An Adaptive Knowledge Transfer Strategy for Evolutionary Dynamic Multi-objective Optimization

Dynamic multi-objective optimization problems (DMOPs) are optimization problems involve multiple conflicting objectives, and these objectives change over time. The challenge in solving DMOPs is how to quickly track the Pareto optimal solution set when the environment changes. Recently, dynamic multi-objective evolutionary algorithms (DMOEAs) combined with transfer learning (TL) have been proven to be promising in solving DMOPs. TL-based DMOEAs showed advantages in reusing historical information and predicting high-quality solutions in the new environment. Various TL techniques have been employed to DMOEAs, which learn and transfer knowledge either in decision space or in objective space to predict the Pareto optimal solutions. However, problems usually have different types of change in decision and objective spaces. A single knowledge learning and transfer strategy may be unsuitable for all types of DMOPs. In this paper, a DMOEA with an adaptive knowledge learning and transfer strategy is proposed to solve DMOPs. It first estimates the change type of the problem when the environment changes, i.e., whether there exists change in decision or objective spaces, and then based on the change type, it adaptively chooses to learn and transfer knowledge in the decision space or objective space or both to generate an initial population that guides the search in new environment. A comprehensive empirical study is conducted to evaluate the performance of the proposed method. The method is compared to six state-of-the-art prediction-based DMOEAs on widely used DMOP benchmarks. Experimental results demonstrate that the proposed method outperforms or achieves comparable results to the compared algorithms on most of the test problems.