Romat: Role-Based Multi-Agent Transformer for Generalizable Heterogeneous Cooperation

12 Pages Posted: 17 May 2023

See all articles by Dongzi Wang

Dongzi Wang

National University of Defense Technology

Fangwei Zhong

Peking University

Muning Wen

Shanghai Jiao Tong University (SJTU)

Minglong Li

National University of Defense Technology

Yuanxi Peng

National University of Defense Technology

Teng Li

National University of Defense Technology

Yaodong Yang

Peking University

Abstract

Multi-task multi-agent systems (MASs) are challenging to model because they involve heterogeneous agents with different behavior patterns that need to cooperate across various tasks. Existing networks for single-agent policies are not suitable for this setting, as they cannot share policies among agents without losing task-specific performance. We propose a novel framework called Role-based Multi-Agent Transformer (RoMAT), which uses a sequence modeling technique and a role-based actor to enable agents to adapt to different tasks and roles in MASs. RoMAT has a modular model architecture, where backbone networks are shared by all agents, but a small part of the parameters (role-based actor) is independent, depending on the agents’ exclusive structures. We evaluate RoMAT on several benchmark tasks and show that it can capture the behavior patterns of heterogeneous agents and achieve better performance and generalization than other methods in both single and multi-task settings.

Keywords: Multi-Agent System, Multi-Task Generalization, imitation learning

Suggested Citation

Wang, Dongzi and Zhong, Fangwei and Wen, Muning and Li, Minglong and Peng, Yuanxi and Li, Teng and Yang, Yaodong, Romat: Role-Based Multi-Agent Transformer for Generalizable Heterogeneous Cooperation. Available at SSRN: https://ssrn.com/abstract=4437059 or http://dx.doi.org/10.2139/ssrn.4437059

Dongzi Wang

National University of Defense Technology ( email )

Fangwei Zhong

Peking University ( email )

No. 38 Xueyuan Road
Haidian District
Beijing, 100871
China

Muning Wen

Shanghai Jiao Tong University (SJTU) ( email )

Minglong Li

National University of Defense Technology ( email )

Yuanxi Peng

National University of Defense Technology ( email )

Teng Li

National University of Defense Technology ( email )

Yaodong Yang (Contact Author)

Peking University ( email )

No. 38 Xueyuan Road
Haidian District
Beijing, 100871
China

Do you have negative results from your research you’d like to share?

Paper statistics

Downloads
81
Abstract Views
197
Rank
541,576
PlumX Metrics