A New Prediction Approach for Dynamic Multiobjective Optimization

This study develops a prediction-based reinitialization approach that is comprised of three components for dynamic multiobjective optimization (DMO). The first component is a controlled translation of the population centroid, which analyzes the successive movement of the population Pareto optimal set (POS) at the end of each problem instance. The second and third components are directional and random variation. In addition, a metric to quantify the variation in the POS that does not fit in a simple translation is proposed. The rationale behind each component is explained and demonstrated in some carefully designed descriptive experiments. Different variants of the proposed strategy are assessed and compared with two recently proposed reinitialization strategies, on an accredited test suite for DMO. A comparison of the numerical results reveals that unlike the random variation, the directional variation operator significantly improves the performance. Overall, our proposed strategy considerably outperforms the other considered strategies, especially when the evaluation budget for each problem instance is limited.