20 October 2025 in IEEE Transactions on Cybernetics
Selecting targets to attack and assigning weapons are among the most critical decisions on the battlefield. The decision problem is represented as a dynamic weapon-target assignment (DWTA) problem. While deep reinforcement learning (DRL) is the state-of-the-art approach for DWTA, previous studies have limitations in three key aspects: 1) representing topological relationships on the battlefield; 2) scalability to increased problem sizes; and 3) performance metric relevance. To overcome these limitations, this study aims to solve the DWTA problem by leveraging DRL and graph neural networks (GNNs), with a novel partially observable Markov decision process (POMDP) design, including graph-based action representation, observation features, and reward design. Experiments are conducted across multiple military domains, including naval and ground combat, comparing the proposed approach with existing heuristic and meta-heuristic methodologies. The effectiveness of the GNN and decision-making pattern is extensively analyzed through comprehensive experimental validation.
30 October 2025 in Journal of Computational Design and Engineering
This paper addresses a critical research gap in applying deep reinforcement learning to job shop scheduling problems by focusing on state feature design, which has been largely treated as an empirical matter. We present three key contributions: (1) a state feature design framework for reinforcement learning application in scheduling domain: dynamics equation (DE) and predicting state value based on nominal reward (PSVNR) perspectives, (2) the concept of ‘undercooked’ lower bounds that balance computational efficiency with bound tightness, and (3) successful integration of graph attention networks, transformer architecture, and the proposed ‘undercooked’ lower bounds into a unified framework. Our approach outperforms existing methods on 81$\%$ of the benchmark data sets, demonstrating exceptional generalizability and scalability. Through extensive experimentation, we explore and analyse the relationship between lower bound tightness and state value prediction accuracy, suggesting a meaningful connection between classical optimization techniques and modern learning-based approaches. The theoretical framework we develop provides clear guidelines for state feature design in reinforcement learning applications to scheduling problems, while our practical implementation shows significant performance improvements over state-of-the-art methods.
30 April 2026 in Engineering Applications of Artificial Intelligence
3 November 2025 in Applied Ocean Research