Communication-Aware Consensus-Based Distributed Task Allocation Based on Asynchronous Reinforcement Learning
Yizhe Cao, Jie Li, Zhen Xiong, Yexun Xi · 2025
Due to the impact of communication protocols and channel resources on algorithm performance, distributed task allocation in multi-agent systems is still challenging. To address the challenges of asynchronous collaboration and communication efficiency in task allocation, this paper proposes a communication strategy learning framework for task allocation. This framework models the action as an adaptive gate for communication between agents, and using the global task conflict change as a shared reward. Due to the fact that agents may fail to act synchronously in real-world scenarios, we propose the Async-VDN ((Asynchronous Value Decomposition Networks) algorithm, which innovatively integrate the Asynchronous Consensus-Based Bundle Algorithm (ACBBA) with off-policy multi-agent reinforcement learning (VDN) for more accurate policy updates by decoupled trajectory collection and value decomposition networks. Experimental results demonstrate stable reward convergence in the scenarios of 15 and 30 agents, and a reduction in required communication bandwidth on the selforganizing network simulation platform, validating the effectiveness of asynchronous MARL frameworks for dynamic task allocation.