Large Populations Are Not Always The Best Choice In Genetic Programming

In genetic programming a general consensus is that the population should be as large as practically possible or sensible. In this paper we examine a batch of problems of combinatory logic, previously successfully tackled with genetic programming, which seem to defy this consensus. Our experimental data gives evidence that smaller populations are competitive or even slightly better. Moreover, hill-climbing appears to exhibit the best performance. While these results are in a way unexpected, theoretical considerations provide a possible explanation in terms of a special constellation rather than a general misconception as to the benefits of large populations or genetic programming as such.