Learning General Graphplan Memos through Static Domain Analysis

Abstract

Graphplan [1] is one of the most efficient algorithms for solving the classical AI planning problems. Graphplan algorithm as originally proposed [1] repeatedly searched parts of the same space during its backward solution extraction phase. This suggested a natural way of speeding up Graphplan’s performance by saving memos from a pre- viously performed search and reusing them to avoid traversing the same part of the search tree later on in the search for the solution of the same problem. Blum and Furst [1] suggested a technique for learning such memos. Kambhampati [2] suggested im- provements on this algorithm by using a more principled learning technique based on explanation-based learning (EBL) from failures. However, these and other [3] learning techniques for Graphplan, learn rules that are valid only in the context of the current problem and do not learn general rules that can be applied to other problems. This paper reports on an analytic learning technique that can be used to learn general memos that can be used in the context of more than a single planning problem.