A Dynamic Constrained Multi-objective Evolutionary Algorithm Based on Graph-Temporal Neural Network

Chenwen Ding, Jiaping Hu, Wenji Li, Hanyuan Zhang, Jiachun Huang, Zhaojun Wang, Biao Xu, Yun Li, Zhun Fan

Keywords: Evolutionary computation, Dynamic constrained multi-objective optimization, Graph convolutional networks.

TL;DR: We present a GCN–GRU–based DCMOEA that learns population topology and historical environmental changes to predict Pareto population migration, and outperforms five baselines on DCP benchmarks.

Abstract: Dynamic constrained multi-objective optimization problems (DCMOPs) are common in engineering applications. Their objectives and constraints change over time, requiring algorithms to adapt quickly and track the dynamic constrained Pareto front continuously. Existing DCMOEAs mainly predict or perturb individual solutions or the centroid of the Pareto set. They fail to exploit the overall spatial structure of the population. Moreover, they insufficiently utilize historical information, making it difficult to capture long-term change patterns of the dynamic environment. To address these issues, this paper proposes a dynamic constrained multi-objective evolutionary algorithm based on graph-temporal neural networks. The algorithm constructs population topology by partitioning subspaces and uses graph convolutional networks (GCN) to extract topological features among subspaces. Subsequently, it employs gated recurrent units (GRU) to learn migration trends of Pareto sets across historical environments, enabling accurate prediction of Pareto set distribution in new environments. Meanwhile, the algorithm integrates memory-based and diversity-based strategies to generate initial populations that balance convergence, feasibility, and diversity. Experimental results on the DCP test suite show that the proposed algorithm outperforms five representative DCMOEAs in most test scenarios, validating its effectiveness in solving DCMOPs.

Submission Number: 106