Empirical investigation of size-based tournaments for node selection in genetic programming

In genetic programming systems, genetic operators must select nodes upon which to act; the method by which they select nodes influences problem solving performance and possibly also code growth. A recently proposed node selection method using "size-based tournaments" has been shown to have potential, but variations of the method have not been studied systematically. Here we extend the ideas of size-based tournaments and test how they can improve problem-solving performance. We consider allowing tournament size to depend on whether we are selecting nodes within "donors" for crossover, "recipients" for crossover, or targets of mutation. We also consider tournaments that bias selection toward smaller trees rather than larger trees. We find that differentiating between donors and recipients is probably not worthwhile and that size 2 tournaments perform near-optimally.