Thesis Summary: Learning Search-Control Knowledge to Improve Plan Quality

Generating good, production-qualityplans is an essential element in transforming planners from research tools into real-world applications, but one that has been frequently overlooked in research on machine learning for planning systems. Most work has been aimed at improving the efficiency of planning (“speed-up learning”) or at acquiring or refining domain knowledge. This thesis focuses on learning search-control knowledge to improve the quality of the plans produced by the planner. Knowledge about plan quality in a domain comes in two forms: (a) a post-facto quality metric that computes the quality (e.g. the execution cost) of a plan, and (b) planning-time decision-control knowledge used to guide the planner towards producing higher-quality plans. The first kind of knowledge is not operational until after a plan is produced, but is exactly the kind typically available, in contrast to the far more complex operational decision-time knowledge. Learning operational quality control knowledge can be seen as a translation of domain knowledge and quality metrics into runtime decision guidance. The full automation of this mapping based on planning experience is the ultimate objective of this thesis. Given a domain theory, a domain-specific metric of plan quality, and problems which provide planning experience, the QUALITY architecture developed in this thesis automatically acquires operational control knowledge that effectively improves the quality of the plans generated. QUALITY can (optionally) learn from human experts who suggest improvements to the plans at the operator (plan step) level. We have designed two distinct domain-independent learning mechanisms to efficiently acquire quality control knowledge. They differ in the language used to represent the learned knowledge, namely control rules and control knowledge trees, and in the kinds of quality metrics for which they are best suited. QUALITY is fully implemented on top of the PRODIGY4.0 nonlinear planner. Its empirical evaluation has shown that the learned knowledge produces near-optimal plans (reducing before-learning plan execution costs up to 96%). Although the learning mechanisms and learned knowledge representations have been developed for PRODIGY4.0, the framework is general and addresses a problem that any planner that treats planning as a constructive decision-making process must confront.