There Is No Turning Back: A Self-Supervised ...
Document type :
Communication dans un congrès avec actes
Title :
There Is No Turning Back: A Self-Supervised Approach for Reversibility-Aware Reinforcement Learning
Author(s) :
Grinsztajn, Nathan [Auteur]
Centre de Recherche en Informatique, Signal et Automatique de Lille - UMR 9189 [CRIStAL]
Scool [Scool]
Ferret, Johan [Auteur]
Pietquin, Olivier [Auteur]
Preux, Philippe [Auteur]
Geist, Matthieu [Auteur]
Centre de Recherche en Informatique, Signal et Automatique de Lille - UMR 9189 [CRIStAL]
Scool [Scool]
Ferret, Johan [Auteur]
Pietquin, Olivier [Auteur]
Preux, Philippe [Auteur]
Geist, Matthieu [Auteur]
Conference title :
Neural Information Processing Systems (2021)
City :
Virtual
Country :
France
Start date of the conference :
2021-12-06
Book title :
Proc. Thirty-fifth Conference on Neural Information Processing Systems
HAL domain(s) :
Informatique [cs]/Apprentissage [cs.LG]
Informatique [cs]/Intelligence artificielle [cs.AI]
Informatique [cs]/Intelligence artificielle [cs.AI]
English abstract : [en]
We propose to learn to distinguish reversible from irreversible actions for better informed decision-making in Reinforcement Learning (RL). From theoretical considerations, we show that approximate reversibility can be ...
Show more >We propose to learn to distinguish reversible from irreversible actions for better informed decision-making in Reinforcement Learning (RL). From theoretical considerations, we show that approximate reversibility can be learned through a simple surrogate task: ranking randomly sampled trajectory events in chronological order. Intuitively, pairs of events that are always observed in the same order are likely to be separated by an irreversible sequence of actions. Conveniently, learning the temporal order of events can be done in a fully self-supervised way, which we use to estimate the reversibility of actions from experience, without any priors. We propose two different strategies that incorporate reversibility in RL agents, one strategy for exploration (RAE) and one strategy for control (RAC). We demonstrate the potential of reversibility-aware agents in several environments, including the challenging Sokoban game. In synthetic tasks, we show that we can learn control policies that never fail and reduce to zero the side-effects of interactions, even without access to the reward function.Show less >
Show more >We propose to learn to distinguish reversible from irreversible actions for better informed decision-making in Reinforcement Learning (RL). From theoretical considerations, we show that approximate reversibility can be learned through a simple surrogate task: ranking randomly sampled trajectory events in chronological order. Intuitively, pairs of events that are always observed in the same order are likely to be separated by an irreversible sequence of actions. Conveniently, learning the temporal order of events can be done in a fully self-supervised way, which we use to estimate the reversibility of actions from experience, without any priors. We propose two different strategies that incorporate reversibility in RL agents, one strategy for exploration (RAE) and one strategy for control (RAC). We demonstrate the potential of reversibility-aware agents in several environments, including the challenging Sokoban game. In synthetic tasks, we show that we can learn control policies that never fail and reduce to zero the side-effects of interactions, even without access to the reward function.Show less >
Language :
Anglais
Peer reviewed article :
Oui
Audience :
Internationale
Popular science :
Non
Collections :
Source :
Files
- https://hal.archives-ouvertes.fr/hal-03454640/document
- Open access
- Access the document
- https://hal.archives-ouvertes.fr/hal-03454640/document
- Open access
- Access the document
- https://hal.archives-ouvertes.fr/hal-03454640/document
- Open access
- Access the document
- document
- Open access
- Access the document
- Reversibility_Aware_Reinforcement_Learning__NeurIPS_.pdf
- Open access
- Access the document