E-MCTS: Deep Exploration by Planning with Epistemic Uncertainty
Keywords: Reinforcement learning, model based, exploration, uncertainty, planning
TL;DR: We incorporate epistemic uncertainty into MCTS to achieve uncertainty estimation in node-value predictions, and use the value uncertainty for deep exploration.
Abstract: Deep model-based reinforcement learning (MBRL) is responsible for many of the greatest achievements of reinforcement learning. At the core of two of the approaches responsible for those successes, Alpha/MuZero, is a modified version of the Monte-Carlo Tree Search (MCTS) planning algorithm, replacing components of MCTS with learned models (of value and/or environment dynamics). Dedicated deep exploration, however, is a remaining challenge of Alpha/MuZero and by extension MCTS-based methods with learned models. To overcome this challenge, we develop Epistemic-MCTS. E-MCTS extends MCTS with estimation and propagation of epistemic uncertainty, and leverages the propagated uncertainty for a novel deep exploration algorithm by explicitly planning to explore. We incorporate E-MCTS into variations of MCTS-based MBRL approaches with learned (MuZero) and provided (AlphaZero) dynamics models. We compare E-MCTS to non-planning based deep-exploration baselines and demonstrate that E-MCTS significantly outperforms them in the investigated deep exploration benchmark.
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
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: 2004
Loading