Deep Reinforcement Learning Based Task-Oriented Communication in Multi-Agent Systems
Guojun He, Mingjie Feng, Yu Zhang, Guanghua Liu, Yueyue Dai, Tao Jiang
Abstract
Driven by the increasing demand for executing intelligent tasks in various fields, multi-agent system (MAS) has drawn significant attention recently. An MAS relies on efficient communication between agents to exchange task-relevant information, so as support cooperative operation. Meanwhile, traditional communication systems are bit-oriented, which neglect the content and task relevance of the transmitted data. Thus, if bit-oriented communication patterns are applied in a MAS, a significant amount of task-irrelevant data would be transmitted, leading to communication resource waste and low operational efficiency. Considering that many emerging MASs are data-intensive and delay-sensitive, traditional ways of communication are unfit for these MASs. Task-oriented communication is a promising solution to deal with this issue, but its application in MAS still faces various challenges. In this article, we propose a task-oriented communication based framework for MAS, aiming to support efficient cooperation among agents. This framework specifies the collection, transmission, and processing of task-relevant information, in which task relevance is fully utilized to enhance communication efficiency. Based on the proposed framework, we then apply deep reinforcement learning (DRL) to implement task-oriented communication, in which a modular design and an end-to-end design for information extraction, data transmission, and task execution are proposed. Finally, the open problems for future research are discussed.