A Self-adaptive Similarity-based Fitness Approximation for Evolutionary Optimization

Evolutionary algorithms used to solve complex optimization problems usually need to perform a large number of fitness function evaluations, which often requires huge computational overhead. This paper proposes a self-adaptive similarity-based surrogate model as a fitness inheritance strategy to reduce computationally expensive fitness evaluations. Gaussian similarity measurement, which considers the ruggedness of the landscape, is proposed to adaptively regulate the similarity in order to improve the accuracy of the inheritance fitness values. Empirical results on three traditional benchmark problems with 5, 10, 20, and 30 decision variables and on the CEC' 13 test functions with 30 decision variables demonstrate the high efficiency and effectiveness of the proposed algorithm in that it can obtain better or competitive solutions compared to the state-of-the-art algorithms under a limited computational budget.