Genetic Algorithm for Curriculum Generation in Multi-Agent Reinforcement Learning
Primary Area: reinforcement learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Reinforcement Learning, Curriculum Learning, Genetic Algorithm, Multiagent Reinforcement Learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: As the deployment of autonomous agents increases in real life, there is an increased interest in extending their usage to competitive environments populated by other robots. Self-play in Reinforcement Learning (RL) allows agents to explore and learn competitive strategies. However, the complex dynamics of multi-agent RL interactions introduce instability in training and susceptibility to overfitting. Several game-theoretic approaches address the latter by generating approximate Nash equilibrium strategies to train against. The challenge of learning a policy in a complex and unstable multi-agent environment, the former, is not yet well addressed. This paper aims to address this issue by using a curriculum learning approach. We introduce curriculum design by a genetic algorithm to the multi-agent domain to more efficiently learn a policy that performs well and is stable at Nash equilibrium. Empirical studies show that our approach outperforms several strong baselines across various competitive two-player benchmarks in continuous control settings.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8862
Loading