Learning while -Solving problems in single agent search: Preliminary results

Abstract

A traditional problem with best first search on hard problems is that space requirements are too large. However, space saving approaches require excessive time. We describe an approach which uses best first search, so as to keep time complexity low; it keeps space needs small because, through a statistical learning process, it builds accurate heuristics while solving problems. Unlike other search paradigms, solutions returned are accompanied with a statistical error assessment. In experiments solution quality is found high and time complexity is several orders of magnitude lower than competing techniques. The approach does not build a single heuristic for the whole domain. Instead, a new heuristic is built for each problem encountered: the learning time for constructing such “problem relevant” heuristics is low.

Deceased