A Dynamic Multi-objective Optimization Algorithm for Black-Box Optimization in Dynamic Irregular Environment

Abstract

This paper proposes a novel initial population prediction algorithm (IPPA) for solving irregular black-box dynamic multi-objective optimization problems. In the phase of environmental change detection, IPPA find representative solutions based on a clustering algorithm and form environment vector to judge whether the environment has changed. In the population initialization stage, the initial population is generated by three mechanisms. First, a feedforward neural network (FNN) is trained with historical population information, and part of the initial population is generated by FNN. Second, when the new environment is similar to the last environment, part of the elite population in the last environment is retained. Finally, some populations are randomly generated to maintain population diversity. Since IPPA does not require the use of explicit objective functions and recent solutions, it can solve black-box dynamic multi-objective optimization problems with drastic and irregularly changing Pareto set. The proposed algorithm is compared with other state-of-the-art algorithms on a number of benchmark problems. Experimental results show that the proposed IPPA algorithm is effective for irregular blackbox dynamic multi-objective optimization problems.