Learning quality-enhancing control knowledge

TL;DR: The goal is to have a system that improves over experience the quality of the plans it generates by acquiring in a fully automated fashion control knowledge to guide the search.

Abstract: Generating production-quality plans is an essential element in transforming planners from research tools into real-world applications. However most research on planning so far has concentrated on methods for constructing sound and complete planners that find a satisficing solution, and on how to find such solution in an efficient way. Similarly most of the work to date on automated control-knowledge acquisition has been aimed at improving the eficiency of planning; this work has been termed “speed-up learning”. Our work focuses on how control knowledge may guide a planner towards better plans, and how such control knowledge can be learned. “Better” may be defined in a domain-dependent way and vary over time. (Perez & Carbonell 1993) contains a detailed taxonomy of plan quality metrics. We have concentrated on metrics related to plan execution cost, expressed as an evaluation function additive on the cost of the individual operators. These functions are linear and do not capture the existence of tradeoffs between different quality factors. Our goal is to have a system that improves over experience the quality of the plans it generates by acquiring in a fully automated fashion control knowledge to guide the search. Figure 1 shows the architecture of the current system, fully implemented on top of the PRODIGY nonlinear planner (Carbonell et al. 1992).