CGP visits the Santa Fe trail: effects of heuristics on GP

GP uses trees to represent chromosomes. The user defines the representation space by defining the set of functions and terminals to label the nodes in the trees, and GP searches the space. Previous research and experimentation show that the choice of the function/terminal set, choice of the initial population, and some other explicit and implicit "design" factors have great influence on both the quality and the speed of the evolution. Such heuristics are valuable simply because they improve GP's performance, or because they enforce some desired properties on the solutions. In this paper, we evaluate the effect of heuristics on GP solving the Santa Fe trail. We concentrate on improving the solution quality, but we also look at efficiency. Various heuristics are tried and mixed by hand, while evaluated with the help of the CGP system. Results show that some heuristics result in very substantial performance improvements, that complex heuristics are usually not decomposable, and that the heuristics generalize to apply to other similar problems, but the applicability reduces with the complexity of the heuristics and the dissimilarity of the new problem to the old one. We also compare such user-mixed heuristics with those generated by the ACGP system which automatically extracts heuristics improving GP performance.