A Novel Machine Learning-Guided Adaptive Optimizer for Dynamic Multi-Objective Problems

Conventional multiobjective optimization algorithms are challenged by high computational costs, slow convergence rates and a lack of adaptability to dynamic conditions in complex problem domains. This paper introduces Adaptive NSGA-III (AdaNSGA-III), a novel and improved version of Non-Dominated Sorting Genetic Algorithm III (NSGA-III). It leverages Gaussian Process Regression (GPR) and Proximal Policy Optimization (PPO) to effectively manage dynamic multi-objective optimization problems. GPR is used as a surrogate model that approximates objective functions, reducing the computational costs associated with costly evaluations. Additionally, PPO-based reinforcement learning adjusts genetic operators like crossover and mutation rates dynamically during evolution to balance exploration and exploitation. Key performance metrics were used to compare AdaNSGA-III against NSGA-III, MOEA/DD, and K-RVEA on real-world benchmark problems (DDMOP2 and DDMOP6), as well as standard test problems (DTLZ1-D, WFG2-D and DTLZ7). AdaNSGA-III achieved satisfactory solutions in fewer generations and outperformed other methods by attaining high hypervolume values and a high level of solution diversity, indicating a superior balance between exploration and exploitation. Statistical tests are used to validate the superiority and efficiency of AdaNSGA-III.