Timezone: »
Batch reinforcement learning (RL) is important to apply RL algorithms to many high stakes tasks. Doing batch RL in a way that yields a reliable new policy in large domains is challenging: a new decision policy may visit states and actions outside the support of the batch data, and function approximation and optimization with limited samples can further increase the potential of learning policies with overly optimistic estimates of their future performance. Some recent approaches to address these concerns have shown promise, but can still be overly optimistic in their expected outcomes. Theoretical work that provides strong guarantees on the performance of the output policy relies on a strong concentrability assumption, which makes it unsuitable for cases where the ratio between state-action distributions of behavior policy and some candidate policies is large. This is because, in the traditional analysis, the error bound scales up with this ratio. We show that using \emph{pessimistic value estimates} in the low-data regions in Bellman optimality and evaluation back-up can yield more adaptive and stronger guarantees when the concentrability assumption does not hold. In certain settings, they can find the approximately best policy within the state-action space explored by the batch data, without requiring a priori assumptions of concentrability. We highlight the necessity of our pessimistic update and the limitations of previous algorithms and analyses by illustrative MDP examples and demonstrate an empirical comparison of our algorithm and other state-of-the-art batch RL baselines in standard benchmarks.
Author Information
Yao Liu (Stanford University)
Adith Swaminathan (Microsoft Research)
Alekh Agarwal (Microsoft Research)
Emma Brunskill (Stanford University)
More from the Same Authors
-
2020 Poster: Policy Improvement via Imitation of Multiple Oracles »
Ching-An Cheng · Andrey Kolobov · Alekh Agarwal -
2020 Spotlight: Policy Improvement via Imitation of Multiple Oracles »
Ching-An Cheng · Andrey Kolobov · Alekh Agarwal -
2020 Poster: Off-policy Policy Evaluation For Sequential Decisions Under Unobserved Confounding »
Hongseok Namkoong · Ramtin Keramati · Steve Yadlowsky · Emma Brunskill -
2020 Poster: FLAMBE: Structural Complexity and Representation Learning of Low Rank MDPs »
Alekh Agarwal · Sham Kakade · Akshay Krishnamurthy · Wen Sun -
2020 Poster: PC-PG: Policy Cover Directed Exploration for Provable Policy Gradient Learning »
Alekh Agarwal · Mikael Henaff · Sham Kakade · Wen Sun -
2020 Poster: Provably Efficient Reward-Agnostic Navigation with Linear Value Iteration »
Andrea Zanette · Alessandro Lazaric · Mykel J Kochenderfer · Emma Brunskill -
2020 Oral: FLAMBE: Structural Complexity and Representation Learning of Low Rank MDPs »
Alekh Agarwal · Sham Kakade · Akshay Krishnamurthy · Wen Sun -
2020 Poster: Safe Reinforcement Learning via Curriculum Induction »
Matteo Turchetta · Andrey Kolobov · Shital Shah · Andreas Krause · Alekh Agarwal -
2020 Spotlight: Safe Reinforcement Learning via Curriculum Induction »
Matteo Turchetta · Andrey Kolobov · Shital Shah · Andreas Krause · Alekh Agarwal -
2019 Workshop: “Do the right thing”: machine learning and causal inference for improved decision making »
Michele Santacatterina · Thorsten Joachims · Nathan Kallus · Adith Swaminathan · David Sontag · Angela Zhou -
2019 Poster: Offline Contextual Bandits with High Probability Fairness Guarantees »
Blossom Metevier · Stephen Giguere · Sarah Brockman · Ari Kobren · Yuriy Brun · Emma Brunskill · Philip Thomas -
2019 Poster: Bias Correction of Learned Generative Models using Likelihood-Free Importance Weighting »
Aditya Grover · Jiaming Song · Ashish Kapoor · Kenneth Tran · Alekh Agarwal · Eric Horvitz · Stefano Ermon -
2019 Poster: Almost Horizon-Free Structure-Aware Best Policy Identification with a Generative Model »
Andrea Zanette · Mykel J Kochenderfer · Emma Brunskill -
2019 Poster: Limiting Extrapolation in Linear Approximate Value Iteration »
Andrea Zanette · Alessandro Lazaric · Mykel J Kochenderfer · Emma Brunskill -
2018 Poster: Representation Balancing MDPs for Off-policy Policy Evaluation »
Yao Liu · Omer Gottesman · Aniruddh Raghu · Matthieu Komorowski · Aldo Faisal · Finale Doshi-Velez · Emma Brunskill -
2018 Poster: On Oracle-Efficient PAC RL with Rich Observations »
Christoph Dann · Nan Jiang · Akshay Krishnamurthy · Alekh Agarwal · John Langford · Robert Schapire -
2018 Demonstration: Automatic Curriculum Generation Applied to Teaching Novices a Short Bach Piano Segment »
Emma Brunskill · Tong Mu · Karan Goel · Jonathan Bragg -
2018 Spotlight: On Oracle-Efficient PAC RL with Rich Observations »
Christoph Dann · Nan Jiang · Akshay Krishnamurthy · Alekh Agarwal · John Langford · Robert Schapire -
2017 Workshop: OPT 2017: Optimization for Machine Learning »
Suvrit Sra · Sashank J. Reddi · Alekh Agarwal · Benjamin Recht -
2017 Poster: Using Options and Covariance Testing for Long Horizon Off-Policy Policy Evaluation »
Zhaohan Guo · Philip S. Thomas · Emma Brunskill -
2017 Poster: Off-policy evaluation for slate recommendation »
Adith Swaminathan · Akshay Krishnamurthy · Alekh Agarwal · Miro Dudik · John Langford · Damien Jose · Imed Zitouni -
2017 Poster: Unifying PAC and Regret: Uniform PAC Bounds for Episodic Reinforcement Learning »
Christoph Dann · Tor Lattimore · Emma Brunskill -
2017 Spotlight: Unifying PAC and Regret: Uniform PAC Bounds for Episodic Reinforcement Learning »
Christoph Dann · Tor Lattimore · Emma Brunskill -
2017 Oral: Off-policy evaluation for slate recommendation »
Adith Swaminathan · Akshay Krishnamurthy · Alekh Agarwal · Miro Dudik · John Langford · Damien Jose · Imed Zitouni -
2017 Tutorial: Reinforcement Learning with People »
Emma Brunskill -
2016 Demonstration: Project Malmo - Minecraft for AI Research »
Katja Hofmann · Matthew A Johnson · Fernando Diaz · Alekh Agarwal · Tim Hutton · David Bignell · Evelyne Viegas -
2016 Poster: Efficient Second Order Online Learning by Sketching »
Haipeng Luo · Alekh Agarwal · Nicolò Cesa-Bianchi · John Langford -
2016 Poster: Contextual semibandits via supervised learning oracles »
Akshay Krishnamurthy · Alekh Agarwal · Miro Dudik -
2016 Poster: PAC Reinforcement Learning with Rich Observations »
Akshay Krishnamurthy · Alekh Agarwal · John Langford -
2015 Workshop: Optimization for Machine Learning (OPT2015) »
Suvrit Sra · Alekh Agarwal · Leon Bottou · Sashank J. Reddi -
2015 Poster: Efficient and Parsimonious Agnostic Active Learning »
Tzu-Kuo Huang · Alekh Agarwal · Daniel Hsu · John Langford · Robert Schapire -
2015 Spotlight: Efficient and Parsimonious Agnostic Active Learning »
Tzu-Kuo Huang · Alekh Agarwal · Daniel Hsu · John Langford · Robert Schapire -
2015 Poster: Fast Convergence of Regularized Learning in Games »
Vasilis Syrgkanis · Alekh Agarwal · Haipeng Luo · Robert Schapire -
2015 Oral: Fast Convergence of Regularized Learning in Games »
Vasilis Syrgkanis · Alekh Agarwal · Haipeng Luo · Robert Schapire -
2014 Workshop: OPT2014: Optimization for Machine Learning »
Zaid Harchaoui · Suvrit Sra · Alekh Agarwal · Martin Jaggi · Miro Dudik · Aaditya Ramdas · Jean Lasserre · Yoshua Bengio · Amir Beck -
2014 Poster: Scalable Non-linear Learning with Adaptive Polynomial Expansions »
Alekh Agarwal · Alina Beygelzimer · Daniel Hsu · John Langford · Matus J Telgarsky -
2013 Workshop: Learning Faster From Easy Data »
Peter Grünwald · Wouter M Koolen · Sasha Rakhlin · Nati Srebro · Alekh Agarwal · Karthik Sridharan · Tim van Erven · Sebastien Bubeck -
2013 Workshop: OPT2013: Optimization for Machine Learning »
Suvrit Sra · Alekh Agarwal -
2012 Workshop: Optimization for Machine Learning »
Suvrit Sra · Alekh Agarwal -
2012 Poster: Stochastic optimization and sparse statistical recovery: Optimal algorithms for high dimensions »
Alekh Agarwal · Sahand N Negahban · Martin J Wainwright -
2011 Workshop: Computational Trade-offs in Statistical Learning »
Alekh Agarwal · Sasha Rakhlin -
2011 Poster: Distributed Delayed Stochastic Optimization »
Alekh Agarwal · John Duchi -
2011 Poster: Stochastic convex optimization with bandit feedback »
Alekh Agarwal · Dean P Foster · Daniel Hsu · Sham M Kakade · Sasha Rakhlin -
2010 Workshop: Learning on Cores, Clusters, and Clouds »
Alekh Agarwal · Lawrence Cayton · Ofer Dekel · John Duchi · John Langford -
2010 Spotlight: Distributed Dual Averaging In Networks »
John Duchi · Alekh Agarwal · Martin J Wainwright -
2010 Poster: Distributed Dual Averaging In Networks »
John Duchi · Alekh Agarwal · Martin J Wainwright -
2010 Oral: Fast global convergence rates of gradient methods for high-dimensional statistical recovery »
Alekh Agarwal · Sahand N Negahban · Martin J Wainwright -
2010 Poster: Fast global convergence rates of gradient methods for high-dimensional statistical recovery »
Alekh Agarwal · Sahand N Negahban · Martin J Wainwright -
2009 Poster: Information-theoretic lower bounds on the oracle complexity of convex optimization »
Alekh Agarwal · Peter Bartlett · Pradeep Ravikumar · Martin J Wainwright -
2009 Spotlight: Information-theoretic lower bounds on the oracle complexity of convex optimization »
Alekh Agarwal · Peter Bartlett · Pradeep Ravikumar · Martin J Wainwright -
2007 Poster: An Analysis of Inference with the Universum »
Fabian H Sinz · Olivier Chapelle · Alekh Agarwal · Bernhard Schölkopf -
2007 Spotlight: An Analysis of Inference with the Universum »
Fabian H Sinz · Olivier Chapelle · Alekh Agarwal · Bernhard Schölkopf