OPTIMAL TUNNELING: A HEURISTIC FOR LEARNING MACROS

Abstract

This paper presents the Optimal Tunneling heuristic for learning macro operators. Optimal Tunneling produces shorter more useful macros than the similar Minimum to Minimum heuristic presented by Iba. Optimal Tunneling is arguably an improvement since its macros (a) best reduce search cost, (b) give the most accurate modification to the search space to make the heuristic function correct, and (c) result in better performance on comparative tests. Optimal Tunneling creates macros that cross exactly the expensive segment of the heuristic function along the current solution path. A water pouring analogy is proposed to illustrate the effect of macros on the cost of search in problem solving.