, Evolutionary design in a multi-agent design environment, Applied Soft Computing 6 (2) (2006) 207–220.

[88] G. Kampis, L. Gulyas, Emergence out of interaction: developing evolutionary technology for design innovation, Advanced Engineer¬ing Informatics 20 (2006) 313–320.

[89] J.F.A. Madeira, H. Rodriguesa, H. Pina, Multi-objective optimiza-tion of structures topology by genetic algorithms, Advances in Engineering Software (36) (2005) 21–28.

[90] A.H.W. Bos, Aircraft conceptual design by genetic-gradient-guided optimization, Engineering Applications of Artificial Intelligence (11) (1998) 377–382.

[91] T.T.H. Ng, G.S.B. Leng, Application of genetic algorithms to conceptual design of a micro-air vehicle, Engineering Applications of Artificial Intelligence (15) (2002) 439–445.

[92] H.S. Kim, S.B. Cho, Application of interactive genetic algorithm to fashion design, Engineering Applications of Artificial Intelligence (13) (2000) 635–644.

[93] J. Arabas, S. Kozdrowski, Applying an evolutionary algorithm to tele-communication network design, IEEE Transactions on Evolu-tionary Computation 5 (4) (2001).

[94] K. Murawski, T. Arciszewski, K. De Jong, Evolutionary computa-tion in structural design, Engineering with Computers (16) (2000) 275–286.

220 K.M. Saridakis, A.J. Dentsoras /Advanced Engineering Informatics 22 (2008) 202–221

[95] D. Eby, R.C. Averill, W.F. Punch III, E.D. Goodman, Optimal design of flywheels using an injection island genetic algorithm, Artificial Intelligence for Engineering Design, Analysis and Manu-facturing (13) (1999) 327–340.

[96] I. Chakraborty, V. Kumar, S.B. Nair, R. Tiwari, Rolling element bearing design through genetic algorithms, Engineering Optimiza-tion (35) (2003) 649–659.

[97] A. Ekart, A. Markus, Using genetic programming and decision trees for generating structural descriptions of four bar mechanisms, Artificial Intelligence for Engineering Design, Analysis and Manu-facturing (17) (2003) 205–220.

[98] C.A. Coello, A.H. Aguirre, Design of combinational logic circuits through an evolutionary multiobjective optimization approach, Artificial Intelligence for Engineering Design, Analysis and Manu-facturing (16) (2002) 39–53.

[99] Z. Car, T. Mikac, Evolutionary approach for solving cell-formation problem in cell manufacturing, Advanced Engineering Informatics 20 (2006) 227–232.

[100] W.S. McCulloch, W.A. Pitts, A logical calculus of the ideas immanent in neural nets, Bull. Math.Biophysics (5) (1943) 115– 133.

[101] N.K. Bose, P. Liang, Neural Network Fundamentals with Graphs, Algorithms and Applications, McGraw-Hill International Editions, 1996.

[102] S.I. Gallant, Neural Network Learning and Expert Systems, The MIT Press, 1993.

[103] S. Potter, M.J. Darlington, S.J. Culley, P.K. Chawdhry, Design synthesis knowledge and inductive machine learning, Artificial Intelligence for Engineering Design, Analysis and Manufacturing (15) (2001) 233–249.

[104] A. Abraham, D. Steinberg, Is neural network a reliable forecaster on earth? A MARS query!, Bio-inspired applications of connectionism, Lecture Notes in Computer Science, vol. 2085, Springer-Verlag, 2001, pp. 679–686.

[105] N. Ivezic, J.H. Garrett Jr., Machine learning for simulation-based support of early collaborative design, Artificial Intelligence for Engineering Design: Analysis and Manufacturing (12) (1998) 123– 139.

[106] Y.L. Hsu, S.G. Wang, C.C. Yu, A sequential approximation method using neural networks for engineering design optimization problems, Engineering Optimization 35 (5) (2003) 489–511.

[107] C.H. Chen, L.P. Khoo, W. Yan, Evaluation of multicultural factors from elicited customer requirements for new product development, Research in Engineering Design (14) (2003) 119–130.

[108] L. Ding, Y. Yue, Novel ANN-based feature recognition incorpo¬rating design by features, Computers in Industry 55 (2) (2004) 197– 222.

[109] T. Yasuda, K. Ohkura, K. Ueda, A homogeneous mobile robot team that is fault-tolerant, Advanced Engineering Informatics 20 (2006) 301–311.

[110] S.V. Barai, Y. Reich, Ensemble modelling or selecting the best model: Many could be better than one, Artificial Intelligence for Engineering Design, Analysis and Manufacturing (13) (1999) 377– 386.

[111] P.M. Granitto, P.F. Verdes, H.A. Ceccatto, Neural network ensembles: evaluation of aggregation algorithms, Artificial Intelli-gence (163) (2005) 139–162.

[112] S.W. Hsiao, H.C. Huang, A neural network based approach for product form design, Design Studies (23) (2002) 67–84.

[113] X. Wang, M.X. Tang, J. Frazer, Creative stimulator: an interface to enhance creativity in pattern design, Artificial Intelligence for Engineering Design, Analysis and Manufacturing (15) (2001) 433– 440.

[114] B. Kosko, Neural Networks and Fuzzy Systems: A Dynamical Systems Approach to Machine Intelligence, Prentice-Hall, Engle-wood Cliffs, 1992.

[115] K.N. Kasabov, Foundation of Neural Networks, Fuzzy Systems and Knowledge Engineering, MIT Press, 1996.

[116] R.J. Koza, Genetic Programming, On the Programming of Com-puters by Means of Natural Selection, The MIT Press, 2000.

[117] O. Cordon, F. Herrera, F. Hoffman, L. Magdalena, Genetic Fuzzy Systems: Evolutionary Tuning and Learning of Fuzzy Knowledge Bases, World Scientific Co. Ltd, 2001.

[118] F.J. Vico, F.J. Veredas, J.M. Bravo, J. Almaraz, Automatic design synthesis with artificial with artificial intelligence techniques, Arti-ficial Intelligence in Engineering (13) (1999) 251–256.

[119] S. Mitra, P. Mitra, S.K. Pal, Evolutionary modular design of rough knowledge-based network using fuzzy attributes, Neurocomputing (36) (2001) 45–66.

[120] M. Sasaki, M. Gen, Fuzzy multiple objective optimal system design by hybrid genetic algorithm, Applied Soft Computing (2/3) (2003) 189–196.

[121] J. Wang, J. Terpenny, Interactive evolutionary solution synthesis in fuzzy set-based preliminary engineering design, Journal of Intelligent Manufacturing (14) (2003) 153–167.

[122] M.H. Lim, W. Ng, Iterative genetic algorithm for learning efficient fuzzy rule set, Artificial Intelligence for Engineering Design, Analysis and Manufacturing (17) (2003) 335–347.

[123] Y. Xiong, S.S. Rao, Fuzzy nonlinear programming for mixed-discrete design optimization through hybrid genetic algorithm, Fuzzy Sets and Systems (146) (2004) 167–186.

[124] M.R. Delgado, F.V. Zuben, F. Gomid, Coevolutionary genetic fuzzy systems a hierarchical collaborative approach, Fuzzy Sets and Systems (141) (2004) 89–106.

[125] Y. Yang, X. Xu, W. Zhang, Design neural networks based fuzzy logic, Fuzzy Sets and Systems (114) (2000) 325–328.

[126] J.L. Marcelin, Evolutionary optimisation of mechanical structures – towards an integrated optimisation, Engineering with Computers (15) (1999) 326–333.

[127] D. Su, M. Wakelam, Evolutionary optimization within an intelligent hybrid system for design integration, Artificial Intelligence for Engineering Design, Analysis and Manufacturing (13) (1999) 351– 363.

[128] Y.S. Yeun, K.H. Lee, Y.S. Yang, Function approximations by coupling neural networks and genetic programming trees with oblique decision trees, Artificial Intelligence in Engineering (13) (1999) 223–239.

[129] C.Y. Tsai, C.A. Chang, Fuzzy neural networks for intelligent design retrieval using associative manufacturing features, Journal of Intelligent Manufacturing (14) (2003) 183–195.

[130] X.F. Zha, Neuro-fuzzy comprehensive assemblability and assembly sequence evaluation, Artificial Intelligence for Engineering Design, Analysis and Manufacturing (15) (2001) 367–384.

[131] K.M. Saridakis, A.J. Dentsoras, Evolutionary neuro-fuzzy model-ling in parametric design, in: First I*PROMS Virtual International Conference on Intelligent Production Machines and Systems, IPROMS 2005, 2005.

[132] S. Rajasekaran, M.F. Febin, J.V. Ramasamy, Artificial fuzzy neural networks in civil engineering, Computers & Structures 61 (2) (1996) 291–302.

[133] P.C. Lu, The application of fuzzy neural network techniques in constructing an adaptive car-following indicator, Artificial Intelli-gence for Engineering Design, Analysis and Manufacturing (12) (1998) 231–242.

[134] M.R. Gholamiana, S.M.T. Fatemi Ghomia, M. Ghazanfarib, A hybrid systematic design for multiobjective market problems – a case study in crude oil markets, Engineering Applications of Artificial Intelligence (18) (2005) 495–509.

[135] N. Nariman–Zadeh, A. Darvizeh, M.H. Dadfarmai, Adaptive neurofuzzy inference systems networks design using hybrid genetic and singular value decomposition methods for modeling and prediction of the explosive cutting process, Artificial Intel¬ligence for Engineering Design, Analysis and Manufacturing 17 (2003) 313–324.

[136] V. Loia, S. Sessa, U.A. Staiano, R. Tagliaferri, Merging fuzzy logic, neural networks, and genetic computation in the design of a

K.M. Saridakis, A.J. Dentsoras / Advanced Engineering Informatics 22 (2008) 202–221 221

decision-support system, International Journal of Intelligent Sys-tems (15) (2000) 575–594.

[137] M. Hamedi, Intelligent fixture design through a hybrid system of artificial neural network and genetic algorithm, Artificial Intelligence Review 23 (2005) 295–311.

[138] X.F. Zha, Soft computing framework for intelligent human– machine system design, simulation and optimization, Soft Comput-ing (7) (2003) 184–198.

[139] L.A. Quiroga, L.C. Rabelo, Learning from examples: a review of machine learning, neural networks and fuzzy logic paradigms, Computers in Engineering 29 (1–4) (1995) 561–565.

[140] D. Dubois, H. Prade, Fuzzy set modelling in case-based reason¬ing, International Journal of Intelligent Systems (13) (1998) 345– 373.

[141] K.S. Pal, S.T. Dillon, S.D. Yeung, Soft Computing in Case Based Reasoning, Springer-Verlag, London, 2001.

[142] R. Mitra, J. Basak, Methods of case adaptation: a survey, International Journal of Intelligent Systems (20) (2005) 627– 645.

[143] M.L. Maher, A.G.S. Garza, A process model for evolutionary design case adaptation, Proceedings of the Fourth International Conference on Artificial Intelligence in Design, Kluwer Academic Publishers, 2000.

[144] L. Machado, F.C. Pereira, C. Grilo, A. Gardoso, A computational model for creative planning, Mind Modelling: A Cognitive Science Approach to Reasoning, Learning and Discovery, Pabst, Lengerich, Germany, 1998, pp. 193–208.

[145] S. Louis, Case injected genetic algorithms for learning across problems, Engineering Optimization 36 (2) (2004) 237–247.

[146] M. Rosenman, Case-based evolutionary design, Artificial Intelli-gence for Engineering Design, Analysis and Manufacturing (14) (2000) 17–29.

[147] N. Wiratunga, S. Craw, R. Rowe, Learning to Adapt for Case-based Design, ECCBR, Springer-Verlag, Berlin, Heidelberg, 2002, pp. 421–435.

[148] S. Craw, Introspective Learning to Build Case-based Reasoning (CBR) Knowledge Containers, MLDM, Springer-Verlag, Berlin, Heidelberg, 2003, pp. 1–6.

[149] K. Aggour, M. Pavese, P. Bonissone, W. Cheetham, SOFT-CBR: a self-optimizing fuzzy tool for case-based reasoning, ICCBR (2003) 5–19.

[150] C.Y. Tsai, C.A. Chang, A two-stage fuzzy approach to feature-based design retrieval, Computers in Industry 56 (5) (2005) 493–505.

[151] S.J. Louis, Genetic learning for combinational logic design, Soft Computing (9) (2005) 38–43.

[152] A. Kraslawski, W. Pedrycz, L. Nystrom, Fuzzy neural network as instance generator for case-based reasoning system-an example of selection of heat exchange equipment in mixing tanks, Neural Computing & Applications (8) (1999) 106–113.

[153] K.M. Saridakis, A.J. Dentsoras, Case-DeSC: a system for case-based design with soft computing techniques, Expert Systems with Applications 32 (2007) 641–657.

[154] K.M. Saridakis, A.J. Dentsoras, P.E.Radel, V.G. Saridakis, N.V. Exintari, Neuro-fuzzy case-based design: an application in structural design, in: 2006 I*PROMS Virtual International Conference on Intelligent Production Machines and Systems, 2006.

[155] K.M. Saridakis, A.J. Dentsoras, Using case-based reasoning and soft computing techniques for the initialization of engineering design optimization, in: 8th Biennial ASME Conference on Engineering Systems Design and Analysis ESDA 2006, Turin, Italy, 2006.

[156] K.M. Saridakis, A.J. Dentsoras, Collaborative parametric design with soft computing techniques, Advanced Engineering Informatics 20 (2006) 379–399.

NEW YORK UNIVERSITY

SCHOOL OF CONTINUING

AND PROFESSIONAL STUDIES

CORE COURSES: TIER III 3

Y64.1095 Real Estate Capital Markets 3

CONCENTRATIONS (Choose one area of concentration)

FINANCE AND INVESTMENT 9

Y64.2300 Real Estate Finance and Investment Analysis 3

Y64.2315 Risk and Portfolio Management 3

Y64.2399 (Capstone Required) Applied Project in Finance and Investment 3

STRATEGIC REAL ESTATE MANAGEMENT 9

Y64.2610 Strategic Real Estate Management 3

Y64.2635 Commercial Lease Analysis 3

Y64.2699 (Capstone Required) Applied Project in Strategic Real Estate Management 3

DIGITAL DESIGN APPLICATIONS FOR REAL ESTATE 15

Students are required to take all 4 courses in discrete and consecutive semesters as designated and approved by a faculty advisor. Note: This concentration has no elective options.

Y64.2705 3 D Production 1—Tool Sets 3

Y64.2710 CG Production Design 3

Y64.2715 Lighting and Rendering 3

Y64.2799 (Capstone Required—Thesis Project) Applied Project in Digital Design 6

ELECTIVES (Two Courses) 3

Students are required to select two courses in any combination of the following: (a) from the list below, (b) from a course(s) within another concentration in this program, or (c) with permission from the program director, from a course(s) from the M.S. in Real Estate Development or M.S. in Construction Management degree programs.

Please note that not all courses are offered every semester.

Y64.3015 Affordable Housing Development 3

Y64.3055 Analyzing REIT Securities 3

Y64.3065 Federal Taxation for Real Estate Investment 3

Y64.3075 Managing Building Systems and Operations 3

Y64.3145 Acquisition Procedures and Analysis Using ARGUS SoftwareTM 3

Y64.3155 Advanced Seminar in Real Estate Finance 3

Y64.3165 Analytical Techniques in Real Estate Investment Analysis 3

Y64.3170 Derivatives and Real Estate Investing 3

Y64.3175 Real Estate Investing in a Distressed Environment 3

Y64.3180 Global Real Estate Markets and Investments 3

Y64.3185 Urban Economic Development 3

Y64.3190 Seminar in Post Catastrophe Reconstruction (PCR) 3

Y64.3400 Seminar in Community Development 3

Y64.3405 Advanced Real Estate Development and Investment Practices 3

Y64.3410 Entrepreneurship and Innovation in Real Estate Development 3

Y64.3500 Deconstructing the Appraisal 3

Total number of required credits 42 Credits

Proceedings of the

2013 IEEE 17th International Conference on

Computer Supported Cooperative Work in Design

(CSCWD)

June 27-29, 2013

Whistler, BC, Canada

Edited by:

Weiming Shen, Weidong Li, Jean-Paul Barthès, Junzhou Luo,

Haibin Zhu, Jianming Yong, Xiaoping Li

ISBN: 978-1-4673-6085-2

IEEE Catalog Number: CFP13797-CDR

permitted to photocopy beyond the limit of U.S. copyright law for private use of patrons those articles in

this volume that carry a code at the bottom of the first page, provided the per-copy fee indicated in the

code is paid through Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. For

other copying, reprint or republication permission, write to IEEE Copyrights Manager, IEEE Operations

Technical inquiries should be sent to:

Dr. Weiming Shen

Phone: +1 519 430-7134, Fax: +1 519 204-0145, E-mail: wshen@ieee.org

Proceedings of the

2013 IEEE 17th International Conference on

Computer Supported Cooperative Work in Design

(CSCWD)

June 27-29, 2013, Whistler, BC, Canada

International Steering Committee

Honorary Chair: Prof. Zongkai Lin (China)

Chair: Prof. Jean-Paul Barthès (France)

Co-Chairs: Prof. Junzhou Luo (China)

Prof. Weiming Shen (Canada)

Secretary: Dr. Jianming Yong (Australia)

Members: Prof. Marcos Borges (Brazil)

Prof. Kuo-Ming Chao (UK)

Prof. Jano Moreira de Souza (Brazil)

Prof. Liang Gao (China)

Prof. Ning Gu (China)

Prof. Anne James (UK)

Prof. Peter Kropf (Switzerland)

Prof. Weidong Li (UK)

Prof. Hwa Gyoo Park (Korea)

Prof. José A. Pino (Chile)

Prof. Yun Yang (Australia)

Prof. Qinghua Zheng (China)

Organizing Committees

General Conference Chair: Prof. Weiming Shen, Canada

Conference Co-Chairs: Prof. William Gruver, Canada

Prof. Anne James, UK

Prof. Jano de Souza, Brazil

Prof. Chengzheng Sun, Singapore

Program Committee Chairs: Prof. Weidong Li, UK

Prof. Jean-Paul Barthès, France

Prof. Junzhou Luo, China

Prof. Haibin Zhu, Canada

Dr. Jianming Yong, Australia

Publication Chair: Prof. Xiaoping Li, China

Treasurer: Dr. Yunjiao Xue, Canada

Preface

Welcome to the 2013 IEEE 17th IEEE International Conference on Computer Supported Cooperative Work in Design (CSCWD 2013).

It is my great pleasure to host the CSCWD conference again in Canada, this time in beautiful Whistler, British Columbia, after 12 years since CSCWD 2001 which was held in London, Ontario.

Design of complex artifacts, systems, processes, and services requires the cooperation of multidisciplinary design teams. The 2013 IEEE 17th International Conference on Computer Supported Cooperative Work in Design (CSCWD 2013) provides a forum for researchers and practitioners involved in different but related domains to confront research results and discuss key problems. The scope of CSCWD 2013 includes the research and development fields of collaboration technologies and applications to the design of processes, products, systems, and services in industries and societies. Such technologies include collaboration methods, mechanisms, protocols, software tools, platforms, and services that support collaboration and coordination among people, software and hardware systems. Related fields of research include collaborative workflows, collaborative virtual environments, collaborative workspaces, collaborative wireless sensor networks, interoperability, ontology and semantics, security and privacy, as well as social aspects and human factors related to collaboration. Application domains include aerospace, automotive, manufacturing, construction, logistics, transportation, power and energy, healthcare, infrastructure, administration, social networks, and entertainment.

The CSCWD 2013 Program Committee has developed an exciting technical program including three keynote speeches by internationally recognized researchers in the related areas and 21 technical sessions with 116 paper presentations. The proceedings contain 116 papers co-authored by about 380 researchers from 27 countries.

I would like to thank all the co-chairs and members of the Program Committee for their efforts in putting together such a comprehensive program. Special thanks are due to the authors who contributed to this conference. This conference would not have been possible without their strong supports.

I would also like to thank all the chairs and members of the Organizing Committees as well as all the volunteers for taking care of all the details to make this conference successful.

Weiming Shen, PhD, FIEEE, P.Eng. CSCWD 2013 Conference Chair

Table of Contents

Keynote Papers

Internet of Things: Recent Advances and Applications 1

Mengchu Zhou,

Trends in Information and Communication Technologies for Construction: Past, Present and Future 2

Thomas Froese

Sharing Summarised Semantic Data Rather Than Just Data 3

Kuo-Ming Chao

Collaboration Methods and Techniques

Inference of Differential Equations by M-MEP for Cement Hydration Modeling 4

Bo Yang, Qingke Zhang, Lin Wang and Yi Li

Solving Two Special Dependency Conflicts in Real-time Collaborative Design Systems 11 Liping Gao, Shuxian Guo, Yuben Zhang, Shanshan Wang, Qingkui Chen and Tun Lu

Dynamically Modeling Users with MODUS-SD and Kohonen's Map 17 Lucas Galete, Milton Ramos, Julio Nievola and Emerson Paraiso

Multi-Granularity Partial Encryption Method of CAD Model 23 Xiantao Cai, Fazhi He, Weidong Li, Xiaoxia Li and Yiqi Wu

Cooperative Discrete Particle Swarms for Multi-mode Resource-constrained Projects 31 Hong Shen and Xiaoping Li

Effort-Based Incentives for Resource Sharing in Collaborative Volunteer Applications 37 Davide Vega, Roc Meseguer, Felix Freitag and Sergio Ochoa

Fuzzy Prospective Scenarios in Strategic Planning in Large-Group Decision 43 Sergio Palma J. Medeiros, Jano de Souza, Vanessa Janni Epelbaum and Maria Gilda P. Esteves

Towards Cross-Platform Collaboration - Transferring Real-Time Groupware to the Browser 49 Matthias Wenzel, Lutz Gericke, Raja Gumienny and Christoph Meinel

A Survey of Languages to Represent Collaboration as a Means of Designing CSCW Facilities in RDL 55 Edson Lucas, Daniel Schneider, Toacy Oliveira and Jano de Souza

Mix4Crowds - Toward a Framework to Design Crowd Collaboration with Science 61 Alexandre P. Uchoa, Maria Gilda P. Esteves and Jano M. de Souza

Collaborative Information Gathering and Recommendation Using Mobile Computing 67 Jonice Oliveira, Andressa Silva and Raphael Franckini

Simulation of CTCS-3 Protocol with Temporal Logic Programming 72

Peng Zhang, Zhenhua Duan and Cong Tian

Supporting Information Exchange Among Software Developers Through The Development Of 78

Collaborative Information Retrieval Utilities

Adam Grzywaczewski, Rahat Iqbal, Anne James and John Halloran

Hybrid Profiling in Information Retrieval 84

Mandeep Pannu, Anne James and Rachid Anane

Automatic Enhanced CDFG Generation based on Runtime Instrumentation 92

Zhongda Yuan, Yuchun Ma, Jinian Bian, and Kang Zhao

Efficient Custom Instruction Generation Based on Characterizing of Basic Blocks 98

Guoqiang Liang, Yuchun Ma, Kang Zhao and Jinian Bian

A Collaborative Strategy for Reliability-based Multidisciplinary Design Optimization 104

Jun Zhang and Bing Zhang

Multidisciplinary Design Optimization under Uncertainties based on BLISS and PMA 110

Jun Zhang and Bing Zhang

Research on Collaborative IT Governance Model Oriented to Business Architecture 116

Shaohua Zhang, Shuang Yang and Jundian Song

Multi-user Mass Satellite Image Collaborative Scheduling Scheme Research 121

Lei Liu, Lin Li and Xiaoping Liu

Agents and Multi-Agent Systems

A Distributed Algorithm for Agent Coalition Formation with Complex Tasks 127

Hao Wang, Jian Cao and Xiaogang Wang

A Muti-Agent Membrane Computing Technique for Conceptual Design 133

Xiyu Liu, Jie Xue and Xiaolin Yu

The Cooperative Reinforcement Learning in a Multi-Agent Design System 139

Hong Liu

An Efficient Approach to Group Role Assignment with Conflicting Agents 145

Haibin Zhu and Luming Feng

Negotiation Environment for Enterprise Interoperability Sustainability 153

Tiago Santos, Carlos Coutinho, Ricardo Goncalves and Adina Cretan

Collaborative Workflows

Design Communication Types in a Buyer-Supplier Relationship 159

Venlakaisa Hölttä, Petri Mannonen and Sampo Teräs

SLAs Detective Control Model for Workflow Composition of Cloud Services 165

Yong Sun, Wenan Tan, Ler Li, Guangzhen Lu and Anqiong Tang

A Provenance-based Solution for Software Selection in Scientific Software Sharing 172

Xing Huang, Tun Lu, Xianghua Ding, Tiejiang Liu and Ning Gu

A Novel BOM based Multi-Resolution Model for Federated Simulation 178

Chun Zhang, Huachao Mao, Gongzhuang Peng and Heming Zhang

Modeling Highly Collaborative Processes 184

Pedro Antunes, Valeria Herskovic, Sergio Ochoa and Jose A. Pino

A Notation for Knowledge-Intensive Processes 190

Joanne Manhães Netto, Juliana B. S. França, Flavia A. Santoro, Fernanda Araujo Baiao and Flavia M.

Santoro

Collaboration Platforms, Software Tools, and Services

ESaaS: A New Software Paradigm for Supporting Higher Education in Cloud Environment 196

Md. Anwar Hossain Masud and Xiaodi Huang

A Collaborative Platform for Facilitating Standard Cell Characterization 202

Azam Beg, Amr Elchouemi and Raahim Beg

An Intents-based Approach for Service Discovery and Integration 207

Cheng Zheng, Weiming Shen and Hamada Ghenniwa

A Framework for Developing Distributed Collaborative Applications Using HTML5 213

Nelson Baloian, Francisco Gutierrez and Gustavo Zurita

An Intelligent Network User Behavior Analysis System Based on Collaborative Markov Model and 221

Distributed Data Processing

Tao Wu, Hui He, Xiqian Gu, Ying Peng, Yi Zhang, Yuntao Zhou and Senzhe Xu

Learning Context to Adapt Business Processes 229

Juliana do E. Santo Carvalho, Flavia M. Santoro, Kate Revoredo and Vanessa Tavares Nunes

Supporting Component Presence Notifications in Software Development 235

Ramon R. Palacio, German Padilla, Alberto L. Morán, Joaquin Cortez and Aurora Vizcaíno

Collaboration Tools for Multi-User CAD 241

Tom Nysetvold and Chia-Chi Teng

iii

Development and Application of Optimization Model for Customized Text Summarization 246

Jiang-Liang Hou and Yong-Jhih Chen

Improving the Support to Decision Making in Medium-sized Urban Emergencies 251

David Suarez, Alvaro Monares, Sergio Ochoa, Jose A. Pino and Manuel Ibarra

Ontology and Interoperability

Ontology-based Dental Implant Connection Patent Analysis 257

Amy J.C. Trappey, Charles V. Trappey, Hsin-Yi Peng and Tong-Mei Wang

An Abstract Model for Identifying Potential Teams and Communities 263

Cesar A. Tacla, Eliane de Bortoli, Gustavo Giménez-Lugo, Jean M. Simão, Josiane Dall'Agnol and Emerson Paraiso

Multi-Domain Multi-Lingual Collaborative Design 269

Laurent Wouters, Yuki Kaeri and Kenji Sugawara

Design and Implementation of Intents User Agent 275

Cheng Zheng, Weiming Shen and Hamada Ghenniwa

ThesIS: A Semantic Interoperability Service for a Middleware Service Oriented Architecture 281

Diego Arize, Rita Suzana P. Maciel and José Maria N. David

An Ontology-Based Agent for Context Aware Software Process Development 287

Josivan Pereira de Souza, Cesar A. Tacla, Franciele Beal, Emerson Paraiso and Gustavo A. Gimenez-Lugo

Collaborative computing (Clouds, Grids, and Web Services)

Towards Minimizing Cost for Composite Data-Intensive Services 293

Lijuan Wang, Jun Shen, Changyan Di, Yan Li and Qingguo Zhou

Evaluations of Heuristic Algorithms for Teamwork-Enhanced Task Allocation in Mobile Cloud-Based 299 Learning

Geng Sun, Jun Shen, Junzhou Luo and Jianming Yong

Bidding Specification Language and Winner Determination for Grid Computing Scheduling 305

Raafat Omar Aburukba, Hamada Ghenniwa and Weiming Shen

Scheduling Parallel Task Graphs on Non-dedicated Heterogeneous Multicluster Platform with Moldable 313 Task Duplication

Jinghui Zhang, Junzhou Luo and Fang Dong

Collaborative Support for Knowledge-Intensive Processes through a Service-based Approach 319

Ednilson Veloso Moura, Flávia Maria Santoro and Fernanda Araujo Baião

A cooperative intrusion detection model based on granular computing 325

Wei Zhang, Shaohua Teng, Xiufen Fu, Jihui Fan, Yi Teng, Haibin Zhu

Collaborative Virtual Environments

Achieving Critical Consistency through Progressive Slowdown in Highly Interactive Multi-Player 332

Online Games

Haifeng Shen and Suiping Zhou

A Configurable Visual Steering Architecture based on 3D Scene Composition 338

Han Wang, Hongming Cai and Lihong Jiang

An Environment to Support the Discovery of Potential Partners in a Research Group 344

Tatiana P. V. Alves, Marcos R.S. Borges and Adriana S. Vivacqua

A Role-Playing-Game Approach to Accomplishing Daily Tasks to Improve Health 350

João Paulo Santos Da Silva, Daniel Schneider, Jano de Souza and Marcio Antelio Da Silva

Multi-view is Useful for More Accurate Understanding of Object in a Virtual Soccer Field 357

Kasumi Tarukawa, Tomoo Inoue and Kenichi Okada

Collaboration Technology Applications in Manufacturing

An Adaptive Intelligent Method for Manufacturing Process Optimization In Steelworks 363

Xia Zhu and Xiaoping Li

Design of E-commerce Development Roadmap for SMEs by DNA Computing Technique 369

Xue Bai and Xiyu Liu

Simulation Study on Collaborative Behaviors in Mass Collaborative Product Development 375

Shuo Zhang, Xiaodong Zhang, Yang Hu and Yingzi Li

Evaluating Part Machining Processes for Low-Carbon and Energy-Efficiency Contexts on Web 380

Pingyu Jiang, Chaoyang Zhang, Lei Zhang, Weidong Li and Peihua Gu

Improved Genetic Algorithm with External Archive Maintenance for multi-Objective Integrated Process 385

Planning and Scheduling

Xiaoyu Wen, Xinyu Li, Liang Gao, Wenwen Wang and Liang Wan

Parametric Design of Turbodrill Bearing Section based on VB and Solidworks 391

Sizhu Zhou and Lun Qin

Study on the Collaborative Design System Based on Skeleton Model for High Speed Train Body 395

Xiaozhen Mi, Feng Wang, Libin Fu and Daliang Zheng

A multilevel Modeling Framework for Semantic Representation of Cloud Manufacturing Resources 400

Ning Liu and Xiaoping Li

Parameters Optimization of a Multi-pass Milling Process based on Imperialist Competitive Algorithm 406 Yang Yang, Xinyu Li and Liang Gao

Constructing Federate Collaboration Ontology by UML Profiles 411

Yanbing Liu, Hongbo Sun, Tianyuan Xiao and Wenhui Fan

SIMP based Topology Optimization of a Folding Wing with Mixed Design Variables 417

Xiaohui Wang, Zhiwei Lin and Renwei Xia

A Simplified Electromagnetism-like Mechanism Algorithm for Tool Path Planning in 5-Axis Flank 422

Milling

Qing Wu, Xinyu Li, Liang Gao and Ying Li

Solving Multi-objective Flexible Job Shop Scheduling with Transportation Constraints using a Micro 427

Artificial Bee Colony Algorithm

Zhuangcheng Liu, Shuai Ma, Yanjun Shi and Hongfei Teng

A Semantic Support to Improve the Collaborative Control of Manufacturing Processes in Industries 433

Sara Bouzid, Corine Cauvet, Claudia Frydman and Jacques Pinaton

Cooperative Service Registries for the Service-based Product Lifecycle Management Architecture 439

Stefan Silcher, Jan Königsberger, Peter Reimann and Bernhard Mitschang

Managing Engineering Analysis Knowledge 447

Hongwei Wang, Hao Qin and Heming Zhang

A Reliability Prediction Method of Process Plan for Aircraft Structural Parts based on Fuzzy 454

Comprehensive Evaluation

Wangwei Chu, Yingguang Li, Wenping Mou, Changqing Liu and Limin Tang

Framework for Inter-operative e-Procurement Marketplace 459

Sudeep Ghimire, Ricardo Jardim-Goncalves, Antonio Grilo and Miguel Beca

A Honey-bee Mating Optimization Approach of Collaborative Process Planning and Scheduling for 465

Sustainable Manufacturing

Xiaoxia Li, Weidong Li, Xiantao Cai and Fazhi He

Toward a Modeling of Human-centered, Rule-based Cooperative Teamwork 471

Xingguang Fu, Marcel Langer and Dirk Söffker

Optimized Distribution of Product Model by 3D CAD Streaming in Networked Collaborative Design 477

Chih-Hsing Chu and Yu-Hsuan Chan

Collaboration Technology Applications in Architecture, Engineering, Construction

Mobile Collaboration Technology in Engineering Asset Maintenance: A Delphi Study 483

Faisal Syafar and Jing Gao

Towards A Computer Mediated Methodology for Collaborative Design during the Early Architectural 489

Design Stages

Marianthi Leon and Richard Laing

A Self-Stabilizing Protocol for Minimal Weighted Dominating Sets in Arbitrary Networks 496

Guangyuan Wang, Hua Wang, Xiaohui Tao and Ji Zhang

Facilitating Meaningful Collaboration in Architectural Design through the Adoption of BIM 502

James Harty and Richard Laing

Study of the Efficiency Improvement in Power Generation from Photovoltaic 509

Ching-Lung Lin, Yuan-Chuen Hwang and Hsiu-Chen Lin

Collaboration Technology Applications in Healthcare

Improving Outpatient Service Quality in Department of Orthopedic Surgery by Using Collaborative 515

Approaches

Ta-Ping Lu, Jui-Tien Shih, Cholada Kittipittayakorn and Geng-Feng Lian

Design of Exergames with the Collaborative Participation of Older Adults 521

Amado Velazquez, Alejandro Hernandez and Sergio F. Ochoa

How the Crowd Can Change the Collaborative Work in Patient Care 527

Carla Pereira, Maria Gilda P. Esteves, Sergio Medeiros, Jano Souza and Marcio Antelio

Audit Recommendation for Privacy Protection in Personal Health Record Systems 533

Zhong Han, Yuqing Sun and Yuan Wang

Research on Developing Clinical Collaborative Communication Systems 539

Hwa Gyoo Park

AOPUT: A Recommendation Framework Based on Social Activities and Content Interests 545

Yingying Deng, Tun Lu, Huanhuan Xia, Dongsheng Li, Tiejiang Liu, Xianghua Ding, Ning Gu

Collaboration Technology Applications in Social Networks and Entertainment

Child Search Framework: A collaborative information retrieval architecture to assist children in the 551

search process

Sandra Regina Rocha Silva and Geraldo B. Xexéo

vii

Sentiment Analysis on Tweets for Social Events 557

Xujuan Zhou, Xiaohui Tao and Jianming Yong

V-ROOM: Virtual Meeting System Trial 563

Phil Thompson, Anne James and Antonios Nanos

Social Media Learning: an Approach for Composition of Multimedia Interactive Object in a 570

Collaborative Learning Environment

Iván Claros and Ruth Cobos

Quantum Ant Colony Algorithm-Based Emergency Evacuation Path Choice 576

Feng Zhang, Min Liu, Zhuo Zhou, Weiming Shen

Designing a System to Capitalize Both Social and Documentary Resources 581

Etienne Deparis, Marie-Helene Abel, Gaëlle Lortal and Juliette Mattioli

Predicting Short Interval Tracking Polls with Online Social Media 587

Ho Leung Li, Vincent T. Y. Ng and Simon C. K. Shiu

Cassino Musical: A Game with a Purpose for Social Recruitment and Measurement of Musical Talent 593 Carlos Gomes, Daniel Schneider, Jano de Souza, and Geraldo Xexéo

Activity-Led Learning Approach and Group Performance Analysis Using Fuzzy Rule-Based 599

Classification Model

Rahat Iqbal, Faiyaz Doctor, Margarida Romero and Anne James

Evaluation of E-Performance System: A Cultural Perspective 607

Abdulaziz Al-Raisi, Saad Amin, Rahat Iqbal and Anne James

Special Session: Coordination of Collaborative Supply Chains & Enterprise Networks

Mathematical Model for Assignment Policies and Information Sharing in a Supply Chain 615

Mansour Rached, Armand Baboli, Zied Bahroun and Jean-Pierre Campagne

A GSP Double Auction for Smart Exchange 621

Wafa Ghonaim, Hamada Ghenniwa and Weiming Shen

An Optimizer of Grey-Genetic Algorithms to Improve the Prediction Efficiency for Taiwan Import and 627 Export Pollution

Chen-Fang Tsai

A Bi-objective Model for Collaborative Planning in Dyadic Supply Chain 633

Hamza Ben Abdallah, Naoufel Cheikhrouhou, Zied Bahroun and Mansour Rached

ISO14051-based Material Flow Cost Accounting System Framework for Collaborative Green 639

Manufacturing

Amy J.C. Trappey, Squall Chun-Yi Wu, Mike F.M. Yeh and Andy Y.F. Kuo

viii

Workshop on Internet of Things and Logistics

Collaborative Material and Production Tracking in Toy Manufacturing 645

Yulian Cao, Wenfeng Li, Wei Song and W. Art Chaovalitwongse

CloudThings: a Common Architecture for Integrating the Internet of Things with Cloud Computing 651

Jiehan Zhou, Teemu Leppänen, Erkki Harjula, Mika Ylianttila, Timo Ojala, Chen Yu, Hai Jin and

Laurence Tianruo Yang

A RFID-based Tracing Service of Waste Electrical and Electronic Equipment 658

Jing Sun, Yinsheng Li and Kuo-Ming Chao

Max-Cut Based Overlapping Channel Assignment for 802.11 Multi-Radio Wireless Mesh Networks 662

Wei Wang, Bo Liu, Ming Yang, Junzhou Luo and Xiaojun Shen

Modeling and Evaluation of the Building Management Framework based on the Castalia WSN 668

Simulator

Giancarlo Fortino, Raffaele Greco and Antonio Guerrieri

A Utility-Oriented Routing Algorithm for Community Based Opportunistic Networks 675

Xiuwen Fu, Wenfeng Li and Giancarlo Fortino

Geo-Localized Messages Irradiation using Smartphones: An Energy Consumption Analysis 681

Daniel Moreno, Sergio F. Ochoa, Rodrigo Santos and Roc Meseguer

A Data Processing Framework for IoT based Online Monitoring System 686

Zhuo Zhou, Min Liu, Feng Zhang, Li Bai, Weiming Shen

An Improved Shortest Path Algorithm Based on Orientation Rectangle for Restricted Searching Area 692

Wenyan Zhou, Qizhi Qiu, Peng Luo and Pei Fang

Indoor Localization of Ubiquitous Heterogeneous Devices 698

Emil Jimenez and Ruizhong Wei

Eavesdropping Attack in Collaborative Wireless Networks: Security Protocols and Intercept Behavior 704 Yulong Zou, Xainbin Wang and Weiming Shen

Author Index 710

DISSERTATIONS LOCATION LIST (January 2017)

A guide to the location of dissertations or theses submitted at Oxford Brookes University

MPhil and PhD theses

All MPhil/PhD theses are kept in the University libraries; the majority of these are held at Headington, but those for subjects based at Wheatley (e.g. Business and Computing) or Harcourt Hill (e.g. Education) are kept at those libraries.

They are all indexed on the Library Catalogue and on Discover, with links to the online version (stored on RADAR) when available.

Brookes MPhils/PhD theses are also listed on the British Library EThOS service, together with information about whether available for download (EThOS takes this information from RADAR). ethos.bl.uk

MSc, MA, undergraduate and all other dissertations

SUBJECT DEGREE/DIPLOMA LOCATION AVAILABILITY FOR

CONSULTATION INDEX

Anthropology Undergraduate Gibbs 1.03 Ask in Programme

Administrator’s Office. Debra Bates G1.03 (x3750)

Students can take up to

5 dissertations to G1.02, the collaborative learning space (with student card as deposit).

Applied Languages BA Tonge Not yet had final cohort

Architecture BA/major studies Resources Centre (AB2.15) Yes: Mon/Wed/Fri 10-2pm Local index

Automotive Engineering BEng/MEng Wheatley Library store Yes – see Library staff Library catalogue

Biology/Biotechnology BSc/MSc Department Request via dissertation supervisor or personal tutor

Biomedical Science BSc/MSc Department Request via dissertation supervisor or personal tutor

Business BA/MBA/MA/MSc Wheatley Library Store;

MSc Dissertations from

2014 are now on RADAR Yes – see Library Staff Library catalogue

Cell & Molecular Biology BSc Department Request via dissertation supervisor or personal tutor

Coaching and Mentoring Studies MA Wheatley Library store Yes – ask at Wheatley Library counter

Communication Networks BSc Wheatley Library Store Yes – ask at Wheatley Library counter Library catalogue

Communication, Media and Culture BA Harcourt Hill Library Yes – selected titles on Short Loan – see Library staff

Computer Science BSc/MSc Wheatley Library store Yes – see Library staff Library catalogue

Construction

Management BSc Resources Centre (AB215) For possible online provision contact your course administrator or dissertation supervisor Yes: Mon/Wed/Fri 10-2pm Printed list

Creative Writing MA Tonge Yes – contact Debbie Wedge (Administrator) (x3748) dwedge@brookes.ac.uk

Development and Emergency Practice MA Resources Centre (AB215) and some online – see: http://architecture.brooke s.ac.uk/research/cendep/ dissertations.html Yes: Mon/Wed/Fri 10-2pm Google doc – see http://architecture. brookes.ac.uk/rese arch/cendep/disser tations.html

Early Childhood Studies BA Department Apply to Mrs Gill Cross, Programme Administrators' Office (BG20), Harcourt Hill (gcross@brookes.ac.uk, x8388)

Early Childhood Studies MA Harcourt Hill Library Yes Library catalogue

Economics see Business

Education Studies BA Department Apply to Mrs Gill Cross, Programme Administrator's Office (BG20), Harcourt Hill (gcross@brookes.ac.uk x8388)

Education MA/MEd Harcourt Hill Library Yes – see Library staff Library catalogue

Electrical/Electronic Engineering BEng Wheatley Library store Yes – apply to Library staff Library catalogue

Electronics BSc Department Yes - limited

English Literature BA Tonge

Some on Brookes Virtual -

English Literature -

Moodle

at https://moodle.brookes

Yes – contact Debbie Wedge (x3748) dwedge@brookes.ac.uk

.ac.uk/course/view.php?id

=12319

English Literature MA Tonge Yes - contact Debbie Wedge (x3748) dwedge@brookes.ac.uk

English Language and Communication BA School of Education – Academic staff Yes – some titles available on Moodle on the ELC course - otherwise contact Prog Leader

Environmental

Assessment and

Management MSc Headington Library Yes Library catalogue and printed list

Environmental Biology BSc Department Request via dissertation supervisor or personal tutor

Environmental

Management and

Technology MSc Headington Library Yes Library catalogue and printed list

Environmental Sciences BSc Department Request via dissertation supervisor or personal tutor Card index (by

name)

Equine Science BSc Department Request via dissertation supervisor or personal tutor

Fine Art BA School of Arts Apply to principal lecturers in Fine Art

French BA Tonge Building A few are available – contact Catherine Harris T3.02 (x3748)

Geography BA Gibbs 1.03 Apply to Programme

Administrator – Debra Bates G1.03 (x3750)

Health Care MSc Headington Library Yes Library catalogue and printed list

Health Care Studies BA/Undergraduate Marston Rd (PAC Office) Yes

Historic Conservation MSc Course Director’s office Yes

History Undergraduate Tonge Building Apply to Programme

Administrator – T3.02 (x4220) Mike Wilson

m.wilson@brookes.ac.uk

History/History of

Medicine MA Tonge Building Very small number (winners of the Detlef-Muhlberger Prize) kept in Headington Library; otherwise request via dissertation supervisor or academic advisor

History of Art BA Tonge Building Apply to Programme

Administrator – T3.02 (x4220) Mike Wilson

m.wilson@brookes.ac.uk

Hospitality Undergraduate Ox Schl of Hospitality Management Apply to Fran Buller

(Departmental Office)

Hospitality MSc (60% and over) Headington Library Yes Library catalogue

Human Biology/Exercise and Health BSc Department Request via dissertation supervisor or personal tutor

Humanities

NB: Course no longer

running in this 'format' MA A small selection of exemplary dissertations kept at Headington Library. Yes Library catalogue and printed list

International Business, Culture & Languages MA Tonge A few are available – contact Catherine Harris (x3748) in T3.02

International Relations BA & MA Gibbs 1.03 (9am – 5pm only Ask in Programme

Administrator’s Office, G1.03 (x3750) Debra Bates

Japanese BA Tonge Building Students do not do dissertations in their final year

Law LLB u/g (single & joint); Gibbs 1.03 – 9am – 5pm only Contact Ray Simpson (x4931, rsimpson@brookes.ac.uk)

Law LLM p/g int’l law Gibbs 1.03 – 9am-5pm only Contact Adrienn Nagy (x4901, anagy@brookes.ac.uk)

Management see Business

Mathematics BSc Wheatley Library Store Yes – see library staff Library catalogue

Mechanical Engineering BEng/MEng Wheatley Library Store Yes- see Library staff Library catalogue

Media Technologies BSc/MSc Wheatley Library store Yes – see Library staff Library catalogue

Mobile and Wireless Communications MSc Department of Computing Very limited. Students should apply to their project supervisor for access

Motorsport Engineering BEng/MSc Wheatley Library store Yes – see Library staff Library catalogue

Music BA/BSc/MA A small number are

available in the RHB. Apply to Music staff.

Nursing/Midwifery BA/BSc Marston Rd, PAC Office OR Module Leader (via Programme Administrator). Adult Nursing dissns available on Dissertation Moodle site Yes

Nutrition/Food Science BSc/MSc Department Request via dissertation supervisor/personal tutor

Occupational Therapy BSc Marston Rd, PAC Office Yes, contact programme administrator

Osteopathy BSc Res Lab, Marston Rd Contact supervisor Yes

Philosophy BA Department of History, Philosophy and Religion Archive Room Yes – contact Programme Administrator

Physiotherapy BSc Marston Rd, Skills Lab Contact Dissertation Module leader No

Planning/Spatial Planning BA /MSc Headington Library EXCEPT for BA Planning 2008 onwards are in Resources Centre (AB215) Yes Library catalogue and printed lists

Politics BA Gibbs 1.03 (9am-5pm only) Ask in Programme

Administrator’s Office (x3750) Debra Bates

Primate Conservation MSc Primate Lab Contact Magda Svensson (x4938)

Project Management BE MSc Resources Centre (AB215). For possible online provision contact your course administrator or dissertation supervisor Yes: Mon/Wed/Fri 10-2pm Google doc

Psychology BSc/BA/selection of p/g Department

Public Health MSc Academic Liaison

Librarians’ Office To view, email

healthcarelibrarians@brookes.ac. uk

Library catalogue

Publishing Undergraduate/Diploma/ MA MA distinction diss'ns/major projects in Library & also those on African theme. All others in Resources Centre AB215 Yes

(NB: Resources Centre open Mon/Wed/Fri 10-2pm) Library catalogue and printed lists.

Real Estate Management BSc (NB: MSc students no longer do diss’n) Resources Centre (AB215) For possible online provision contact your course administrator or dissertation supervisor Yes: Mon/Wed/Fri 10-2pm Printed list

Religion and Theology Undergraduate and

Distance Learning Department of History, Philosophy and Religion Archive Room Yes – contact Programme Administrator

Social Work BA Marston Rd, PAC Office Yes, contact Programme Administrator No

Sociology BA Student Collaborative Working Space Gibbs 1.02 Ask in Programme

Administrator’s Office (x3750) Debra Bates

Spanish (minor) Undergraduate None available

Sport & Exercise Science BSc/MSc Department Request via dissertation supervisor or personal tutor

Sport, Coaching and

Physical Education BSc Department of Sports and Health Sciences Yes – contact programme administrator.

Sustainable Building MSc Resources Centre (AB215) Yes: Mon/Wed/Fri 10am-2pm

Technology Management BSc Department Yes - limited

Theology MA Harcourt Hill Library Yes – see Library staff Library catalogue

Tourism Undergraduate Ox Schl of Hospitality Management Apply to Fran Buller

(Departmental Office)

Urban Design MA Resources Centre (AB215) Yes: Mon/Wed/Fri 10-2pm

Organization Design Optimization Using

Genetic Programming

Bijan KHosraviani1, Raymond E. Levitt1, and John R. Koza2

1 Stanford University

Department of Civil and Environmental Engineering

Stanford, California 94305

{bijan, Ray.Levitt}@stanford.edu

2 Stanford University

Stanford Biomedical Informatics Program, Department of Medicine

Stanford, California 94305

koza@stanford.edu

Abstract. This paper describes how we use Genetic Programming (GP) tech-niques to help project managers find near optimal designs for their project or¬ganizations. We use GP as a postprocessor optimizer for the project organiza¬tion design simulator Virtual Design Team (VDT). Decision making policy and individual/sub-team properties, activity assignments and percentage alloca¬tion for each activity are varied by GP, and the effect on quality and duration of the project is compared via a fitness function. The solutions found by GP com¬pare favorably with the best human generated designs.

1 Introduction and Overview

In the complex and rapidly changing business environment of the early 21st Century, designing an effective and optimized organization for a major project is a daunting challenge. Project managers have to rely on their experience and/or trial-and-error to come up with organizational designs that best fit their particular projects. This tradi¬tional method of project organization design is very costly. Based on Tatum’s empiri¬cal research, managers adapt personal experience as the primary process in organiza¬tional structuring. They repeat successes, avoid failures, and make adjustments as required by project situation (Tatum, 1983). The Virtual Design Team (VDT) simula¬tion system, based on the information processing theories of Galbraith (1977) and March and Simon (1958), was a successful attempt to develop an analysis tool for project organization design (Jin & Levitt, 1996). VDT now enables project managers to model and analyze project organizations before implementing them in practice. After extensive ethnographic research in engineering organizations to calibrate its pa¬rameters, VDT can predict the schedule, cost and quality performance for a user-specified organization and work process.

However, like the analysis tools that support many engineering design processes, VDT has no inherent ability to improve or optimize current designs automatically. The user must experiment in a “What if?” mode with different alternatives in an at

tempt to find better solutions that can mitigate the identified risks for a given project configuration. Based on her or his expertise, the user must set up the model, run the simulator, analyze the output, make changes to the input, and repeat these steps until an acceptable output is achieved. VDT relies on the expertise of the human user, and offers no guarantee of optimality. The problem has many degrees of freedom, so the search space for better solutions is vast, and exploring it manually is daunting.

In this paper, we demonstrate how we have designed and used a post-processor for VDT that uses genetic programming, an evolutionary computing method, to generate near optimal project organization designs.

2 Motivation and Points of Departures

Over the past 50 years, optimizers have been successfully developed and deployed for a variety of analysis tools aimed at predicting the behavior of physical systems such as structures, engines, or semiconductors. These optimizers, in conjunction with ma¬ture and extensively validated analysis tools, have enhanced the productivity of engi¬neers by orders of magnitude, and have expanded the range and enhanced the quality of products created in many fields of technology. In contrast, organizational analysis tools that can be used by managers to predict performance outcomes of alternative organizational configurations have only begun to emerge over the past decade. Start¬ing with the Virtual Design Team research in the mid-1990s and the pioneering work of Burton and Obel (2004) in the late 1990s, there are now several agent-based com-puter models and rule-based diagnostic tools that help managers analyze candidate organizational configurations for a given set of task requirements and environmental constraints. However, to the best of our knowledge not much work has been done that can claim to optimize organization designs for real world organizations. One of the attempts we are aware of is a fuzzy multicriteria framework for the comparison of alternative organization structures of post corporations (Kujacic & Bojovic, 2003.)

Some attempts have been made in the past at developing a post-processor for VDT; however, those were of limited power and generality. For example, William Hewlett (2000) designed a rule-based “expert system” post processor for VDT that analyzes the outputs of a VDT simulation, and recommends small, incremental changes in the design of modeled organizations. Hewlett’s post processor was tested in a design charrette on a group of Stanford students; it showed that they were able to create better organizations when they used the post-processor than without it. How¬ever, there were many limitations of this initial post-processor. First, the post-processor did not solve for the optimal organization; it was only a small piece of an optimization strategy. It primarily focused on team sizes and suggested reallocating personnel between teams. Thus, after running the VDT simulator, a user had to take advice suggested by the post-processor, make changes in the original design, run the simulator again, and observe whether the optimization was beneficial. Then the user had to repeat this optimizing loop until the desired output was achieved. As a result, this process was an exhaustive, never ending search. Second, although this process was shown to be beneficial for some students with less project management design experience, it provided less benefit for more experienced managers.

An ongoing research effort by Michael Murray, another PhD student in the De¬partment of Civil and Environmental Engineering at Stanford, is beginning to address a few selected aspects of the organization design optimization problem. The focus of Murray’s research, like Hewlett’s, is on the scheduling and resources of the project organization. This optimization tool combines operations research techniques (linear programming and branch and bound search) with artificial intelligence techniques (constraint propagation (Baptiste et al., 2001) and heuristic search (Cheng & Smith, 1994)). The tool optimizes the macro resource sizing and scheduling to eliminate the most serious backlogs for project participants while respecting project priorities. Murray (2002) also conducted a brief investigation of the application of genetic algo¬rithms to engineering design project scheduling problems.

During the last few years, evolutionary computational methods have been used to optimize various kinds of systems in ways that rival or exceed human capabilities. For example, GP has produced optimization results for a wide variety of problems in¬volving automated synthesis of controllers, circuits, antennas, genetic networks, and metabolic pathways (Koza et al., 2003). Prof. John H. Miller (2001) and his group at Carnegie Mellon University have done similar work, in terms of evolving organiza¬tions, but for simpler structures than our proposed research. In their research, they show that simple adaptive mechanisms allow for the creation of superior organiza¬tional structures. In addition, they conclude that, while they do not have proofs of op¬timal structures, the genetic algorithm was designed to solve difficult, nonlinear prob¬lems, and thus the structures that emerge from the algorithm should contain valuable hints about optimal form.

2.1 Virtual Design Team (VDT)

The Virtual Design Team (VDT) is a project organization modeling and simulation tool that integrates organizational and process views of strategic, time-critical pro¬jects. The vision behind VDT is to provide a method and tool to design an organiza¬tion the way an engineer designs a bridge, that is, by first creating and analyzing a virtual model, and then implementing the organization that has predictable capabili¬ties and known limits.

Using VDT a user can develop a case study of his project or projects and run simu¬lations to predict project outcomes. Simulations also identify organizational risks to development quality, schedule, and cost. Software simulation helps the user to set up, monitor, and troubleshoot a large project or a program of projects successfully. By al¬tering the VDT model components, a modeler can experiment with different solutions to determine which one meets his program quality, cost, and scheduling objectives.

Using VDT’s a graphical interface, project managers design the organizational structure—its size, the number of people in the group, and its topology—who reports to whom. The project manager also graphically assigns one or more activities for each individual within the group, as well as the dependencies between the activities.

The user sets other organization attributes such as skill levels of each actor (indi¬vidual or subteam) and decision making policies. The skill level of each actor can be set to low, medium, or high. The higher the skill level of individuals, the faster the task gets done, and the lower the rate of exceptions generated. Decision making poli¬cies include centralization, formalization and matrix strength. Centralization reflects

whether decisions are made by senior management positions or decentralized to first level supervisor or worker positions. Formalization is the relative degree to which communication among positions takes place through formal meetings and memos vs. informally. Matrix strength models the “degree of collocation” of the various special-ists in an organization by setting the probability workers will attend to communica-tions. The above decision-making policies can also be set to low, medium, or high.

Fig. 1. User Interface of the VDT Simulator - Each project participant fills a position in the project organizational hierarchy and works on one or more activities. The organizational structure and the interdependence between activities define coordination requirements among individuals

Once the above attributes and topologies are set, the Monte Carlo discrete event simulation can be run (usually 50-100 trials is sufficient) and the model produces a set of output as shown in Fig. 2. The user can then manually adjust the input parame-ters to obtain the desired output.

Fig. 2. Sample VDT Outputs - Gant charts, quality risks, and person backlogs are among number of graphical outputs that VDT can produce. Gant Chart displays project, tasks, and milestone in rows with duration-represented bars. Quality Risk shows the task or projects at greatest risk of exception-handling failures. actor backlog shows the backlog for each person in the model, which indicates predicted overload of positions over time

3 Statement of the Problem

Once we implemented the first version of our postprocessor optimizer, we applied it to a case study that has been used for several years in a project management course taught at Stanford. The results produced by our GP were then compared against the best solution discovered by student groups and senior project manager groups over the last 6 years. In this case study, student and project manager groups are given a biotech plant project organization and asked to modify some of the individual/sub-team attributes and organizational policy structure in order to reduce the project schedule duration as much as possible, while maintaining acceptable levels of quality risk.

4 Methodology

The method used in this paper is similar to what has been used in designing an im-proved version of Astrom-Hagglund PID controller (Koza, 2003). Instead of redes¬igning a project organization from scratch, we used an existing design done in the tra¬ditional human generated way and tried to adjust different attributes, so the final outcomes of the project could be improved. In order to do this, the genetic transform¬ing tree produces a solution that in fact is an instruction of what, in the given project organization, needs to be changed and by how much.

4.1 The Representation

The remainder of this section explains the setup for the genetic programming tree and how it was used to produce set of solution to the given problem. The standard genetic programming tableau appears in Table 1.

Table 1. Tableau for the project organization design optimization problem

Objective: Find the changes need to be made to the current project or

ganization in order to reduce the project simulated dura

tion, reduce cost and improve quality of the final outcome

Terminal Set P1, P2, P3, P4, P5, P6, P7, CFM

Function Set Up, Down, Same, FTE, Assign, Aloc

Fitness Cases 15 total – 1 for simulation duration, 1 for FTE, 13 for each activities

Raw Fitness SPD + TFTE * FTEW +  (FRIi * FRIWi + PRIi * PRIWi + CRi * CRWi) (see section 4.4 Fitness Evaluation)

Standardized Fitness Same as raw fitness

Parameters Population size M = 3000

Maximum number of generations, G = 100

Crossover = 90% Mutation = 3% Reproduction = 7%

Success Predict None – search for the shortest simulation duration with the given quality and FTE constraints

4.2 Function and Terminal Sets

Terminal set P1 through P7 represents actors in the group. CFM stands for Centrali¬zation, Formalization and Matrix Strength. Functions Up, Down, Same can have dif¬ferent meaning depending on the Terminals that they connect to and whether there are FTE, Assign or Aloc functions in between. For example, the function FTE increases or decreases the number of FTE for each actor depending on the number of Up/Down functions preceding it in the genetic Tree. The Assign function assigns an activity to an actor, and Aloc specifies percentage allocation for each activity.

A combination of above function and terminal sets transforms the initial project organization suggested by a project manager to a near optimal one. Figure 3 shows a sample of a transforming tree produced by this genetic operation. In this configura¬tion, for example, the skill attributes of P3 (person 3) in the organization structure changes based on the type of its parent and grandparent nodes. So, in this case, P3’s first skill level (e.g. Project Management) remains the same, her second skill level (e.g. Software Engineering) increases, and her third skill level (e.g. Design Coordina¬tion) decreases. In the sample tree below CFM’s parents and grandparents are Same, Up, and Down. So, in this case, the genetic program tree suggests that centralization should remain the same, formalization should increase by one level, and matrix strength should decrease to optimize the overall project outcome.

Fig. 3. Sample of a Transforming Genetic Tree. Program trees created by genetic opera-tions modify the structure and attributes of a project organization. The genetic tree above tran¬sforms an organization design (not shown here) proposed by a project manager to a near opti¬mal one

4.3 Genetic Tree Constraints

Since FTE, Assign , and Aloc functions can only appear next to the bottom of the ge¬netic tree (i.e., Terminal sets should be the only children of them), constraints were added to the ECJ parameter files to enforce such limitations.

4.4 Fitness Evaluation

In each successive generation, evaluation of the fitness function is calculated on three primary inputs. First, the total simulation duration, or number of days to complete the project. Second, the quality risk values including communication risk, functional risk, and project risk. Third, the total FTEs, since there was a constraint for the maximum number allocated to the project. The formula 1 below shows how weighting factors are applied to these inputs to calculate the total fitness value. Note that the weighting factors are designed such that the fitness function heavily penalizes an increase in quality risk or FTEs.

SPD + TFTE * FTEW + = (FRI

i 1 i * FRIWi + PRIi * PRIWi + CRi * CRWi) (1)

Where

• SPD = Simulated Project Duration

• TFTE = the Total FTE added

• FTEW = FTE Weight ( if TFTE > 3.0 => equal 1000 otherwise 1)

• FRIi = Functional Risk Index for activity i

• FRIWi = FRI weight for activity i (if FRIi > 0.5 => equal 1000 otherwise 1)

• PRIi = Project Risk Index for activity i

• PRIWi = PRI weight for activity i (if PRIi > 0.5 => equal 1000 otherwise 1)

• CRi = Communication Risk for activity i

• CRWi = CR weight for activity i (if CRi > 0.5 => equal 1000 otherwise 1)

• M = maximum number of activities

4.5 Software/Hardware Used

The runs reported in this paper were designed based on the standard genetic pro-gramming paradigm as defined in Koza (1992). The problem was coded in Java using the ECJ 10, the Evolutionary Computation and Genetic Programming System by Sean Luke. The runs were executed on a PC with a Pentium 4 - 2.8GHz processor.

5 Results

We divided our case study experiment into two phases. In Phase I, we defined a sim¬plified GP. In this process, we varied only the levels of the actors’ skills. Then, we compared the results found by the GP with the known optimal solution. In Phase II, we kept skill levels constant, and varied the number of Full Time Equivalent FTEs (i.e., human resources) added to different positions. We also varied organizational policy attributes such as the levels of centralization, formalization and matrix strength and the assignment of activities to actors using GP. We then compared the GP results against the best solution found by previous student and manager groups. In the next two sections, we discuss the findings of this experiment.

5.1 Varying Actors’ Skill Levels

There are seven positions (actors) in this project organization, and each one of these positions has two to eight different skills. The skills range from biotechnology to de¬sign coordination to mechanical/electrical, etc. There are a total of 29 skills for all seven positions. Each one of these skills can be set to three levels of low, medium, and high. Therefore, the total number of combinations that one could try to find an optimal solution exhaustively is 329 = 6.8 * 1013. Thus, the sample space is vast and an exhaustive search is infeasible.

It should be obvious that the more skilled the actors, the faster the tasks get done, and the fewer the exceptions (i.e., when an actor requires additional information or a decision to complete part of a task, or the actor generates an error that may need cor¬recting.). In this case where we are not concerned about cost, the optimal solution would be when the skill levels of all actors are set all to high. Knowing the above fact, in one scenario we set skill levels of all actors to “high”, ran the VDT simulation and compared the results with the base results where we had the skill levels of all ac¬tors set to “medium”. At the base level, we found that the simulated schedule end was March 28, 2001, and when we set all skill levels to “high”, the project duration was reduced by 69 days and the simulation showed that the project schedule end would be Jan 17, 2001. Then, we ran the simulation again using the suggested solution by GP and we found identical results as when all skill levels were set at high.

5.2 Varying Actors’ FTEs and Organization’s Policy

In Phase II, we allowed the GP to vary the assignment of activities to actors, percent¬age allocation for each activity, the Full Time Equivalent’s (FTE) of each actor in 0.5 FTE increments, and organizational policy properties such as levels of centralization, formalization and matrix strength using GP.

The best individual found by GP in generation 21, and it is shown below in a lisp-type format:

(Up (Down (Same (Same P5 P4) (Down (Down P1 P5) (Up (FTE P0) (Up (Down (Up (FTE P0) (Down P5 P5)) (Up (FTE P1) (Up (FTE P0) (Same P3 P6)))) (FTE P5))))) (Up (Same (Same (Down (Up (Up (Assign P0) (FTE P1)) (Same (Up (Same (Down (FTE P4) (FTE P0)) (Down (FTE P2) (Up (Up P6 (Up (Up P0 (FTE P1)) (FTE P4))) (FTE P1)))) (Up (FTE P4) (Assign P4))) (Up (Up (Up (FTE P5) (FTE P5)) (FTE P4)) (Up (FTE P0) (Up (Assign P0) (Same P5 P4)))))) (Up (FTE P5) (Aloc P0))) P2) (FTE P0)) (Same (Same (Down (Up (Up (Assign P0) (Same P5 P4)) (Same (Up (Same (Up (Assign P0) (Up (Assign

P1) (Assign P0))) (Aloc P1)) (Up (FTE P4) (Assign P4))) (Up (Up (Up (FTE P5) (FTE P5)) (FTE P4)) (Up (FTE P0) (Up (Assign P0) (Same P5 P4)))))) (Up (FTE P5) (Aloc P0)))

P2) (FTE P0)))) (FTE P4))

Then, we compared the results with the best results obtained by more than 40 teams of students and managers over the past six years.

Fig. 4. Comparison of Gant Charts before (Left) and after (Right) Evolutionary Process. GP reduced end date from Feb 20, 2001 to Dec 5, 2000. This GP solution is better than the best solution (Dec 7) found by >40 student and manager teams for this problem over the last 6 years

The best individual found by GP in generation 21 beats the best human-discovered solution by 2 days. The best human solution reduced project completion from Feb 20, 2001 to Dec 7, 2000; the GP-suggested solution reduced the project end date to Dec 5, 2000. This is shown in Figure 4 above. In addition the quality risks such as com¬munication risks were improved as shown in figure 5 below.

Fig. 5. Comparison of Quality Risks Before (Left) and After (Right) Evolutionary Process. Originally 7 out of 14 activities had quality risks higher than acceptable 0.5 thresholds (orange bars). With the suggested organizational changes, quality risks for all activities improved

6 Discussion of Results

We compared the solution produced by our genetic programming methodology both against the theoretical optimal solution and the best human generated solution. In both cases, the results were promising.

In the first case, as mentioned above, the outcome results found by GP were iden¬tical with the optimal case. However, interestingly, the suggested solution found by

GP was not identical to the optimal solution. (i.e., there were multiple solutions that yielded identical optimal outcome.) Unlike the optimal case scenario, GP did not have to set all skill levels to “high”. In fact, there were situations where the levels of some actors’ skills were reduced from “medium” to “low”, and still the outcome matched the optimal solution. For example, the “General” skill of the “Structural De¬sign Sub-team” was reduced to “low”, and the “Mechanical” skill of the “Construc¬tion PM” was kept at “medium”. This showed that GP could generate different solu¬tions to a problem, so that a project manager can better decide which solutions to pick. For example, different solutions suggested by GP could advise that we could reduce skill levels of certain actors and increase skill level of others. So the project manager has the choice between different alternatives to pick a solution that better fits her/his specific project and available resources.

In the second case, where GP was allowed to vary the number of FTEs added to actors, the assignment of activities to actors, percentage allocation for each activity, and the policy attributes of the organization project, the solution found by the evolu¬tionary process surpass the best Student/Manager (S/M) solution ever found. The number of FTEs and where they were added in GP solution matched exactly the best S/M solution. The S/M solution had suggested for swapping two activities between two actors. GP solution suggested an additional reassignment of an activity and change in percentage allocation of a couple of activities, in addition to the swaps sug¬gested by the S/M solution. Also the S/M solution had suggested increasing both formalization and matrix strength by one level, whereas the GP solution suggested adding a level only to matrix strength.

As shown in the results section, GP has been able to improve the final outcome in both cases and meet the given constraints. The greatest improvement is seen in the project schedule, where the simulated duration was reduced by 77 days, followed by substantial improvement in communication, functional and project quality risks.

We also considered looking at the trend of improvements through different genera¬tions. Figure 6 below shows that the greatest improvement of the best individual fit¬ness value was made between generations 1 and 6 (solid pink line). This figure also displays the mean fitness improvements during different generations (yellow dotted line). As shown, although the fitness value of best individual is not improved from generation 21, the mean fitness value has improved. This means that GP has produced more individuals within a generation that match or are near the solution found by the best individual. This can be an opportunity for a project manager to select a solution that better matches her/his specific project needs.

Several trials with different population sizes and different mutation and crossover rates were made, but these changes did not affect the final outcome significantly. Also, in one case, the number of FTE children was changed to two instead one and that also did not affect the results greatly.

Although the transforming genetic tree in its current form was shown to produce improved results, it might not be the most efficient. The tight dependencies of func¬tion set FTE, Assign, and Aloc on their preceding nodes can cause some deficiency during the crossover operations, where only part of a branch of one individual is swapped with another. Another factor is that when the recurrence of actors occurred, the very last actor to the right of the tree was used for the skill levels and FTEs. Thus, after many generations, the right side of the genetic tree was more active than the left side.

Fig. 6. Improvement of Fitness Value Generations 1 Through 30 – Although the best fit-ness value (solid line) does not improve after generation 21 (best solution found), the mean fit¬ness value is improving. This means there are more alternative solutions that a manager can pick among the best solutions

7 Conclusions

This research has made successful first steps towards the optimization of project or¬ganization designs. Instead of redesigning the project organization from scratch, a human generated design was used as a baseline. Several input attributes such as deci¬sion making policies, individual / sub-team properties, activity assignments, and ac¬tors’ attention allocation—were adjusted using genetic programming to evolve the project organization design against a fitness function representing the goals for the project. The effects of the evolutionary process on simulated project duration and quality risks were noticeable. We compared the results produced by the GP with those generated by humans. Our GP post processor for VDT beats the best human trial-and-error performance of > 40 teams for this realistic problem.

8 Future Research

This work is only the beginning use of genetic programming in optimizing organiza¬tional designs. Much time on this project was spent on writing the code to translate the transforming genetic tree and connecting the results to the VDT simulator. Now that the preliminary work is done, the next step is to add different organizational at¬tributes to the genetic operation and see the effect on the final outcome. The topol¬ogy, communication, and other individual and team properties—such as who reports

to whom, team experience, and application experience—are some of the other pa-rameters that could be added to this model. Eventually, the evolutionary post-processor should be integrated within the VDT model.

9 Acknowledgments

We would like to thank the developer of ECJ (A Java-based Evolutionary Computa¬tion and Genetic Programming Research System) Sean Luke, for providing the easy-to-use programming environment that was used for this project. We would also like to thank Mark Ramsey, for helping us to connect the genetic tree to the VDT model.

References

1. Baptiste P., Le Pape, C., Nuijten, W.: Constraint-Based Scheduling: Applying Con-straint Programming to Scheduling Problems, Kluwer Academic Publishers, Boston, MA. (2001)

2. Burton, R., Obel, B.: Strategic Organizational Diagnosis and Design: Developing Theory for Application, second edition, Academic Publishers, Boston, MA (2004)

3. Cheng-Chung C., Stephen F. S.: Generating feasible schedules under complex metric constraints. In Proceedings of the Twelfth National Conference on Artificial Intelli¬gence (AAAI-94), volume 2, pages 1086--1091, Seattle, Washington, USA, AAAI Press/MIT Press (August 1994)

4. Galbraith J.R.: Organizational Design, Reading, MA: Addison-Wesley (1977)

5. Hewlett W.R.: Design and Experimental Results of a Post-Processor of VitéProject, B.S. Honors Thesis, Symbolic Systems Program, Stanford University (2000)

6. Jin Y., Levitt RE. The Virtual Design Team: A computational Model of Project Or-ganizations. Computational and Mathematical Organization Theory; 2(3):171-196 (1996)

7. Koza J.R., Keane M.A, Streeter M.J., Mydlowec W., Yu J., Lanza G.: Genetic Pro-gramming IV: Routine Human-Competitive Machine Intelligence, Kluwer Academic Publishers (2003)

8. Koza, John R.: Genetic Programming: On the Programming of Computers by Means of Natural Selection. Cambridge, MA: The MIT Press (1992)

9. Kujacic, M., Bojovic N.: Organizational Design of Post Corporation Structure Using Fuzzy Multicriteria Decision Making, Computational & Mathematical Organization Theory, 9, 5-18, Kluwer Academic Publishers, Manufactured in the Netherlands (2003)

10. March, J.G., Simon H.A.: Organizations. New York: John Wiley and Sons (1958)

11. Miller J. H.: Evolving Information Processing Organizations – In Dynamics of Or-ganizations Computational Modeling and Organization Theories: edited by Alessan¬dro Lomi and Erick R. Larsen, MIT Press / AAAI Press, Ch-10 pp 307-327 (2001)

12. Murray, M.: On the Application of Genetic Algorithms to Scheduling Engineering Design Projects. In Koza, John R. (editor). Genetic Algorithms and Genetic Pro-gramming at Stanford 2002. Stanford, CA: Stanford University Bookstore (2002)

13. Tatum C.B.: Decision Making in Structuring Construction Project Organizations. Ph.D. Dissertation Thesis (1983)

Incorporating Side Information into Recurrent Neural Network Language

Models

Cong Duy Vu Hoang

University of Melbourne

Melbourne, VIC, Australia

vhoang2@student.unimelb.edu.au

Gholamreza Haffari

Monash University

Clayton, VIC, Australia

gholamreza.haffari@monash.edu

Abstract

Recurrent neural network language models (RNNLM) have recently demonstrated vast potential in modelling long-term dependen¬cies for NLP problems, ranging from speech recognition to machine translation. In this work, we propose methods for conditioning RNNLMs on external side information, e.g., metadata such as keywords or document title. Our experiments show consistent improve¬ments of RNNLMs using side information over the baselines for two different datasets and genres in two languages. Interestingly, we found that side information in a foreign lan¬guage can be highly beneficial in modelling texts in another language, serving as a form of cross-lingual language modelling.

1 Introduction

Neural network approaches to language modelling (LM) have made remarkable performance gains over traditional count-based gram LMs (Bengio et al., 2003; Mnih and Hinton, 2007; Mikolov et al., 2011). They offer several desirable characteristics, includ¬ing the capacity to generalise over large vocabular¬ies through the use of vector space representation, and – for recurrent models (Mikolov et al., 2011) – the ability to encode long distance dependencies that are impossible to include with a limited context windows used in conventional gram LMs. These early papers have spawned a cottage industry in neu¬ral LM based applications, where text generation is a key component, including conditional language models for image captioning (Kiros et al., 2014; Vinyals et al., 2015) and neural machine translation

Trevor Cohn

University of Melbourne

Melbourne, VIC, Australia

t.cohn@unimelb.edu.au

(Kalchbrenner and Blunsom, 2013; Sutskever et al., 2014; Bahdanau et al., 2015).

Inspired by these works for conditioning LMs on complex side information, such as images and for-eign text, in this paper we investigate the possibility of improving LMs in a more traditional setting, that is when applied directly to text documents. Typi-cally corpora include rich side information, such as document titles, authorship, time stamp, keywords and so on, although this information is usually dis-carded when applying statistical models. However, this information can be highly informative, for in-stance, keywords, titles or descriptions, often in-clude central topics which will be helpful in mod-elling or understanding the document text. We pro-pose mechanisms for encoding this side informa¬tion into a vector space representation, and means of incorporating it into the generating process in a RNNLM framework. Evaluating on two corpora and two different languages, we show consistently sig¬nificant perplexity reductions over the state-of-the-art RNNLM models.

The contributions of this paper are as follows:

1. We propose a framework for encoding struc-tured and unstructured side information, and its incorporation into a RNNLM.

2. We introduce a new corpus, the RIE corpus, based on the Europarl web archive, with rich annotations of several types of meta-data.

3. We provide empirical analysis showing consis-tent improvements from using side information across two datasets in two languages.

2 Problem Formulation & Model

We first review RNNLM architecture (Mikolov et al., 2011) before describing our extension in §2.2.

2.1 RNNLM Architecture

The standard RNNLM consists of 3 main layers: an input layer where each input word has its embedding via one-hot vector coding; a hidden layer consisting of recurrent units where a state is conditioned recur-sively on past states; and an output layer where a target word will be predicted. RNNLM has an ad-vantage over conventional n-gram language model in modelling long distance dependencies effectively.

In general, an RNN operates from left-to-right over the input word sequence; i.e.,

ht = RU (xt ht1) = f W (hh)ht1 + W (ih)xt + b(h)

xt+1  softmax W (ho)ht + b(o) ;

where  () is a non-linear function, e.g., tanh, ap-plied element-wise to its vector input; ht is the cur-rent RNN hidden state at time-step ; and matrices W and vectors b are model parameters. The model is trained using gradient-based methods to optimise a (regularised) training objective, e.g. the likelihood function. In principle, a recurrent unit (RU) can be employed using different variants of recurrent struc-tures such as: Long Short Term Memory (LSTM) (Hochreiter and Schmidhuber, 1997), Gated Recur-rent Unit (GRU) (Cho et al., 2014), or recently deeper structures, e.g. Depth Gated Long Short Term Memory (DGLSTM) – a stack of LSTMs with extra connections between memory cells in deep layers (Yao et al., 2015). It can be regarded as being a generalisation of LSTM recurrence to both time and depth. Such deep recurrent structure may capture long distance patterns at their most general. Empirically, we found that RNNLM with DGLSTM structure appears to be best performer across our datasets, and therefore is used predominantly in our experiments.

2.2 Incorporating Side Information

Nowadays, many corpora are archived with side in-formation or contextual meta-data. In this work, we

Figure 1: Integration methods for auxiliary information, e: a) as input to the RNN, or b) as part of the output softmax layer.

argue that such information can be useful for lan-guage modelling (and presumably other NLP tasks). By providing this auxiliary information directly to the RNNLM, we stand to boost language modelling performance.

The first question in using side information is how to encode these unstructured inputs, y, into a vector representation, denoted e. We discuss several meth-ods for encoding the auxiliary vector:

BOW additive bag of words, e = t yt, and

average the average embedding vector,



t yt, both inspired by (Hermann



and Blunsom, 2014a);

bigram convolution with sum-pooling,

e = t tanh (yt1 + yt) (Hermann and Blunsom, 2014b); and

RNN a recurrent neural network over the word se-quence (Sutskever et al., 2014), using the final hidden state(s) as e.

From the above methods, we found that BOW worked consistently well, outperforming the other approaches, and moreover lead to a simpler model with faster training. For this reason we report only results for the BOW encoding. Note that when using multiple auxiliary inputs, we use a weighted combi-nation, e = i W (ai)e(i).

The next step is the integration of e into the RNNLM. We consider two integration methods: as input to the hidden state (denoted input), and con-nected to the output softmax layer (output), as shown in Figure 1 a and b, respectively. In both cases, we compare experimentally the following in-tegration strategies:

add adding the vectors together, e.g., using xt + e as the input to the RNN, such that

Method test (en) test (fr)

5-gram LM 55.7 38.5

LSTM 40.3 28.5

DGLSTM 36.4 25.4

output+mlp+h 33.3 24.0

Table 4: Perplexity scores based on the sampled RIE dataset. +h: topic headline.

line) of a textual utterance. We provided an empir¬ical analysis of various ways of injecting such in¬formation into a distributed representation, which is then incorporated into either the input, hidden, or output layer of RNNLM architecture. Our ex¬perimental results reveal consistent improvements are achieved over strong baselines for different datasets and genres in two languages. Our future work will investigate the model performance on a closely-related task, i.e., neural machine translation (Sutskever et al., 2014; Bahdanau et al., 2015). Fur¬thermore, we will explore learning methods to com¬bine utterances with and without the auxiliary side information.

Acknowledgements

Cong Duy Vu Hoang was supported by full schol¬arships of the University of Melbourne, Australia. Dr Trevor Cohn was supported by the ARC (Future Fellowship).

References

D. Bahdanau, K. Cho, and Y. Bengio. 2015. Neural Machine Translation by Jointly Learning to Align and Translate. In Proceedings of International Conference on Learning Representations (ICLR 2015), September.

Yoshua Bengio, R´ejean Ducharme, Pascal Vincent, and Christian Janvin. 2003. A Neural Probabilistic Lan¬guage Model. The Journal of Machine Learning Re¬search, 3:1137–1155.

Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bah-danau, and Yoshua Bengio. 2014. On the Proper¬ties of Neural Machine Translation: Encoder–Decoder Approaches. In Proceedings of SSST-8, Eighth Work¬shop on Syntax, Semantics and Structure in Statisti¬cal Translation, pages 103–111, Doha, Qatar, October. Association for Computational Linguistics.

Kenneth Heafield. 2011. KenLM: Faster and Smaller Language Model Queries. In Proceedings of the EMNLP 2011 Sixth Workshop on Statistical Machine

Translation, pages 187–197, Edinburgh, Scotland, United Kingdom, July.

K. M. Hermann and P. Blunsom. 2014a. Multilingual Distributed Representations without Word Alignment. In Proceedings of International Conference on Learn¬ing Representations (ICLR 2014), December.

Karl Moritz Hermann and Phil Blunsom. 2014b. Multi¬lingual Models for Compositional Distributed Seman¬tics. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Vol¬ume 1: Long Papers), pages 58–68, Baltimore, Mary¬land, June. Association for Computational Linguistics.

Geoffrey E Hinton. 2002. Training Products of Experts by Minimizing Contrastive Divergence. Neural com¬putation, 14(8):1771–1800.

Sepp Hochreiter and Jurgen Schmidhuber. 1997. Long Short-Term Memory. Neural Comput., 9(8):1735– 1780, November.

Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent Continuous Translation Models. In Proceedings of Empirical Methods in Natural Language Processing (EMNLP 2013).

Ryan Kiros, Ruslan Salakhutdinov, and Rich Zemel. 2014. Multimodal Neural Language Models. In Pro¬ceedings of the 31st International Conference on Ma¬chine Learning (ICML-14), pages 595–603.

T. Mikolov, S. Kombrink, A. Deoras, and J. H. Burget, L.and Cernocky. 2011. RNNLM - Recurrent Neural Network Language Modeling Toolkit. In 2011 IEEE Workshop on Automatic Speech Recognition & Under¬standing (ASRU). IEEE Automatic Speech Recogni¬tion and Understanding Workshop, December.

Andriy Mnih and Geoffrey Hinton. 2007. Three New Graphical Models for Statistical Language Modelling. In Proceedings of the 24th International Conference on Machine Learning, pages 641–648.

R. Pascanu, C. Gulcehre, K. Cho, and Y. Bengio. 2013. How to Construct Deep Recurrent Neural Networks. ArXiv e-prints, December.

Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Se¬quence to Sequence Learning with Neural Networks. In Advances in Neural Information Processing Sys¬tems (NIPS 2014), pages 3104–3112.

Oriol Vinyals, Alexander Toshev, Samy Bengio, and Du-mitru Erhan. 2015. Show and Tell: A Neural Image Caption Generator. In The IEEE Conference on Com¬puter Vision and Pattern Recognition (CVPR), June.

Frank Wilcoxon. 1945. Individual Comparisons by Ranking Methods. Biometrics Bulletin, 1 (6):80–83, Dec.

K. Yao, T. Cohn, K. Vylomova, K. Duh, and C. Dyer. 2015. Depth-Gated LSTM. ArXiv e-prints, August.

Planning and uilding nit

eb: tt :// .education.ie

Guidance on Procuring Consultants for

Small Works

1st Edition, Mar 2012 (Revision 1 Aug 2012)

Rev 1 Page 10 Clause 6.2 - in to DoES ebsite for Conditions of Engagement amended

endi E Para 1 in to DoES ebsite for Conditions of Engagement amended

1. INTRODUCTION 2

1.1. PURPOSE OF DOCUMENT 2

1.2. DEFINITIONS 2

1.3. SCOPE OF PROFESSIONAL ADVICE 2

1.4. APPROPRIATE CONSULTANTS FOR TYPE OF WORK 2

2. TECHNICAL REPORTS 3

2.1. APPROPRIATE CONSULTANTS FOR TECHNICAL REPORTS 3

2.2. PROCUREMENT OF CONSULTANTS FOR TECHNICAL REPORTS 3

3. CONSULTANTS FOR BUILDING PROJECTS 4

3.1. GENERAL 4

3.2. ADVERTISING REQUIREMENTS. 5

4. SEEKING TENDER SUBMISSIONS: 5

4.1. ASSESSING THE TENDER SUBMISSIONS: 5

4.2. VALIDATING THE PREFERRED TENDERER’S SUITABILITY 6

5. PROJECT SUPERVISOR (DESIGN) PROCESS 7

5.1. LEGAL OBLIGATIONS 7

5.2. PROCUREMENT OF PSDP FOR A BUILDING PROJECT 8

5.3. EVIDENCE OF COMPETENCY 8

5.4. ALTERNATIVE PROCUREMENT PROCESS FOR PSDP 9

6. APPOINTMENT OF CONSULTANT(S) 9

6.1. LETTERS OF REGRET 9

6.2. CONDITIONS OF ENGAGEMENT 9

6.3. SCHEDULE A AND B 10

6.4. CONTRACT SIGNINGS 10

6.5. MAINTAINING RECORDS 10

APPENDIX A - FORM OF TENDER FOR SMALL WORKS 11

APPENDIX B - TENDER PROPOSAL FORM 13

APPENDIX C - LETTER OF INTENT 14

APPENDIX D - DESIGN APPOINTEE - HEALTH AND SAFETY DECLARATION 16

APPENDIX E - SCHEDULES A AND B CONDITIONS OF ENGAGEMENT 18

uidance on Procuring Consultants for Small or s 1st Edition Marc 2012

1. Introduction

1.1. Purpose of Document

is document sets out t e rocedures to be follo ed for t e Engagement of Consultant(s) for Small

Works including Summer or s and Emergency or s ro ects and s ould be read in con unction

it t e dministrative Procedures governing t e articular Sc eme.

or details of t e terms and conditions of a articular sc eme refer to t e documentation issued it t e relevant sc eme.

1.2. Definitions

or t e ur ose of t is document t e term Sc ool ut ority means t e oard of Management or ot er body legally entrusted it t e management of t e Sc ool.

Design eam refers to all t e rofessional advisors engaged by t e Sc ool ut ority for t e ro ect 1.3. Scope of Professional advice

e engagement of a ro riate rofessional advice is essential for bot t e preparation of the initial technical reports, and all other technical advice for construction projects. These are 2 separate appointments.

en engaging rofessional advice for t e re aration of a re ort, t e Sc ool ut ority must a oint t e relevant consultant for t at tas only and must not enter into any commitments regarding an overall

a ointment or fees for or s ot er t an t e re aration of t e re ort and must ma e t is clear to t e Consultant before t e a ointment is made.

For all construction projects (including refurbishment & repairs) it is a legal requirement to appoint a competent person to act as Project Supervisor (Design Process [PSDP].

or most ro ects it an estimated construction value of less t at €500,000 incl. V a full design team is not normally re uired. nless t e ro ect is articularly com le and difficult, t e a ointment of a single Consultant it relevant advice from ot er disci lines may be a ro riate.

et er for t e re aration of a re ort or for rofessional services including design and contract su ervision, t e a ro riate consultants to be a ointed and t e range of a ro riate rofessional advice ill vary from ro ect to ro ect.

Refer to ec nical Re orts and Engaging Consultants for a Pro ect.

o ensure t at t e Sc ool ut ority obtains value for money and ee s control of costs at all stages t roug t e ro ect it is recommended t at t e Consultants sco e of or s includes buying in t e advice

of a Registered Quantity Surveyor, o ill rovide inde endent building cost control services.

1.4. Appropriate Consultants for type of work

• dditional ccommodation 200m2 bot ermanent and re-fabricated , or ro ects it a substantial element of Design. Pro ects it a substantial element of design include ro ects ere t ere are im lications for t e future develo ment of t e sc ool e.g. a ma or Pro ect is in rc itectural Planning or is envisaged it in t e medium term or ro ects it a conservation element.

o e consultant must be a Registered rc itect not an Engineer, C artered rc itectural ec nologist or ec nician, uilding Surveyor or Quantity Surveyor

• Straig t for ard additional ccommodation 200m2 bot ermanent and re-fabricated

o e consultant may be a Registered rc itect, Civil/Structural Engineer, C artered

rc itectural ec nologist, or uilding Surveyor but not a Mec anical Electrical Engineer,

Quantity Surveyor, or ec nician

• Re airs/remedial or s related to t e fabric of t e building e.g. ma or roof re airs, indo s, toilet

De artment of Education and S ills, Planning and uilding nit

Pg 2 of 22

uidance on Procuring Consultants for Small or s 1st Edition Marc 2012

refurbis ment/u -grade, H S or s etc

o or or s greater t an €500,000 or com le ro ects, t e consultant must be a Registered rc itect

o or or s it out articular com le ity less t an €500,000, t e consultant may be a Registered rc itect, Civil/Structural Engineer, C artered rc itectural ec nologist, or uilding Surveyor but not a Mec anical Electrical Engineer, Quantity Surveyor, or ec nician

• Site or s, drainage, structural integrity of t e building

o If t e ma or element of t e or com rises site or s or drainage e.g. car- ar ing, traffic management, effluent treatment etc , or to t e structural integrity of t e building e.g. structural defects, concrete s alling etc t e consultant must be a Civil/Structural Engineer not an rc itect, M E Engineer, C artered rc itectural ec nologist or ec nician, uilding Surveyor or Quantity Surveyor

• Mec anical or Electrical or s

o If t e re ort refers to Mec anical or Electrical or s e.g. oiler, Heating installation, Electrical Installation, ire alarm etc a uilding Services Engineering Mec anical

Electrical Consultancy Practice it a C artered Engineer at director/ management level, ualified in t e s ecific disci line note general mec anical engineering ualifications are not

a ro riate for electrical ro ects to com lete final uality assurance c ec s and sign off,

must be directly a ointed. o ot er consultant is acce table for t is or and no subcontracting of t is or is ermitted.

e consultant must ave a ro riate rofessional Qualifications, Professional Indemnity Cover and revious e erience in consultancy service of a similar nature. See also Validating t e referred tenderer s suitability.

2. TECHNICAL REPORTS

2.1. Appropriate Consultants for Technical Reports

e Sc ool ut ority s ould refer to t e documentation issued it t e relevant De artment Circular for advice on t e sco e of t e tec nical re ort re uired.

When engaging professional advice for the preparation of a report, the School Authority must select an appropriate consultant (an individual consultant or a consultancy firm) for the type of work being recommended. Refer to Appropriate Consultants for type of work above.

e Sc ool ut ority must a oint t e relevant consultant for t at tas only.

e Sc ool ut ority must not enter into any commitments regarding an overall a ointment or fees for

or s ot er t an t e re aration of t e re ort and must ma e t is clear to t e Consultant before t e

a ointment is made.

en a Sc ool ut ority engages rofessional advice to assist in t e re aration of t e initial tec nical re ort, t e cost must be met in full out of t e sc ool s o n resources.

e a ointment of a Pro ect Su ervisor Design Process is not normally re uired for t e re aration of

a re ort.

2.2. Procurement of Consultants for Technical Reports

e ee for t e ec nical Re ort must be a lum -sum fee e V .lum -sum fee means a fi ed rice in Euros and must include t e cost of buying in a ro riate rofessional sub- consultants.

e Sc ool ut ority is re uired to:

• Consider any Consultant being ualified in t e a ro riate disci line o as e ressed an

interest in tendering for t e ro ect e.g. by riting or emailing t e sc ool, and ere t at Consultant meets t e minimum standard re uired, t at Consultant must be included on t e list of

firms from om uotes are to be obtained. ere is no u er limit on t e number of Consultants ermitted to submit uotes.

• See a minimum of at least 5 quotes in riting or by email from suitable consultants so t at a minimum of 3 ritten uotations ill be received

De artment of Education and S ills, Planning and uilding nit

Pg 3 of 22

uidance on Procuring Consultants for Small or s 1st Edition Marc 2012

o enable t e Sc ool ut ority to assess ic firm to engage it is recommended t at t e follo ing information be re uested from t e Consultants under consideration for t e or :

• Evidence of e erience in ro ects of a similar nature eit er in t e re aration of re orts or carrying out t e or , including e erience in estimating costs.

• V inclusive all-in lum -sum fee including buying-in ot er services as re uired, lus all

e enses

• list of t e relevant ot er services ere re uired, including Quantity Surveying Services ic

ill be boug t in, and confirmation t at t ose service roviders ill be ualified rofessionals in t e relevant disci line See belo

• brief summary of t e sco e of or covered by t e above lum -sum fee and a timescale for t e delivery oft ere ort

If a consultancy firm is under consideration for a ointment t e Sc ool ut ority s ould also see confirmation of a ro riate Professional Qualifications, Professional Indemnity Insurance and Em loyer s iability Insurance. See also Validating t e referred tenderer s suitability

If fe er t an 5 firms submit tenders, t e Contracting ut ority s ould only roceed it an a ard if it considers t at t ere as been genuine com etition and t at t e tender being considered for acce tance re resents value-for-money.

If t ere is t e ossibility t at t ere could be interest from anot er urisdiction e.g. ort ern Ireland it is

im ortant to ensure t at t e rinci les of rans arency and on-Discrimination under t e reaty of Rome are observed.

or a tec nical re ort as above a formal contract is not re uired. sim le letter of a ointment

summarising t e agreed sco e of or s ould suffice. e level of fee in Euros, not s ould be confirmed in t e letter of a ointment and must relate to t e re aration of t at re ort only.

3. Consultants for Building Projects

3.1. General

en funding as been a roved and a Sc ool ut ority engages rofessional advice to carry out a

building ro ect t e cost of suc rofessional services including V all e enses must be met in full out oft e a roved grant.

Sc ool ut orities are not allo ed to a oint a consultant for t e ro ect ot er t an for t e initial ec nical Re ort until funding for t e ro ect as been a roved.

e a ointment of an a ro riate consultant for t e ro ect i.e. t e design and contract su ervision

stage is a se arate a ointment rocess to t e engagement of a consultant for t e ec nical Re ort.

Sc ool ut orities are not allo ed to e tend t e a ointment of t e consultant for t e ec nical Re ort

to include t e design and contract su ervision stages. se arate tender rocess must ta e lace.

For all construction projects it is a legal requirement to appoint a competent person to act as Project Supervisor (Design) Process. This is a separate appointment with a separate fee that should reflect the nature and complexity of the project. [see Project Supervisor (Design) Process]

or all ro ects t e engagement of a ro riate rofessional advice is essential. or most small or s

ro ects a full design team is not re uired. nless t e ro ect is articularly com le and difficult, t e

a ointment of a single consultant may be a ro riate as long as relevant advice from ot er disci lines

e.g. Quantity Surveyor, Structural Engineer, M E Engineer, rc itect or ot er relevant disci line is

included as art oft e consultants service.

en engaging rofessional advice for a building ro ect including tem orary accommodation or re airs , t e Sc ool ut ority must select an a ro riate consultant an individual consultant or a

consultancy firm for t e ty e of or being recommended. Refer to ro riate Consultants for ty e of or above.

e ee for t e consultancy a ointment must be a lum -sum fee e V . lum -sum fee means a

fi ed rice in Euros and must include t e cost of buying in t e relevant advice of ot er disci lines including Quantity Surveying Services .

De artment of Education and S ills, Planning and uilding nit

Pg 4 of 22

uidance on Procuring Consultants for Small or s 1st Edition Marc 2012

3.2. Advertising Requirements.

e follo ing re uirements a ly:

• en t e aggregated estimated value total fees of all consultancy a ointments for a ro ect is

greater than €50,000 including V t e tender o ortunity must be advertised on

.e enders.gov.ie using t e O en Procedure and using t e tender documentation available on

t e De artment s ebsite at .education.ie under Planning/ uilding nit ec nical uidance.

• en t e aggregated estimated value total fees of all consultancy a ointments for a ro ect is below €50,000 (including VAT) t e consultancy a ointments do not ave to be advertised on .e enders.gov.ie.

• If t ere is t e ossibility of cross-border interest you must lace an advertisement in .e enders.gov.ie even if t e aggregated estimated value total fees is belo €50,000

including V .art from t at t e rocedure for a ointing consultants is t e same as belo

4. Seeking Tender Submissions:

e Sc ool ut ority is re uired to:

• Consider any Consultant being ualified in t e a ro riate disci line o as e ressed an

interest in tendering for t e ro ect e.g. by riting or emailing t e sc ool, and ere t at

Consultant meets or declares t at t ey meet t e minimum standard re uired, t at Consultant must be included on t e list of firms from om tender submissions are to be obtained.

• ere is no u er limit on t e number of Consultants ermitted to submit tender submissions.

• See a minimum of at least 5 tender submissions in riting or by email from suitable consultants so t at a minimum of 3 tender submissions in riting or by email ill be received

including V .art from t at t e rocedure for a ointing consultants is t e same as belo

o see tenders t e Sc ool ut ority s ould send a covering email/letter and t e follo ing documents to

all t e consultants o ave re lied to t e e enders advertisement or ave contacted t e sc ools

is ing to be included on t e ender ist:

• co y of t e ec nical Re ort and t e a roval letter from t e De artment,

• co y of t e orm of ender at endi , and

• co y oft e ender Pro osal orm, endi

e covering email/letter s ould include any relevant information suc as t e si e, nature and sco e of

t e ro ect and t e time and date for return of Quotes/ ender Submissions allo at least 2 ee s .

It s ould also say t at tender submissions must be in t e format described in t e ender Pro osal orm, e tender submissions returned by t e consultants s ould com rise:

• e form of tender and declaration at endi signed by a member of t e firm

• e ender Pro osal format endi signed by a member of t e firm including:

o e Pro ect Service i.e. a summary of t e service to be rovided on t is ro ect including a

list of t e relevant ot er services re uired including Quantity Surveying Services ic ill

be boug t in, and confirmation t at t ose service roviders ill be ualified rofessionals in t e relevant disci line See belo

o brief commentary on y t e firm is com etent to carry out t e or s and t e s ills t ey ill bring to t e ro ect and t e e erience t ey ave of delivering ro ects of a similar si e and nature

o Ho t e firm ill deliver t e ro ect including o t e ro ect ill be com leted on time and it in budget, and in t e case of additional accommodation o it ill be carried out it out adversely affecting future e ansion

o ot er information is necessary or s ould be considered

4.1. Assessing the Tender Submissions:

en t e tender time and date as e ired, o en all submissions.

ere are 4 a ard criteria i Pro ect Service, 2 Com etency of firm, 3 Pro ect Delivery and 4 Price.

De artment of Education and S ills, Planning and uilding nit

Pg 5 of 22

uidance on Procuring Consultants for Small or s 1st Edition Marc 2012

4.2. Validating the preferred tenderer’s suitability

o you need to c ec t e ualifications, e erience, Healt and Safety Com etence, insurances, ta clearance of t e referred tenderer.

ou do t is by riting a etter of Intent to t e firm in uestion as ing for t e relevant evidence see

tem late at endi C .

en t e information is returned c ec to see t at it is com lete and t at t e firm does meet t e necessary standard you must see t e original current C2/ a clearance certificate - a copy won’t do .

e re uired standard for t e ualifications of t e erson ro osed for t e ro ect is as follo s:

• ll consultants must ave a relevant degree/di loma ualification

• rc itects must be Registered rc itects under t e uilding Control ct 2007. e Panel

establis ed by t e Minister for t e Environment and ocal overnment in 1996 for t e ur ose of

registration is no longer a licable. All Architects must be registered.

• uilding Surveyors or Quantity Surveyors must be Registered uilding Surveyors or Registered Quantity Surveyors under t e uilding Control ct 2007 and

• Engineers must ave an a ro riate degree ualification and be eligible for members i of t e IEI/ CEI.

• ec nologists must be members oft e C artered Institute of rc itectural ec nologists.

ec nician o is not a full member not associated or affiliated or eligible for full members i of

any recogni ed Institute RI I, SCSI, CEI, EI, CI or e uivalent is not eligible.

• E uivalent ualifications from anot er E member state and members i of an e uivalent E rofessional body ill also be deemed acce table in com liance it E Directive 2005//36/EC.

De artment of Education and S ills, Planning and uilding nit

Pg 6 of 22

uidance on Procuring Consultants for Small or s 1st Edition Marc 2012

o e minimum e uivalent ualifications re uirement for rc itects is rc itects olding ualifications under E Directive 85/384/EEC or E uivalent . E uivalent ere means a erson olding a ualification e uivalent to a ualification for t e ur oses of t e Directive.

o o com ly it E Directive 85/384/EEC, rc itects must ave education and training leading to di lomas, certificates and ot er evidence of formal ualifications t roug courses of studies at university level concerned rinci ally it arc itecture and balanced bet een

t e t eoretical and ractical as ects. e total lengt of education and training s all be a minimum of eit er four years of full time studies at a university or com arable educational establis ment, or at least si years of study at a university or com arable educational establis ment of ic at least t ree must be full time suc education and training s all be concluded by successful com letion of an e amination of degree standard. See also:

tt ://euro a.eu.int/

If t e firm does not rovide t e necessary evidence or does not meet t e re uired standard e.g. if t e erson ro osed for t e ro ect as not suitably ualified you may e clude t at firm and roceed to t e ne t ig est scoring tenderer.

en you ave confirmed t at t e referred bidder meets t e necessary re uirements as above , you

are no ready to a oint t at firm. Refer to ointment of Consultants for o to com lete t e rocess

and re are a contract Conditions of Engagement and Sc edules for signature by bot arties.

astly, if fe er t an five firms submit tenders in riting or by email, t e Sc ool ut ority s ould only

roceed it an a ard if it considers t at t ere as been genuine com etition and t at t e tender being considered for acce tance re resents value-for-money. If t ere is t e ossibility t at t ere could be interest from anot er urisdiction e.g. ort ern Ireland it is im ortant to ensure t at t e rinci les under t e reaty of Rome of rans arency and on-Discrimination are observed.

5. PROJECT SUPERVISOR (DESIGN) PROCESS

5.1. Legal Obligations

For all construction projects (including refurbishment & repairs) it is a legal requirement to appoint a competent person to act as Project Supervisor (Design) Process [PSDP].

is is a se arate a ointment it a se arate fee t at s ould reflect t e nature and com le ity of t e ro ect.

s art of, or in addition to, t e general duties of a Pro ect Su ervisor Design Process under t e Healt Safety and elfare at or Construction Regulations 2006 to 2010 and subse uent regulations, t e sco e of t e or s s ould include t e follo ing:

• n assessment of all relevant safety issues

• Pre aration of a Preliminary Healt Safety Plan

• ssisting in t e assessment of t e com etence of t e Contractors to act as Pro ect Su ervisor Construction Stage and

• Pre aration and assembly of t e Safety file based on information su lied by ot ers t e Consultant and t e Contractor

It is not a re uirement t at t e consultant engaged to carry out t e rc itectural or Engineering services

s ould also be a ointed as PSDP. nless t e Sc ool ut ority is satisfied t at t e firm under

consideration is com etent to act as PSDP, it s ould not a oint t at firm to t e role of PSDP. Only

t ose firms o ave demonstrated com etence s ould be considered for a ointment.

It is a requirement of the Health Safety and Welfare at Work (Construction) Regulations that the Client, (in this case the School Authority) satisfies itself that the individual engaged to act as PSDP is competent to carry out that role.

e a ointment of a Pro ect Su ervisor Design Process is not normally re uired for t e re aration of

an initial ec nical Re ort.

De artment of Education and S ills, Planning and uilding nit

Pg 7 of 22

uidance on Procuring Consultants for Small or s 1st Edition Marc 2012 5.2. Procurement of PSDP for a building project

irst roceed it t e tender com etition for t e main Consultant. en you are satisfied t at t e referred bidder meets t e minimum re uirements See Validating t e referred tenderer s suitability , but rior to t at a ointment refer to t e form of tender for t e referred bidder:

5.3. Evidence of Competency

e t you must verify t e referred bidder s com etence. is is a legal re uirement see above and is very im ortant.

If the School Authority appoint a PSDP without first verifying the PSDP’s competence, the School Authority is guilty of a criminal offence.

o be suitable to act as PSDP on a ro ect an individual must bot be a are of is/ er res onsibilities

under t e Safety, Healt and elfare at or ct 2005 and t e Safety Healt elfare at or Construction Regulations 2006 or subse uent legislation/regulations and ave t e resources and com etency to carry t em out.

• Com etency can be demonstrated by rofessional ualifications and e erience. ile H S training is im ortant, t e more im ortant re uirement is rofessional ualifications and e erience in t e relevant rofessional disci line

• H S ualifications it out e erience in t e construction rocess are not sufficient.

• or ro ects it out s ecial H S considerations, t e signed declaration at endi D lus rofessional ualifications in a relevant disci line and e erience of t e erson ro osed for t e role of PSDP may be sufficient evidence of com etency. Evidence of aving fulfilled t e role of PSDP in a satisfactory manner on revious ro ects ould rovide additional reassurance.

• If t e ro ect as s ecial H S considerations evidence of e ternally accredited H S training for t e role of PSDP may also be re uired. Ot er evidence t at may be re uested includes:

o a co y of t eir current general ealt and safety olicy

o an outline of t e firm s management organisational structure it regard to allocation of duties, delegation of res onsibilities etc., in relation to Healt and Safety

o Co ies of standard forms used for ris assessments as art of t eir duties under t e Safety, Healt and elfare at or ct 2005

o rrangements for continuing rofessional develo ment

• Details of t e firm s rocedures for disseminating information and u -to-date develo ments on ealt and safety issues

nder Validating t e referred tenderer s suitability you ill be riting a letter of intent to t e Consultant em late . is etter contains a section on t e role of PSDP and describes t e evidence of

com etency re uired. en t e evidence is returned, ma e sure t at:

De artment of Education and S ills, Planning and uilding nit

Pg 8 of 22

uidance on Procuring Consultants for Small or s 1st Edition Marc 2012

• e H S Declaration includes t e role of PSDP or t at a se arate signed PSDP declaration is included

• ere enforcement actions, legal roceedings, accidents, fatalities or incidents ave been recorded, evidence must be rovided t at ade uate measures ave been ut in lace by t e

a licant firm to address any deficiencies in t eir H S rocedures.

• e erson to act as PSDP as a ro riate rofessional ualifications and e erience bot in is/ er rofessional disci line and in fulfilling t e role of PSDP.

en you ave confirmed t at t e referred bidder meets t e necessary re uirements as above , you

are no ready to a oint t at firm bot as consultant and as PSDP. Refer to ointment of Consultants

for o to com lete t e rocess and re are a contract Conditions of Engagement and Sc edules for signature by bot arties.

5.4. Alternative procurement process for PSDP

e Sc ool ut ority at all times may c oose to engage a se arate PSDP not being t e main Consultant for t e ro ect . If t e Sc ool is es to engage a se arate PSDP t ey s ould ma e it clear to consultants tendering t at t e role of PSDP is not art of t e tender com etition.

E ually, if t e referred bidder to fulfil t e role of Consultant for t e ro ect does not is to be

considered for t e role of PSDP, or is not considered by you to meet t e standard re uired, or if a reasonable rice for t e service cannot be negotiated it t e referred bidder, t en you may c oose to rocure a PSDP by ot er means.

• irst get t e el of your Consultant et er actually a ointed or still a referred bidder in finding suitable firms to tender for t e role. Suc firms may include s ecialist H S firms or t ose o ere unsuccessful tenderers for t e role of main consultant.

• e t rite or e-mail t e selected list giving details of t e ro ect and as for a rice for t e role of

PSDP and a declaration of com etency for t e role. s it ot er consultants you need to send

t em enoug information to describe t e sco e of or a brief cover note and a co y of t e

ec nical Re ort and a roval letter from t e De artment s ould suffice.

• Most firms ill rovide bac -u evidence of com etency it t eir tender submission but t is is not essential at t is stage.

• Select t e firm it t e best rice for t e ro ect c ec t at t ere are no onerous terms and conditions and if no evidence of com etency as included it t e uote rite to t at firm re uesting suc evidence See Evidence of Com etency above

en you are satisfied as to t e com etency of t e referred firm you are no ready to a oint t at firm

as PSDP. Refer to ointment of Consultants for o to com lete t e rocess and re are a contract Conditions of Engagement and Sc edules for signature by bot arties.

6. APPOINTMENT OF CONSULTANT(S)

6.1. Letters of Regret

en t e a arently successful tender as been selected follo ing recei t of tenders, and before t e Sc ool ut ority ma e an a ointment you must send letters of regret to t e unsuccessful candidates informing t em t at t ey ave not been successful.

ou must allo a reasonable eriod of time say 16 calendar days from t e date of issue of t e regret

letters before you confirm t e successful consultants a ointments. is eriod is obligatory to allo time for t e unsuccessful candidates to uery t e reasons for t eir lac of success

6.2. Conditions of Engagement

Once t e letters of regret ave issued and a reasonable eriod of time as ela sed t e Sc ool ut ority are ready to a oint t e Consultant and/or PSDP. o do t is, it is necessary to e ecute a formal agreement it t at consultant called t e Standard Conditions of Engagement for Consultancy Services

ec nical Services ic is available on t e eb at .education.ie under School

Planning/Building Technical guidance Appointment of Consultants.

e Standard Conditions of Engagement for Consultancy Services ec nical Services sets out t e general terms and conditions of contract.

De artment of Education and S ills, Planning and uilding nit

Pg 9 of 22

uidance on Procuring Consultants for Small or s 1st Edition Marc 2012

6.3. Schedule A and B

Sc edule and to t e Conditions of Engagement ic are attac ed to t e Contract set out t e

ro ect s ecific re uirements. e Sc ool ut ority must com lete Sc edule and before issuing t e

Conditions of Engagement to t e Consultant for signature.

o el t e Sc ool ut ority fill in Sc edule and a guide to com leting t e Sc edule is attac ed at endi E. e Sc ool ut ority s ould use it to el t em com lete t e actual form

In general any items not ig lig ted are not ca able of c ange in t e actual electronic form. Items in green, are guidance to you on at to do but do not form art of t e sc edule. Items in ello are t e moving arts oft e Sc edule bot Parts and . In most instances t ey s ould not be c anged unless t e green guidance recommends it.

e Sc edule bot and s ould al ays be com leted by t e Contracting ut ority and not t e Consultant.

e Sc edule must reflect and include t e sco e of or for ic t e consultant submitted is fee and

any s ecial arrangements for e am le any additional services t at t e consultant is including for t e lum -sum fee made at t at time.

6.4. Contract Signings

Once t e Sc edule as been re ared and attac ed to t e Conditions of Engagement, t ey s ould be sent to t e Consultant for signature toget er it a letter of acce tance. Once t e consultant as signed t e contract, e/s e can start or on t e ro ect

6.5. Maintaining records

e im ortance of maintaining an efficient system for ee ing records cannot be overstated. t any time

t roug out t e ro ect, t e De artment and/or in t e case of an accident t e Healt Safety ut ority may is to audit t e records of t e oard of Management to ensure t at ro er rocedures are being ad ered to and t at vital information is being retained.

e oard of Management s ould set u t eir o n filing system, one t at recognises t e different areas of a building ro ect e.g. Design eam a ointments, inancial, submissions from t e consultant, general corres ondence, minutes of meetings, etc .

De artment of Education and S ills, Planning and uilding nit

Pg 10 of 22

Appendix A Form of Tender for Small Works

Appendix A - Form of Tender for Small Works

Form of Tender for Small Works for Construction Consultancy Services

Pro ect

using t e

Standard Conditions of Engagement for Consultancy Services ec nical

D aoine aisle

e ave e amined and understand t e Conditions of Engagement for Consultancy Services ec nical,

t e com leted Sc edules and ere a ro riate , and any ot er tender information su lied, all as amended by any su lemental information, for t e above contract. e offer to rovide and com lete t e Services re uired on t e terms of and in conformity it t e documents referred to in t e receding

aragra using fully designed tender documents develo ed in full com liance it t e De artment of Education and S ills tec nical guidance documents for t e fi ed rice lum sum fee of:

Insert amount in Euro e cl V ords or numerals acce ted

as ad usted in accordance it t e contract.

e amounts set out in t is orm of ender includes t e full sco e of services re uired to com lete t e

ro ect in a rofessional manner including buying-in all relevant ot er consultancy disci lines , e enses, and any intellectual ro erty rig ts re uired. Costs associated it fulfilling t e role of PSDP are not included in t e above tender sum. e confirm t at no furt er fee ad ustments ill a ly it t e e ce tion of substantive Client c anges.

e is to be considered for t e role of PSDP and declare t at our firm is com etent to carry out t at role. es/ o

In consideration of your roviding us it t e contract documents, e agree not to it dra t is offer until t e later of:

60 days after t e end of t e last day for submission of t is ender e iry of at least 21 days ritten notice to terminate t is ender given by us.

our acce tance of t is ender it in t at time ill result in t e Contract being formed bet een us.

e agree t at you are not bound to acce t t e lo est or any tender you may receive. e furt er agree t at if any contract formed by acce tance of t is ender is determined to be void, voidable, unenforceable, or ineffective, any damages for ic you may be liable ill not e ceed t e amount t at

ould ave been ayable under Clause 14.29 of t e Conditions of Engagement on termination under clause 14.9 of t e Conditions.

e declare t at e meet t e minimum re uirements for a ard stated belo and confirm t at e ill if

re uested by t e Sc ool ut ority in riting or by email rovide t e re uired evidence of com liance it in 7 days. e furt er ac no ledge and acce t t at in t e event t at our firm cannot or does not

De artment of Education and S ills, Planning and uilding nit

Pg 11 of 22

Appendix A Form of Tender for Small Works

rovide all of t e re uired evidence to t e satisfaction of t e Sc ool ut ority, it in t e time eriod re uired, t e firm s ender ill be deemed to be invalid and ill be e cluded from furt er consideration.

Minimum re uirements for a ard:

• e or ill be carried out by a member of our staff to be nominated rior to a ard being fully

ualified in t e relevant disci line for t e categories of or stated in t e ender Documents and DoES uidance on Procuring Consultants for Small or s 1st Edition

• 3 e am les of or by t e firm of a similar nature si e and com le ity it in t e last 7 years

• Current Professional Indemnity Insurance minimum €1.0m and e cess not to e ceed 1.5 , Public liability Insurance €6.5m and Em loyer s iability Insurance €13.0m

• Current C2/ a clearance certificate

• Signed Healt and Safety Declaration of Com etency as Designer as at DoES uidance on

Procuring Consultants for Small or s 1st Edition, endi C

Minimum re uirements for a ard for PSDP for ro ects it out s ecial H S considerations:

• e or ill be carried out by a member of our staff to be nominated rior to a ard being fully

ualified and e erienced in t e rovision of rofessional services for construction ro ects o as fulfilled t e role of PSDP in a satisfactory manner on revious construction ro ects

• Signed Healt and Safety Declaration of Com etency as PSDP as at DoES uidance on

Procuring Consultants for Small or s 1st Edition, endi C

dditional re uirements for a ard for PSDP at t e discretion of t e Sc ool ut ority for ro ects, ere

in t e vie of t e Sc ool ut ority s ecial H S considerations do a ly:

• Evidence of e ternally accredited H S training for t e role of PSDP

• a co y of t eir current general ealt and safety olicy

• an outline of t e firm s management organisational structure it regard to allocation of duties, delegation of res onsibilities etc., in relation to Healt and Safety

• Co ies of standard forms used for ris assessments as art of t eir duties under t e Safety, Healt and elfare at or ct 2005

• rrangements for continuing rofessional develo ment

• Details of t e firm s rocedures for disseminating information and u -to-date develo ments on ealt and safety issues

Is sinne, le meas

ame of enderer:

Signature of aut orised erson : Princi al or Director

ame of aut orised erson and

osition in firm:

Date:

De artment of Education and S ills, Planning and uilding nit

Pg 12 of 22

Appendix B Tender Proposal Form

Appendix B - Tender Proposal Form

Tender Proposal Form for Consultancy Services for small projects (Maximum 2 Pages)

Project Title:

Consultancy Discipline:

Name & address of Tendering firm:

Telephone:

Authorised Representative:

Telephone:

e minimum standard for a ass mar on eac of t e follo ing criteria is 40 . enderers o fail to get a ass mar on any one criterion ill be e cluded

Criterion 1 – Summary of the service to be provided [max 25 marks]

is s ould be a summary of t e service to be rovided on t is ro ect including a list of t e relevant ot er

services re uired including Quantity Surveying Services ic ill be boug t in, and confirmation t at t ose

service roviders ill be ualified rofessionals in t e relevant disci line

Criterion 2 – Competency of firm to carry out work [max 25 marks]

is s ould be a brief commentary on y t e firm is com etent to carry out t e or s and t e s ills t ey ill

bring tot e ro ect

Criterion 3 – Project Delivery [max 20 marks]

Ho t e firm ill deliver t e ro ect including o t e ro ect ill be com leted on time and it in budget, and

in t e case of additional accommodation o it ill be carried out it out adversely affecting future e ansion

Signed on behalf of

ame of enderer

Signature of aut orised erson

ame of aut orised erson and osition in firm

Date:

De artment of Education and S ills, Planning and uilding nit

Pg 13 of 22

Appendix C Letter of Intent

Appendix C - Letter of Intent

Date

o ame and address of tenderer

Re: Construction Consultancy Services for itle of Pro ect e.g. Re lacement Roof to loc , sc ool

Sub ect to Contract/Contract Denied

D aoine aisle

I refer to your tender dated for t e above contract in t e amount of € e V .

I rite to inform you t at e intend to issue a etter of cce tance to you sub ect to t e recei t of all t e

follo ing items it in t e re uired time eriod of 7 days. Please be advised t at if you cannot or do not

rovide all of t e re uired evidence to t e satisfaction of t e Sc ool ut ority, it in t e above time eriod your tender ill be deemed to be invalid and ill be e cluded from furt er consideration.

• e name of t e ualified erson to underta e t e or

• Details of t e ualifications of t e nominated erson as above sufficient to demonstrate

com liance it DoES uidance on Procuring Consultants for Small or s 1st Edition and any

re uirements stated in t e ender Documents

• 3 e am les of or by t e firm of a similar nature si e and com le ity it in t e last 7 years

• Evidence of Current Professional Indemnity Insurance minimum €1.0m and e cess not to e ceed 1.5 , Public liability Insurance €6.5m and Em loyer s iability Insurance €13.0m

• Current C2/ a clearance certificate, and

• a Signed Healt and Safety Declaration of Com etency as Designer as at DoES uidance on

Procuring Consultants for Small or s 1st Edition, endi D

If t e tenderer as indicated e/s e is es to be considered for t e role of PSDP t e follo ing additional re uirements s ould be included Delete t is note

• e name of t e ualified erson to underta e t e role of PSDP may be t e same as above

• Details of t e ualifications and e erience of t e nominated erson sufficient to demonstrate

com etency to fulfil t e role of PSDP. is s ould include details of construction ro ects ere

t e nominated erson fulfilled t e role of PSDP in a satisfactory manner

• Signed Healt and Safety Declaration of Com etency as PSDP as at DoES uidance on Procuring Consultants for Small or s 1st Edition

• Evidence t at t e firm s Professional Indemnity Insurance is endorsed to include PSDP services

If t e tenderer as indicated e/s e is es to be considered for t e role of PSDP and, in t e vie of t e Sc ool ut ority t e ro ect as s ecial H S considerations e.g. sbestos , some or all of t e follo ing may be included Delete t is note

• Evidence of e ternally accredited H S training for t e erson nominated to fulfil role of PSDP

• a co y of t e firm s current general ealt and safety olicy

• an outline of t e firm s management organisational structure it regard to allocation of duties, delegation of res onsibilities etc., in relation to Healt and Safety

De artment of Education and S ills, Planning and uilding nit

Pg 14 of 22

Appendix C Letter of Intent

• Co ies of standard forms used for ris assessments as art of t eir duties under t e Safety, Healt and elfare at or ct 2005

• rrangements for continuing rofessional develo ment

• Details of t e firm s rocedures for disseminating information and u -to-date develo ments on ealt and safety issues

If any of t e above listed items is not rovided it in 7 days of t e date of t is letter, e may roceed to a ard t e contract to anot er tenderer.

ard of t e contract ill also be conditional on oard of Management a roval and t e a roval of funding for t is ro ect by t e De artment of Education and Science.

is is not t e etter of cce tance. e Em loyer as not acce ted your tender. Please return a co y of t is letter ac no ledging recei t as indicated belo .

Is mise, le meas

Signed: On be alf of t e Em loyer

e ac no ledge recei t of t is letter on date

Signed: On be alf of t e tenderer

De artment of Education and S ills, Planning and uilding nit

Pg 15 of 22

Appendix D Design Appointee - Health and Safety Declaration

Appendix D - Design Appointee - Health and Safety Declaration

Re: Pro ect ame

e name of a licant firm

ro osing to act as Designer/PSDP on t e above Pro ect ereby declare t e follo ing: e t e above stated firm are members of, or eligible to be a member of t e

In t e case of Design Services t e RI I/ CEI/SCS/IEI/ Ot er . In t e case of PSDP as a s ecial s ill t e

ISO/OSH/RI I/ CEI/SCS/IEI/ Ot er .

being t e relevant rofessional institution for t e above stated Design/PSDP Consultancy service.

ame is res onsible for Healt Safety Management it in t e ractice.

e confirm t at eac member of staff is a are of is/ er res onsibilities under t e Safety, Healt and elfare

at or ct 2005 and t e Safety Healt elfare at or Construction Regulations 2006.

DESI

In articular as designers e are a are of and ill ta e into account t e general rinci les of revention as

enumerated belo en carrying out design or associated it t e ro ect and underta e to liaise it ,

communicate and coo erate it t e PSDP in is role.

GENERAL PRINCIPLES OF PREVENTION APPLICABLE TO DESIGNER AND PSDP

The purpose of the General Principles of Prevention is to provide a framework within which design and detailing issues can be assessed.

GENERAL PRINCIPLES OF PREVENTION

1. e avoidance of ris s

2. e evaluation of unavoidable ris s

3. e combating of ris s at source

4. e ada tation of or to t e individual, es ecially as regards t e design of laces of or , t e c oice of

or e ui ment and systems of or , it a vie to alleviating monotonous or and or at a

redetermined rate and to reduce t eir effect on ealt

5. e ada tation of t e or lace to tec nical rogress

6. e re lacement of dangerous articles, substances or systems of or by non dangerous articles,

substances or systems of or .

7. e giving to collective rotective measures of riority over individual rotective measures.

8. e develo ment of an ade uate revention olicy in relation to safety, ealt and elfare at or , ic

ta es account of tec nology, organisation of or , or ing conditions, social factors and t e influence of

factors related to t e or ing environment.

9. e giving of a ro riate training and instruction to em loyees.

DESI ER

e are a are as designers of our obligations under Section 17 2 of t e Safety Healt elfare at or ct

2005 to ensure so far as is reasonably racticable t at t e ro ect

a is designed and is ca able of being constructed to be safe and it out ris to ealt ,

b can be maintained safely and it out ris to ealt during use, and

c com lies in all res ects, as a ro riate, it t e relevant statutory rovisions

e confirm t at all staff ave received, read and ill a ly t e Safety, Healt and elfare at or eneral

De artment of Education and S ills, Planning and uilding nit

Pg 16 of 22

Appendix D Design Appointee - Health and Safety Declaration

De artment of Education and S ills, Planning and uilding nit

Pg 17 of 22

Appendix E Schedules A and B Conditions of Engagement

Appendix E - Schedules A and B Conditions of Engagement

GUIDANCE ON COMPLETING THE SCHEDULES TO THE STANDARD CONDITIONS OF ENGAGEMENT FOR CONSULTANCY SERVICES (TECHNICAL)

en t e Sc ool ut ority is ready to a oint a Consultant See Section 5 , it is necessary to e ecute a formal agreement it t at consultant called t e Standard Conditions of Engagement for Consultancy Services

available on t e eb at .education.ie under School Planning/Building Technical guidance

Appointment of Consultants or .construction rocurement.gov.ie.

e follo ing guidance s ould be used to el t e Sc ool ut ority com lete Sc edules and to t e rear of t e Conditions. The Schedules should always be completed by the School Authority and not the Consultant. e Sc edules s ould reflect and include t e sco e of or for ic t e consultant submitted is fee and any s ecial arrangements for e am le any additional services t at t e consultant is including for t e lum -sum fee made at t at time. In general any items not ig lig ted are not ca able of c ange in t e actual electronic form.

Items in green are guidance to you on at to do but do not form art of t e sc edule. Items in ello are t e

moving arts oft e Sc edule bot Parts and . In most instances t ey s ould not be c anged unless t e green guidance recommends it.

SCHEDULE A: CONTRACT PARTICULARS

1. APPOINTMENT

1 Client, Consultant, Contact Details

Client

ame normally Sc ool oard of Management

ele one enter details

Mobile enter details

a enter details

Email enter details

Client’s Representative

Mobile s above amend if different

a s above amend if different

Email s above amend if different

Consultant

Mobile enter details

a enter details

Email enter details

Consultant’s Representative

Mobile s above amend if different

a s above amend if different

Email s above amend if different

De artment of Education and S ills, Planning and uilding nit

Pg 18 of 22

Appendix E Schedules A and B Conditions of Engagement

2 PROJECT:

Enter rief descri tion of t e or s e.g. ll as er attac ed sco e of or

7 Whole, parts, of other documents included in the Contract

2. PERFORMANCE

12 Consultant has no authority to make – any C ange Order it an e tra value above:

C ange Orders in any t ree mont eriod it a cumulative e tra value above:

any C ange Order causing or contributing to a reduction in safety, uality, usefulness, of t e Pro ect. not a licable do not c ange t is

18 Insurance types, terms do not c ange figures in yello

Em loyers iability for deat , in ury, to em loyees € 13.0m one rom start to com letion of t e

Services.

Insurance of lans,

documents € n/a one rom start to com letion of t e

Services.

4. PROGRESS, PERIODS

6 Total Performance Period

e otal Performance Period is Enter otal Pro ect duration from a ointment of consultant until substantial com letion defects liability eriod say 6 mont s 3 mont s Client loat days starting on t e day t e Parties made t e Contract.

7 COORDINATION

5 Facilities from the Client

De artment of Education and S ills, Planning and uilding nit

Pg 19 of 22

Appendix E Schedules A and B Conditions of Engagement

6 Client’s resident staff

11 Team Leader

e Consultant is team leader.

10. PAYMENTS

4 Interest

e rate of interest ayable on Client s rongful deduction is 5 .a.

13. INTELLECTUAL PROPERTY, DOCUMENTS

11 Transfer

ere is not transfer to t e Client instead of licence.

14 Licence

Client may use Consultants design etc. for

Individual ro ects: /a

y es of ro ect: /a

15 Fees [if any]

e only fees ayable by t e Client for its rig ts under t is clause are: n/a

18 Publicity

Consent to ublicity is re uired from t e client.

14. TERMINATION

29 Payment where Client terminates at will do not c ange figures in yello

ere t e Client terminates Services at ill alt oug t e Pro ect is continuing, t e Consultant is entitled to

10 of t e difference bet een t e ees ayable under clause 14.26 it out any a ortionment under 14.27 , and t e total fee t at ould, as estimated at termination, ave been ayable on com letion of t e

Services for t e last Stage in or after ic t e termination occurred.

16. DISPUTES

1 Initial resolution method

not a licable

4 Nominator

In default of agreement, an arbitrator, conciliator or ot er for 16.1, ill be nominated at t e re uest of eit er

arty by t e follo ing erson s : President Royal Institute of rc itects of Ireland amend to IEI President if

consultant is Engineer

5 Rules

e a licable Rules are t e follo ing ublis ed for use it t ese Standard Conditions of Engagement:

In t e case of rbitration, t e rbitration rules are t e Public or s and Services rbitration Rules 2008.

De artment of Education and S ills, Planning and uilding nit

Pg 20 of 22

Appendix E Schedules A and B Conditions of Engagement

SCHEDULE B: CONSULTANT’S SERVICES AND FEES

CONSULTANT’S STAGE SERVICES

e Consultants a ointment is for Stage i to iii as tabled belo .

Do not c ange

PSDP SERVICES

Performance of all t e duties of Pro ect Su ervisor for t e Design Process as tabled belo is not amend if re uired included in t e Services as tabled belo and t e Stage ees .

TOTAL FEE

Ignore Stages iv and v or enter n/a .

PROJECT SUPERVISOR FOR THE DESIGN PROCESS SERVICES

ll t e duties of Pro ect Su ervisor for t e Design Process according to t e Safety, Healt and elfare Construction Regulations 2006,

not it standing any inconsistent Contract contents.

Performance period

ile re uired for t ese Services

sub ect to any later a ointment, earlier

clause 12 ermination, by t e Client.

o additional

ayments

De artment of Education and S ills, Planning and uilding nit

Pg 21 of 22

Appendix E Schedules A and B Conditions of Engagement

TIME CHARGES

e follo ing ime C arges a ly to Stage Services if Sc eduled , sus ension 4.22 , Client s C anges clause 11 .

Grade

Em loyers Re resentative

Pro ect Manager

Eac Senior rc itect/Engineer

Pro ect rc itect/Engineer n/a

Cler of or s / Resident Engineer em loyed directly by Consultant

ssistant Cler of or s / Resident Engineer n/a

MANAGEMENT SERVICES

e follo ing management services are included as art of t e sco e of service and t e overall fees. e agreed fee is for t e or com lete including management services as belo re uired to com lete t e ro ect. Do not change any of the items highlighted in yellow unless you have agreed specific services with the consultant.

[Included in Stage Services as relevant, and in the Total Fee and Stage fees]

s re uired to com lete ro ect satisfactorily

De artment of Education and S ills, Planning and uilding nit

Pg 22 of 22

Semi-Supervised Feature Importance

Evaluation with Ensemble Learning

Hasna Barkia*, Haytham Elghazel* and Alex Aussem*

*Universit´e de Lyon, 69000, Lyon, France; Universit´e de Lyon 1,

Laboratoire GAMA, EA 4608, 69622 Villeurbanne.

Email: hasna.barkia@etu.univ-lyon1.fr, haytham.elghazel@univ-lyon1.fr, alex.aussem@univ-lyon1.fr

http://gama.univ-lyon1.fr/MLKD

Abstract—We consider the problem of using a large amount of unlabeled data to improve the efficiency of feature selection in high dimensional datasets, when only a small set of labeled examples is available. We propose a new semi-supervised feature importance evaluation method (SSFI for short), that combines ideas from co-training and random forests with a new permutation-based out-of-bag feature importance measure. We provide empirical results on several benchmark datasets indicating that SSFI can lead to significant improvement over state-of-the-art semi-supervised and supervised algorithms.

Keywords-Feature Selection, Semi-Supervised Learning, En-semble Method, Co-training, Bagging, Random Subspaces Method

I. INTRODUCTION

The identification of relevant subsets of random variables among thousands of potentially irrelevant and redundant variables is a challenging topic of pattern recognition re¬search that has attracted much attention over the last few years. In supervised learning, feature selection algorithms use only information from labeled data to find the relevant subsets of variables, i.e., those that conjunctively prove useful to construct an efficient classifier from data. It enables the classification model to achieve good or even better solutions with a restricted subset of features [1], [2], [3]. However, in many real-world applications, the amount of labeled data is very limited and it becomes difficult to iden¬tify and remove the redundant and irrelevant variables from the feature set, especially in high dimension. This situation arises naturally in many real-world applications, where large amount of data can be collected cheaply and automatically, but when manual labeling of samples remains extremely time consuming and/or cannot be taken for granted. In this case, unsupervised feature selection methods could be envisaged to exploit the information conveyed by the a large amount of unlabeled training data [4], [5], [6], [7]. Broadly speaking, the feature selection in unsupervised learning aims at finding relevant subsets of variables that produce ”natural” groupings by grouping ”similar” objects together based on some similarity measure. Clearly, the combination of both paradigms (supervised and unsupervised) allows the merging of sophisticated semi-supervised approaches that

can handle both labeled and unlabeled data. The problem of semi-supervised feature selection has attracted a great deal of interest recently and its effectiveness has already been demonstrated in many applications [8], [9], [10], [11].

On the other hand, databases have increased many fold in recent years. Important recent problems (i.e., DNA data in biology) often have the property that there are hundreds or thousands of features, with each one containing only a small amount of information. A single learner is known to produce very bad results as the learning algorithms break down with high dimensional data. Ensemble learning paradigms train multiple component learners and then combine their output results. Ensemble techniques are considered as an effective solution to overcome the dimensionality problem and to im¬prove the robustness and the generalization ability of single learners, and therefore has been a hot topic during the past years. Although considerable attention has been given on the problem of constructing an accurate and diverse ensemble committee for supervised and unsupervised learning, and using this committee to estimate the feature importance [12], [6], [5], little attention has been given to exploiting the power of ensemble with a view to identify and remove the irrelevant features in a semi-supervised setting.

The way internal estimates are used to measure variable importance in the Random Forests (RF) paradigm [12] have been influential in our thinking. In this study, we show that these ideas are also applicable to semi-supervised feature selection. We propose a novel semi-supervised feature im¬portance evaluation method termed SSFI as a shorthand. The algorithm ranks features through an ensemble frame¬work, in which a feature’s relevance is evaluated by its predictive accuracy using both labeled and unlabeled data. SSFI combines both data resampling (bagging) and random subspace strategies for generating an ensemble learner using a co-training style algorithm. A combination of these two main strategies for producing ensemble of classifiers leads to exploration of distinct views of inter-pattern relationships. Once each ensemble member is obtained, an extension of the RF permutation importance measure [12], using the labeled and unlabeled data together, is proposed to measure feature’s relevance. A ranking of all features is finally obtained with

respect to their relevances in all obtained semi-supervised classifiers.

The rest of the paper is organized as follow: Section 2 reviews recent studies on semi-supervised feature selection and ensemble methods. Section 3 introduces the SSFI frame¬work and describes how variable importance used in RF can be extended in semi-supervised context by using both labeled and unlabeled data. Experiments using relevant high-dimensional benchmarks and real datasets are presented in Section 4.

II. RELATED WORK

In this section, we briefly review the semi-supervised feature selection and semi-supervised ensemble learning approaches that appeared recently in the literature.

A. Semi-Supervised feature selection

The key for designing an effective semi-supervised feature selection algorithm is to develop a framework, under which the relevance of a feature can be evaluated by both labeled and unlabeled data in a natural way. Recently, several studies have focused on semi-supervised feature selection. Like in supervised and unsupervised FS, these methods can be divided into three categories, depending on how they interact with the learning algorithm: filter, wrapper and embedded approaches. Filter methods discover the relevant and redundant features through analyzing the correlation and dependence among features without involving any learning algorithms [10], [11]. The most common filter strategies are based on feature ranking. Feature ranking is a relaxed version of feature selection which ranks all features with re¬spect to their relevances and chooses the top ranked features as the working feature vector manually. Therefore, feature ranking can be viewed as a kind of flexible feature subset selection approach. Feature ranking has been well studied for semi-supervised classification. Zhao et al. [10] proposed a semi-supervised feature ranking algorithm, referred to as Sselect, based on the spectral graph theory. Their method first constructs a neighborhood graph using original data, and then evaluates each feature vector by transforming it into a cluster indicator and checking whether it is consistent with label information. It has demonstrated promising results on some benchmark datasets. In [11], a semi-supervised feature selection algorithm, called Locality Sensitive Discriminant Feature (LSDF) was proposed. Unlike Fisher score which makes use of only labeled data points and Laplacian score which makes use of only unlabeled data points, the proposed algorithm makes use of both labeled and unlabeled data points. It tries to discover both geometrical and discrimi¬nant structure in the data. using two graphs, i.e., within-class graph and between-class graph. The within-class graph connects data points which share the same label or are sufficiently close to each other, while the between-class graph connects data points which are sufficiently close to

each other but have different labels. The importance of the features is characterized by its degree of preserving these graph structures. Specifically, a feature is considered as ”good” if at this dimension nearby points, or points sharing the same label, are close to each other, while points with different labels are far apart. However, the presence of a large amount of irrelevant features often leads to inexact neighborhood mapping and causes both aforementioned methods to fail [8].

On the other hand, wrapper methods perform a search in the space of feature subsets, guided by the outcome of the learning model. Typically, a criterion is firstly defined for evaluating the quality of a candidate feature subset and wrapper approaches aim to identify a feature subset such that the learning algorithm trained on this feature subset can achieve the optimal value of the predefined criterion. In [9], a forward search based semi-supervised feature ranking method is proposed. It uses the mechanism of random selection on unlabeled data to form new training sets, and the most frequently selected feature, using supervised sequential forward search strategy, is added to the result feature subset in each iteration. In this method, the subset of features derived from the random training sets used may not be adequate, but once the feature is chosen, it will never be eliminated.

In contrast to filter and wrapper approaches, the search for an optimal subset of features with embedded methods is built into the model construction making these techniques specific of a given learning algorithm. Recently, Zenglin et al. [13] proposed a semi-supervised feature selection method that works in an embedded way. The feature selection process is integrated to the semi-supervised classifier by taking ad¬vantage of manifold regularization. In the proposed method, an optimal subset of features is identified by maximizing a performance measure that combines classification margin with manifold regularization. The manifold regularization in the proposed feature selection method assures that the decision function is smooth on the manifold constructed by the selected features of the unlabeled data.

B. Semi-supervised ensemble learning

Semi-supervised learning has been widely applied in many real-world application domains such as medical di-agnosis, fraud detection and pattern recognition. Semi-supervised learning methods are used in order to make use of unlabeled data in addition to the labeled data for better classification. According to the feature spaces used, semi-supervised learning (SSL) algorithms can be divided into single-view and multiple-view algorithms. One of the most successful single-view algorithms is the Self-Training algorithm in which a single classifier is initially trained using a small amount of labeled data. Then it adds the most confident unlabeled data incrementally into the la¬beled dataset and retrains the underlying classifier with

the augmented training set. On the other hand, Co-training is one of the most attractive multi-view SSL algorithms. Introduced by Blum and Mitchell in [14], in Co-training two classifiers are initially trained using two redundant and independent sets of features (views). Then in each further iteration, each classifier classifies the unlabeled examples, adds the examples about which it is most confident into the training set. The aim is that the most confident examples with respect to one classifier can be informative with respect to the other. Although co-training has emerged as a powerful method in some fields, the requirement on two sufficient and redundant attribute subsets is too strong to be met in many real-world applications. Therefore, many extensions of co-training have been proposed in the literature to deal with this problem. The proposed alternatives are generally ensemble-based and differ on the strategy they used to generate component classifiers. Methods for constructing ensembles include manipulation of the training samples by resampling (bootstrap aggregation or bagging) [15], [16], [17], [18], [19], [20] or using random subspaces [15], [19], [21], [22].

In [18], an ensemble co-training style method named Co-Forest is proposed. It extends the co-training paradigm by incorporating a well-known ensemble learning algorithm named Random Forest [12] to tackle the problems of how to determine the most confident examples to label and how to produce the final hypothesis. Co-Forest uses bootstrap sample data from training set and trains random trees. At each iteration each random tree is reconstructed by newly selected examples for its concomitant ensemble. Further¬more, [16] gives another extension of the usage of RF to semi-supervised learning problems. In order to incorporate unlabeled data, the main idea consists to use the predicted labels of the unlabeled data as additional optimization vari¬ables. The authors in [16] perform an iterative determinis¬tic annealing-style training algorithm maximizing both the multi-class margin of labeled and unlabeled samples.

Another ensemble semi-supervised learning approach is given by the work in [15] named Co-training By Committee (CoBC). In this work, an ensemble of diverse classifiers is used instead of redundant and independent views. The com¬mittee of diverse accurate classifiers is initially constructed by using a successful ensemble learning algorithms: Bagging or random subspace method. At each iteration and for each classifier, a subset of unlabeled examples are drawn ran¬domly from the whole unlabeled dataset and classified using the concomitant ensemble. The most confident examples to label are then determined and the committee members are retrained using their updated training sets.

It should be noted that all extensions of Co-training that requires bootstrapping may need a lot of labeled samples in order to be successful. For high dimensional datasets,the classifiers trained on small bootstrapped data samples using single feature view may face the ”large p, small n problem” (the size of the training set is much smaller than the number

of dimensions in the feature vector) and, thus, may cause an overfitting problem.

As a solution, random subspace methods (RSM) are one of the successful methods used for producing an ensemble of classifiers and dealing with high dimensional datasets. RASCO [21] algorithm combines the ideas of Co-training and random subspace methods. Instead of using two feature subspaces, it uses random feature splits in order to train different classifiers. The unlabeled data samples are labeled and added to the training set based on the combination of decisions of the classifiers trained on different feature splits. The intuition behind this is that each classifier can com-plement another one. RASCO has been shown to perform better than Co-training method. In [22], instead of totally random feature subspaces, the authors propose to produce relevant random subspaces by means of drawing features with probabilities proportional to their relevances measured by the mutual information between features and class labels. The results obtained on different datasets show that the proposed algorithm, termed as Rel-RASCO, outperforms both RASCO and Co-training methods.

Another similar semi-supervised learning approach to RASCO, that uses support vector machines, was proposed to be used for content based image retrieval [19]. Authors in [19] propose to use bagging and random subspace strategy in the same framework since they are especially effective when the original classifier is not very stable and can generate more diversified classifiers.

III. THE METHOD

In this section, we discuss our semi-supervised feature importance evaluation method, that combines ideas from co-training and RF with a new permutation-based out-of-bag feature importance measure.

A. Committee construction

As discussed before, the most important condition for a successful ensemble learning method is to combine models which are different from each other, i.e. that make error on different training examples. Thus, to maintain diversity between committee members, we have employed two strate¬gies. Firstly, a well known ensemble method named RSM, is employed to face the curse of dimensionality problem by constructing multiple classifiers each one trained on different subset of examples projected on a smaller feature set RSMi. Secondly, the diversity is further maintained, by applying the bootstrapping method. The formal description of our approach is given in Algorithm 1. Given a set of labeled training examples L, and a set of unlabeled training examples U, described over the input space F = f1,..., fp, our approach constructs a committee according to the following steps.

First, as described in the steps from 3 to 11 of the Algorithm 1, the initial committee is constructed as follows:

Algorithm 1 SSFI(L, U, F, K, N, n, maxiter, BaseLearn)

Require:

set of labeled training examples (L), set of unlabeled training examples (U), input space (F = f1, ... , fp), number of classes (K), committee size (N), sample size (n), maximum number of iterations (maxiter) and base learning algorithm (BaseLearn)

1: Get the class prior probabilities, PrkKk=1

2: Set the class growth rate, nk = n  Prk where k = 1, ... , K

Initial committee construction H

3: H = 

4: for i = 1 : N do

5: RSMi = randomly draw m features from F

6: Libag = bootstrap sample from L projected onto RSMi

7: Uibag = bootstrap sample from U projected onto RSMi

8: Lioob = LLi bag, Uioob = UUibag

9: hi = BaseLearn(Libag)

10: H = H hi

11: end for

Committee refinement using SSL ensemble method

12: t = 1

13: repeat

14: for each hi H do

15: 7ri=SelectConfidentExamples(i, H, Uibag,nkKk=1)

16: Libag = Libag  7ri, Uibag = Uibag7ri

17: hi = BaseLearn(Libag)

18: end for

19: t = t + 1

20: until (t > maxiter OR no committee member changes)

Feature relevance estimate

21: imp = 0

22: for each hi H do

23: [Oidata, Oilabel, Oi ccnf] =BuildOOBMatrix(i, H, Lioob, Uioob, K)

24: for each f  RSMi do

25: randomly permute the values of f over the Oi daraexamples to form Oiperm

26: for each x  Oiperm do

27: if (hi(x) = Oilabel(x)) then

28: imp(f) = imp(f) + Oiconf (x)

29: end if

30: end for

31: end for

32: end for

33: rank the features f according to imp(f)

34: return F and imp

For each committee member hi, Libag and Uibag are selected with replacement, from L and U respectively, and projected over RSMi, a feature subspace with m randomly selected features (m < p). Then, each component hi is constructed according to a given baseLearner based on its corresponding labeled training examples Libag.

Second, according to the steps from 12 to 20 in the Algo¬rithm 1, the co-training method trains each hi, by asking a subset of the concomitant classifiers to label examples from Uibag for it, then a set 7ri, which consists of the nk most confident examples assigned to each class k, is removed from Uibag, and incrementally added into Libag. Then a new hi is retrained over the augmented set Libag. A formal description is given in the Algorithm 2, to describe how the most confident examples are selected.

The co-training steps are repeated until a maximal num-ber of iteration is reached or the committee is no longer changing.

B. Confidence Measure

An important factor that affects the performance of any Co-Training style algorithm is how to measure the confi-dence about the labeling of an unlabeled example which determines its probability of being selected. An inaccurate confidence measure leads to adding mislabeled examples to the labeled training set which leads to performance degradation during the SSL process.

In the Algorithm 2, a formal description is given, to explain how the most confidant examples are selected. In order to improve the accuracy of a committee member hi, its unlabeled examples, Uibag will be labeled by the other components. More specifically, for a given unlabeled example x, let Hx be the concomitant ensemble of hi, which contains only members where x is out of bag. In order to guarantee the consistence of the learning process and an accurate labeling for unlabeled data, we have chosen to label a given unlabeled example x, only by the members hj of its corresponding Hx. Thus, a given example x, in a given iteration t, will have the same label for all the committee members hi H, where x is  Uibag.

Then, for each unlabeled example x  Uibag, each commit¬tee member hj  Hx, will label it, in order to generate the class probability distribution for the given x. Then a majority voting method is applied over Hx, in order to attribute the final class label of x: As described in the Algorithm 3, each classifier from Hx is asked to label x, in order to generate the class probability distribution for the given x. Then the class which receives the maximal votes, is assigned to the example x, with a label confidence equal to the degree of agreement on the labeling, i.e. the number of classifiers that agree on the label assigned by Hx.

C. Out-of-bag based feature relevance measure

In our approach, the Random subspace method is com-bined to bootstrapping. Actually, in each bootstrapped la¬beled and unlabeled set, almost 33% are left oob, i.e., they are not used for the construction of the corresponding model. We refer to them as Uioob and Lioob. Thus, these patterns can be used to estimate non biased feature relevancies. The first step consists to build the Out Of Bag information structure Oi = [Oi data, Oi label, Oiccnf] as described in the Algorithm 4. For each classifier, we select the well predicted instances from Lioob and Uioob using hi to form the set Oi. Clearly, for the labeled examples, an example is well predicted, if the class label given by hi corresponds to the real label. Its label confidence is set to 1. For the unlabeled examples, the right label is unknown. Also, the key idea is to assume that an unlabeled example x is “well labeled” by hi, if the label given by hi is the label given by the majority vote given by the committee Hx. In that case, its label confidence will be set to the degree of agreement for winning label among the members of Hx. Second, the values of the fth feature in the Oidata, are randomly permuted to form Oiperm, and hi is used to predict the label of the new Out Of Bag patterns. The procedure is repeated for every feature f f1, ... , fp. At the end of the run, the sum of all the example’s confidence for which the predicted label in the Oi permdiffers from the initial predicted label in the initial Oidata, is computed. The latter value is averaged over N, i.e., the committee size. The resulting value is taken as the importance of the feature f. The key idea in our approach is the use of label’s confidence in the evaluation of the feature importance. So the unlabeled examples play an important role in the feature importance evaluation.

Algorithm 2 SelectConfidentExamples(i, H, Uibag, nkKk=1)

Require:

a committee (H), current committee member index (i), pool of unlabeled examples (Uibag), growth rate (nkK k =1) and number of classes (K)

1: 7ri = 0)

2: for each x  Uibag do

3: Hx=hjHxUjoob

4: [label(x), conf (x)] =MeasureConfidence(x, Hx, K)

5: end for

6: Rank the examples in Uibag by decreasing order of confidence and select the nk most confident examples for each class k

7: for each x  Uibag do

8: if (x is selected as confident) then

9: 7ri = 7rix, label(x)

10: end if

11: end for

12: return 7ri

Algorithm 3 MeasureConfidence(x, Hx, K)

Require:

an unlabeled training example (x), a committee of classifiers for which x is out-of-bag (Hx) and number of classes (K)

1: Apply Hx to generate the class probability distribution for x as P (x) = {pk(x) : k = 1, ... , K}

2: conf (x) = max1<k<KP (x)

3: label(x) = arg max1<k<KP (x)

4: return conf (x) and label(x)

Algorithm 4 BuildOOBMatrix(i, H, Lioob, Uioob, K)

Require:

a committee (H), current committee member index (i), out-of-bag labeled examples of hi (Lioob), out-of-bag unlabeled examples examples of hi (Uioob) and number of classes (K)

1: Oi data = 0, Oi label = 0, Oi conf = 0

2: for each x  Lioob do

3: if (hi(x) == Lioob(x)) then

4: Oidata = O idata  {x}

5: Oilabel(x) = hi(x)

6: Oiconf (x) = 1

7: end if

8: end for

9: for each x  Uioob do

10: Hx={hjH|xUjoob}

11: [label(x), conf (x)] =MeasureConfidence(x, Hx, K)

12: if (hi(x) == label(x)) then

13: Oidata = Oidata  {x}

14: Oilabel(x) = hi(x)

15: Oiconf (x) = conf (x)]

16: end if

17: end for

18: return Oidata, Oilabel and Oiconf

D. Why should our approach work

There are several advantages with the proposed method. First, SSFI will outperform RF when the available labeled training set is small. RF relies on the available training data for encouraging diversity. So if the size of the training set is small as for semi supervised setting, then the diversity among the ensemble members will be limited. Consequently, the ensemble error reduction will be small. SSFI incre¬mentally adds newly-labeled examples to the training set. Therefore, it can improve the diversity and the average error of ensemble members constructed by RF and then improve the feature ranking paradigm. Second, since SSFI uses a diverse ensemble creation method, the measure of feature importance based on ensemble is more accurate than using a single classifier. Third, It is also worth mentioning that

Table I

THE DATASETS USED IN THE EXPERIMENTS

Datasets # patterns # features # classes Reference

Baseshock 1993 4862 2 [24]

Colon 62 2000 2 [25]

Leukemia 73 7129 2 [26]

Madelon 2598 500 2 [23]

Orlraws 100 10304 10 [24]

Ovarian 54 1536 2 [27]

Pcmac 1943 3289 2 [24]

SMK-CAN 187 19993 2 [24]

Toxicology 171 5748 2 [24]

Warpar10P 130 2400 10 [24]

the way feature importance measure is performed, in our approach differs, from the feature importance measure in RF as well as its recent extensions: Co-forest [18] and semi-supervised random forest [16]. In Co-forest, the variable importance measure can not be estimated from OOB samples since the bootstrap sample used to train each random tree is discarded after the first iteration. In semi-supervised RF, OOB data are all labeled. However, since the amount of labeled data is very small, the diversity of oob data is not sufficient. The out-of-bag estimates are biased as they depend on too few data.

IV. EXPERIMENTS

In this section, we provide empirical results on several benchmark and real high-dimensional datasets and com-pare SSFI against over state-of-the-art semi-supervised and supervised algorithms feature ranking algorithms. SSFI is compared with three other feature selection methods : (1) Breiman’s supervised random forests (RF) [12] taken as our gold standard ensemble supervised feature ranking approach, (2) the wrapper-type Forward Semi-Supervised Feature Se¬lection (FwFS) [9], and (3) the filter-type Semi-Supervised Feature Selection via Spectral Analysis (sSelect) [10]. Ten benchmark and real labeled datasets, mostly selected from the UCI Machine Learning Repository [23], and from ASU feature selection Repository [24], were used to assess the performance of SSFI. They are described in Table I. We se¬lected these datasets as they contain thousands features with comparatively much smaller sample size (e.g., Leukemia, Toxicology, Orl10p, ovarian, colon and SMK-CAN) and are thus good candidates for feature selection. Most of these datasets have already been used by other authors for testing the performance of their feature selection algorithms [5], [10], [9], [6].

A. Evaluation framework

To make fair comparisons, the same experimental settings in [10] was adopted here for sSelect approach, i.e., a neighborhood graph with a neighborhood size of 10, and the λ value is set to 0.1. For FwFS, we set sizeFS = 10, SamplingRate = 0.5, Sampling Times = 10, f nsteps =

6 and startf n = 5, as suggested by the authors in [9]. RF and SSFI are tuned similarly. The number of features per bag is p. The committee size N is computed using the following formula:

log(0.01) log(1  1/p) .

This formula ensures that each feature is drawn ten times at a confidence level of 0.01. Furthermore, the number of iterations maxiter and the sample size n in our approach are set to 10, and 1, respectively. As we have to compare our approach with RF that uses decisiontree, the treefit Matlab implementation of decision tree is used as the base classifier in FwFS and SSFI for fair comparison. For each dataset, experimental results are averaged over 10 runs. At each run, the whole dataset is splitted (in a stratified way) into a training partition with 2/3 of the observations and a test partition with the remaining 1/3 observations. Training set is further splitted into labeled and unlabeled datasets. As in [9], [10], the labeled sample set L consists of randomly selected 3 patterns per class, and the remaining patterns are used as unlabeled sample set U.

In order to assess the quality of a feature subset ob-tained with the aforementioned semi-supervised procedure, we train a classifier (a decision tree) on the whole labeled training data and evaluate its accuracy on the test data. The latter is taken as the score for the feature subset. The details of the evaluation framework are shown in Algorithm 5. As mentioned above, the process specified in Algorithm 5 is repeated 10 times. The obtained accuracy is averaged and used for evaluating the quality of the feature subset selected according to each algorithm. In Figure 1, we plotted the accuracies of the above four approaches against the 10 most important features.

B. Results

Figure 1 shows the plots for accuracy vs. different numbers of selected features. As may be observed, SSFI outperforms the other three methods by a noticeable margin, especially on BaseHock, Leukemia, Madelon, PC-Mac and Toxicology, and sSelect performs the worst. SSFI seems to combine more efficiently the labeled and unlabeled data for feature evaluation and it shows promise for scaling to larger domains in a semi-supervised way in view of the good performance on Leukemia, SMK-CAN and Orlraws. More importantly, SSFI outperforms RF on most datasets except on Warpar10P. However, we would like to mention that the labeled training examples in Warpar10P contains 30 examples for this dataset, which is an important amount of data (25% of the whole dataset). As expected, we a general trend that is observed in Figure 1 is that the more features we select, the better accuracy we achieve. Again, this is not surprising. However, it is worth mentioning that the accuracy of SSFI generally increases swiftly at the beginning

Algorithm 5 Feature Evaluation Framework

1: for each dataset X do

2: build a randomly stratified partition (Tr, Te), from X where Tr = 23.X and Te = 31.X;

3: Generate labeled data L by randomly sampling from Tr 3 instances per class;

4: U = TrL;

5: SFsSelect = Apply sSelect with L + U;

6: SFFwFS = Apply FwFS with L + U;

7: SFSSFI = Apply SSFI with L + U;

8: SFRF = Apply RF with L;

9: for i = 1 to 10 step 1 do

10: Select top i features from SFsselect, SFFwFS, SFSSFI and SFRF;

11: TrsSelect = "SFsS elect(Tr);

12: TrFwFS = "SFF w F S (Tr);

13: TrSSFI = "SFSSF I (Tr);

14: TrRF = "SFRF (Tr);

15: Train the Baselearner using TrsSelect, TrFwFS, TrSSFI and TrRF and record accuracy obtained on Te;

16: end for

17: end for

(the number of selected feature is small) and slows down at the end. These experiments suggest that SSFI ranks the features properly and that a classifier can achieve a very good classification accuracy with the top 5 features while the other methods require more features to achieve comparable results. Note also that the high computational complexity of FwFS is a major drawback with large dimensional data. In addition, the accuracy of FwFS often tend to decrease as more features are included.

Fore sake of completeness, we also averaged the accuracy for different numbers of selected features. The averaged accuracies between SSFI and the other methods over the top 10 features are depicted in Table II. Again, as may be observed, SSFI clearly outperforms RF, Sselect and Fw-semiFS by a noticeable margin, on all datasets except for Warpar10P where RF works the best. SSFI is significantly better then all three approaches (p < 0.02) according to the Wilcoxon signed-rank test (unsufficient number of datasets and classifiers to apply the Friedman test) [28]. Finally, these experiments confirm the ability of the proposed permutation feature importance measure to rank the relevant features ac¬curately, compared to a powerful fully supervised approach like RF, by exploiting efficiently the information from the unlabeled data.

V. CONCLUSION

We discussed a new semi-supervised feature importance evaluation method, called SSFI, combining ideas from co

Table II

ACCURACY AVERAGED OVER THE 10 MOST IMPORTANT FEATURES

data SSFI FwFS RF sSelect

Basehock 0.6080 0,5193 0,5445 0,5064

Colon 0,5645 0,5404 0,5327 0,5431

Leukemia 0,7707 0,6396 0,6950 0,6246

Madelon 0,5580 0,5024 0,5076 0,5008

Orlraws 0,6952 0,6200 0,6563 0,6145

Ovarian 0,8372 0,7444 0,7583 0,6327

Pcmac 0,5953 0,5312 0,5120 0,5123

SMKCAN 0,5671 0,5339 0,5408 0,5275

Toxicology 0,4777 0,3656 0,4008 0,3435

Warpar10P 0,4432 0,3952 0,4876 0,2966

training and random forests with a new permutation-based out-of-bag feature importance measure. Both labeled and

unlabeled out-of-bag instances were used to evaluate the relevance of the features. Empirical results on ten benchmark datasets indicated that SSFI lead to significant improvement

over state-of-the-art semi-supervised algorithms. More im¬portantly, SSFI was shown to outperform Random Forests

on several datasets in terms of feature selection accuracy. The method also shows promise to deal with very large domains. Future substantiation through more experiments on biological databases containing several thousands of variables are currently being undertaken.

REFERENCES

[1] I. Guyon and A. Elisseeff, “An introduction to variable and feature selection.” Journal of Machine Learning Research, vol. 3, pp. 1157–1182, 2003.

[2] J. Hua, W. Tembe, and E. Dougherty, “Performance of feature-selection methods in the classification of high-dimension data,” Pattern Recognition, vol. 42, pp. 409–424, 2009.

[3] Y. Saeys, I. Inza, and P. Larra˜naga, “A review of feature se¬lection techniques in bioinformatics,” Bioinformatics, vol. 23, pp. 95–116, 2007.

[4] J. Dy and C. Brodley, “Feature selection for unsupervised learning,” Journal of Machine Learning Research, vol. 5, pp. 845–889, 2004.

[5] H. Elghazel and A. Aussem, “Feature selection for unsu-pervised learning using random cluster ensembles,” in Pro-ceedings of the 10th IEEE International Conference on Data Mining (ICDM 2010), 2010, pp. 168–175.

[6] Y. Hong, S. Kwong, Y. Chang, and R. Qingsheng, “Un-supervised feature selection using clustering ensembles and population based incremental learning algorithm,” Pattern Recognition, vol. 41, no. 9, pp. 2742–2756, 2008.

[7] P. Mitra, A. Murthy, and S. Pal, “Unsupervised feature selec¬tion using feature similarity,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, no. 3, pp. 301– 312, 2002.

[8] Y. Cheng, Y. Cai, Y. Sun, and J. Li, “Semi-supervised feature selection under logistic i-relief framework,” in Proceedings of the 19th International Conference on Pattern Recognition (ICPR 2008), 2008, pp. 1–4.

[9] J. Ren, Z. Qiu, W. Fan, H. Cheng, and P. S. Yu, “Forward semi-supervised feature selection,” in Proceedings of the 12th Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD 2008), 2008, pp. 970–976.

[10] Z. Zhao and H. Liu, “Semi-supervised feature selection via spectral analysis,” in Proceedings of the Seventh SIAM International Conference on Data Mining, 2007, pp. 641– 646.

[11] J. Zhao, K. Lu, and X. He, “Locality sensitive semi-supervised feature selection,” Neurocomputing, vol. 71, no. 10-12, pp. 1842–1849, 2008.

[12] L. Breiman, “Random forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001.

[13] Z. Xu, I. King, M. R. Lyu, and R. Jin, “Discriminative semi-supervised feature selection via manifold regularization,” IEEE Transactions on Neural Networks, vol. 21, no. 7, pp. 1033–1047, 2010.

[14] A. Blum and T. M. Mitchell, “Combining labeled and unla-beled sata with co-training,” in Proceedings of the 11th An-nual Conference on Computational Learning Theory (COLT 1998), 1998, pp. 92–100.

[15] M. F. A. Hady and F. Schwenker, “Combining committee-based semi-supervised learning and active learning,” Journal of Computer Science and Technology, vol. 25, no. 4, pp. 681– 698, 2010.

[16] C. Leistner, A. Saffari, J. Santner, and H. Bischof, “Semi-supervised random forests,” in Proceedings of the 12th Inter¬national Conference on Computer Vision (ICCV 2009), 2009, pp. 506–513.

[17] Z.-H. Zhou and M. Li, “Tri-training: Exploiting unlabeled data using three classifiers,” IEEE Transactions on Knowledge and Data Engineering, vol. 17, no. 11, pp. 1529–1541, 2005.

[18] M. Li and Z. H. Zhou, “Improve computer-aided diagnosis with machine learning techniques using undiagnosed sam-ples,” IEEE Transactions on Systems, Man, and Cybernetics, vol. 37, no. 6, pp. 1088–1098, 2007.

[19] D. Tao, X. Tang, X. Li, and X. Wu, “Asymmetric bagging and random subspace for support vector machines-based relevance feedback in image retrieval,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 28, no. 7, pp. 1088–1099, 2006.

[20] M.-L. Zhang and Z.-H. Zhou, “Exploiting unlabeled data to enhance ensemble diversity,” in Proceedings of the 10th IEEE International Conference on Data Mining (ICDM 2010), 2010, pp. 619–628.

[21] J. Wang, S. Luo, and X. Zeng, “A random subspace method for co-training,” in Proceedings of the International Joint Conference on Neural Networks, IJCNN 2008, 2008, pp. 195– 200.

[22] Y. Yaslan and Z. Cataltepe, “Co-training with relevant random subspaces,” Neurocomputing, vol. 73, no. 10-12, pp. 1652– 1661, 2010.

[23] C. Blake and C. Merz, “Uci repository of machine learning databases,” 1998.

[24] Z. Zhao, F. Morstatter, S. Sharma, S. Alelyani, and A. Anand, “Feature selection,” 2011.

[25] A. Ben-Dor, L. Bruhn, A. Laboratories, N. Friedman, M. Schummer, I. Nachman, U. Washington, U. Washington, and Z. Yakhini, “Tissue classification with gene expression profiles,” Journal of Computational Biology, vol. 7, pp. 559– 584, 2000.

[26] T. Golub, Slonim, D.K., P. Tamayo, C. Huard, M. Gaasen-beek, J. Mesirov, and H. Coller, “Molecular classication of cancer: Class discovery and class prediction by gene expression monitoring,” Science, vol. 286, pp. 531–537, 1999.

[27] M. Schummer, W. V. Ng, and R. E. Bumgarnerd, “Compar-ative hybridization of an array of 21,500 ovarian cdnas for the discovery of genes overexpressed in ovarian carcinomas,” Gene, vol. 238, no. 2, pp. 375–385, 1999.

[28] J. Demsar, “Statistical comparisons of classifiers over multiple data sets,” Journal of Machine Learning Research, vol. 7, pp. 1–30, 2006.

The Choice of Stochastic Process in Real Option

Valuation II: Selecting Multiple Factor Models

Luiz de Magalh˜aes Ozorio

Department of Management, Ibmec Business School,

Rio de Janeiro, Brazil

email: lmozorio@ibmecrj.br

Pavel V. Shevchenko

CSIRO Mathematics, Informatics and Statistics,

Sydney, Australia

email: Pavel.Shevchenko@csiro.au

Carlos de Lamare Bastian-Pinto

Department of Management, Ibmec Business School,

Rio de Janeiro, Brazil

email: carbastian@gmail.com

Draft paper, 5 July 2013

The Choice of Stochastic Process in Real Option Valuation II: Selecting Multiple Factor Models

Abstract

The stochastic process choice plays a central role in real option valuation and it can have an impact not only on the project value but also on the investment rule. The first studies on real options used one-factor models such as Geomet¬ric Brownian Motion (GBM) and Mean Reversion Models (MRM) to represent uncertainties in the valuation modeling. Selecting the most appropriate model is not always a trivial issue, and besides statistical tools, in general, theoretical considerations are taken for this task. In order to generate more realistic models, in the last decades many authors have presented papers proposing the combi¬nation of different kinds of stochastic processes creating multiple factor models. Although these models can be more realistic, the task of selecting among many multiple factor models is more difficult than in case of one-factor models. This paper discusses the choice of multiple factor models in real options valuation, and the main statistical tools and theoretical considerations that can be used for this task.

Key words: stochastic process, real options valuation, multiple factor models,

model choice

JEL codes: C15, C53, G13, M21.

1 Introduction

The choice of a stochastic process is an issue of great relevance in the assets valuation modeling aiming to represent uncertainties related to investments. In the case of real options it can have an impact not only on the project value but also on the investment rule (Dixit & Pindyck 1994; Schwartz 1997).

Earlier studies on financial options (Black & Scholes 1973; Cox, Ross & Rubin¬stein 1979) and real options (Brennan & Schwartz 1985; McDonald & Siegel 1985 and 1986; Paddock, Siegel, & Smith 1988) assumed Geometric Brownian Motion (GBM) for the underlying asset. For valuation of commodities, it is common to use Mean Reversion Models (MRM) (Bhattacharya 1978; Brennan & Schwartz 1985; Dixit & Pindyck 1994), assuming that commodity price might behave randomly in short term but tends to converge to an equilibrium level in the long run reflecting the marginal cost of production.

The task of determining the most appropriate process for the underlying stochastic variables is usually not a trivial question and, in some cases, analysts realize that these uncertainties have elements of more than one type of process. In order to generate more realistic models, several authors proposed models with multiple factors that com¬bine different kinds of processes (Schwartz 1997; Pindyck 1999; Dias & Rocha 1999; Schwartz & Smith 2000). Although the use of these models might increase the accuracy to the asset valuation, it is not straightforward to select the appropriate multi factor model. Moreover, real options typically have a feature that allows exercise at any time before maturity. Thus, valuation of real options requires numerical methods. In the case of one or two-factor models, one can use the finite difference or tree methods to solve corresponding partial differential equations. However, in the case of models with more than two factors, one has to resort to special Monte Carlo approaches such as the least square Monte Carlo method suggested in Longstaff and Schwartz (2001).

This paper discusses the choice of stochastic process in real option valuation and useful tests and considerations to resolve this task: Dickey Fuller Test, Variance Ratio Test, Q-Q plots, autocorrelation, likelihood ratio test, Akaike information criterion, Bayesian information criterion, Bayes factors, direct calculation of model probabili¬ties, deviance information criterion, out-of-sample or cross validation, in-sample errors, hedging errors and sensitivity analysis. We also present a set of empirical examples using real datasets and some simple real option applications in order to discuss the effect of stochastic process choice in the analysis.

The paper is structured as follows. Section 2 presents a bibliographical revision of stochastic processes focusing on multiple factor models applied to real option analysis. In Section 3 we describe some statistical tools that can be used to select the model.

Section 4 presents some empirical applications, and we conclude in Section 5.

2 Multiple Factor Models in Real Options Theory

GBM is the most popular stochastic model used not only in financial derivatives (Black & Scholes 1973; Cox, Ross & Rubinstein 1979) but also in real options analysis (Bren¬nan & Schwartz 1985; McDonald & Siegel 1985, 1986; Paddock, Siegel, & Smith 1988). Its mathematical simplicity that allows to obtain analytical solutions for asset valua¬tion and small number of parameters to calibrate can be considered the main reasons to explain its popularity. In other situations, when the uncertainties in prices depend on an equilibrium level, such as in case of commodities and interest rates, it is debated if the use of GBM would be appropriate (Bhattacharya 1978; Brennan & Schwartz 1985; Dixit & Pindyck 1994). In case of commodities such as oil, copper, sugar and ethanol, it is usual to assume that the price is driven by a mean reversion component, which makes the prices to behave randomly in short term and converge to the equilibrium level associated to the marginal cost of production in the long run.

The task of selecting GBM or MRM in order to represent the main uncertainties involved in a valuation is usually not a trivial issue. Besides statistical diagnostic tools such as Dickey Fuller Test (Dixit & Pindyck 1994) and Variance Ratio Test (Pindyck 1999), some theoretical considerations such as the economic theory and lifetime of assets (Ozorio, Bastian-Pinto & Brand˜ao 2012) can be useful. Nevertheless, in many cases, analysts realize that these uncertainties have elements of more than one process. This has motivated many authors to consider multiple factor models.

One of the pioneer works that presented a multiple factor model was Merton (1976), where GBM and Poisson process are mixed as

= (α  λk) dt + σ dz + dqt. (1)

Hereafter, St is the price of a financial asset at time t; dz is a Wiener increment; and qt is a Poisson process with the mean number of events per unit time λ and percentage change in price 0  1 if the Poisson event occurs. The uncertainty about the size and direction of the jump is represented by random variable 0  1 with the mean k = E[0  1]. That is, dqt is a Poisson process increment that takes a value 0  1 with probability λdt and zero with probability 1  λdt, i.e. E[dqt] = λk dt. The model parameters α and σ represent the drift and volatility parameters respectively. Here, dqt, dz and 0 are assumed to be independent. This model was used for stocks where the effect of common news in the stock prices is represented by GBM while the effect of rare events corresponds to a Poisson jump. Using a compensated Poisson process where

the jumps are non-systematic and the size of jumps is from a lognormal distribution, Merton (1976) derived a closed form formula for European options.

Another work that presents a multiple factor model with jump diffusion process is Dias & Rocha (1999), where the authors proposed a combination of Poisson process and MRM to represent the stochastic behavior of oil prices in real options valuation as

= [71( S  St)dt  λk]dt + adz + dqt, (2)

where 71 is a speed of reversion parameter; S¯ is the equilibrium level to which the process reverts in the long run. Here, E [dSt/St] = 71( S  St)dt because k = E[ϕ  1]. In this model, similarly to Merton (1976), the common news would cause marginal changes in oil prices, whereas abnormal events (such as crisis, wars and economic booms) would cause discrete jumps. The jumps can be systematic, which does not allow to obtain a risk neutral portfolio, or non-systematic, which allows the use of contingent claims.

There are also many papers proposing the combination of MRM and GBM in or¬der to represent the stochastic behavior of commodity prices; see Gibson & Schwartz (1990), Schwartz (1997), Pindyck (1999), and Schwartz & Smith (2000). These papers claim that besides MRM factor, price processes of some commodities may also have a stochastic upward trend factor. In practical terms, this trend factor would tend to in¬crease the equilibrium level to which the process reverts in the long run as time passes. These increases would have additional motivations to momentary mismatches of sup¬ply and demand (captured by MRM) and they would be caused by the progressive exhaustion of natural resources and incremental costs related to new requirements of environmental laws, among other issues. At the same time the improvements in the ex¬ploration and production technologies can impose a downward trend of the commodity prices.

Gibson & Schwartz (1990) proposes a two-factor model for pricing financial and real assets contingent on the price of oil, in which the factors are the spot price of oil that follows a GBM and the instantaneous convenience yield that follows a MRM. As explained by the authors, “the notion of convenience yield, viewed as a net “dividend” accruing to the owner of physical commodity at the margin, has already proven to drive the relationship between future and spot prices of many commodities.” Nevertheless, in order to justify the assumption of stochastic convenience yield the authors postulate that “the theory of storages posits an inverse relationship between the level of inven¬tories and the net convenience yield which suggests that a constant convenience yield assumption will only hold under very restrictive assumptions.” In Gibson & Schwartz (1990) model, it is assumed that the spot price of oil St and the net convenience yield

δt follow a joint diffusion process

dSt/St = µdt + σdz, (3)

dδt = κ(α  δt)dt + σdz,

where µ is the drift parameter of oil price process; σ is the volatility parameter of oil price process; dz and dz are Wiener increments; α is the equilibrium level to which the convenience yield process reverts in the long run; κ is the speed of reversion parameter of convenience yield; σ is the volatility parameter of convenience yield process. Here, dz and dz are correlated and E[dzdz] = ρdt, where ρ denotes the correlation coeffi¬cient between the two Wiener increments. In order to calibrate the parameters of the MRM (convenience yield process) and the correlation the authors used the seemingly unrelated regression model in conjunction with unrestricted regression model which was used to calibrate the parameters of GBM (oil spot price process). The risk pre¬mium of convenience yield was estimated exogenously using computational numerical techniques.

Schwartz (1997) considered three different models for commodity prices with appli-cations to commodity derivatives and commodity production assets valuation. The first model is one-factor model where the log of the price is a MRM. The second model has two stochastic factors and it is similar to the model developed by Gibson & Schwartz (1990) where the convenience yield follows a MRM while the price of commodity fol¬lows a GBM. In the third model, in addition to the stochastic commodity price and convenience yield, the author also considered the risk free rate as stochastic (following a MRM). The corresponding model for risk neutral porcesses is

dSt = (rt  δt)Stdt + σStdz*,

dδt = κ(α*  δt)dt + σdz*,

drt = a(m*  rt)dt + σdz*,

E[dz*dz*] = ρdt, E[dz*dz*] = ρdt, E[dz*dz*] = ρdt, (4)

where a is the speed of reversion of the risk free rate; m* is the equilibrium level to which the risk free rate reverts in the long run. The difficulty with implementation of the commodity price models is that some factors are not directly observable. Often, even the spot price of the commodities is not observable. Other kinds of variables, such as the instantaneous convenience yield, are even harder to estimate. On the other hand, futures prices of commodities are negotiated in many currencies and easy to observe. Schwartz (1997) derived state-space representations of the proposed models and applied Kalman filtering approach with the maximum likelihood method to estimate model parameters for copper, oil and gold. The analysis presented a strong evidence of mean

reversion component in copper and oil prices, but not in gold prices. It also points that the investments tend to be delayed if the mean reversion component is neglected in the real option analysis.

Pindyck (1999) analysed the price behavior of oil, coal and natural gas using 127-year time series and proposed, based on the historical performance, alternative ways for stochastic modeling of these commodity prices. The author comments that it would be ideal to be able to explain the price behavior of these commodities in structural terms by the movement of supply and demand in the market and the variables which determine them. Nevertheless he ponders that the structural models are not appro¬priately applicable to long term forecasts due to the difficulty related to explanatory variable forecasts of the models. As a result, the long run forecasts of energy commod¬ity prices are made many times assuming that these prices grow on fixed taxes in real terms, in order to reflect the depletion of these natural resources reserves. Typically in such cases, additionally to the drift, stochastic shocks are incorporated in order to reflect the future prices uncertainties, which in practical terms would correspond to assumption a random walk with drift or a GBM for the prices. Alternatively, in many cases it is assumed that in short run the prices may wander randomly due to the mo¬mentary pressures of supply and demand, but in the long run they tend to converge to their production marginal cost, which would mean that prices follow a MRM. The identification and choice of the process which best represents the price behavior of such commodities have serious implications in the project valuation, mainly in cases when the real options are being considered in the projects. The Variance Ratio Tests applied to the price series suggested the presence of mean reversion components, despite the difficulty in rejecting the unit root (which would be similar to difficulty in rejecting the GBM). Therefore, at first, the author proposes a mean reversion model where the mean has a quadratic deterministic trend in order to incorporate the increase in time of the production marginal cost of commodities

dxt = [ry(xt  α  αt  αt) + α + 2αt]dt + σdz. (5)

Here xt is the log of commodity price; α, α, α are the parameters of the quadratic trend of the log prices; and σ is the volatility parameter. Later, the author extended the model so that the level and slope could fluctuate stochastically, and proposed Kalman filter as the adequate approach to the parameter calibration.

Schwartz & Smith (2000) proposed a model with two stochastic factors Xt and fit (correlated and unobservable) to describe the behavior of commodity prices. The sum of these factors forms the log of commodity prices ln St = Xt+fit. The first factor, Xt, is a MRM with null mean reflecting the short run deviations of prices, caused by momentary

mismatches of supply and demand of the commodities. The second factor, ξt, represents the long run tendency of prices, influenced by the progressive exhaustion of natural resources and incremental costs related to new requirements of environmental laws, among other issues. Differently from other multiple factor models, this one does not consider the stochastic convenience yield factor, nevertheless the authors commented that it is equivalent to Gibson & Schwartz (1990) with an appropriate calibration. Specifically, the evolution of χt and ξt is given by the following model

dχt = κχtdt + σxdzx,

dξt = µ~dt + σ~dz~, (6)

E[dz~dzx] = ρdt,

where κ is the mean reversion parameter of short deviations; σx is the volatility pa¬rameter of the short run changes in prices; µ~ is the drift parameter of the long run price tendency; σ~ is the volatility parameter of the long run price tendency; ρ is the correlation parameter of the two factor increments. The authors estimated the param¬eters by fitting futures prices of commodities using state-space approach with Kalman filter method.

3 Methods for selecting stochastic processes

Different approaches have been suggested for valuation of real options. The real op¬tion analysis is more complex than standard option pricing in financial markets. The difficulties come from the facts such as the asset underlying the option may not be a tradable asset; the investment project can have controllable or uncontrollable cash-flows; the project can be or cannot be actively managed; for overview of these issues, see e.g. Sick and Gamba (2005). In general, the evolution of the underlying asset in real time is modelled by some stochastic process (referred to as real process) but eval¬uation of the fair price of financial derivatives driven by this underlying is done under the risk adjusted process (referred to as risk neutral process). Roughly speaking, there are two types of models for pricing derivatives: mark-to-market models and spot price models. In mark-to-market models, the modeller fits the risk neutral process to match exactly a set of market instruments traded today such as today’s prices of futures and vanillas. In spot price models, we fit both the real and risk neutral processes to the historical data (e.g. observed futures and vanilla options over some period of time), i.e. we do not fit exactly the market instruments exactly on any specific trading date.

It is important to note that one can assume different mark-to-market risk neutral processes that will match some liquid instruments exactly but will produce different

prices for illiquid derivatives. Also, the model parameters estimated to match say futures curve for a specific trading date will have to change for another trading date due to the change in futures curve. Without entering into further debate, modelling of both the real process and risk neutral process of the underlying are important for real option valuation and we believe that spot models are more appropriate. In this section, after some discussion on spot price and mark-to-market models we present methods that can be used for model diagnostic, model selection and model assessment.

3.1 Mark-to-market models

A simple example of mark-to-market model is geometric Brownian motion with time dependent drift and volatility, i.e. risk neutral process

dSt/St = µ(t)dt + σ (t)dzt, (7)

where zt is the standard Brownian motion, µ(t) is calculated to match futures curve, σ (t) is calculated to match vanilla options. Specifically, for this model, the futures price at t = 0 with maturity T is F (0, T) = S exp(T µ(τ)dτ) and thus

1 d ln(F (0, t)/S)

µ(t) =

= α(µ(t)  ln St)dt + σdzt, (9)

 ln F(0, t) σ

µ(t) = + ln F(0, t) + 4 (1  et). t

This model will match today’s market futures curve exactly because the mean reverting level in spot price is a function of time derived from the futures curve F (0, t). Vanilla prices can be easily calculated using Black-Scholes formula with the variance replaced by σ(1  exp(2α(T  t)))/α because ln ST is normally distributed and thus vanilla options prices can be used easily to estimate σ and α.

Note that time dependent drift µ(t) derived from the current futures curve F (0, t) cannot be used for another trading date due to the change in futures curve between trading dates in real time. Generally speaking, traditional statistical approach cannot be used to validate this type of model because there is no observation/measurement

errors between observed and model predicted prices. However, one can validate hedging strategy, i.e. calculate difference between replication portfolio and instrument (hedging error) using historically observed prices and compare the models using, for example, the mean squared error for the hedging error.

3.2 Spot price models

In general, in spot price models, we assume that spot price St is some function of underlying state variables t that can be observable or not observable. Then we have to assume a stochastic processes for t; these are typically modelled as continuous time Ito processes in real time

dY i

t = µi(t, t)dt + σi(Yt, t)dzi, (10)

where µi(Yt, t) and σi(Yt, t) are the drift and volatility of Y i

t that can be functions

of the underlyings t and time t, and E[dzidzj] = ρijdt. The risk neutral process which is used to value derivatives (i.e. options, futures, etc) driven by Yt is obtained from no-arbitrage considerations. In general, it can be written as

dY i

t = (µi(t, t)  λi(Yt, t)σ(Yt, t))dt + σ(Yt, t)dzi, (11)

where λi(Yt, t) is the risk premium that can be function of Yt and time t, and E[dzi dzj ] = ρijdt. One can consider adding Poisson jumps to the above processes; also stochastic volatility can be one of the unobserved factors.

Two-factor model. The well known Schwartz and Smith (2000) two-factor model for commodity futures assumes that the log spot price of a commodity is ln St = ξt + χt, where χt is unobservable short-term deviation in prices and ξt is an unobservable long¬term equilibrium price level with the following real processes

dχt = κχtdt + σdz,

dξt = µdt + σdz, (12)

E[dWdW] = ρdt.

One can add a seasonality component f (t) so that ln St = ξt + χt + f (t). Then, the corresponding risk neutral process used to value futures and options is

dχt = (κχt  λ)dt + σdz,

dξt = (µ  λ)dt + σdz, (13)

E[dzdz] = ρdt,

where λx and λ~ are the risk premia that typically assumed to be constant but in general can be functions of χt and ξt. If the risk premia are linear functions of state variables, then the price of the future contracts with maturity T is

F0,T = E*[ST] = exp(ξ0 + B(T)χ0 + A(T )), (14)

where expectation is calculated under the risk neutral process. A(T ) and B(T) are simple functions of time. Given that log spot is normally distributed, A(T ) and B(T) functions are easily calculated; for details see Schwartz and Smith (2000).

Remark 3.1 It is important to note that the real process can be mean reverting while risk neutral is not mean reverting and vice versa.

Three factor model. A popular extension of the above two-factor model is adding extra factor. Namely allowing drift of the long term factor to be stochastic itself

dχt = κχtdt + σxdzx,

dξt = µtdt + σ~dz~,

dµt = γ(µ~  µt)dt + σµdzµ,

E[dzxdz~] = ρx~dt, E[dzxdzµ] = ρxµdt, E[dzµdz~] = ρµ~dt. (15)

Risk neutral processes are obtained by including risk premia into the drift terms. If risk premia are linear functions of the state variables then the price of the future contracts with maturity T becomes

F0,T = E*[ST] = exp(BJT)ξ0 + Bx(T)χ0 + Bµ(T)µ0 + A(T )), (16)

where expectation is calculated under the risk neutral process and all functions of time A(T ), Bµ(T), B~(T) and Bx(T) are easily calculated in closed form.

Multi factor affine models. In general, if the stochastic risk neutral model for the underlying variables t = (Y (,)

t , ... , Y (M)

t ) is exponentially affine model, i.e. drifts and

covariances in (11) are linear functions of t, and log spot price is a linear function with respect to t, then the futures price can always be calculated as

FT,t = E*[ST] = exp(B,(T  t)Yt(,) +... + BM(T  t)Y (M)

t + A(T  t)),

where functions of time A(T  t), B,(T  t), ... , BM(T  t) are calculated from the

system of ODEs.

State-space representation. In general, spot price multi factor models can be for¬mulated as a state-space model

t = g(t, t); state equation, (17)

 = h(t, ϵt); space/measurement equation, (18)

where g(•) and h(•) are some functions,  = (Xt,, ... , Xt,n) are observations on the trading date t (e.g. futures, vanillas, etc), t and ϵt are the vectors of serially independent normally distributed errors with zero mean and some covariances. In the case of the above two/three factor models, state and space equations are linear in Yt and in error terms and can be written in the form

t =  + t + t, (19)

ln  = t + t + ϵt, (20)

where  = (Ft,T1, ... , Ft,Tn) are observed futures prices at trading date t;  and ϵt are M dimensional vectors and  is M × M matrix; t and ϵt are n-dimensional vectors and t is n × M matrix. Using Kalman filter procedure, one can calculate the density of the observed data (so-called likelihood) and fit the model using frequentist of Bayesian inference methods as described in the following sections; for application examples, see e.g. Schwartz & Smith (2000), Schwartz (1997). For a detailed discussion of state-space models and Kalman filter, see Harvey (1989). In the case of nonlinear relationships and non-Gaussian errors, one can try nonlinear Kalman filter or particle filter Monte Carlo methods; see Peters et al (2012)

The modeller should choose the model (i.e. the number of factors and model pa¬rameters for the risk neutral and real processes) and fit the model parameters to the observed data. In general, fitting can be done using the frequentist or Bayesian ap¬proaches. Using calibration results for different models, the user can make the model choice based on standard statistical criteria. Note that here, we aim to validate both the real process and risk neutral process of the underlying variables.

3.3 Frequentist approach

Fitting model parameters using data via the frequentist approach is a classical problem described in many textbooks. Under the frequentist approach a modeller says that parameters are fixed while their estimators have associated uncertainties that typically converge to zero when a sample size increases. The most popular approach to fit the parameters of the assumed model is the maximum likelihood method. Given the

model parameters  = (θ, θ,... , θK), assume that the joint density of data X = (X, X, ... , Xn) is f (x|). Then the likelihood function is the joint density f (x|) considered as a function of parameters , formally defined as

ℓ"() = f(x|). (21)

The maximum likelihood estimators MLE = () of the parameters  are the values of these parameters that maximize the log-likelihood function ln ℓ"(). Under the suitable regularity conditions, as the sample size increases, MLEs converge to the true value and are distributed from the K-variate normal distribution NK(•) as

Here, [()] is the inverse matrix of the expected Fisher information matrix whose matrix elements are typically approximated by the observed information matrix

( )km =  1 n  ln ℓ"() θkθm 00-. (23)

That is, standard errors (and covariances between errors) of MLE are estimated by covariance matrix n(). For precise details on regularity conditions and proofs, see Lehmann (1983, Theorem 6.2.1 and 6.2.3); these can also be found in many other books. Though very useful and widely used, these asymptotic approximations are usually not accurate enough for small samples, that is the distribution of parameter errors can be materially different from normal and MLEs may have significant bias.

Typically, maximisation of the likelihood (or minimisation of some distances in other methods) must be done numerically. Popular numerical optimisation algorithms include simplex method, Newton methods, expectation maximisation (EM) algorithm, and simulated annealing.

Example 3.1 Consider geometric Brownian motion real process

dSt/St = µdt + σdWt, (24)

observed at discrete equally spaced times t, t, ... , tn i.e.

ln Si = ln Si + (µ  σ/2)δt + σ δtϵi,

where δt  ti  ti and ϵi are iid from the standard normal distribution. Then the likelihood of return data Ri = ln(Si/Si), i = 1, ..., N is

Maximizing ln ℓ(µ, θ) with respect to µ and σ gives the following MLEs

Example 3.2 As an another example, consider the mean reverting real time process

dSt = (ω  θSt)dt + σdWt,

observed at discrete equally spaced times t, t, ... , tn, i.e.

Si = Siρ + θ (1  ρ) + vϵi, v = 1 2θ(1  ρ).

Then the likelihood of Si, i = 1, ... , N is

ℓ(ω, θ, σ) oc N i 1 v  

 ϵ

exp i .

2v

Maximizing ln ℓ(ω, θ, σ) with respect to (ω, θ, σ) gives the following MLEs

 



Si Si  N SiSi,

i i i

 



Si SiSi  S i Si,

i i i i

det = Si Si  N Si.

i i i



Using the above estimators, calculate ˆv =  i ˆϵi and finally θ ˆ =  ln ˆρ/dt, ω ˆ =

ˆµˆθ /(1  ˆρ), and σ = 2ˆθˆv/(1  ˆρ).

For a general multifactor models the likelihood is not so easy to calculate. However, in the case of state variables from Gaussian distribution and linear measurement equa¬tion, i.e. linear state-space model (19), the likelihood can be calculated using Kalman filter recursion. In particular, using Kalman filter procedure, one can calculate the probability density function of  for given 1, i.e. f (|1). Then the likelihood of all data is

ℓ1: () = T f(Ft|Ft1), (25)

t

where  are models parameters (drift, volatility, correlations and risk premia). Once the likelihood function is calculated, the parameters can be estimated using maximum likelihood method with numerical optimization. One can also use Bayesian approach with Markov chain Monte Carlo (MCMC) methods described in the next section. In the case of nonlinear relationships and non-Gaussian errors, one can try nonlinear Kalman filter or particle filter Monte Carlo methods, see Peters et al (2012), but this goes beyond the purpose of this paper.

3.4 Bayesian approach

There is a broad literature covering Bayesian inference and its applications; for a good introduction, see e.g. Berger (1985). In the Bayesian approach, both data and parameters are considered to be random. A convenient interpretation is to think that parameter is a random variable with some distribution and the true value (which is deterministic but unknown) of the parameter is a realisation of this random variable. Consider a random vector of data X = (X, X, ... , Xn) whose density, for a given vector of parameters 9, is f (x|9). Then the joint density of the data and parameters is

f(x, 9) = f(x|9)π (9) = π (9|x)f(x), (26)

where π (9) is the density of parameters (a so-called prior density); π (9|x) is the density of parameters given data X = x (a so-called posterior density); f(x, 9) is the joint density of the data and parameters; f (x|9) is the density of the data given parameters 9, i.e. it is a likelihood function ℓx(9) = f (x|9); f (x) is the marginal density of X. If π (9) is continuous, then f (x) =  f (x|9)π (9)d9 and if π (9) is a discrete, then the integration should be replaced by a corresponding summation.

Using (26), the posterior density can be calculated as

π (9|x) = f (x|9)π (9)/f (x) a f (x|9)π (9). (27)

Here, f (x) plays the role of a normalisation constant and the posterior can be viewed as a combination of prior knowledge (contained in π (9)) with information from the data (contained in the likelihood f (x|9)). Using the posterior π (9|x), one can easily construct a probability interval for 9 to contain the true value with the required prob¬ability, which is the analogue for confidence intervals under the frequentist approach. Sometimes the posterior density can be calculated in closed form, but in general, one should use Gaussian approximation or MCMC methods.

Gaussian Approximation for Posterior. For a given data realisation X = x,

denote the mode of the posterior π (9|x) by 

9. If the prior is continuous at 9, then a

Gaussian approximation for the posterior is obtained by a second-order Taylor series expansion around 9:

Under this approximation, π (9|x) is a multivariate normal distribution with the mean 9 and covariance matrix

. (29)

e=~

In the case of improper constant priors, this approximation is comparable to the Gaus¬sian approximation for the MLEs (22). Also, note that in the case of constant priors, the mode of the posterior and the MLE are the same. This is also true if the prior is uniform within a bounded region, provided that the MLE is within this region.

Once the posterior density 7r(0 |x) is found, for given data X, one can define point estimators of 0. The mode and mean of the posterior are the most popular point estimators. The median of the posterior is also often used as a point estimator for 0.

Sometimes there is no prior knowledge about the model parameter 9, or we would like to rely on data only and avoid an impact from any subjective information. In this case we need a noninformative prior (sometimes called vague prior) that attempts to represent a near-total absence of prior knowledge. A natural noninformative prior is the uniform density

7r(0 ) oc const for all 0. (30)

If parameter 0 is restricted to a finite set, then this 7r(0 ) corresponds to a proper uniform distribution. However, if the parameter 0 is not restricted, then a constant prior is not a proper density (since  f (0)d0) = oc). Such a prior is called an improper prior. It is not a problem to use improper priors as long as the posterior is a proper distribution.

With respect to multi-factor models, once the likelihood is derived (e.g. the Kalman filter likelihood for linear models (25)), then one can use the Bayesian approach. Pos¬terior cannot be found in closed form but various MCMC methods can be used to get samples from the posterior. The easiest to implement is Metropolis-Hastings algorithm which is a universal algorithm used to generate a Markov chain {9(1), 9(2), ... } with a stationary distribution 7r(9|x). It has been developed by Metropolis et al (1953) in mechanical physics and generalised by Hastings (1970) in a statistical setting. Given a density 7r(9|x), known up to a normalisation constant, and a conditional proposal density q(9*|9), the method generates {9(1), 9(2), ... } using the following algorithm.

1. Initialise 9(l=0) with any value within a support of 7r(9|x);

2. For l = 1, ..., L

(a) Set 9(l) = 9(l1);

(b) Generate a proposal 9* from q(9*|9(l));

 

1, π(9*|x)q(9l|9*)

p(9l, 9*) = min , (31)

π(9l|x)q(9*|9l)

i.e. simulate U from the uniform distribution function U(0, 1) and set 9l = 9* if U < p(9l, 9*). Note that the normalisation constant of the posterior does not contribute here;

3. Next l (i.e. do an increment, l = l + 1, and return to step 2).

3.5 Model diagnostic checking

Once the model parameters are calibrated, the model assumptions should be checked. Typical assumptions include independence and normality of the model residuals. For example, for linear state space model (19), one should check that error terms εt and ϵt are serially independent and are from the standard normal distribution. For this task, the following statistical methods are often used.

• Unit root testing. In general, unit root test is defined for autoregressive process of the order k. Here, for simplicity we consider k = 1, i.e. the model

xt = αxt + δt,

where δt are serially independent normal variables. Unit root testing is testing of null hypothesis that α = 1. If |α| < 1, then time series xt is stationary. If null hypothesis is rejected, then we can estimate α by some fitting procedure such as maximum likelihood. There are many tests for unit root such as Dickey-Fuller test and augmented Dickey Fuller tests. One can also perform Bayesian inference approach and estimate the posterior for α.

• Q-Q plots. It is also common to check quantiles of the sample (y-coordinate) against model assumed quantiles (x-coordinate), this (x, y) plot is referred to as Q-Q plot. If model assumption is correct then the points of the plot should be close to x = y line. For example, if we check that x, ... , xN are from the standard normal distribution, then we plot the sample order statistics x, ... , xN against the quantiles of the standard normal distribution yi = FN((i  0.5)/N), i = 1, ... , N. If the assumption of normality is valid then correlation R between yi and xi should be close to one. Formal percentage points for the R for samples from normal distribution is given in Shapiro and Francia (1972). For example for N = 200, Pr[R < 0.987] = 0.05. One can also apply popular goodness of fit tests including Kolmogorov-Smirnov, Anderson-Darling and chi-square tests.

• Autocorrelation. In addition to visual inspection of the model residuals, it is useful to monitor the serial correlation. For a given sample x, ... , xN, the autocorrelation at lag k is estimated as

Nk

_ACF(k) = 1 

Nσ

i (θi  µ)(θik  µ), (32)

where µ and s are the mean and variance of a sample x, ... , xN. Of course, it is biased estimate because we use sample estimators µ and s, and model parameter point estimators to calculate the residuals, but for large N and consistent point estimators it will converge to the true autocorrelation. It is possible to estimate the variance of the autocorrelations due to the finite sample. For example, if xi are iid, then for large N, ACF(1), ... , ACF(m) are iid normal variables with zero mean and variance 1/N. That is, autocorrelations should be within bounds

±1.96/ N with 0.95 confidence. This is usually used to check if residuals are iid. It is also a good idea to check absolute residuals. Often, returns in financial time series exhibit small autocorrelations while absolute returns have significant autocorrelations. This is typically an indication of time dependent volatility and can be removed by GARCH models for volatility if required depending on time horizon of the model use.

3.6 Model Selection

Given several competing models that passed diagnostic check, the modeller should decide which model to be used. Here, one can use the following procedures depending on the calibration approach selected to fit the model. Typically, under the frequentist approach, the modeller takes likelihood ratio tests and Akaike Information Criterion; under the Bayesian approach, the modeller often calculates the Bayes factors and Deviance Information Criterion. These are briefly described below.

• In-sample errors Once the statistical model parameterized by 9 is fitted to a data sample, we can calculate the difference between model predicted and ob¬served values in the data sample. Typically one calculates the following quantities between observations Xt and predicted values Xpred

t = E[Xt|X, . . ., Xt;9]:

– squared correlation coefficient/coefficient of determination/R-squared (R),

(Corr[Xt, Xpred

t ]); larger value indicates better model;

– mean squared error (MSE), E[(Xt  Xpred

t )]; lower value indicates better

model;



– Root mean square error (RMSE), E[(Xt  Xpred

t )]; lower value indicates

better model;

– mean absolute percentage error, E[|Xt  Xpred

t |/Xt]; lower value indicates

better model.

Here, the model parameters estimated 9 are obtained using the full dataset XT.

• Likelihood Ratio test. It is a statistical test comparing two models. The test statistic for the null model with parameters 9 and alterative model with 9is

 

LR = 2 ln ℓ( 9)/ℓ( 9) , (33)

where ℓ(9) and ℓ(9) are the likelihoods of the models. The distribution of statistic LR is chi-squared distribution with degrees of freedom mm, where mand m are the number of parameters in the null model and alternative model respectively. The models should be nested, i.e. more complex model can be reduced to a simpler model via constraints on the parameters.

• Akaike Information Criterion (AIC). It is a measure of the relative goodness of fit of a statistical model introduced by Akaike (1983):

AIC = 2m  2ℓ(9), (34)

where 9 = (θ, ... , θm) are model parameters, m is the number of parameters, and ℓ(9) is likelihood function for the data maximized at 9. Point estimators 9 are the maximum likelihood estimators. The best model within a set of K candidate models for the data corresponds to the smallest AIC* = min(AIC, ... , AICK). Note that AIC penalises for the increase in the number of parameters while re¬wards for goodness of fit. The quantity exp((AIC*AICk)/2) can be interpreted as the relative likelihood of the ith model. It looks similar to the likelihood ratio test if the number of parameters in the candidate models is the same but note that likelihood ratio test is used for nested models. To account for the number of observations N used to fit the model, the criteria is adjusted as

AIC = 2m  2ℓD( 9) + 2m(m + 1) N  m  1. (35)

• Bayesian Information Criteria. Bayesian information criterion (BIC) or Schwarz criterion is a criterion for model selection among a finite set of mod¬els. It is closely related to AIC and is an asymptotic result for the data from the exponential family distribution. Formally it is given by

BIC = 2 ln ℓD( 9) + m ln N. (36)

The model with the lower value of BIC is the one to be preferred. Note that there is no requirement for compared models to be nested.

• Bayes factors. Consider a model M with parameter vector 9. The model likelihood with data x can be found by integrating out the parameter 9

7r(x|M) = 7r(x|9, M)7r(9|M)d9, (37)

where 7r(9|M) is the prior density of 9 in M. Given a set of K competing models (M1, ... , MK) with parameters 9[1], ... , 9[K] respectively, the Bayesian alterna¬tive to traditional hypothesis testing is to evaluate and compare the posterior probability ratio between the models. Assuming we have some prior knowledge about the model probability 7r(Mi) (if no knowledge is available one can assign equal probabilities to the models), we can compute the posterior probabilities for all models using the model likelihoods

7r(x|Mi) 7r(Mi)

7r(Mi|x) = K

k =1 7r(x|Mk) 7r(Mk). (38)

Consider two competing models M1 and M2, parameterised by 9[1] and 9[2] re-spectively. The choice between the two models can be based on the posterior model probability ratio, given by

7r(M1|x) 7r(M2|x) 7r(x|M1) 7r(M1) 7r(M1) 7r(M2)B12, (39)

7r(y|M2) 7r(M2)

where B12 = 7r(x|M1)/7r(x|M2) is the Bayes factor, the ratio of the posterior odds of model M1 to that of model M2. As shown by Lavine and Scherrish (1999), an accurate interpretation of the Bayes factor is that the ratio B12 captures the change of the odds in favour of model M1 as we move from the prior to the poste¬rior. A Bayes factor B12 > 10 is considered strong evidence in favour of M1. Kass and Raftery (1995) give a detailed review of the Bayes factors. Typically, the integral (37) required by the Bayes factor is not analytically tractable, and sam¬pling based methods must be used to obtain estimates of the model likelihoods. There are quite a few methods in the literature for direct computation of the Bayes factor or indirect construction of the Bayesian model selection criterion, both based on MCMC outputs. The popular methods are direct estimation of

the model likelihood thus the Bayes factor; indirect calculation of an asymptotic approximation as the model selection criterion; and direct computation of the posterior model probabilities, as discussed below. Also, given MCMC samples from the posterior distribution obtained through MCMC, there is a reciprocal importance sampling estimator (RISE) proposed in Gelfand and Dey (1994) to approximate the model likelihood that can be regarded as a generalization of the harmonic mean estimator suggested by Newton and Raftery (1994).

• Direct calculation of model probabilities. Accurate estimation of the re¬quired posterior distributions usually involves development of a Reversible Jump MCMC framework. This type of Markov chain sampler is complicated to develop and analyse. It goes beyond the scope of this paper but interested reader can find details in Green (1995). In the case of small number of models, Congdon (2006) suggests to run a standard MCMC for each model separately and use the obtained MCMC samples to estimate 7r(Mk|). It was adopted in Peters, et al (2009) for modelling claims reserving problem in the insurance. Using the Markov chain results for each model, in the case of equiprobable nested models, this procedure calculates the posterior model probabilities 7r(Mi|) as

L

7r(Mi|) = L

l  

f |Mi, l

i

K  , (40)

j f |Mj, l

j

where l

i is the MCMC posterior sample at Markov chain step l for model Mi,

f (|Mi, l

i ) is the joint density of the data  given the parameter vector l

i for

model Mi, and L is the total number of MCMC steps after burn-in period.

• Deviance Information Criterion. For a dataset  =  generated by the model with the posterior density 7r(|x), define the deviance

D() = 2 ln 7r(x|) + C, (41) where the constant C is common to all candidate models. Then the deviance information criterion (DIC) is calculated as

DIC = 2E[D()|X = x]  D(E[|X = x])

the data; the smaller this is, the better the fit. The difference E[D(9)| = ]  D(E[9| = ]) can be regarded as the effective number of parameters. The larger this difference, the easier it is for the model to fit the data. The DIC criterion favours the model with a better fit but at the same time penalises the model with more parameters. Under this setting the model with the smallest DIC value is the preferred model. DIC is a Bayesian alternative to Akaike’s information criterion. For more details on the above-mentioned criteria, see e.g. Robert (2001, Chapter 7).

• Model averaging. If competing models have significant probabilities, then the modeller may choose to average the results across the models (weighted by model probabilities) instead of selecting the best model. Here, one can use the model probabilities implied by the above discussed AIC or Bayes factors.

3.7 Model assessment

Once the best model is selected using the above described statistical criteria, the mod¬eller should perform the assessment of the selected model. Typically it involves esti¬mating the prediction error on new data.

• Out-of-sample or cross validation. Cross-validation is a technique for as¬sessing the accuracy of the prediction of the fitted statistical model. It involves partitioning a sample of data into complementary subsets, performing the anal¬ysis on one subset and validating the analysis on the other subset. To reduce variability, multiple rounds of cross-validation are performed using different par¬titions, and the validation results are averaged over the rounds. Often, one step prediction are calculated. Typically one calculates the following quantities be

tween observations Xt and predicted values Xpred t = E[Xt|X, ... , Xt]: – squared correlation or pseudo R: (Corr[Xt, Xt pred]); – mean quadratic error: E[(Xt  Xt pred)];

– mean absolute percentage error: E[|Xt  Xpred

t |/Xt].

Here, we assume that the model is parameterized by 9 estimated using one data subset and the above prediction errors are calculated for another data subset.

• Hedging errors. Given that risk neutral process used for the pricing of real option is based on assumption of replication portfolio, it is worthwhile to check the hedging errors (difference between the option/future and replicated portfolio) for observed data realization. This is especially critical for mark-to-market models.

• Sensitivity. The modeller should check the sensitivity of the model output (i.e. price of real option) to the inputs. For example, it is typical to observe that for a long investment project, short-term factor is not important and pricing can be done using a model for a long term factor only.

4 Empirical Applications

The data we use to test the models are weekly observations of futures prices of crude oil, contract CL traded on the New York Mercantile Exchange (NYMEX), from 23 November 1990 till 10 May 2013. For illustration example we use the first 12 contracts available for each observation date although for most of the dates there are about 20 contracts and for some dates there are more than 70 contracts. We fit the following two models (one-factor and two-factor models) for the log spot price of a commodity ln St.

• Model 1 (Two-factor model). Assume that ln St = ξt + χt, where χt is unobservable short-term deviation in prices and ξt is an unobservable long-term equilibrium price level with the following real processes

dχt = βχtdt + σdz,

dξt = (µ  γξt)dt + σdz, (43)

E[dWdW] = ρdt.

Corresponding risk neutral processes (used to value futures and options) are

dχt = (βχt  λ)dt + σdzx,

dξt = (µ  γξt)dt + σdz~, (44)

E[dzxdz~] = ρdt.

This is just a well known Schwartz and Smith (2000) two-factor model extended to have mean reversion in a long term factor and more general (linear in state variables) risk premia. Under this risk neutral process, the price of the future contract at time t with maturity at time T can be easily calculated as

Ft,T = E*[ST] = exp(B(T  t)ξt + B(T  t)χt + A(T  t)); (45)

B(τ) = exp(γτ); B(τ) = exp( βτ);

σσρ

+ (1  e

γ + β

• Model 2 (One-factor model). Here we assume that there is only one-factor, i.e. ln St = ξt, that follows real and risk neutral processes

dξt = (µ  γξt)dt + σdz, (46)

dξt = (µ  %)dt + σdz~,

correspondingly. Then, the price of the futures contract is

Ft,T = E*[ST] = exp(B(T  t)ξt + A(T  t)); (47)

that can be obtained from (45) by setting χt = λ = β~ = σ = ρ = 0.

A statistically sound method to fit the above models is the Kalman filter procedure typically used in the academic literature. In this paper, we choose more simple, fast and easy to implement procedure suggested in Cortazar and Schwartz (2003). In the case of crude oil futures, this procedure produces results that are not materially different from the Kalman filter results. Of course the advantage of the Kalman procedure is that it calculates the model likelihood which is used to get the point estimates of the parameters, confidence intervals for the estimates and can be used to apply formal model selection criteria such as Akaike criteria. Under the simplified procedure, we have to resort to in-sample and out-of-sample tests. The fitting procedure estimates risk neutral mean reversion parameters (%γ, ~), volatilities (σ, σ) and correlation ρ using historical covariances between futures contracts via nonlinear least square method. Then, risk neutral drift parameters (λ, µ) are estimated by nonlinear least square method minimizing

[ln F(ti, Tj)  ln Fobs(ti, Tj)]

i j

where unobservable factors ξti and χti for each trading date ti are calculated in closed

form by least square method minimizing,

[ln F(ti, Tj)  ln Fobs(ti, Tj)].

Finally, obtained time series for ξt and χt are used to estimate real process parameters (µ, γ, β, σ, σ, ρ) by maximum likelihood method for mean reversion process; see Ex¬ample 3.2. The calibration for the one-factor model (Model 2) is easily obtained from this procedure by setting χt = λ =β =β = σ = ρ = 0. Calibration results are summarized in Table 1 and Table 2. The percentage root mean square error (RMSE) between model and market log prices (across different contracts and total across all

contracts and trading dates) is significantly smaller for Model 1 indicating that two-factor model is much superior than one-factor model. Other model selection statistics (R and AIC) are also in favour of Model 1; see Table 1. Figure 1 shows prediction error for some contracts, estimated long and short factors, and predicted value for the 12th contract (F12) versus time.

Model 1 (Two-Factor Model)

= 6.06, µ = 7.62, y = 1.69, ug = 0.31, ux = 0.20, p = 0.26, µ = 1.04, = 6.73, y = 0.23

RMSE=0.45%, R2=0.999, AIC=-25175

Model 2 (One-Factor Model)

µ = 7.27, y = 1.62, ug = 0.29, µ = 0.69, y = 0.15

RMSE=1.65%, R2=0.99, AIC=-18509

Table 1: Parameter estimates for two-factor model (Model 1) and one-factor model (Model 2). RMSE is the percentage root mean square error between model and market log prices across all contracts and all trading dates in the dataset.

F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12

Model 1

0.6% 0.7% 0.5% 0.4% 0.4% 0.3% 0.2% 0.2% 0.2% 0.3% 0.5% 0.6%

Model 2 4.2% 1.8% 0.8% 0.6% 0.7% 0.8% 0.8% 0.9% 1.0% 1.2% 1.4% 1.7%

Table 2: The percentage root mean square error (RMSE) between model and market log prices across different contracts. F1 corresponds to the 1st available contract, F2 to the 2nd available contract, etc.

5 Conclusions

We considered the use of multiple factor models for valuation of real options. The choice of underlying stochastic model is certainly important for valuation of real options especially for projects with long lifetime. There are many statistical tools that can help to resolve this issue. In this paper, we reviewed and discussed methods of model selection, model assessment and model diagnostics.

Prediction error

Figure 1: Fitting results for two-factor model (Model 1).

References

[1] AKAIKE, H. Information measure and model selection. Bull. Int. Statist. Inst. 50 (1983), 277–290.

[2] BERGER, J. O. Statistical Decision Theory and Bayesian Analysis, 2nd ed. Springer, New York, 1985.

[3] BHATTACHARYA, S. Project valuation with mean-reverting cash flow streams. Journal of Finance 33, 5 (1978), 1317–1331.

[4] BLACK, F., AND SCHOLES, M. The pricing of options and corporate liabilities. Journal of Political Economy 81, May-June (1973), 637–659.

[5] BRENNAN, M. J., AND SCHWARTZ, E. S. Evaluating natural resource invest¬ments. The Journal of Business 58, 2 (1985), 135–157.

[6] CLEWLOW, L., AND STRICKLAND, C. Valuing energy options in a one factor model fitted to forward prices. Preprint (1999).

[7] CONGDON, P. Bayesian Statistical Modelling, 2nd ed. John Wiley & Sons, Ltd, 2006.

[8] CORTAZAR, G., AND SCHWARTZ, E. S. Implementing a stochastic model for oil futures prices. Energy Economics 25 (2003), 215238.

[9] COX, J., ROSS, S., AND RUBINSTEIN, M. Option pricing: A simplified approach. Journal of Financial Economics 7 (1979), 229–264.

[10] DIAS, M. A. J., AND ROCHA, K. Petroleum concessions with extendible options using mean reversion with jumps to model oil prices. Paper presented at the 3rd Real Option Conference.

[11] DIXIT, A., AND PINDYCK, R. Investment under uncertainty. Princeton Univer¬sity Press, Princeton, NJ, 1994.

[12] GELFAND, A., AND DEY, D. Bayesian model choice: Asymptotic and exact calculations. Journal of the Royal Statistical Society, series B 56 (1994), 501–514.

[13] GIBSON, R., AND SCHWARTZ, E. S. Stochastic convenience yield and the pricing of oil contingent claims. Journal of Finance 45, 3 (1990), 959–976.

[14] GREEN, P. Reversible jump MCMC computation and Bayesian model determi¬nation. Biometrika 82 (1995), 711–732.

[15] HARVEY, A. C. Forecasting, structural time series models, and the Kalman filter. Cambridge University Press, 1989.

[16] HASTINGS, W. Monte Carlo sampling methods using Markov chains and their applications. Biometrika 57 (1970), 97–109.

[17] LAVIN, M., AND SCHERRISH, M. Bayes factors: what they are and what they are not. The American Statistician 53 (1999), 119–122.

[18] LEHMANN, E. L. Theory of Point Estimation. Wiley & Sons, New York, 1983.

[19] LONGSTAFF, F. A., AND SCHWARTZ, E. S. Valuing American options by sim¬ulation: a simple least-square approach. The Review of Financial Studies 14, 1 (2001), 113–147.

[20] MCDONALD, R., AND SIEGEL, D. Investment and the valuation of firms when there is an option to shut down. International Economic Review June (1985), 331–49.

[21] MCDONALD, R., AND SIEGEL, D. The value of waiting to invest. The Quartely Journal of Economics 101 (1986), 707–728.

[22] MERTON, R. C. Option pricing when underlying stock returns are discontinuous. Journal of Financial Economics 3 (1976), 125–144.

[23] METROPOLIS, N., ROSENBLUTH, A. W., ROSENBLUTH, M. N., TELLER, A. H., AND TELLER, E. Equations of state calculations by fast computing machines. Journal of Chemical Physics 21 (1953), 1087–1091.

[24] NEWTON, M., AND RAFTERY, A. Approximate Bayesian inference by the weighted likelihood bootstrap. Journal of the Royal Statistical Society, series B 56 (1994), 1–48.

[25] OZORIO, L. M., BASTIAN-PINTO, C. L., BAYDIA, T. N., AND BRAND˜AO, L. E. T. Revers˜ao a` m´edia com tendˆencia e op¸c˜oes reais na siderurgia. Revista Brasileira de Finan¸cas 10 (2012), 1–27.

[26] PADDOCK, J. L., SIEGEL, D. R., AND SMITH, J. L. Option valuation of claims on real assets: The case of offshore petroleum leases. Quarterly Journal of Economics 103 (1988), 479–508.

[27] PETERS, G. W., BRIERS, M., SHEVCHENKO, P., AND DOUCET, A. Calibration and filtering for multi factor commodity models with seasonality: Incorporating

panel data from futures contracts. Methodology and Computing in Applied Prob-ability (2012). DOI 10.1007/s11009-012-9286-7.

[28] PETERS, G. W., SHEVCHENKO, P. V., AND W¨UTHRICH, M. V. Model un¬certainty in claims reserving within Tweedie’s compound Poisson models. ASTIN Bulletin 39, 1 (2009), 1–33.

[29] PINDYCK, R. S. The long-run evolution of energy prices. Energy Journal 20, 2 (1999), 1–27.

[30] ROBERT, C. P. The Bayesian Choice. Springer, New York, 2001.

[31] SCHWARTZ, E., AND SMITH, J. E. Short-term variations and long-term dynamics in commodity prices. Management Science 46,, 7 (2000), 893–911.

[32] SCHWARTZ, E. S. The stochastic behavior of commodity prices: Implications for valuation and hedging. Journal of Finance 52, 3 (1997), 923–973.

[33] SICK, G., AND GAMBA, A. Some important issues involving real options: An overview. Preprint (2005).

Jasa Konsultasi Disertasi Keuangan (1) dan Perbankan S3, hubungi WA 0821 2230 7021

Selasa, 09 Agustus 2022

Penelitian doktoral bidang knowledge management: Retrieval and learning of design knowledge

Tidak ada komentar:

Posting Komentar