Timezone: »
Reinforcement learning with function approximation can be unstable and even divergent, especially when combined with off-policy learning and Bellman updates. In deep reinforcement learning, these issues have been dealt with empirically by adapting and regularizing the representation, in particular with auxiliary tasks. This suggests that representation learning may provide a means to guarantee stability. In this paper, we formally show that there are indeed nontrivial state representations under which the canonical SARSA algorithm is stable, even when learning off-policy. We analyze representation learning schemes that are based on the transition matrix of a policy, such as proto-value functions, along three axes: approximation error, stability, and ease of estimation. In the most general case of a defective transition matrix, we show that a Schur basis provides convergence guarantees, but is difficult to estimate from samples. For a fixed reward function, we find that an orthogonal basis of the corresponding Krylov subspace is an even better choice. We conclude by empirically demonstrating that these stable representations can be learned using stochastic gradient descent, opening the door to improved techniques for representation learning with deep networks.
Author Information
Dibya Ghosh (Google)
Marc Bellemare (Google Brain)
More from the Same Authors
-
2021 : Why Generalization in RL is Difficult: Epistemic POMDPs and Implicit Partial Observability »
Dibya Ghosh · Jad Rahme · Aviral Kumar · Amy Zhang · Ryan P. Adams · Sergey Levine -
2022 : Distributionally Adaptive Meta Reinforcement Learning »
Anurag Ajay · Dibya Ghosh · Sergey Levine · Pulkit Agrawal · Abhishek Gupta -
2022 : Distributionally Adaptive Meta Reinforcement Learning »
Anurag Ajay · Dibya Ghosh · Sergey Levine · Pulkit Agrawal · Abhishek Gupta -
2022 Poster: Distributional Hamilton-Jacobi-Bellman Equations for Continuous-Time Reinforcement Learning »
Harley Wiltzer · David Meger · Marc Bellemare -
2022 Poster: Offline RL Policies Should Be Trained to be Adaptive »
Dibya Ghosh · Anurag Ajay · Pulkit Agrawal · Sergey Levine -
2022 Oral: Offline RL Policies Should Be Trained to be Adaptive »
Dibya Ghosh · Anurag Ajay · Pulkit Agrawal · Sergey Levine -
2022 Spotlight: Distributional Hamilton-Jacobi-Bellman Equations for Continuous-Time Reinforcement Learning »
Harley Wiltzer · David Meger · Marc Bellemare -
2021 Social: RL Social »
Dibya Ghosh · Hager Radi · Derek Li · Alex Ayoub · Erfan Miahi · Rishabh Agarwal · Charline Le Lan · Abhishek Naik · John Martin · Shruti Mishra · Adrien Ali Taiga -
2019 Poster: Statistics and Samples in Distributional Reinforcement Learning »
Mark Rowland · Robert Dadashi · Saurabh Kumar · Remi Munos · Marc Bellemare · Will Dabney -
2019 Oral: Statistics and Samples in Distributional Reinforcement Learning »
Mark Rowland · Robert Dadashi · Saurabh Kumar · Remi Munos · Marc Bellemare · Will Dabney -
2019 Poster: The Value Function Polytope in Reinforcement Learning »
Robert Dadashi · Marc Bellemare · Adrien Ali Taiga · Nicolas Le Roux · Dale Schuurmans -
2019 Poster: DeepMDP: Learning Continuous Latent Space Models for Representation Learning »
Carles Gelada · Saurabh Kumar · Jacob Buckman · Ofir Nachum · Marc Bellemare -
2019 Oral: The Value Function Polytope in Reinforcement Learning »
Robert Dadashi · Marc Bellemare · Adrien Ali Taiga · Nicolas Le Roux · Dale Schuurmans -
2019 Oral: DeepMDP: Learning Continuous Latent Space Models for Representation Learning »
Carles Gelada · Saurabh Kumar · Jacob Buckman · Ofir Nachum · Marc Bellemare -
2017 : Panel Discussion »
Balaraman Ravindran · Chelsea Finn · Alessandro Lazaric · Katja Hofmann · Marc Bellemare -
2017 : Marc G. Bellemare: The role of density models in reinforcement learning »
Marc Bellemare -
2017 Poster: Count-Based Exploration with Neural Density Models »
Georg Ostrovski · Marc Bellemare · Aäron van den Oord · Remi Munos -
2017 Talk: Count-Based Exploration with Neural Density Models »
Georg Ostrovski · Marc Bellemare · Aäron van den Oord · Remi Munos -
2017 Poster: A Laplacian Framework for Option Discovery in Reinforcement Learning »
Marlos C. Machado · Marc Bellemare · Michael Bowling -
2017 Poster: A Distributional Perspective on Reinforcement Learning »
Marc Bellemare · Will Dabney · Remi Munos -
2017 Poster: Automated Curriculum Learning for Neural Networks »
Alex Graves · Marc Bellemare · Jacob Menick · Remi Munos · Koray Kavukcuoglu -
2017 Talk: A Laplacian Framework for Option Discovery in Reinforcement Learning »
Marlos C. Machado · Marc Bellemare · Michael Bowling -
2017 Talk: A Distributional Perspective on Reinforcement Learning »
Marc Bellemare · Will Dabney · Remi Munos -
2017 Talk: Automated Curriculum Learning for Neural Networks »
Alex Graves · Marc Bellemare · Jacob Menick · Remi Munos · Koray Kavukcuoglu