Abstract The evolution of intelligent systems has spotlighted the significance of cooperative behavior among autonomous agents,particularly in dynamic environments where strategic interactions and real-time decision-making are crucial. In line with thescope of Frontiers in Computer Science, which emphasizes intelligent systems, artificial intelligence, and distributed computing,this study investigates advanced coordination strategies in multi-agent systems, aiming to optimize collective performance inreal-time tactical environments. Traditional models of multi-agent cooperation—be they centralized or decentralized—facechallenges in scalability, adaptability, and real-time communication under constrained conditions. Centralized models, whileoptimal in action planning, often falter under large-scale agent scenarios due to bottlenecks in processing and communication.Decentralized approaches, although more scalable, struggle with policy alignment and information asymmetry, leadingto suboptimal global behavior. A novel optimization framework, termed Hierarchical Policy Synchronization (HiPS), isproposed and incorporated into the Cooperative Decision-Making Network (CDMN) to tackle these limitations. This approachintroduces hierarchical decomposition, where agents are organized into coordination groups with synchronization protocols.Through local policy aggregation, inter-group meta-policy alignment, and a reinforcement learning objective regularized byfeedback consistency, the model achieves high scalability, robustness to agent failure, and efficient coordination in uncertainenvironments. Empirical evaluations validate that HiPS significantly improves convergence speed and cooperative efficacycompared to conventional decentralized strategies. By integrating adaptive communication and coordination-aware learning,our method aligns with the journal’s focus on intelligent algorithms and real-world applications of cooperative AI.
Ying Ji (Tue,) studied this question.