A Mixture of Prediction Strategies for Dynamic Multi-objective Evolutionary Optimization

Dynamic multi-objective optimization problems (DMOPs) exist in different fields, such as metal rolling, resource allocation, and so forth. DMOP requires Evolutionary algorithms (EA) to track Pareto-optimal solution (PS) or/and Pareto-optimal front (PF) in a dynamic environment. Therefore, it is crucial to present high performance EA to solve DMOPs. Many prediction-based algorithms solely construct single predictor based on historical information, which could result in unsatisfactory performance in complicated DMOPs. To overcome this limitation, this paper proposes a mixture of prediction strategies combined with decomposition-based static algorithm (MOEA/D), which is named MOEA/D-MPS. MOEA/D-MPS presents a center-based prediction and an individual-based prediction. The former adopts a first-order difference method to estimate the evolutionary moving distance of population by center point, the latter selects a second-order difference method to predict the new position of each individual. The predicted individuals are mixed up as the solution set for new environment. To testify the performance of MOEA/D-MPS, this paper selects nine classical DMOP benchmarks. Compared to three state-of-the-art algorithms, the experimental results indicate that MOEA/D-MPS is sufficient to solve different kinds of DMOPs.