Accelerate Distributed Task Allocation via Learning Cooperative Communication Policy
Zhen Xiong, Zhou Yu, Yexun Xi, Jie Li · 2025
Task allocation is essential for the collective efficiency of multi-agent systems. Distributed approaches are more suitable for large-scale fleets, where agents determine the optimal assignment via distributed communication and local optimization. However, existing methods often neglect the resource constraints and channel conflicts at the medium access control (MAC) layer of the communication network, yielding the well-known problems of hidden node and channel contention. Thus, this work proposes a learning framework for communication-aware task allocation policies. Features such as Bron- Kerbosch (BK), Value of Message (VoM), and channel access quality are incorporated into the agent observations. Actions are modeled as adaptive gating mechanisms for inter-agent communication. Moreover, a counterfactual reward is proposed to evaluate global task allocation value, assignment consistence and communication conflicts. Via Multi-Agent Proximal Policy Optimization (MAPPO), the learned policy significantly outperforms methods such as MTD3, QMIX, and VDN across numerous metrics, including allocation value, learning speed, and communication quality. Notably, applications of the proposed framework to five classical allocation algorithms all show consistent and drastic improvement in coordination efficiency, which validates its effectiveness and necessity. Finally, the self-aware communication policy is trained on a server that simulates a real distributed network environment, significantly reducing the communication bandwidth required by the task allocation algorithm.