Multi-agent systems show considerable potential for applications in many fields, including information retrieval, Web assistants, and information trading. In these domains, the application can be structured so that the agents only read data, and therefore agent interaction does not create synchronization or robustness problems.

If multi-agent systems are to be applied to enterprise resource planning problems or production planning and control problems, however, then robustness and synchronization become of paramount importance. A multi-agent production planning and control system that cannot cope with the breakdown of production machinery is that much less useful. This book describes how one can construct robust multi-agent systems in domains such as production planning and control.

What it means for such a system to be robust is that the system as a whole executes its actions correctly under normal operations, and that in the event of a disturbance, it reaches a well-defined state. This is achieved by the introduction of an underlying middleware that provides formal guarantees of robustness, independently of the original structure of the multi-agent system. Indeed, only an approach of this kind can work: the internal workings of some of the agents may be inaccessible to the programmer; for example, transaction logs for database accesses may not be available.

Nagi describes his approach to robustness in the context of a production planning and control system. The basis for the middleware is the interpolation of execution agents between the planning agents and the agent transaction managers, which are responsible for the actual execution of a transaction on the distributed data at one location. The transaction agent is responsible for such tasks as recovering from an aborted transaction and requesting a new plan made necessary by the failure of a local agent.

The book contains 11 chapters. The introduction defines the scope of the work described, and the organization of the rest of the book. Chapter 2 presents the application scenario, and discusses the sources from which the disturbances to be handled by the system may arise. Chapter 3 is an overview of agents and planning. This includes a survey of existing standardized agent infrastructures. Chapter 4 addresses transaction processing, with particular emphasis on recovery, and what can be expected in the way of recovery. Chapter 5 describes the transactional agent approach and the proposed system architecture. Chapter 6 describes the fundamental data structure: the global precedence graph that serves as the basis for recovery. This graph is obtained as the union of the precedence graphs for the individual transactions, along with a collection of event, condition, and action (ECA) rules that relate actions at one node to actions at another. These ECA rules provide for communication among the corresponding execution agents, thus modeling the global dependencies.

Chapter 7 considers mechanisms that guarantee robustness via cycle detection and logging within the execution agent middleware. Chapter 8 describes the interaction of the executing agent with the planning and transaction agents. Chapter 9 discusses a model used to simulate the system; the results obtained in the simulation studies appear in chapter 10. The work is summarized in chapter 11, along with a discussion of future work. An appendix illustrates a prototype for an implementation of the system.

As the author himself says, the system is too complex to support a clean mathematical description. The result is that at times, the intertwining of the logical development and the example of a production planning and control system get in the way of clarity of description, and the reader needs to step back a moment to engage with the author’s ideas. For example, the functions of the execution agent are described in the course of several pages, rather than in the text including a preparatory listing of functions that would help to delineate the scope of the execution agents. Also, because the system is described almost exclusively in the context of a manufacturing system, it is hard to determine just how widely the middleware might be applicable. The book provides an interesting approach to robustness, while making it clear that much work remains to be done in this area.