As agent technologies move from the laboratory into practical applications, industrial practitioners need handbooks that integrate and organize the essential research results. These two volumes attempt this task along lines that are very similar to each other in terms of their overall vision, which has three points: agents are a merger of five basic technologies; implementers should build on an integrating architecture; and a discipline of agent design can be developed by extending object-oriented techniques. The books differ widely in how they work out the details of this vision.
Both books view agents as the natural merger of the same five base technologies: expert systems, object-oriented programming, neural networks, fuzzy logic, and genetic algorithms. This vision (the source of the word “hybrid” in Khosla and Dillon’s title) reflects a sound instinct, that engineers should be ready to deploy multiple methods as needed to solve the problem at hand. It also reflects a common emphasis on coarse-grained agents that aspire to individual intelligence. Neither volume seems aware of the considerable potential of communities of fine-grained agents, with limited cognitive resources, to provide intelligent system-level behavior. Both volumes offer extensive reviews of the five base technologies, with pointers to additional literature.
The integration of these technologies is a serious technical challenge that Knapik and Johnson do not address. They survey research work in these domains broadly, if somewhat unevenly; introduce the capabilities of each technology in isolation; and suggest some conditions under which each might be useful. The main value of this section of their work is in the references to primary sources.
In contrast, Khosla and Dillon develop a detailed taxonomy of ways that these technologies can be combined, and they provide numerous examples. Fusion systems hard-wire one technology into another. Transformation systems apply multiple technologies in series, using one to preprocess input for use by another. Combination systems apply different technologies, in parallel, to different system functions or aspects of the problem-solving process. Associative systems combine the other three approaches to integration.
Widespread application of a software technology in industry requires the existence of accepted architectures and infrastructures that enforce good practice. Such resources permit systems to be configured by people who are not researchers.
Knapik and Johnson review a variety of infrastructures, including OpenDoc, ActiveX, CORBA, DCE, and several research architectures. The discussions are not closely integrated with one another. As with the material on base technologies, the value of the book is more in calling readers’ attention to several available alternatives than in expounding the tradeoffs among them.
Khosla and Dillon present a detailed account of their own Intelligent Multi-Agent Hybrid Distributed Architecture (IMAHDA). This complex proposal incorporates all of the basic technologies and integration mechanisms described in the first part of the book, but it lacks the conceptual clarity and elegance that can make an architecture compelling and useful. At times it seems to diverge from the vision of an agent as the integration of these technologies. For example, the authors propose separate agents to handle fuzzy logic, expert systems, neural networks, and knowledge-based reasoning, rather than merging these modalities in each agent. The architecture is so all-inclusive that it is unlikely to provide software engineers with a clear guide to best practice.
Both volumes recognize that agents are extensions of software objects and advocate extension of object-oriented design methods to handle the additional functionality of agents.
In keeping with their emphasis on enumerating alternatives, Knapik and Johnson approach the design question by presenting a list of issues to be kept in mind, including requirements analysis; the relation of agents to people; different levels of intelligence; security; end-user programmability; communications; and mobility. This last section is largely a reproduction of a General Magic white paper that might be more appropriate in the discussion of agent architectures or infrastructures. Knapik and Johnson discuss implementation toolkits, focusing on Smalltalk, Java, and Telescript (again with generous reproduction of General Magic documents), and review a number of domains in which agents have been or might be applied.
Khosla and Dillon work out the implications of their IMAHDA architecture in the context of a specific industrial application, alarm processing in an electrical power distribution system. The amount of detail suggests that the report is based on an actual industrial implementation. Their account of the design and implementation decisions involved in this system is an integrated narrative that highlights many of the tradeoffs and alternatives they consider.
In sum, Knapik and Johnson provide a catalog of technologies, organized roughly around base technologies, frameworks, and design and implementation alternatives, with little overall integration. The general tone of the book is informal, giving the impression that readers can easily build their own agents by mixing appropriate selections of the ingredients. Both authors are practicing software engineers and certainly understand the complexities of application software, but the book does not convey this depth. It will be useful as a collection of pointers to a wide range of interesting systems and technologies. However, as the authors note in the preface, new technologies have appeared since they wrote the book, and some of the technologies they describe are now obsolete. The book includes a helpful list of acronyms, a comprehensive integrated bibliography of sources cited in the text, and a reasonably detailed index.
From a software engineering perspective, Khosla and Dillon offer a more integrated view of how to design and construct agent-based software. Unfortunately, the volume gives no evidence of copyediting. Every page includes serious textual errors that at best distract the reader and at worst make the authors’ intent unintelligible. These lapses include sentence fragments stranded between paragraphs, with no apparent home in the text; multiple repeated lines of text; figures printed over their own legends; and even sentences in which an earlier wording has been abandoned in favor of another, but has not been removed from the text. The severity and frequency of these lapses makes the valuable insights the book promises almost completely inaccessible. The index is brief, and references are grouped by chapter rather than being combined at the end of the book.
