Elitism-based transfer learning and diversity maintenance for dynamic multi-objective optimization

Abstract

In handling dynamic multi-objective optimization problems (DMOPs), transfer learning driven methods have received considerable attention for finding a high-quality initial population with good convergence and diversity performance to adapt to the new environment. However, they commonly suffer from loss of population diversity and high computational consumption. Therefore, this study proposes a hybrid method combining elitism-based transfer learning and diversity maintenance to efficiently identify a high-quality initial population in response to environmental changes. An elite selection mechanism is developed to select elite individuals from the memory pool when the environment changes. Subsequently, an elitism-based transfer learning method is proposed to predict individuals by leveraging the knowledge from the selected elite individuals, thereby improving the computational efficiency and quality of the solutions. Subsequently, a random diversity maintenance strategy is developed to generate diverse individuals within the regions where the predicted individuals are located to defy the loss of diversity in the population. Finally, the generated diverse and predicted individuals are merged to form an initial population to adapt to the new environment. The experimental results have demonstrated the competitiveness of the proposed algorithm for most DMOP test instances in terms of convergence, diversity, and computational efficiency.