The Statistical Learning of Accurate Heuristics

Heuristics used by search algorithms are usually composed of more primitive functions which we call "features". A method for combining features is presented which is based on linear regression with true distance to goal as the dependent variable. Our method also provides a probabilistic estimate of the solution error produced when the combination of features is used as a heuristic by A'. The accuracy of the heuristics learned is demonstrated in the TSP and sliding tile domains. The high quality solutions returned and the considerable savings in nodes expanded in both domains show effectiveness as well as generality of the method.

• Type: Conference
• Year: 1993
• Where: IJCAI
• Authors: Anna Bramanti-Gregor, Henry W. Davis
• Keywords: Dependent Variable, Heuristics, IJCAI 1993, IJCAI 2007, Primitive Functions