HeurEAKA – A New Approach for Adapting GAs to the Problem Domain

Abstract

We propose a new approach for the efficient development of effective heuristic problem solvers for combinatorial problems. Our approach is based on Genetic Algorithms (GA) and addresses the known problem of allowing the efficient adaptation of a general purpose GA to a given problem domain. The adaptation is done by building a knowledge base that controls part of the GA, i.e. the fitness function and the mutation operators. The knowledge bases are built by a human who has at least a reasonable intuition of the search problem and how to find a solution. The human monitors the GA and intervenes when he/she feels that the GA produces individuals which have only a small chance of leading to an acceptable solution or the human helps by providing rules describing how to generate an individual with high chances of success.

We use an incremental knowledge acquisition approach based on Nested Ripple Down Rules. We provide initial experiments on an industrially relevant domain of channel routing in VLSI design. Industrial algorithms have been developed over decades in this domain. Our results so far are extremely encouraging, as we managed to solve some benchmark problems with a relatively small knowledge base in conjunction with a general purpose GA.