Deep Reinforcement Learning and the Deadly Triad
Document type :
Pré-publication ou Document de travail
Title :
Deep Reinforcement Learning and the Deadly Triad
Author(s) :
van Hasselt, Hado [Auteur]
DeepMind [London]
Doron, Yotam [Auteur]
DeepMind [London]
Strub, Florian [Auteur]
Sequential Learning [SEQUEL]
Université de Lille
Centre de Recherche en Informatique, Signal et Automatique de Lille - UMR 9189 [CRIStAL]
DeepMind [London]
Hessel, Matteo [Auteur]
DeepMind [London]
Sonnerat, Nicolas [Auteur]
DeepMind [London]
Modayil, Joseph [Auteur]
DeepMind [London]
DeepMind [London]
Doron, Yotam [Auteur]
DeepMind [London]
Strub, Florian [Auteur]
Sequential Learning [SEQUEL]
Université de Lille
Centre de Recherche en Informatique, Signal et Automatique de Lille - UMR 9189 [CRIStAL]
DeepMind [London]
Hessel, Matteo [Auteur]
DeepMind [London]
Sonnerat, Nicolas [Auteur]
DeepMind [London]
Modayil, Joseph [Auteur]
DeepMind [London]
HAL domain(s) :
Informatique [cs]/Réseau de neurones [cs.NE]
Informatique [cs]/Intelligence artificielle [cs.AI]
Informatique [cs]/Intelligence artificielle [cs.AI]
English abstract : [en]
We know from reinforcement learning theory that temporal difference learning can fail in certain cases. Sutton and Barto (2018) identify a deadly triad of function approximation, bootstrapping, and off-policy learning. ...
Show more >We know from reinforcement learning theory that temporal difference learning can fail in certain cases. Sutton and Barto (2018) identify a deadly triad of function approximation, bootstrapping, and off-policy learning. When these three properties are combined, learning can diverge with the value estimates becoming unbounded. However, several algorithms successfully combine these three properties, which indicates that there is at least a partial gap in our understanding. In this work, we investigate the impact of the deadly triad in practice, in the context of a family of popular deep reinforcement learning models - deep Q-networks trained with experience replay - analysing how the components of this system play a role in the emergence of the deadly triad, and in the agent's performanceShow less >
Show more >We know from reinforcement learning theory that temporal difference learning can fail in certain cases. Sutton and Barto (2018) identify a deadly triad of function approximation, bootstrapping, and off-policy learning. When these three properties are combined, learning can diverge with the value estimates becoming unbounded. However, several algorithms successfully combine these three properties, which indicates that there is at least a partial gap in our understanding. In this work, we investigate the impact of the deadly triad in practice, in the context of a family of popular deep reinforcement learning models - deep Q-networks trained with experience replay - analysing how the components of this system play a role in the emergence of the deadly triad, and in the agent's performanceShow less >
Language :
Anglais
Collections :
Source :
Files
- http://arxiv.org/pdf/1812.02648
- Open access
- Access the document
- 1812.02648
- Open access
- Access the document