Two-Space Prediction Approach Based Evolutionary Dynamic Multiobjective Optimization

To tackle dynamic multiobjective optimization problems (DMOPs), researchers have developed many prediction approaches based on dynamic multiobjective evolutionary algorithms (DMOEAs). Howecer, most methods only predict in a single decision or objective space. Most existing prediction methods are generally utilized the historical Pareto optimal solutions (POS) obtained in the decision space to predict the initial population. However, changes in DMOP may happen simultaneously in both decision and objective spaces. Therefore, predicting only in decision space may not provide accurate estimates of changes in DMOP. In response to this issue, a two-space based prediction strategy (TSP) is proposed, which is adapted to environmental changes. First, a denoising autoencoder search model is built, which is predicted from the perspective of decision space. Then, directly predict individuals in the objective space. After that, an improved inverse model is designed, which is to map the predicted individuals to the decision space. Finally, new solutions predicted based on decision and objective spaces are gathered together to form the initial population. The experimental results display that TSP is effective in handling most DMOP test cases.