Tomas De la Rosa, Angel Garcia Olaya, Daniel Borrajo
Currently, among the fastest approaches to AI task planning we find many forward-chaining heuristic planners, as FF. Most of their good performance comes from the use of domain-independent heuristic functions, together with efficient search techniques. When analysing their performance, most of the time is spent precisely on computing the heuristic value of nodes. The goal of this paper is to present a way of reducing the number of calls to the heuristic function, and, therefore, the time spent on finding a solution. We use a case-based reasoning approach that automatically acquires domain-dependent typed sequences (cases) from some training problems. Then, the learned cases are used to recommend to each search node which of its successors to evaluate first. Experimental results in several competition domains show the advantages of the approach.
Subjects: 3.1 Case-Based Reasoning; 1.11 Planning
Submitted: Feb 10, 2007