Constraint-based reasoning by E. C. Freuder and A. K. Mackworth (Eds), MIT Press, Cambridge, MA, 1994, pp 403, £31.50/$47.25, ISBN 0-262-56075-5

exchange and multi-agent planning. The need for communication and interaction in cooperative multi-agent communities is identified. Two examples of industrial applications are then introduced for use throughout the book: particle acceleration and electricity transportation. Chapter 2 describes an initial agent architecture called GRATE, in which generic layers of knowledge for cooperation, situation assessment and control are incorporated. GRATE is interfaced to domain-specific systems via the control module. The situation assessment module decides whether activities are local or social, whether to honour a requested social activity and how to respond to new information. The cooperation module is only responsible for managing the agent's social activities and interacting with other agents. The GRATE architecture is then applied to the two industrial applications given in Chapter 1. The particle accelerator example is particularly interesting in that two real, stand alone expert systems are transformed into cooperating agents under the control of GRATE with few modifications to the expert systems and none to GRATE itself. Simulated domain systems are used in the electricity transportation application. Improvements in various aspects are reported in both cases. Chapter 3 explains a theoretical model for joint social actions, called the Joint Responsibility model. A novel mechanism for coordination is first introduced based on the notions of commitments and conventions. Conventions are used for monitoring agent commitments, which may be withdrawn in a dynamic environment. These conventions define the conditions under which comments may be dropped and what actions to take in such circumstances. A formal language based on modal and dynamic logic is then described for the formal representation of beliefs, goals, actions and commitments from both individual and social points of view. A mental state of Joint Responsibility is finally proposed for coordinated actions together with an associated social convention, known as the Joint Responsibility Social Convention. The use of the Joint Responsibility model is demonstrated in the electricity transportation example. Chapter 4 presents a computational architecture for the Joint Responsibility model. A specific realisation of the architecture is devised in a refined version of GRATE, called GRATE*. GRATE* overcomes the shortcomings of GRATE in that it can deal with unpredictable and dynamic changes in an agent community since it incorporates the Joint Responsibility Social Convention. GRATE* is finally evaluated against the original GRATE in terms of coherence, computational cost and communication cost. Chapter 5 gives a summery of the book's contributions to the DAI community, discusses desirable enhancements to the GRATE systems and the Joint Responsibility model, and expresses a vision that the next generation of DAI systems shall have an explicit knowledge level for cooperation. The book is extremely well-written in its clarity and structure. It offers a substantial amount of theoretically and practically valuable ideas, concepts and techniques for use in multi-agent systems. In addition, it covers general DAI material very well. Therefore the book is highly recommendable to DAI researchers, practitioners, as well as readers who have a general interest in multi-agent systems.