A LSTM Assisted Prediction Strategy for Evolutionary Dynamic Multiobjective Optimization

Abstract

Dynamic multiobjective optimization problems (DMOPs) are widely spread in real-world applications. Once the environment changes, the time-varying Pareto-optimal solutions (PS) are required to be timely tracked. The existing studies have pointed out that the prediction based mechanism can initialize high-quality population, accelerating search toward the true PS under the new environment. However, they generally ignore the correlation between decision variables during the prediction process, insufficiently predict the future location under the complex problems. To solve this issue, this paper proposes a long short-term memory (LSTM) assisted prediction strategy for solving DMOPs. When an environmental change is detected, the population is divided into center point and manifold. As for center point, historical ones are utilized to train LSTM network and predict the future one. Subsequently, the manifold is estimated by Gaussian model in terms of two past ones. In this way, an initial population is generated at the new time by combining the predicted center point and manifold. The intensive experimental results have demonstrated that the proposed algorithm has good performance and computational efficiency in solving DMOPs, outperforming the several state-of-the-art dynamic multiobjective evolutionary algorithms.