Abstract
We propose a new program evolution method named PORTS (Program Optimization by Random Tree Sampling) which is motivated by the idea of preservation and control of tree fragments in GP (Genetic Programming). We assume that to recombine genetic materials efficiently, tree fragments of any size should be preserved into the next generation. PORTS samples tree fragments and concatenates them by traversing and transitioning between promising trees instead of using subtree crossover and mutation. Because the size of a fragment preserved during a generation update follows a geometric distribution, merits of the method are that it is relatively easy to predict the behavior of tree fragments over time and to control sampling size, by changing a single parameter. From experimental results on RoyalTree, Symbolic Regression and 6-Multiplexer problem, we observed that the performance of PORTS is competitive with Simple GP. Furthermore, the average node size of optimal solutions obtained by PORTS was simple than Simple GP's result.
