An online-learning-based evolutionary many-objective algorithm

Abstract When optimizing many-objective problems (MaOP), the same strategy might behave differently when facing problems with different features. Therefore, obtaining problem features helps to obtain high-quality solutions. However, in practice, the problem features are unknown during the optimization process. In this case, learning to adjust strategies to match the problem features is a challenging work. In this paper, a learning-based algorithm is proposed, aimed to enhance the generalization ability. On the basis of a decomposition-based many-objective optimization framework, a learning automaton (LA) is included in the algorithm. The LA adjusts the evolutionary strategies of the algorithm to adapt to the problem characteristics, according to the feedback information during the optimizing procedure. An external archive is employed to store the Pareto non-dominant solutions. Based on the external archive, a reference vector adjustment strategy is designed to enhance the capability of solving problems with a degenerate or discrete Pareto front (PF). To validate the performance of the proposed algorithm, a comparison experiment is conducted on a novel authority test suite. Five state-of-the-art algorithms are selected as peer algorithms. The results of the experiment indicate that the proposed algorithm obtains satisfactory performance in determining the convergence and the approximation of the PF.