Nearest neighbor regression for evolutionary dynamic multiobjective optimization

Dynamic multi-objective optimization problems (DMOPs), characterized by time-varying objectives and constraints, demand algorithms capable of rapid adaptation while balancing convergence and diversity. Existing evolutionary algorithms are frequently subjected to significant online training costs due to their utilization of data-driven prediction models. A novel hybrid framework, MOEA/D-KNN, is proposed, integrating K-Nearest Neighbor (KNN) regression with the MOEA/D methodology. Within this framework, historical data is leveraged by KNN to dynamically predict Pareto-optimal solutions, enabling rapid adaptation to environmental changes. Simultaneously, the problem is decomposed by MOEA/D to facilitate an effective search strategy. Comprehensive empirical evaluation was conducted on standard DMOP benchmarks across diverse dynamic scenarios. MOEA/D-KNN is demonstrated to outperform state-of-the-art algorithms, particularly in managing abrupt and frequent environmental changes. Machine learning prediction is successfully bridged with evolutionary optimization through this approach, offering a robust and efficient solution for dynamic multi-objective challenges.