.png){kind=icon}
.png){kind=icon}
Return
Transformer Based Models for Offline Multi-agent ReinforcementLearning
.png){kind=icon}
PRE
AI
DOI：10.1007/978-3-032-07638-0_1.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Multi-Agent Reinforcement Learning (MARL) has shown great potential in solving complex decision-making problems, but traditional approaches often require extensive online interactions, making them computationally expensive and sample inefficient. Recent advances in transformer-based architectures, particularly Decision Transformers (DTs), offer an alternative paradigm by enabling sequential decision-making from offline datasets. While DTs have demonstrated success in single-agent RL, their effectiveness in multi-agent scenarios remains an open question. In this paper, we explore the application of DTs for offline MARL in game scenarios using the StarCraft Multi-Agent Challenge (SMAC) environment. We train a DT on an offline dataset of expert trajectories and evaluate its performance in an online environment, demonstrating that transformer-based models can effectively learn multi-agent policies from offline data, capturing long-term dependencies and strategic behaviors. Our study provides valuable insights into the feasibility of DTs for MARL in game scenarios, contributing to the growing field of transformer-based RL.
Keywords:
Multi-Agent Reinforcement Learning
Offline Reinforcement Learning
Decision Transformers
StarCraft II