Search Trajectory Network-Enhanced Multi-Objective Dynamic Algorithm Configuration
Robbert Reijnen, Zaharah Allah Bukhsh, Hoong Chuin Lau, Yaoxin Wu, Yingqian Zhang · Frontiers in artificial intelligence and applications · 2025
Deep reinforcement learning (DRL) has emerged as an effective technique for dynamic algorithm configuration, particularly in evolutionary computation, enabling adaptive parameter updates during algorithmic execution. DRL-based methods have shown broad applicability across different problem domains and are designed to configure algorithms without problem-specific information, making them highly transferable across problem variants and scalable to different problem sizes. This paper proposes a novel graph neural network-based approach that learns representations of Search Trajectory Networks (STNs) to track the convergence behavior of multiple objectives and dynamically reconfigures multi-objective evolutionary algorithms during execution. By capturing how solutions evolve and interact over time, the STN-based state representation enables real-time insight into convergence, diversity, and their trade-offs, facilitating more informed and adaptive configuration decisions. Extensive experiments indicate that our method outperforms the state-of-the-art DRL-based algorithm configuration methods. It also demonstrates good scalability to large problem instances and effectiveness in real-world optimization problems, which are often computationally expensive to tune.