An agent-based approach to the diagnosis of faults in complex physical systems is developed. These systems range from manufacturing processes to computer networks. Such systems follow general rules of behavior, but are susceptible to faults. Because these systems are complex or distributed, we need special tools and techniques to diagnose their faults based on observations of anomalous behavior.
The book introduces the diagnosis problem, considers logic programming techniques for diagnosis, develops a declarative language to specify strategies to control diagnostic reasoning, and, finally, applies a special kind of agent to diagnosis.
Diagnosis is an interesting problem area, a microcosm of artificial intelligence (AI). Solutions to the diagnosis problem engage many of the major themes of AI--search, knowledge representation, and learning, as well as case-based, logical, and probabilistic inference to deal with incomplete information. Schroeder focuses on logic-based representation and reasoning. Model-based diagnosis is about constructing a model of the system being examined so that the fault being diagnosed is simply a succinct explanation of the observed (faulty) behavior.
Model-based diagnosis accords well with logic programming. Simply put, the laws governing the system’s behavior are stated as integrity constraints; the specific configuration of the system and the observations of its behavior are stated as facts. For components that can fail, their correct behavior is assumed by default (expressed as weak negation in logic programming). Anomalous behavior manifests itself as a contradiction, which can be removed by changing some of the assumptions about the components being correct. Diagnosis algorithms described in this book can find the “right” set of faults, the right set usually being one that assumes the least upheaval or malfunctioning in the system. In general, diagnosis is computationally intractable. Various strategies can be employed to control the complexity; the author shows how a range of strategies can be declaratively specified in a form of modal temporal logic.
The book treats agents as entities with states described in mental terms, such as knowledge and beliefs, represented using formal logic, and reasoned about in logic programming. Diagnosis agents execute diagnosis algorithms of the type developed earlier in the book. In addition, they communicate with each other to exchange observations and diagnoses. Using agents for diagnosis has several advantages. For example, agents can naturally correspond to different components of a system being diagnosed and can make local observations and diagnoses about the components they are observing.
In the author’s approach, the value added by agents over traditional diagnosis algorithms is in their exchange of information. This is expressed well in the framework of vivid agents, an interesting approach related to logic programming that has general applicability beyond diagnosis problems. The agents carry out a rudimentary form of coordination (reminiscent of traditional distributed computing) by collecting results from their peers.
Unfortunately, despite the title of this book, its specific coverage of agents is minimal. Agents are introduced in the last technical chapter, which is only 18 pages long. Readers expecting to learn about autonomous agents in general might be disappointed. The coordination mechanisms discussed are weak. For example, the main technique seems to be to count messages until an expected number has been received, and then to initiate voting, which is handled procedurally. More sophisticated negotiation techniques exist and would be suitable for distributed diagnosis, if, for example, the agents had conflicting diagnoses. Although the coverage of agents builds on the general approach for diagnosis, it has no obvious relationship to the previous (somewhat longer) chapter on declarative strategies for diagnosis. Schroeder could have strengthened this work by integrating control strategies with his discussion of agents, and by having the agents reason about their collective strategies.
Although the book is based on the author’s doctoral dissertation, it includes an extensive survey of model-based diagnosis and logic programming treatments of diagnosis and intelligent agents. (Most of Schroeder’s own research contributions are presented in the chapters on control strategies and agents.) One of this book’s strengths is that it makes the survey accessible to graduate students and researchers, who may not be familiar with the combination of topics covered here, especially model-based diagnosis, which remains somewhat of a niche area within AI.
The book is well written. Although it could benefit from more extensive results on agents, its brevity is another strength. I recommend it highly to readers with an interest in logic programming and agents; even if they are not interested in diagnosis per se, they will learn much from it.
