Learning steppingstones for problem solving

Abstract

Most classic artificial-intelligence domains require satisfying a set of Boolean constraints. Real-world problems require finding a solution that meets a set of Boolean constraints and performs well on a set of real-valued constraints. In addition, most classic domains are static while domains from the real world change. In the present work, the authors demonstrate that SteppingStone, a general learning problem solver, is capable of solving problems with these characteristics. SteppingStone heuristically decomposes a problem into simpler subproblems, and then learns to deal with the interactions that arise between the subproblems. In lieu of an agreed-upon metric for problem difficulty, significant problems which are difficult for both people and programs are used as good candidates for evaluating progress. Consequently, the domain of logic synthesis from VLSI design is used to demonstrate SteppingStone's capabilities.<>