Selectively Generalizing Plans for Problem-Solving

Abstract

Problem solving programs that generalize and save plans in order to improve their subsequent performance inevitably face the danger of being overwhelmed by an ever-increasing number of stored plans. To cope with this problem, methods must be developed for selectively learning only the most valuable aspects of a new plan. This paper describes MORRIS, a heuristic problem solver that measures the utility of plan fragments to determine whether they are worth learning. MORRIS generalizes and saves plan fragments if they are frequently used, or if they are helpful in solving difficult subproblems. Experiments are described comparing the performance of MORRIS to a less selective learning system.