Understanding the Transfer of High-Level Reinforcement Learning Skills Across Diverse Environments
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, representation learning, multi-task reinforcement learning, transfer learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: A large number of reinforcement learning (RL) environments are available to the research community. However, due to differences across these environments, it is difficult to transfer skills learnt by a RL agent from one environment to another. For this transfer learning problem, a multitask RL perspective is considered in this paper, the goal being to transfer the skills from one environment to another using a single policy. To achieve such goal, we design an environment agnostic policy that enables the sharing of skills. Our experimental results demonstrate that: (a) by training on both desired environments using standard RL algorithms, the skills can be transferred from one environment to another; (b) by changing the amount of data that the RL algorithm uses to optimize the policy and value functions, we
show empirically that the transfer of knowledge between different environments is possible, and results in learning tasks with up to 84% fewer gradient update steps. This study takes an important step towards enabling more effective transfer of skills by learning in multitask RL scenarios across diverse environments by designing skill-sharing, sample-efficient RL training protocols.
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: 3692
Loading