Deep variational reinforcement learning for POMDPs
Many real-world sequential decision making problems are partially observable by nature, and the environment model is typically unknown. Consequently, there is great need for reinforcement learning methods that can tackle such problems given only a stream of incomplete and noisy observations. In this...
| প্রধান লেখক: | , , , , |
|---|---|
| বিন্যাস: | Conference item |
| প্রকাশিত: |
Journal of Machine Learning Research
2018
|
| _version_ | 1826293943406428160 |
|---|---|
| author | Igl, M Zintgraf, L Le, T Wood, F Whiteson, S |
| author_facet | Igl, M Zintgraf, L Le, T Wood, F Whiteson, S |
| author_sort | Igl, M |
| collection | OXFORD |
| description | Many real-world sequential decision making problems are partially observable by nature, and the environment model is typically unknown. Consequently, there is great need for reinforcement learning methods that can tackle such problems given only a stream of incomplete and noisy observations. In this paper, we propose deep variational reinforcement learning (DVRL), which introduces an inductive bias that allows an agent to learn a generative model of the environment and perform inference in that model to effectively aggregate the available information. We develop an n-step approximation to the evidence lower bound (ELBO), allowing the model to be trained jointly with the policy. This ensures that the latent state representation is suitable for the control task. In experiments on Mountain Hike and flickering Atari we show that our method outperforms previous approaches relying on recurrent neural networks to encode the past. |
| first_indexed | 2024-03-07T03:37:59Z |
| format | Conference item |
| id | oxford-uuid:bced68dd-f0de-4c9a-a0b2-be6b1dd6ca8b |
| institution | University of Oxford |
| last_indexed | 2024-03-07T03:37:59Z |
| publishDate | 2018 |
| publisher | Journal of Machine Learning Research |
| record_format | dspace |
| spelling | oxford-uuid:bced68dd-f0de-4c9a-a0b2-be6b1dd6ca8b2022-03-27T05:28:05ZDeep variational reinforcement learning for POMDPsConference itemhttp://purl.org/coar/resource_type/c_5794uuid:bced68dd-f0de-4c9a-a0b2-be6b1dd6ca8bSymplectic Elements at OxfordJournal of Machine Learning Research2018Igl, MZintgraf, LLe, TWood, FWhiteson, SMany real-world sequential decision making problems are partially observable by nature, and the environment model is typically unknown. Consequently, there is great need for reinforcement learning methods that can tackle such problems given only a stream of incomplete and noisy observations. In this paper, we propose deep variational reinforcement learning (DVRL), which introduces an inductive bias that allows an agent to learn a generative model of the environment and perform inference in that model to effectively aggregate the available information. We develop an n-step approximation to the evidence lower bound (ELBO), allowing the model to be trained jointly with the policy. This ensures that the latent state representation is suitable for the control task. In experiments on Mountain Hike and flickering Atari we show that our method outperforms previous approaches relying on recurrent neural networks to encode the past. |
| spellingShingle | Igl, M Zintgraf, L Le, T Wood, F Whiteson, S Deep variational reinforcement learning for POMDPs |
| title | Deep variational reinforcement learning for POMDPs |
| title_full | Deep variational reinforcement learning for POMDPs |
| title_fullStr | Deep variational reinforcement learning for POMDPs |
| title_full_unstemmed | Deep variational reinforcement learning for POMDPs |
| title_short | Deep variational reinforcement learning for POMDPs |
| title_sort | deep variational reinforcement learning for pomdps |
| work_keys_str_mv | AT iglm deepvariationalreinforcementlearningforpomdps AT zintgrafl deepvariationalreinforcementlearningforpomdps AT let deepvariationalreinforcementlearningforpomdps AT woodf deepvariationalreinforcementlearningforpomdps AT whitesons deepvariationalreinforcementlearningforpomdps |