Communication Adaptive Multi-Task Coupled Multi-Agent Reinforcement Learning Framework

Despite the impressive success achieved in various domains, multi-agent reinforcement learning (MARL) is still faced with the problem of incomplete information and how to deal with complex decision-making tasks with multi-task coupled. The information fusion algorithm ensures the consistency of the global state information by integrating local information to cope with information loss and communication interruption. The majority of prior work only considers the fusion prediction of single state information, while the continuously changing state information in real scenarios is relatively unexplored. Furthermore, traditional multi-agent reinforcement learning algorithms can only handle a single task, leaving open the important question of how to adapt to complex scenarios with multiple subtasks coupled. This paper proposes a novel framework called CAMM (Communication Adaptive Multi-Task Coupled Multi-Agent Reinforcement Learning Framework) to solve the above problems. CAMM provides a distributed fusion prediction method for continuously changing state information to obtain a unified global state for each agent, which leverages the continuity of state changes. CAMM designs a multi-task coupled multi-agent hierarchical reinforcement learning framework, which makes it feasible to decouple complex decision-making tasks into separate decision solutions for each task. As a result, CAMM can be easily combined with existing MARL algorithms and any feasible solution algorithm for each task. Experimental results show that CAMM significantly predicts and fuses state information, ensures the consistency of global state, and achieves comparable performance with the best solution in each subtask in multi-task decision-making environments.