A Generative Adversarial Networks Model Based Evolutionary Algorithm for Multimodal Multi-Objective Optimization

The key to solving multimodal multi-objective optimization problems is to achieve good diversity in the decision space. However, the existing algorithms usually adopt the reproduction operation based on random mechanism, which do not make full use of the distribution features of promising solutions in the population, resulting in the defects of the diversity of the obtained Parteo optimal solution sets. In order to solve the above problem, this paper proposes a multimodal multi-objective optimization evolutionary algorithm (MMOEA) based on generative adversarial networks (GANs). Specifically, we firstly design a classification strategy to distinguish good solutions from poor solutions. The solutions in the population are classified as real samples and fake samples by non-dominated selection sorting based on special crowding distance, and the training data of GANs are obtained. Secondly, a GANs-based offspring generation method is proposed. Through the adversarial training of GANs, the generator can simulate the distribution of promising solutions in the population and generate offspring with good diversity. Thirdly, an environment selection strategy based on GANs is constructed. By sorting the classification probability of the solutions output by the discriminator, the population are selected and updated. Finally, the proposed algorithm is compared with seven other competitive multimodal multi-objective optimization evolutionary algorithms on the CEC 2019 test suite and a real-word problem, and experimental results indicate its superior performance.