RoMAT: Role-based multi-agent transformer for generalizable heterogeneous cooperation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Multi-task multi-agent systems (MASs) present unique challenges due to heterogeneous agents and diverse task requirements.
• Existing single-agent policy networks struggle with policy sharing in MASs, leading to performance degradation.

Purpose of the Study:

• To propose a novel framework, Role-based Multi-Agent Transformer (RoMAT), for effectively modeling heterogeneous agents in MASs.
• To enable agents to adapt to different tasks and roles within a cooperative MAS setting.

Main Methods:

• RoMAT utilizes a sequence modeling approach combined with a role-based actor architecture.
• The framework features a modular design with shared backbone networks and independent role-based actors tailored to agent structures.

Main Results:

• RoMAT successfully captures the complex behavior patterns of heterogeneous agents in MASs.
• The proposed method demonstrates superior performance and generalization capabilities compared to existing approaches in both single and multi-task scenarios.

Conclusions:

• RoMAT offers an effective solution for modeling cooperative heterogeneous agents in multi-task environments.
• The role-based actor mechanism allows for flexible adaptation and improved performance in MASs.