Targeted Multi-Agent Communication with Deep Metric Learning

novel targeted multi-agent communication model based on deep metric learning (DMLTarMAC) is proposed in this paper. The nonlinear relationship between the internal state of an agent and the received message is described by the deep metric learning (DML) module in DMLTarMAC. Compared with the scheme using a linear relationship, DMLTarMAC can improve the accuracy and effectiveness of the receiver's attention. In order to reveal the advantage of the proposed DMLTarMAC, it is evaluated in cooperative and competitive multi-agent tasks with different difficulty levels and environment settings. The experimental results show that DMLTarMAC outperforms the benchmarks, especially in challenging settings. Furthermore, the ablation experiments demonstrate that agents' communication and behavior strategies are effective and intuitive.