Feasibility-guided momentum prediction for dynamic constrained multi-objective optimization
•A subspace co-evolution framework with momentum-based centroid prediction is proposed.•Historical centroid trends are smoothed to predict future search directions adaptively.•A feasibility-guided perturbation strategy enhances robustness in dynamic constraints.•Extensive experiments show superior p...
Saved in:
| Published in: | Expert systems with applications Vol. 299; p. 129978 |
|---|---|
| Main Authors: | , , , , , , |
| Format: | Journal Article |
| Language: | English |
| Published: |
Elsevier Ltd
01.03.2026
|
| Subjects: | |
| ISSN: | 0957-4174 |
| Online Access: | Get full text |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | •A subspace co-evolution framework with momentum-based centroid prediction is proposed.•Historical centroid trends are smoothed to predict future search directions adaptively.•A feasibility-guided perturbation strategy enhances robustness in dynamic constraints.•Extensive experiments show superior performance in convergence and adaptability.
Dynamic constrained multi-objective optimization problems (DCMOPs) pose significant challenges due to the simultaneous changes in objectives and constraints, which lead to shifting feasible regions and disrupt the search process. Most existing dynamic constrained multi-objective evolutionary algorithms (DCMOEAs) emphasize convergence speed or diversity maintenance but fail to effectively balance feasibility preservation with adaptive prediction of search directions. To address these challenges, we propose a novel feasibility-guided momentum prediction (FGMP) algorithm. FGMP employs a subspace co-evolution framework that partitions the decision space into multiple subspaces and maintains historical centroids for each subspace. A momentum-based center prediction (MCP) strategy is introduced to capture evolutionary trends by smoothing centroid trajectories, thus providing reliable guidance for future search directions. In addition, a feasibility-guided perturbation strategy (FGPS) enhances feasibility by repairing predicted centroids with nearby feasible solutions or retaining unrepaired ones to preserve potentially valuable information. Extensive experiments on the DCP benchmark suite demonstrate that FGMP outperforms four state-of-the-art algorithms in terms of convergence, feasibility maintenance, and adaptability, highlighting its robustness in complex and rapidly changing environments.
[Display omitted]
Dynamic constrained multi-objective optimization problems (DCMOPs) are challenging due to the simultaneous change in both the objectives and constraints, often leading to dynamically changing feasible regions and disrupting the search progress. Existing dynamic constrained multi-objective evolutionary algorithms (DCMOEAs) mostly focus on convergence speed or diversity maintenance but struggle to balance feasibility preservation with adaptive search direction prediction under highly dynamic environments. To address these issues, we take the feasibility of the solutions simultaneously into consideration during the prediction process, thus two strategies, which include a momentum-based center prediction (MCP) strategy and a feasibility-guided perturbation strategy (FGPS), are introduced. A novel algorithm, feasibility-guided momentum prediction (FGMP), is proposed to deal with DCMOPs. FGMP divides the decision space into multiple subspaces and maintains historical centroids for each subspace. MCP exploits the trends of historical centroids in the decision space by smoothing their trajectories to predict future search directions, enabling guided and stable initialization after environmental changes. Meanwhile, FGPS improves the feasibility of predicted centroids by repairing them with reference to nearby feasible solutions, while retaining infeasible ones that cannot be repaired to preserve useful search information and enhance robustness. Extensive experiments on the DCP benchmark suite demonstrate that FGMP outperforms four state-of-the-art algorithms in convergence, feasibility maintenance, and adaptability, showing robust performance even in complex and rapidly changing environments. |
|---|---|
| ISSN: | 0957-4174 |
| DOI: | 10.1016/j.eswa.2025.129978 |