Graph Learning-based Dynamic Multi-objective Optimization Algorithm

Dynamic Multi-objective Optimization Problems (DMOPs) involve multiple conflicting objective functions that evolve over time. Solving DMOPs requires Dynamic Multiobjective Optimization Algorithms (DMOAs) capable of tracking the dynamic evolution of the Pareto-optimal Front (POF) in real time. In rec...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:2025 7th International Conference on Data-driven Optimization of Complex Systems (DOCS) S. 283 - 288
Hauptverfasser: Wang, Chi, Song, Wei
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 19.08.2025
Schlagworte:
Accuracy
Benchmark testing
Computational modeling
Dynamic multi-objective optimization
Evolutionary computation
Graph Learning
Heuristic algorithms
Optimization
Partitioning algorithms
Prediction algorithms
Prediction model
Predictive models
Time series analysis
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Dynamic Multi-objective Optimization Problems (DMOPs) involve multiple conflicting objective functions that evolve over time. Solving DMOPs requires Dynamic Multiobjective Optimization Algorithms (DMOAs) capable of tracking the dynamic evolution of the Pareto-optimal Front (POF) in real time. In recent years, prediction-based DMOAs have demonstrated significant potential in solving DMOPs. However, most existing methods heavily rely on a limited number of Pareto-optimal Sets (POS) from past environments, hindering accurate modeling of the POS's temporal migration patterns, and most approaches model and predict each subspace independently, neglecting the spatial correlations and interdependencies among them. Additionally, the uneven distribution of individuals in the objective space may leave some subspaces sparsely populated, further limiting the algorithm's search capability. Faced with these challenges, this paper proposes a graph learning-based dynamic multi-objective optimization algorithm, termed GroupDMOEA. The algorithm begins by partitioning the objective space into multiple subspaces and constructing a graph structure based on the adjacency relationships among them. Graph Convolutional Network (GCN) is then employed to model and predict the migration trends of subspace centers, enabling a more comprehensive characterization of the dynamic evolution of the POS. To address the issue of missing individuals in certain subspaces, Group-DMOEA incorporates a local perturbationbased reconstruction mechanism, which utilizes historical information from neighboring subspaces to complete the sparse regions. We compare the proposed Group-DMOEA with six state-of-the-art DMOAs on fourteen benchmark test problems, and the experimental results demonstrate the excellent performance of Group-DMOEA in handling DMOPs
DOI:10.1109/DOCS67533.2025.11200665