Learning search control knowledge to improve plan quality.

My thesis work falls into the areas of AI planning and machine learning. Previous work in these areas focused on making planning efficient and overlooked the need to generate good, production-quality plans. We found this an essential step for transforming planners from research tools into real-world applications. In the thesis I have developed a learning algorithm to automatically acquire control knowledge that guides the planner towards better solutions based on problem solving experience. The planner's domain theory is extended with a plan quality evaluation function. The learner interacts with a human domain expert who suggests improvements to the plans. My system interprets this feedback and automatically operationalizes it into effective planning control knowledge. In the thesis I have fully implemented the learning algorithm, and demonstrated it in several domains including a large process planning domain obtaining significant plan quality improvement. A longer summary is available in Postscript.