An ensemble learning based prediction strategy for dynamic multi-objective optimization

Abstract

Prediction strategies are widely-used in dynamic multi-objective evolutionary algorithms (DMOEAs). However, the characteristics of the environmental changes are different and only use one single prediction model cannot react to the changes effectively. The mismatching of the changes and prediction models may make the predicted results inaccurate and unstable. To overcome this shortage, an ensemble learning based prediction strategy (ELPS) is proposed in this paper to help algorithms re-initialize a new population after a change is detected. There are four base prediction models in ELPS, i.e., linear prediction model (LP), knee point-based autoregression model (KP-AR), population-based autoregression model (P-AR) and random re-initialization model (RND). Once a change happens, these four base prediction models are trained by the historical information with ensemble learning and a strong prediction model can be constructed on these four base prediction models. The final re-initialized population is generated by this strong prediction model to react to the new environment. With the help of ELPS, the re-initialized population can adapt different environmental changes and improve the performance on prediction accuracy and robustness. The experimental results show that, compared with other state-of-the-art prediction strategies on benchmark test suite, ELPS has better performance on dealing with dynamic multi-objective optimization problems.