Improving DNSGA-II based on Transfer Learning of Population Feature

Dynamic multi-objective optimization problems (DMOPs) are common in real-world applications. To effectively address these problems, algorithms are required to maintain solution diversity and quickly adapt to environmental changes. Transfer learning is a promising approach for tackling dynamic multi-objective optimization. In this paper, a transfer learning evolutionary optimization method based on population distribution characteristics is proposed for solving DMOPs. First, mutual information is used to calculate the correlation between the optimal solutions from two consecutive environments before the environmental change. Then, the distribution characteristics of the highly correlated population individuals are extracted as valuable knowledge for the transfer process. This information is employed to guide the generation of new population individuals, which are then randomly introduced into the initial population generated by the DNSGA-II-A algorithm in the new environment, forming an improved initial population. Experimental results demonstrate that, the proposed DMOEA-TLF exhibits significant advantages in convergence and diversity in most cases.