A methodology for programming production systems and its implications on parallelism

Abstract

The performance of large production rule systems suffers due to the large amounts of computation required to run them. In addition, the programming styles of individuals primarily accustomed to conventional programming has adversely affected the maintainability of the resulting systems (i.e., the systems often have high software complexity). The parallel execution of production systems has been studied in order to address the performance issues. Preliminary results have been disappointing; production systems have been observed to contain only moderate to low levels of parallelism. This thesis demonstrates, however, that by programming production systems as sets of independent rules, and macrorules and table-driven rules, and by creating constrained copies of culprit rules, production systems are both more maintainable because of their reduced software complexity and more parallelizable because of a resulting improved distribution of work. In addition, characteristics of production systems which can affect the amount of performance improvement provided by these techniques are identified.