Reviewed by
Victor Korobitsin
Computer Science Department
Omsk State University
Omsk, Russia.
The book is composed of eight chapters. The first one focuses on applying knowledge-based technologies in describing and developing multi-agent systems. The second chapter focuses on the state of the art, with an analysis of different agent models and architectures. Architectures and approaches are classified into hierarchical and layered, reactive and real-time, heterogeneous, co-operative and social. The third part focuses on knowledge engineering. Important milestones and research projects are discussed, starting with Newell's knowledge level hypothesis from 1982. Description of the CommonKADS approach and related projects also constitute a major part of this chapter. The fourth part focuses on describing the CoMoMAS methodology and on setting out the generic models that are proposed to describe multi-agent systems at a conceptual and implementation-independent level. The fifth chapter illustrates the knowledge engineering environment that has been developed for the validation of the CoMoMAS methodology. The sixth chapter addresses the question of validating the conceptual agent model at an operational level. The seventh chapter presents a case study and details the reverse engineering of the architecture for a mobile robot. The eighth chapter summarises the advantages, unresolved problems and possible applications of the conceptual modelling methodologies suggested by the authors.
The general background for the book is presented. The perspectives of knowledge modelling, knowledge representation and knowledge acquisition are discussed. Finally, a book outline is offered.
This chapter presents a short overview of important issues in the domain of multi-agent systems with a focus on European research projects. The author compares the following architectures: ARCHON, ATLANTIS, ATOME, CIRCA, IMAGINE, InteRRap, PALADIN and REAKT. His purpose is to illustrate the core ideas that have motivated the design of a huge number of agent architectures and agent models. A set of representative multi-agent approaches illustrates different research topics which range from agent communication, co-ordination and negotiation, via distributed and centralised control, to reactive and cognitive architectures, heterogeneous systems and layered agent models.
This chapter presents an overview of historic milestones and recent research projects in the knowledge engineering domain. Newell's knowledge level hypothesis (Newell 1982) was the corner stone for several research projects which described systems on a conceptual and abstract level. This view has the advantage of generic and reusable knowledge structures for system description. Nevertheless, researchers are then confronted with a question about how to operationalise these structures and the models built from them.
The approaches proposed by the PROTEGE (Musen 1989) and MACAO (Aussenac-Gilles and Matta 1994) systems, for example, argue for the automatic creation of a skeleton for an architecture, which has to be refined and updated by the knowledge engineer or domain expert. The MIKE approach uses a formal language for knowledge representation that has restricted expressive power but is operational. Designers must make a choice between the expressive power of a representation language and the possibility of proving its correctness and consistency. A semi-formal representation like the conceptual modelling languages of KADS (Schreiber et al. 1993) is better comprehended by domain experts and contains more explanatory information but its verification is significantly more complex.
The development of a knowledge-based system is illustrated by several projects and theoretical discussions as a constructive process, which builds a conceptual description of the system using model templates, rewrite rules, a generalised directive model and model libraries. Thereby, knowledge acquisition is a cyclic process, which is closely linked to the previous constructive design process: new knowledge is acquired and integrated into the current model which is evaluated and refined and thus new knowledge is required.
The author chooses to apply the CommonKADS model set to the domain of distributed problem solving rather than to a multi-agent system.
This chapter presents the set of conceptual models for the description of multi-agent systems at the knowledge level already discussed.
The agent has a central position in the system and is endowed with knowledge and competence. Social aspects of agents are represented in a system model, which describes the multi-agent society, i.e. architecture and organisation. The identification of the agents is based on the task model, which identifies and describes the expertise tasks to be solved by the system; it also includes the development tasks describing important evolutionary processes of interaction. The expertise and co-operation models provide a view of planning and reasoning competence for the agent as well as reactive and co-operative skills. The design model describes (as in CommonKADS) the design requirements and decisions important for implementation.
The author describes the model based approach for constructing knowledge-based systems. In particular, the model set for multi-agent system development is described. The model is presented (under the conceptual modelling approach) in terms of model content, model rationale and model states. The terminology for domain knowledge modelling (as introduced by the conceptual modelling language) is illustrated. A few examples that illustrate the CML specification for describing the agents, tasks and methods are provided. The classification of agents by their individual and social abilities and knowledge is discussed. The choices of components for an agent model based on this classification are also illustrated in a diagram and described. The model rationale is presented in terms of an expertise model that includes the task knowledge, problem-solving knowledge and reactive knowledge. The main features of the task problem-solving method are provided.
In addition, the task model, co-operation model, system model and design model are described. Formal definitions are given for the main components of multi-agent models. For example
DEFINITION 3 (SOCIETY) A society S is a set of roles which identify the position which individual agents ai ∈ W can play, S={r1,...,rn}. The set A represents the agents being part of the society.
A distinctive feature of multi-agent modelling is the reuse of knowledge during development. The elements of a model library are thus a generic model, a modelling component and modelling operators. The organisation of these model libraries in CoMoMAS is shown.
Model-driven knowledge acquisition is presented in terms of two concepts: local and global. The local acquisition cycle is inspired by the knowledge acquisition cycle of the ACKnowledge project.
The global acquisition cycle is defined and described. This chapter outlines the steps in developing a complete set of models and representing the conceptual description of a multi-agent system. The agent is determined by its social, co-operative, reactive and problem-solving competence. Agents are organised into a society and have certain roles within it. This organisation or architecture is modelled in the system model. Initial agent models are based on the results from competence analysis and functional analysis. These models are refined after identifying the organisational structure and co-operation knowledge/competence.
In addition, the problem-solving process within a typical agent is illustrated. The process includes matching, selection, application and evaluation steps.
Finally, a comparison of CoMoMAS with MAS engineering methodologies is presented. The comparison is carried out with KSM methodology, DESIRE, MAS-CommonKADS, BDI agent models (as a sort of object-oriented methodology), Agent UML (Unified Modelling Language), MaSE and GAIA.
Chapter 5 is devoted to the CoMoMAS knowledge engineering toolkit that allows a user to create a set of conceptual agent models. The CoMoMAS knowledge engineering environment is composed of four different modules. The first modules acquires knowledge, the second one models and represents this knowledge, the third one allows the knowledge engineer to build conceptual models and the fourth one transforms conceptual agent models into a programming language. An additional module has been implemented for the validation of the agent architectures in simulation (Figure 1).
The CoMoMAS Acquire and Modeller modules are realised by the KADSTOOL toolkit. KADSTOOL is a knowledge engineering environment for the construction of expertise models following the CommonKADS approach.
The CoMoMAS Constructor module builds the conceptual model set based on the knowledge delivered by the CoMoMAS Modeller module in the form of knowledge kernels. The Constructor module covers the operations for model composition, which were defined in the local and global knowledge acquisition cycles.
The CoMoMAS Coder module has the task of translating the conceptual descriptions into a programming language and also validates the conceptual agent models. The Coder module uses and maintains a library of functions which are linked to the inference steps within the inference structures describing the competence of an agent.
All modules have the graphical user interface developed using the GINA interface builder.
The validation of conceptual agent models, once they have been transformed into program constructs, is realised within the simulation environment MICE, the Michigan Intelligent Co-ordination Experiment. The original version of MICE has been extended and adapted to fit the particular requirements of validating different types of conceptual agent models.
The extended MICE test bed provides a graphical user-interface with functions for a simplified analysis and a comprehensive evaluation of experiments. The extension of MICE incorporates two categories of agent: reactive and cognitive.
The MICE has been integrated within the Nomad200 simulation toolkit that supplies the simulation of the Nomad200 robot.
Chapter 7 contains a case study describing a multi-agent system in terms of conceptual models. The author shows how these models are constructed and subsequently transformed into executable code. The chapter illustrates this by way of two examples. The first example presents the adaptive intelligent systems model AIbot, which was developed in the research group of B. Hayes-Roth at the University of Stanford. The second example illustrates the CoNomad agent model, which has been developed within the research group of Jean-Paul Haton at the INRIA/LORIA research lab during the CoMoMAS project. Both agent models have been developed for the control of a Nomad200 mobile robot.
A summary and discussion complete the book.
This book is a good reference for familiarisation with the conceptual approach to designing multi-agent systems. The author considers the different architectures of agent models and knowledge acquisition systems. The comparison of the CoMoMAS engineering environment with other similar systems is a good addition to the main contents. Describing the CoMoMAS methodology is the most important part of the book. It contains the description of a formalised approach for the model development process. This part of the book will be very interesting for researchers in Artificial Intelligence as a comprehensive review. The description of the CoMoMAS engineering environment will also be useful for practical engineers. Enough detailed description of the architecture of the environment is provided to serve as a suitable guide for the process of developing models using CoMoMAS. The case study supplied in the book is a simple and useful way of studying the conceptual modelling approach in action. The book will thus be useful for students as well.
ReferencesAUSSENAC-GILLES, N. and N. Matta 1994. Making a method of problem solving explicit with MACAO. International Journal of Human Computer Studies, 40:193-219.
MUSEN, M. 1989. Automated Generation of Model-Based Knowledge Acquisition Tools. Morgan-Kaufmann, San Mateo, CA.
NEWELL, A. 1982. The knowledge level. Artificial Intelligence , 18:87-127.
SCHREIBER, G., J. Breuker, B. Biedeweg and B. Wielinga, editors, 1993. KADS: A Principled Approach to Knowledge-Based System Development. Academic Press, London.
Return to Contents
of this issue
© Copyright Journal of Artificial Societies and Social Simulation, 2005