Résumé
We propose the Bayes-UCBVI algorithm for reinforcement learning in tabular, stage-dependent, episodic Markov decision process: a natural extension of the Bayes-UCB algorithm by Kaufmann et al. (2012) for multi-armed bandits. Our method uses the quantile of a Q-value function posterior as upper confidence bound on the optimal Q-value function. For Bayes-UCBVI, we prove a regret bound of order Oe(√H3SAT) where H is the length of one episode, S is the number of states, A the number of actions, T the number of episodes, that matches the lower-bound of Ω(√H3SAT) up to poly-log terms in H, S, A, T for a large enough T. To the best of our knowledge, this is the first algorithm that obtains an optimal dependence on the horizon H (and S) without the need of an involved Bernstein-like bonus or noise. Crucial to our analysis is a new fine-grained anti-concentration bound for a weighted Dirichlet sum that can be of independent interest. We then explain how Bayes-UCBVI can be easily extended beyond the tabular setting, exhibiting a strong link between our algorithm and Bayesian bootstrap (Rubin, 1981).
| langue originale | Anglais |
|---|---|
| Pages (de - à) | 21380-21431 |
| Nombre de pages | 52 |
| journal | Proceedings of Machine Learning Research |
| Volume | 162 |
| état | Publié - 1 janv. 2022 |
| Modification externe | Oui |
| Evénement | 39th International Conference on Machine Learning, ICML 2022 - Baltimore, États-Unis Durée: 17 juil. 2022 → 23 juil. 2022 |
Empreinte digitale
Examiner les sujets de recherche de « From Dirichlet to Rubin: Optimistic Exploration in RL without Bonuses ». Ensemble, ils forment une empreinte digitale unique.Contient cette citation
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver