Attentive Relational State Representation for Intelligent Joint Operation Simulation

Renlong Chen, Ling Ye, Shaoqiu Zheng, Yabin Wang, Peng Cui, Ying Tan

Abstract: In the multi-agent task, due to the constant changes in the location and state of each agent, the information considered by each agent when making decisions is also constantly changing. This makes it difficult to model cooperatively among agents. Previous methods mainly used average embedding to model feature aggregation. However, this aggregation has the problem of losing permutation invariance or excessive information loss. The feature aggregation method based on attentive relational state representation establishes an insensitive state representation to permutation and problem scale. In our experiments on Intelligent Joint Operation Simulation, experimental results show that attentive relational state representation improves the baseline performance.