Knowledge-Guided Communication Preference Learning Model for Multi-agent Cooperation

In partially observable scenarios such as distributed multi-agent systems with limited perceptual ranges, information sharing among agents is particularly crucial. Existing research, however, mostly focuses on broadcast communication, which not only depends heavily on bandwidth efficiency, it also causes information redundancy that may have a detrimental impact on collaboration. To solve these problems, we present a novel communication model, named Preference-oriented Multi-agent Communication (PoMaC) which enables agents to learn a communication preference to choose valuable agents with whom to share information. Specifically, communication preference is defined as the communication probability, which represents the degree of influence between agents' policies. It is jointly generated by the preference network and prior knowledge, which constrains the communication selection space and improves the efficiency of network learning. After communication, the cooperative policy is generated by fusing shared messages from selected communication partners. The experimental results demonstrate that the proposed model outperforms the state-of-the-art methods in communication efficiency and results in improvement in team cooperation.