Learning Episodes for Optimization

Abstract

Real-world problems often require optimizing a solution for various constraints. Application systems for many of these optimization problems already exist. These systems efficiently encode a core set of knowledge for solving problems in their domain. Where these systems fail is at dealing with the exceptions to the core knowledge. By learning to recognize these exceptions, the quality of the solutions can be further improved. In this paper we describe EASe, a method for learning these exceptions. EASe uses search on simple problems to learn episodes where improvement on a solution was possible. These are then reapplied to improve the quality of more complex solutions. We demonstrate EASe with empirical results from experiments with a set of benchmark problems from the logic synthesis domain.