While multi-agent reinforcement learning algorithms have attracted many research interests,very few algorithms in the field were deployed in real-world scenarios due to their uninterpretablilty and sample inefficiency in the training process. In this work, we propose an algorithm to use meta-strategy as regulators to train multiagent deep reinforcement learning agents to account for these challenges. We also propose several approaches to solve for meta-strategies, including linear program based approaches and shortest cycle based approaches. Through experiments, we discuss the effectiveness of incorporating meta-strategy in reinforcement learning.

43 pages

Thesis Committee:
Fei Fang (Chair)
Zico Kolter

Srinivasan Seshan, Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu