A Generative Adversarial Network Based Prediction Strategy for Evolutionary Dynamic Multiobjective Optimization

Recently, dynamic multi-objective evolutionary optimization has attracted much research attention, and the prediction based strategy has been proved to be an effective way for dynamic multi-objective optimization evolutionary algorithms (DMOEAs) to track the changing Pareto set (PS). Recently, some strategies have been proposed to use the information from both the historical and new environments to predict the population in the new environment. However, their performance in complex new environments is limited, as they partially utilize information from the new environment. Generative adversarial networks (GANs) have been proven to be an efficient generative model, which can capture the relationships between high-quality solutions in the new environment and estimate the data distribution more accurately. In this paper, we propose a GAN based prediction strategy (GANPS) to generate more high-quality solutions that can adapt to the new environment. In GANPS, a PS estimation method is employed to improve the quality of samples (solutions) for GAN training at first, which helps the GAN better extract the information from historical and new environment. Then GANPS utilizes the trained generator to re-initialize the population, and a model reuse strategy is further designed to reduce training costs and the algorithm’s response time. Experimental results on a set of benchmark test suites show that GANPS could outperform other state-of-the-art reaction strategies in most cases and has better change tracking capability in dynamic environments.