TMAC: a Transformer-based partially observable multi-agent communication method.

Area of Science:

• Artificial Intelligence
• Multi-Agent Systems
• Reinforcement Learning

Background:

• Effective communication is vital for coordinating multiple agents in complex environments.
• Partial observability and limited communication ranges hinder multi-agent collaboration.
• Existing methods struggle with sophisticated coordination tasks.

Purpose of the Study:

• To propose a novel algorithm for enhancing communication in partially observable multi-agent systems.
• To improve agents' ability to extract features and generate effective messages for collaboration.
• To address the challenges posed by limited observations and communication ranges.

Main Methods:

• Developed the Transformer-based Partially Observable Multi-Agent Communication (TMAC) algorithm.
• Incorporated a self-message fusing module for multi-source feature integration.
• Utilized Transformer architecture for improved feature extraction and output message generation.

Main Results:

• TMAC demonstrated improved performance in multi-agent communication tasks.
• Achieved 6% and 10% performance gains over baseline algorithms in survival and StarCraft Multi-Agent Challenge (SMAC) environments, respectively.
• Validated effectiveness in scenarios with limited local observations and neighbor-only communication.

Conclusions:

• The proposed TMAC algorithm significantly enhances multi-agent collaboration through improved communication.
• TMAC offers a robust solution for partially observable environments with communication constraints.
• The method shows promise for advancing research in coordinated multi-agent systems.