An Expensive Many-Objective Optimization Algorithm Based on Hierarchical Uncertainty Weighted Ensemble and Adaptive Filling Strategy

Abstract

To address the challenges of surrogate model inaccuracy, imbalanced sample distribution, and difficulties in balancing exploration and exploitation posed by expensive evaluation costs and high-dimensional objective spaces in expensive many-objective optimization problems, this paper proposes a hierarchical integrated dropout-assisted many-objective evolutionary algorithm (HEDA-MOEA). Firstly, the algorithm constructs a dual-model integrated prediction framework that combines the global approximation capability of dropout neural networks with the local fitting advantages of influence point-insensitive kriging models, achieving adaptive fusion through uncertainty weighting. Secondly, a training sample selection strategy is designed based on cosine distance. When integrated with insensitive kriging modeling, this strategy effectively mitigates the impact of outliers and enhances sample diversity. Finally, a phase-adaptive multi-criteria selection mechanism is proposed to dynamically balance the weights of uncertainty, diversity, and convergence according to the search process, enabling precise control of exploration and exploitation behaviors. The experimental results demonstrate that the proposed HEDA-MOEA outperforms the other five algorithms on problems involving 3 to 10 objectives.