Too much knowledge hurts: Acceleration of genetic programs for learning heuristics

Abstract

Among many other applications, evolutionary methods have been used to develop heuristics for several optimization problems in VLSI CAD in recent years. Although learning is performed according to a set of training benchmarks, it is most important to generate heuristics that have a good generalization behaviour and hence are well suited to be applied to unknown examples. Besides large runtimes for learning, the major drawback of these approaches is that they are very sensitive to a variety of parameters for the learning process. In this paper, we study the impact of different parameters, like e.g. stopping conditions, on the quality of the results for learning heuristics for BDD minimization. If learning takes too long, the developed heuristics become too specific for the set of training examples and in that case results of application to unknown problem instances deteriorate. It will be demonstrated here that runtime can be saved while even improving the generalization behaviour of the heuristics. © Springer-Verlag 2001.