1988 Reports

A Methodology for Programming Production Systems and its Implications on Parallelism

Pasik, Alexander J.

Production systems have been studied as a language for artificial intelligence programming for over a decade. The flexibility of a programming paradigm which allows for loosely structured, independent rules to represent knowledge is attractive. Unfortunately, two seemingly independent phenomena have hindered the ability to take full advantage of production systems. First, the performance of large production systems suffers due to the large amounts of computation required to run them. Second, the programming styles of individuals primarily accustomed to conventional programming has adversely affected the maintainability and performance of the resulting systems. The parallel execution of production systems has been studied in order to address the performance issues. Preliminary results have been interpreted pessimistically; production systems have been observed to contain only moderate to low levels of parallelism. By investigating the issue of programming style, however, it will be shown that the apparent lack of large-scale or massive parallelism is an artifact of this problem. Indeed, a set of programming guidelines and tools will be presented which yield more maintainable, understandable, and parallelizable production systems. Is there a programming methodology or environment which will allow for the development of more maintainable and parallelizable production systems? This work will attempt to demonstrate that using a combination of several techniques, resulting production systems will more appropriately conform with the theory which supports their use. Production systems are not appropriate for encoding all problem solving tasks. They are appropriate when there is a clear separation of explicit control knowledge, tabular knowledge, and pattern-directed knowledge. This classification has been presented by many researchers in the field, often in order to advocate their separation. The issue has been addressed from a knowledge representation standpoint: here it will be one of several issues which, when addressed properly, will result in systems with improved performance in addition to their more adequate representation of the knowledge. Substantially more paral1elism can be extracted from these systems. In this regard, the techniques complement parallel match algorithms which provide the first step in the solution for mapping production systems onto parallel architectures. The techniques are table-driven rules, creating constrained copies of culprit rules, multiple rule firing, and combining rule chains. These methods are combined into a new way of viewing production system execution. Rather than assuming the sequentiality of production systems and trying to extract parallelism explicitly, the systems are assumed to be implicitly parallel and all necessarily sequential aspects are explicitly defined.