Browse Books

Reviewer: H. Van Dyke Parunak

Driven by the growth of the Internet and the increased number of computerized information applications, software agents have become an essential component of many real-world systems. Some agents (such as Web searching agents) are designed to function in isolation, but many systems require the interoperation of different agents. Such interoperation can be achieved by constructing agents specifically for the purpose of working together, following rigidly defined standards, but this approach requires the reengineering of existing information system components that are otherwise completely adequate for their intended functions. The authors take a different approach. They encourage the development of multi-agent systems by integrating new purpose-built agents with existing code that has been agentified by providing it with a wrapper so that it can communicate with its peers. For them, any body of code that provides useful functions can serve as an agent. The main burden of the book is to develop a formal mathematical language for describing and reasoning about such a heterogeneous society of agents, with the objective of providing system engineers with the tools they need to bring such agent communities to life. Chapter 1 motivates agent architectures in general by describing three applications that serve as running examples throughout the book: a personalized department store application that enables a Web site to interact adaptively and proactively to customers, an autopilot application for controlling aircraft and avoiding controlled flight into terrain, and a supply chain example. It outlines ten desiderata for agent systems and provides an overview of the book. Chapter 2 describes the authors’ interactive Maryland platform for agents collaborating together (IMPACT) architecture, which includes both agents (with the wrappers that enable legacy systems to join the community) and servers that provide infrastructure services. IMPACT’s fundamental view of an agent is as a program providing certain computational services. Chapter 3 is an attempt to provide an open service description language that would enable agents developed by different people to communicate with one another about the services they offer to or request from one another. The language is based on semantic hierarchies of nouns and verbs, and the chapter devotes a great deal of attention to the notion that the semantic distance between two words can be estimated by their separation in such a hierarchy. The authors recognize the importance of an open service description language by discussing it early in the book, but their proposed solution leaves much to be desired. The problem of deciding whether one agent’s reference to “cup” is close to another agent’s reference to “mug” is far more complex than locating these strings in a tree. A far greater problem, not addressed at all, is the mapping between either string and an abstract concept. Those experienced with such real-world examples as the databases of parts used in automotive companies know that even among different divisions of the same company, one not only has different labels for the same part, but also different parts referred to in different places by the same name. The construction of a formal logical system for describing agents begins in chapter 4 with methods for describing legacy data and software. Any body of code can be described as a triple, specifying the set of all data types maintained by the code, a set of predefined functions that give external processes access to the data objects managed by the code, and a set of type composition operations that can generate new types. This definition permits specification of an agent’s state, and of a code call atom, a construct that can express any access to an agent in terms of a Boolean predicate. This apparatus enables the definition of integrity constraints and the linking of service descriptions and code calls. Chapter 5 describes the IMPACT server and the protocol by which users communicate with it. IMPACT servers manage the service description language hierarchies, provide yellow pages and white pages look-up, and help manage connections, among other functions. Chapter 6 develops the logical formalism to define an agent program, which is based on a modal logic drawing from previous work in deontic and praxiological logic. The resulting language permits the specification of constraints on an agent’s actions. A major contribution of this chapter is the definition of a range of agent semantics, each defining different relations among an agent’s program, its state, action constraints, and underlying integrity constraints on its state. Chapters 7 to 10 extend this language to permit agents to reason about one another (chapter 7), specification of temporal aspects of actions (chapter 8), reasoning about uncertainty in their knowledge of the world and of each other (chapter 9), and provision of security (chapter 10). Chapter 11 derives basic complexity results for the reasoning that agents do as a function of their underlying semantics. In the light of these complexity results, implementers will want to ensure that common agent operations are efficiently computable, and chapter 12 describes a class of agent programs that satisfy this requirement, and discusses how the IMPACT development environment supports constructing such agents. Chapter 13 outlines an example application in the domain of logistics. In general, the formalism is intended to be applied in constructing the wrapper that enables a legacy program to participate in an agent-based system. An important question is how much effort must be invested to construct this wrapper, compared with reprogramming the application directly. Chapter 13 offers reassurance in this regard by exhibiting the programming required for some sample legacy-based agents. To judge by these examples, most of the logical machinery developed in the book is handled automatically by the IMPACT agent development environment. Chapter 14, the conclusion, reviews the IMPACT approach against the desiderata identified in chapter 1 and desiderata proposed by other researchers. An appendix provides detailed code calls and actions for the running examples. The book includes a bibliography of over 300 entries through 1999 and an index, although a number of key technical terms used throughout the book do not appear in the index. The book is one of the most comprehensive, coherent discussions of a formal logical approach to agents that is available. It draws together a wide range of logical formalisms into an integrated system in service of the design and analysis of agent systems. In spite of the practical flavor of the applications considered, its main focus is theoretical, and it will appeal mainly to readers who are comfortable with detailed formalisms and extended logical constructions. Online Computing Reviews Service