Improving Competitive Differential Evolution using Automatic Programming

In this paper, we automatically improve the competitive differential evolution algorithm through automatic programming. The improved algorithm outperforms the original for over 73% of the 50-dimensional CEC 2014 problems and is worse for less than 17% of the problems when comparing using a Wilcoxon rank-sum test. The evolutionary automatic programming system ADATE that is used in this paper systematically searches for better programs by evaluating millions of candidate programs. The candidates are graded by first evaluating on a small training set consisting of five synthetic optimization problems, with well performing candidates being evaluated more extensively on a larger and more computationally expensive validation set with 100 problems. Thus, we use one evolutionary algorithm to rewrite the source code of another evolutionary algorithm. The results show that the techniques introduced in this paper are capable of improving the heuristics of contemporary numerical optimization algorithms.