Learning Plan Rewriting Rules

Considerable work has been done to automatically learn domain-specific knowledge to improve the performance of domain independent problem solving systems. However, most of this work has focussed on learning search control knowledge-knowledge that can be used by a problem solving system during search to improve its performance. An alternative approach to improving the performance of domain independent systems is by using rewriting rules. These are the rules that can be used by a problem solving system after generating an initial solution in order to transform it into a higher quality solution. This paper reviews various approaches that have been suggested for automatically learning rewriting rules, analyses them, and suggests novel algorithms for learning plan rewriting rules. The ability to produce high quality plans is essential if AI planners are to be applied to the real world planning problems. Machine learning for planning suggests automatically learning domain specific information that can be used by the AI planners to produce high quality solutions. Considerable planning and learning research has been devoted to learning domain specific search control rules to improve planning performance (planning efficiency and plan quality). These rules improve planning performance by guiding a planner towards higher quality plans during planing. An alternative technique is planning by rewriting (Ambite ~ Knoblock 1997) that suggests first generating an initial plan using a planner and then using a set of rewrite-rules to transform it into a higher quality plan. However, automatically learning such rules has been a challenging problem. Previously, I designed a system called REWRITE (Upal 1999; 2000) that automatically learned plan rewrite rules by comparing two planning search trees (one search tree leading to a lower quality plan and the other search tree leading to a higher quality plan). However, REWRITE’s performance with respect to planning efficiency was poor (Upal gz Elio 2000) because was unable to learn good rewrite rules. In (Upal 2000) we show that the problem lies with REWRITE’s strategy of learning rules in one context (i.e., the context Copyright (~)2001, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. of search nodes) and then applying them in a different context (i.e., the context of the complete plans). (Upal 2000) also presents evidence to show that learning search control rules from search trees is a better option than learning rewrite rules by comparing planning traces. Ambite et al. (Ambite, Knoblock, & Minton 2000) suggest a similar technique for learning rewrite rules. However, their system (called Pbr) learns by comparing two completed solutions to a planning problem; one of higher quality and other of lower quality. Pbr’s learning component simply stores those steps of the better quality solution that are not present in the lower quality plan as replacing steps, and it stores those steps of the lower quality solution that are not present in the higher quality solution as the to-be-replaced steps and to-be-replaced constraints on the steps. Pbr also ranks its rules according to goodness. It prefers the rules that are smaller. Next I present the background material on planning by rewriting followed by a review of the previously suggested algorithms and suggest new algorithms for learning plan rewriting rules. Planning by Rewriting The basic idea of rewriting can be traced back to the work on graph and number rewriting (Baadr &: Nipkow 1998). The idea of using rewrite rules for AI planning was introduced by Ambite et al. (Ambite & Knoblock 1997) who referred to it as planning by rewriting. A planning by rewriting system consists of two component: a planning component for generating the initial plans, and a rewriting component that can rewrite these plans to transform them into higher quality plans. Since, by definition, planning by rewriting systems spend extra time in plan-rewriting, planning by rewriting can be expected to be useful in those planning domains in which generating suboptimal plans is significantly more efficient than generating optimal quality plans. Interestingly, many AI planning domains such as Blocksworld and the logistics transportation domain exhibit these properties (Ambite & Knoblock 1997). A rewrite rule consists of two equivalent sequences of actions such that one of them can be replaced by the 412 FLAIRS-2001 From: FLAIRS-01 Proceedings. Copyright © 2001, AAAI (www.aaai.org). All rights reserved. to-be-replaced: actions: {drive-truck(Truck, From, To), drive-truck(Truck, To, From)) replacing: actions: {} Figure h A rewrite rule for the logistics transportation domain. other. Figure 1 shows a rewrite rule from the logistics transportation domain consisting of two sequences of actions replacing and to-be-replaced. Given an initial suboptimal plan produced by the planning component, the task of the rewriting component is to delete the tobe-replaced sequence of actions from the initial plan and add the replacing sequence of actions to it. In order to better understand the rewriting process, it is useful to view a viable plan for a problem as a graph in which actions correspond to vertices and constraints on the actions (such as casual-link and the ordering constraints) correspond to the edges between the action. The rewriting process can be understood as deleting a subgraph and replacing it with another subgraph. Consider the