Learning and Planning in Complex Action Spaces

Many important real-world problems have action spaces that are high-dimensional, continuous or both, making full enumeration of all possible actions infeasible. Instead, only small subsets of actions can be sampled for the purpose of policy evaluation and improvement. In this paper, we propose a general framework to reason in a principled way about policy evaluation and improvement over such sampled action subsets. This sample-based policy iteration framework can in principle be applied to any reinforcement learning algorithm based upon policy iteration. Concretely, we propose Sampled MuZero, an extension of the MuZero algorithm that is able to learn in domains with arbitrarily complex action spaces by planning over sampled actions. We demonstrate this approach on the classical board game of Go and on two continuous control benchmark domains: DeepMind Control Suite and Real-World RL Suite.

[1]  Sergey Levine,et al.  Advantage-Weighted Regression: Simple and Scalable Off-Policy Reinforcement Learning , 2019, ArXiv.

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

[3]  Richard S. Sutton,et al.  Learning to predict by the methods of temporal differences , 1988, Machine Learning.

[4]  Nicolas Le Roux,et al.  An operator view of policy gradient methods , 2020, NeurIPS.

[5]  Jian Sun,et al.  Identity Mappings in Deep Residual Networks , 2016, ECCV.

[6]  Guy Lever,et al.  Deterministic Policy Gradient Algorithms , 2014, ICML.

[7]  Alan Fern,et al.  Dream and Search to Control: Latent Space Planning for Continuous Control , 2020, ArXiv.

[8]  Nir Levine,et al.  An empirical investigation of the challenges of real-world reinforcement learning , 2020, ArXiv.

[9]  Demis Hassabis,et al.  Mastering Atari, Go, chess and shogi by planning with a learned model , 2019, Nature.

[10]  John Rust.  Using Randomization to Break the Curse of Dimensionality , 1997 .

[11]  Geoffrey E. Hinton,et al.  Layer Normalization , 2016, ArXiv.

[12]  Demis Hassabis,et al.  Mastering the game of Go with deep neural networks and tree search , 2016, Nature.

[13]  H. Jaap van den Herik,et al.  Progressive Strategies for Monte-Carlo Tree Search , 2008 .

[14]  Yunhao Tang,et al.  Discretizing Continuous Action Space for On-Policy Optimization , 2019, AAAI.

[15]  Ruben Villegas,et al.  Learning Latent Dynamics for Planning from Pixels , 2018, ICML.

[16]  Frank Hutter,et al.  Fixing Weight Decay Regularization in Adam , 2017, ArXiv.

[17]  Alec Radford,et al.  Proximal Policy Optimization Algorithms , 2017, ArXiv.

[18]  Mohammad Norouzi,et al.  Dream to Control: Learning Behaviors by Latent Imagination , 2019, ICLR.

[19]  Rémi Coulom,et al.  Efficient Selectivity and Backup Operators in Monte-Carlo Tree Search , 2006, Computers and Games.

[20]  Yoshua Bengio,et al.  Probabilistic Planning with Sequential Monte Carlo methods , 2018, ICLR.

[21]  Michal Valko,et al.  Monte-Carlo Tree Search as Regularized Policy Optimization , 2020, ICML.

[22]  D. Rubin,et al.  The calculation of posterior distributions by data augmentation , 1987 .

[23]  Junhyuk Oh,et al.  A Self-Tuning Actor-Critic Algorithm , 2020, NeurIPS.

[24]  Demis Hassabis,et al.  A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play , 2018, Science.

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

[26]  Sergey Levine,et al.  Trust Region Policy Optimization , 2015, ICML.

[27]  Jackie Kay,et al.  Local Search for Policy Iteration in Continuous Control , 2020, ArXiv.

[28]  Noam Brown,et al.  Superhuman AI for heads-up no-limit poker: Libratus beats top professionals , 2018, Science.

[29]  Byron Boots,et al.  Information Theoretic Model Predictive Q-Learning , 2020, L4DC.

[30]  Martin A. Riedmiller,et al.  Imagined Value Gradients: Model-Based Policy Optimization with Transferable Latent Dynamics Models , 2019, CoRL.

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

[32]  Kevin Waugh,et al.  DeepStack: Expert-level artificial intelligence in heads-up no-limit poker , 2017, Science.

[33]  Nicolas Heess,et al.  Hierarchical visuomotor control of humanoids , 2018, ICLR.

[34]  Peng Wei,et al.  Continuous Control for Searching and Planning with a Learned Model , 2020, ArXiv.

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

[36]  Yuval Tassa,et al.  MuJoCo: A physics engine for model-based control , 2012, 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems.

[37]  H. Francis Song,et al.  V-MPO: On-Policy Maximum a Posteriori Policy Optimization for Discrete and Continuous Control , 2019, ICLR.

[38]  Yishay Mansour,et al.  A Sparse Sampling Algorithm for Near-Optimal Planning in Large Markov Decision Processes , 1999, Machine Learning.

[39]  Jürgen Schmidhuber,et al.  Long Short-Term Memory , 1997, Neural Computation.

[40]  Richard Evans,et al.  Deep Reinforcement Learning in Large Discrete Action Spaces , 2015, 1512.07679.

[41]  Catholijn M. Jonker,et al.  A0C: Alpha Zero in Continuous Action Space , 2018, ArXiv.

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