Multilayer perceptron-based offspring prediction model for constrained multi-objective optimization

Constrained multi-objective optimization problems generally have both multiple constraint violations and conflicting objective functions. Some of them not only have sparse feasible regions, but also are difficult to converge. For these problems, the evolutionary operators used in traditional constrained multi-objective evolutionary algorithms (CMOEAs) are difficult to generate solutions with ideal quality. Therefore, this paper proposes a multilayer perceptron-based offspring prediction model for constrained multi-objective optimization (MOPCMO). Specifically, an evolutionary direction guidance strategy is designed that utilizes historical populations as training data to train a multilayer perceptron, which guides the evolution of the population by predicting and generating offspring, thereby improving the overall evolutionary efficiency of the algorithm. In addition, as the population iterates, evolutionary direction guidance strategy adaptively transforms the training data of multilayer perceptron. Finally, the multilayer perceptron is intermittently updated and uses an evolutionary direction guidance strategy to generate promising offspring, guiding the algorithm to achieve efficient search. Compared with seven state-of-the-art CMOEAs on 33 benchmark test problems and 8 engineering application problems, MOPCMO achieves excellent performance.