Browse Books

Reviewer: Adina Magda Florea

Miranker investigates the applicability of parallel computation to rule-based systems. He describes the architecture and software systems of the DADO machine, a parallel tree-structured computer designed to improve the performance of large production systems, and TREAT, a new match algorithm for implementing the OPS5 rule-based programming language on such a machine. The first two chapters contain background information about production systems and the OPS5 language as well as a taxonomy of parallel computers. Chapter 3 details algorithmic aspects of production system interpreters, particularly as they relate to parallelism, from three perspectives: low-level matching, rule partitioning, and synchronization. The chapter focuses on low-level matching and describes the RETE match algorithm, considered to be the best algorithm for the sequential execution of production systems. Chapter 4 presents the TREAT match algorithm in detail; its development is motivated by the inadequacies of the parallel versions of existing algorithms, in particular the RETE match, to a DADO machine. The author shows how TREAT incorporates database query optimization techniques. He presents an implementation of OPS5 using a sequential version of TREAT and concludes with a comparison of the performances of the TREAT and RETE algorithms for five benchmarks. Chapter 5 presents the DADO system architecture, a distributed-memory, medium-grained, tree-connected, parallel processor that displays good properties from the perspective of production systems requirements. It discusses the DADO software layers with special emphasis on PPL/M, a parallel superset of the PL/M language. Chapter 6 is an overview of some parallel AI machines classified by the knowledge representation method they are intended to support. The author includes some criticism of each of them. Chapter 7 presents the implementation of the parallel version of TREAT on the DADO machine, with supporting design decisions. The author highlights the capacity of TREAT to effectively induce parallelism in the associative matching of individual rules. The chapter concludes with the performance evaluation of TREAT on the DADO machine. The book is an exposition of the research embedded in the author's doctoral dissertation. Its central contribution is the design and implementation of the TREAT algorithm. Miranker provides a clear and well-structured presentation of the design decisions and lucid quantitative and qualitative evaluations of the results. He gives insights into multiple constraints and requirements of hardware and software solutions for implementing parallel production systems and compares TREAT with several related approaches; this comparison is the best feature of the book. Another virtue of the presentation is its clear-cut organization and conciseness; the conciseness and density of ideas may sometimes turn into a drawback for readers lacking the necessary background. The approach of using the database area as a source of both the terminology for explaining production rules and the optimization techniques for improving the performance of TREAT highlights the growing connection between knowledge-based systems and DBMSs. The worst feature of the document is its title, which indicates only a part of the topics presented; this drawback is easily eliminated when reading the work. Another weakness of the book is the heterogeneous level of presentation. Topics related to the author's research are presented in depth and can be fully understood only by persons with a background in AI, while other topics, such as fifth-generation computers, are treated at a general level that is superfluous for such readers. Nonetheless, the book succeeds in offering the reader a good perspective on the author's seminal research results and represents a challenge to the AI research community. The book is of primary interest to parallel computer designers and knowledge-based systems implementors, but will also appeal to knowledge engineers and database system programmers. It is of special interest for the AI specialist who wants to fill the gap between now classical sequential AI systems and parallel computer architectures. A background in databases and knowledge representation methods is required. Consulting additional material about the OPS5 language and the RETE match algorithm will help in understanding the issues covered. The physical form is pleasant and easy to read. No index is included but, due to the brevity of the book, none is required. The references are good, covering a wide area of related material, but including the year of publication in the text would have been a better decision. The material could be a good starting point for a graduate course on AI system architectures, although this was not the author's intent.