مدل سازي ارتباط موثر در سيستم هاي چند عاملياز طريق رويكردهاي مبتني بر گراف

Multi-Agent Reinforcement Learning (MARL) is recognized as a prominent approach for solving complex an‎d distributed problems. Nevertheless, achieving efficient coordination among a large number of intelligent agents remains a fundamental challenge. Current models often rely on ”all-to-all” communication paradigms; an approach that not only imposes a heavy computational burden but also leads to the propagation of unnecessary an‎d irrelevant information, resulting in model confusion during the learning process. Conversely, methods that delegate the task of discovering communication structures entirely to neural networks tend to exhibit unstable performance an‎d generate noisy outputs in complex environments. Furthermore, disregarding causal dependencies among agent decisions an‎d persisting with simultaneous decision-making hinders the formation of precise, coordinated strategies. To address these limitations, this research introduces a novel architecture titled ”Factor Prioritized Transformer” (FPT). The core innovation of this model lies in integrating intelligent communication structuring with dynamic sequential decision-making. In this architecture, agent interactions are modeled within a bipartite graph framework, where agents are organized into overlapping groups represented by intermediary nodes called ”factors.” This structure ensures that information flow is channeled exclusively through these factors, thereby reducing complexity an‎d preventing the formation of redundant connections. At the decision-making layer, this study moves beyond the simultaneous paradigm by employing a novel sequential action generation mechanism. Utilizing probabilistic sampling an‎d assigning credit scores to each agent’s observations, this mechanism dynamically determines which agent possesses critical information at any given moment an‎d should be prioritized in decision-making. Consequently, subsequent agents can adjust their actions with awareness of the decisions made by key agents. The efficacy of the proposed method was rigorously eva‎luated in Google Research Football (GRF) an‎d StarCraft II (SMAC) simulation environments, which require high levels of cooperation. Empirical results demonstrate that the proposed model, achieving a 92% win rate in challenging scenarios an‎d significant improvements in final scores, exhibits far superior an‎d more stable performance compared to stateof- the-art algorithms. These findings substantiate that combining a coherent, controlled communication structure with a priority-based decision-making process is a key solution for overcoming scalability an‎d noise challenges in multi-agent systems.

5

167423