[PDF][PDF] Scalable and sample-efficient multi-agent imitation learning

W Jeon, P Barde, D Nowrouzezahrai… - Proceedings of the …, 2020 - researchgate.net
Proceedings of the Workshop on Artificial Intelligence Safety, co …, 2020researchgate.net
Multi-agent generative adversarial imitation learning (MAGAIL) is a recent approach that
extends single-agent GAIL to problems in multi-agent imitation learning. While MAGAIL
shows promising results on cooperative and competitive tasks, it requires agent-
environment interactions during training, which may reduce sample efficiency in practice.
Moreover, MAGAIL was validated empirically on only a handful of agents, and its scalability
to larger numbers of agents remains a question. We propose a multi-agent imitation learning …
Abstract
Multi-agent generative adversarial imitation learning (MAGAIL) is a recent approach that extends single-agent GAIL to problems in multi-agent imitation learning. While MAGAIL shows promising results on cooperative and competitive tasks, it requires agent-environment interactions during training, which may reduce sample efficiency in practice. Moreover, MAGAIL was validated empirically on only a handful of agents, and its scalability to larger numbers of agents remains a question. We propose a multi-agent imitation learning algorithm that addresses these issues. Specifically, we apply multi-agent actor-critic (MAAC) and multi-agent attention-actor-critic (MAA2C)–off-policy multi-agent reinforcement learning (MARL) approaches–in the MARL imitation learning inner loop, as opposed to MACK–the onpolicy MARL method used in MAGAIL. We then model centralized and decentralized discriminators to evaluate whether a given behavior results from agent or expert actions, defining reward functions for the MARL inner loop. We demonstrate that our method scales more effectively, and more sampleefficient, than MAGAIL. We also demonstrate that imitation learning with decentralized discriminators is robust, performing surprisingly well for a large number of agents compared to its centralized counterpart.
researchgate.net
以上显示的是最相近的搜索结果。 查看全部搜索结果

Google学术搜索按钮

example.edu/paper.pdf
搜索
获取 PDF 文件
引用
References