Analysis of Complete-plan Comparison Strategy for Learning Plan Rewriting Rules

Abstract:  Planning by Rewriting is an attractive approach for improving the performance of domain-independent planning while maintaining high-quality plan output. However, performance has been a problem for rewrite planning systems which do not have the ability to choose the best rewrite rules to use. This paper presents the empirical results of using complete-plan comparison to learn rules for a rewrite planner. A rule rating algorithm called effect similarity is used to determine the best rewrite rules, and this system's performance is compared to the performance of the planner when using unrated rules. Also discussed is the effect that keeping only a small number of the best rules has on planning time and plan quality.