Parallelization of Goal-Driven, Production Systems on Hypercube Machines in a C Environment.

摘要

Production systems are widely used in artificial intelligence to capture the notion of expertise in modeling expert systems. Production systems are computationally intensive programs spending most of the execution time in their MATCH or recognise phase. Efforts have been made by the research in this dissertation to minimize the production system's execution time by optimizing the MATCH phase. Goal oriented deterministic production systems are commonly used for robotics applications and formed the main class of production systems that were studied in this dissertation. The main motivation for the research was to provide a better MATCH algorithm and use the multiprocessing capabilities of existing parallel computer hardware. The dissertation realizes these goals by transforming a traditional production system's scalar equivalence operations into C arithmetic hashing function to generate an indexing variable for the switch-case construct of the C language. Partitioning of the working memory into homogeneous blocks and distributing production memory over the multiprocessors enhanced the MIMD operation of the production system. A scheme is formulated and implemented to identify a few key condition elements that may be used as an indexing variable and reduce the number of condition elements used in the MATCH phase. The complete translation from OPS5 code to C and the implementation scheme is presented in this dissertation. Various issues regarding the distribution of the inference engine over the multiprocessor environment and other related synchronization topics for distributed systems are covered in the dissertation. A detailed description of the parallel computer's simulator is also provided in the dissertation. The dissertation identifies other research topics and problems related to parallelization of production systems, the most significant being the ability to incorporate LEARNING in production systems by using one or all of the idle processors that are waiting for the active processor to complete it's activities.