A knee-guided prediction approach for dynamic multi-objective optimization

Although dynamic multi-objective optimization problems dictate the evolutionary algorithms to quickly track the varying Pareto front when the environmental change occurs, the decision maker in the loop still needs to select a final optimal solution among a large number of candidate solutions before and after the environmental change. Most designs focus on searching for a well-distributed Pareto front which inadvertently demand excessive computational burden during the evolutionary process. In this paper, we propose a novel knee-guided prediction evolutionary algorithm (KPEA) which maintains non-dominated solutions near knee and boundary regions, in order to reduce the burden of maintaining a large and diversified population throughout the evolution process. When a change is detected, this design relocates the knee and boundary solutions based on the movement of the global knee solution in the new environment. In this way, this algorithm incurs a lower computational cost, allowing the evolutionary algorithm to converge quickly. In order to test the performance of the proposed algorithm, five popular dynamic multi-objective evolutionary algorithms (DMOEAs) are compared with KPEA based on two newly proposed metrics. The experimental results validate that the proposed algorithm effectively and efficiently converges to the global knee solution under the changing environments.