Regularized Behavior Value Estimation

Offline reinforcement learning restricts the learning process to rely only on logged-data without access to an environment. While this enables real-world applications, it also poses unique challenges. One important challenge is dealing with errors caused by the over-estimation of values for state-action pairs not well-covered by the training data. Due to bootstrapping, these errors get amplified during training and can lead to divergence, thereby crippling learning. To overcome this challenge, we introduce Regularized Behavior Value Estimation (R-BVE). Unlike most approaches, which use policy improvement during training, R-BVE estimates the value of the behavior policy during training and only performs policy improvement at deployment time. Further, R-BVE uses a ranking regularisation term that favours actions in the dataset that lead to successful outcomes. We provide ample empirical evidence of R-BVE’s effectiveness, including state-of-theart performance on the RL Unplugged ATARI dataset. We also test R-BVE on new datasets, from bsuite and a challenging DeepMind Lab task, and show that R-BVE outperforms other state-ofthe-art discrete control offline RL methods.

[1]  Pieter Abbeel,et al.  Towards Characterizing Divergence in Deep Q-Learning , 2019, ArXiv.

[2]  Thorsten Joachims,et al.  MOReL : Model-Based Offline Reinforcement Learning , 2020, NeurIPS.

[3]  Matthew W. Hoffman,et al.  Distributed Distributional Deterministic Policy Gradients , 2018, ICLR.

[4]  Jiayu Zhou,et al.  Ranking Policy Gradient , 2019, ICLR.

[5]  Pierre Geurts,et al.  Tree-Based Batch Mode Reinforcement Learning , 2005, J. Mach. Learn. Res..

[6]  S. Sathiya Keerthi,et al.  Efficient algorithms for ranking with SVMs , 2010, Information Retrieval.

[7]  Stefan Schaal,et al.  2008 Special Issue: Reinforcement learning of motor skills with policy gradients , 2008 .

[8]  Joelle Pineau,et al.  Benchmarking Batch Deep Reinforcement Learning Algorithms , 2019, ArXiv.

[9]  Sergio Gomez Colmenarejo,et al.  RL Unplugged: Benchmarks for Offline Reinforcement Learning , 2020, ArXiv.

[10]  Demis Hassabis,et al.  Mastering the game of Go without human knowledge , 2017, Nature.

[11]  Jimmy Ba,et al.  Adam: A Method for Stochastic Optimization , 2014, ICLR.

[12]  Tom Schaul,et al.  Prioritized Experience Replay , 2015, ICLR.

[13]  Mohammad Norouzi,et al.  An Optimistic Perspective on Offline Reinforcement Learning , 2020, ICML.

[14]  Sergey Levine,et al.  Stabilizing Off-Policy Q-Learning via Bootstrapping Error Reduction , 2019, NeurIPS.

[15]  Richard S. Sutton,et al.  Reinforcement Learning: An Introduction , 1998, IEEE Trans. Neural Networks.

[16]  Tor Lattimore,et al.  Behaviour Suite for Reinforcement Learning , 2019, ICLR.

[17]  Dale Schuurmans,et al.  Striving for Simplicity in Off-policy Deep Reinforcement Learning , 2019, ArXiv.

[18]  Lantao Yu,et al.  MOPO: Model-based Offline Policy Optimization , 2020, NeurIPS.

[19]  S. Levine,et al.  Offline Reinforcement Learning: Tutorial, Review, and Perspectives on Open Problems , 2020, ArXiv.

[20]  Sergio Gomez Colmenarejo,et al.  Acme: A Research Framework for Distributed Reinforcement Learning , 2020, ArXiv.

[21]  Martin A. Riedmiller,et al.  Batch Reinforcement Learning , 2012, Reinforcement Learning.

[22]  Dean Pomerleau,et al.  ALVINN, an autonomous land vehicle in a neural network , 2015 .

[23]  Jürgen Schmidhuber,et al.  A possibility for implementing curiosity and boredom in model-building neural controllers , 1991 .

[24]  Yuval Tassa,et al.  Continuous control with deep reinforcement learning , 2015, ICLR.

[25]  Marc G. Bellemare,et al.  A Distributional Perspective on Reinforcement Learning , 2017, ICML.

[26]  Rémi Munos,et al.  Observe and Look Further: Achieving Consistent Performance on Atari , 2018, ArXiv.

[27]  Craig Boutilier,et al.  ConQUR: Mitigating Delusional Bias in Deep Q-learning , 2020, ICML.

[28]  Shane Legg,et al.  Human-level control through deep reinforcement learning , 2015, Nature.

[29]  Sergey Levine,et al.  D4RL: Datasets for Deep Data-Driven Reinforcement Learning , 2020, ArXiv.

[30]  Michail G. Lagoudakis,et al.  Least-Squares Policy Iteration , 2003, J. Mach. Learn. Res..

[31]  Rémi Munos,et al.  Recurrent Experience Replay in Distributed Reinforcement Learning , 2018, ICLR.

[32]  Sergey Levine,et al.  Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor , 2018, ICML.

[33]  Steven J. Bradtke,et al.  Linear Least-Squares algorithms for temporal difference learning , 2004, Machine Learning.

[34]  Martin A. Riedmiller.  Neural Fitted Q Iteration - First Experiences with a Data Efficient Neural Reinforcement Learning Method , 2005, ECML.

[35]  Jakub W. Pachocki,et al.  Dota 2 with Large Scale Deep Reinforcement Learning , 2019, ArXiv.

[36]  Tom Schaul,et al.  Rainbow: Combining Improvements in Deep Reinforcement Learning , 2017, AAAI.

[37]  Gregory N. Hullender,et al.  Learning to rank using gradient descent , 2005, ICML.

[38]  Jakub W. Pachocki,et al.  Learning dexterous in-hand manipulation , 2018, Int. J. Robotics Res..

[39]  Doina Precup,et al.  Off-Policy Deep Reinforcement Learning without Exploration , 2018, ICML.

[40]  S. Levine,et al.  Conservative Q-Learning for Offline Reinforcement Learning , 2020, NeurIPS.

[41]  Natasha Jaques,et al.  Way Off-Policy Batch Deep Reinforcement Learning of Implicit Human Preferences in Dialog , 2019, ArXiv.

[42]  Nando de Freitas,et al.  Critic Regularized Regression , 2020, NeurIPS.

[43]  Seyed Kamyar Seyed Ghasemipour,et al.  EMaQ: Expected-Max Q-Learning Operator for Simple Yet Effective Offline and Online RL , 2020, ICML.

[44]  Marc G. Bellemare,et al.  The Arcade Learning Environment: An Evaluation Platform for General Agents (Extended Abstract) , 2012, IJCAI.

[45]  Tie-Yan Liu,et al.  Ranking Measures and Loss Functions in Learning to Rank , 2009, NIPS.

[46]  Ronald J. Williams,et al.  Simple Statistical Gradient-Following Algorithms for Connectionist Reinforcement Learning , 2004, Machine Learning.

[47]  Herke van Hoof,et al.  Addressing Function Approximation Error in Actor-Critic Methods , 2018, ICML.