.png){kind=icon}
.png){kind=icon}
Return
TMAC: a Transformer-based partially observable multi-agent communication method
.png){kind=icon}
OA
AI
DOI：10.7717/peerj-cs.2758.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Effective communication plays a crucial role in coordinating the actions of multiple agents. Within the realm of multi-agent reinforcement learning, agents have the ability to share information with one another through communication channels, leading to enhanced learning outcomes and successful goal attainment. Agents are limited by their observations and communication ranges due to increasingly complex location arrangements, making multi-agent collaboration based on communication increasingly difficult. In this article, for multi-agent communication in some partially observable scenarios, we propose a Transformer-based Partially Observable Multi-Agent Communication algorithm (TMAC), which improves agents extracting features and generating output messages. Meanwhile, a self-message fusing module is proposed to obtain features from multiple sources. Therefore, agents can achieve better collaboration through communication. At the same time, we performed experimental verification in the surviving and the StarCraft Multi-Agent Challenge (SMAC) environments where agents had limited local observation and could only communicate with neighboring agents. In two test environments, our method achieves an improvement in performance 6% and 10% over the baseline algorithm, respectively. Our code is available at https://gitee.com/xs-lion/tmac.
Keywords:
Multi-agent reinforcement learning
Communication
Attention mechanism
Journal
IF:
2.5
Papers:
3.4K
Citations:
6.9K
Organization
No organization information available