Distributed artificial intelligence (DAI) is a subfield of artificial intelligence that deals with concurrency. Multi-agent systems are a specialty within DAI that deal with concurrency at the highest levels, presupposing a community of computational agents that can coordinate their efforts to solve many problems in a domain.
Jennings reports on experiences of such a system in the realm of industrial problems, which includes process control and fault diagnosis in power distribution systems. He considers how such systems support a rich behavioral repertoire, the unexpected changes in the environment, and the integration of preexisting computer systems under cost constraints that preclude re-implementation.
The book focuses on a set of preexisting expert systems for diagnosing electrical distribution systems. Before the implementation of the multi-agent approach, human operators had to move information back and forth manually among the several available expert systems. Now a wrapper around each of the expert systems can convert it into an agent that can interact with colleagues directly and without human intervention.
Chapter 1, “Introduction,” provides the background information for both DAI and industrial applications. Chapter 2, “Initial Experiences with Industrial Multi-agent Systems,” describes a system called GRATE that was constructed to deal with such applications. GRATE reveals how a common set of wrappers can be applied to different domains, thus reducing the effort necessary to integrate preexisting expert systems within each domain. However, in a few cases, the community integrated through GRATE does not act purposefully and coherently. Chapter 3, “Representing Joint Actions,” traces these shortcomings to a lack of any explicit representation of joint actions by the individual GRATE agents. Thus they cannot reason adequately about their collaboration with one another. Jennings reviews several research models for representing joint actions, and subsequently extends the work of Phil Cohen and his colleagues to produce the joint responsibility model.
Chapter 4, “Implementing Responsible Agents,” describes a system called GRATE* that includes an implementation of the joint responsibility model and offers an empirical assessment that compares GRATE* with a model of implicit cooperation based on GRATE and a model of selfish problem solvers that deliberately refrain from any coordinating activity. Chapter 5, “Conclusions and Future Directions,” suggests the three directions in which the impact of this research can go: enhancements to GRATE*; extensions to the formal model of joint responsibility; and recognition of cooperation knowledge as a qualitatively new level of computer system description, beyond the previously recognized levels of the implementation platform, logical formalisms, and the knowledge held by an individual agent.
Jennings persuasively shows how progress in basic scientific issues (namely, the notion of a cooperative knowledge level as a previously unrecognized layer of computational capability) can be commercially relevant.
The book will be required reading for both the DAI community and forward-thinking industrial engineers. Two minor qualifications should be kept in mind, however. First, while GRATE and GRATE* are considerably more industrial than most DAI implementations, this volume leaves some important questions unanswered. What platform supports the application? Many environments in use in academic research would not be entrusted to control an industrial plant. What are the real-time characteristics of the system? Both expected and guaranteed response times are critical in assessing whether a system can meet the needs of a given problem. How can the system be validated? Users want assurance that unexpected system behavior will not result in personal or property damage. How difficult is the system to maintain? Information systems frequently become paralyzed when they cannot be kept synchronized with physical changes in the environment.
Second, GRATE and GRATE* are fairly coarse-grained agent architectures with small populations that divide the problem into agents on the basis of preexisting expert systems. The architecture of GRATE and GRATE* fits well into the preexisting systems that they support, but it may be too complex for an application decomposed around physical entities rather than abstract functions. This distinction is particularly important for the performance experiments in chapter 4. The category of selfish problem solvers included in these experiments responds to various proposals for agents whose behavior is driven by a local value function maintained by the agent itself. The agents’ selfishness motivates them to couple perceptively to dissipational fields that their actions generate in the environment, and thus actually leads to new modes of communication that unselfish agents would not manifest.
Even with these two fairly technical qualifications, Jennings’s book is an excellent report on an important research effort that deserves wide circulation. It includes two brief appendices (a short introduction to the formal languages in which the joint responsibility model is constructed, and a listing of the generic rules that implement the model in GRATE*), a brief index, and a bibliography of over 190 books and articles extending to 1994, the year of publication.
