Bayesian Bellman operators
We introduce a novel perspective on Bayesian reinforcement learning (RL); whereas existing approaches infer a posterior over the transition distribution or Q-function, we characterise the uncertainty in the Bellman operator. Our Bayesian Bellman operator (BBO) framework is motivated by the insight t...
| Main Authors: | , , |
|---|---|
| Format: | Conference item |
| Language: | English |
| Published: |
NeurIPS
2022
|
| _version_ | 1797106510564687872 |
|---|---|
| author | Fellows, M Hartikainen, K Whiteson, S |
| author_facet | Fellows, M Hartikainen, K Whiteson, S |
| author_sort | Fellows, M |
| collection | OXFORD |
| description | We introduce a novel perspective on Bayesian reinforcement learning (RL); whereas existing approaches infer a posterior over the transition distribution or Q-function, we characterise the uncertainty in the Bellman operator. Our Bayesian Bellman operator (BBO) framework is motivated by the insight that when bootstrapping is introduced, model-free approaches actually infer a posterior over Bellman operators, not value functions. In this paper, we use BBO to provide a rigorous theoretical analysis of model-free Bayesian RL to better understand its relationship to established frequentist RL methodologies. We prove that Bayesian solutions are consistent with frequentist RL solutions, even when approximate inference is used, and derive conditions for which convergence properties hold. Empirically, we demonstrate that algorithms derived from the BBO framework have sophisticated deep exploration properties that enable them to solve continuous control tasks at which state-of-the-art regularised actor-critic algorithms fail catastrophically.
|
| first_indexed | 2024-03-07T07:03:31Z |
| format | Conference item |
| id | oxford-uuid:9e50316f-7e83-4236-9e49-4546bbe03871 |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-03-07T07:03:31Z |
| publishDate | 2022 |
| publisher | NeurIPS |
| record_format | dspace |
| spelling | oxford-uuid:9e50316f-7e83-4236-9e49-4546bbe038712022-04-05T07:29:04ZBayesian Bellman operatorsConference itemhttp://purl.org/coar/resource_type/c_5794uuid:9e50316f-7e83-4236-9e49-4546bbe03871EnglishSymplectic ElementsNeurIPS2022Fellows, MHartikainen, KWhiteson, SWe introduce a novel perspective on Bayesian reinforcement learning (RL); whereas existing approaches infer a posterior over the transition distribution or Q-function, we characterise the uncertainty in the Bellman operator. Our Bayesian Bellman operator (BBO) framework is motivated by the insight that when bootstrapping is introduced, model-free approaches actually infer a posterior over Bellman operators, not value functions. In this paper, we use BBO to provide a rigorous theoretical analysis of model-free Bayesian RL to better understand its relationship to established frequentist RL methodologies. We prove that Bayesian solutions are consistent with frequentist RL solutions, even when approximate inference is used, and derive conditions for which convergence properties hold. Empirically, we demonstrate that algorithms derived from the BBO framework have sophisticated deep exploration properties that enable them to solve continuous control tasks at which state-of-the-art regularised actor-critic algorithms fail catastrophically. |
| spellingShingle | Fellows, M Hartikainen, K Whiteson, S Bayesian Bellman operators |
| title | Bayesian Bellman operators |
| title_full | Bayesian Bellman operators |
| title_fullStr | Bayesian Bellman operators |
| title_full_unstemmed | Bayesian Bellman operators |
| title_short | Bayesian Bellman operators |
| title_sort | bayesian bellman operators |
| work_keys_str_mv | AT fellowsm bayesianbellmanoperators AT hartikainenk bayesianbellmanoperators AT whitesons bayesianbellmanoperators |