Learning Rationales to Improve Plan Quality for Partial Order Planners

May 1999

Proceedings of the Twelfth International Florida Artificial Intelligence Research Society Conference (FLAIRS 1999)

Volume

Proceedings of the Twelfth International Florida Artificial Intelligence Research Society Conference (FLAIRS 1999)

All Papers

Plan rationale has been variously defined as "why the plan is the way it is," and as "the reason as to why the planning decisions were taken." The usefulness of storing plan rationale to help future planning has been demonstrated by several types of case-based planners. However, the existing techniques are unable to distinguish between planning decisions that, while leading to successful plans, may produce plans that differ in overall quality, as defined by some quality metric. We outline a planning and learning system, PIPP, that applies analytic techniques to learn plan-refinement control rules that partial-order planners can use to produce better quality plans. Quality metrics are assumed to be variant: whether a particular plan refinement decision contributes to a better plan is a function of the planning context. Preliminary results indicate that the overhead for applying these techniques to store and then use these control rules is not very large, and the outcome is the ability to produce better quality plans with in a domain. Techniques like this are useful in domains in which knowledge about how to measure plan qualityis available and the quality of the final plan is more important than the time taken to produce it.