CS394R: Reinforcement Learning: Theory and Practice -- Fall 2016: Resources Page

Resources for Reinforcement Learning: Theory and Practice

Week 0: Class Overview, Introduction

  • Slides from week 0: Peter's; Scott's.

  • Week 1: Multi-armed Bandits and MDPs

  • Slides from Tuesday: pdf.
  • Gradient Bandit Slides from Thursday: pdf.
  • Ch.3 Slides from Thursday: pdf.
  • The one from Shivaram Kalyanakrishnan: pdf.
  • Readings from a past version of the course:
  • Sections 1, 2, 4, and 5 and the proof of Theorem 1 in Section 3. The proof of Theorem 3 and the appendices are optional.
    UCB: Finite-time Analysis of the Multiarmed Bandit Problem
    Peter Auer, Nicolo Cesa-Bianchi, and Paul Fischer
  • Sections 1, 2, 3.1, 4, and 5. The details of the proof (Sections 3.2-3.4) are optional.
    Thompson Sampling: an asymptotically optimal finite-time analysis
    Emilie Kaufmann, Nathaniel Korda, and Remi Munos
  • Csaba Szepesvari's banditalgs.com.
  • Vermorel and Mohri: Multi-Armed Bandit Algorithms and Empirical Evaluation.
  • Shivaram Kalyanakrishnan and Peter Stone: Efficient Selection of Multiple Bandit Arms: Theory and Practice. In ICML 2010. Here are some related slides.
  • An RL reading list from Shivaram Kalyanakrishnan.
  • An Empirical Evaluation of Thompson Sampling
    Olivier Chapelle and Lihong Li
    NIPS 2011

  • Week 2: Dynamic Programming and Monte Carlo Methods

  • Slides from Tuesday: pdf
  • Slides from Thursday: pdf
  • A paper on "On the Complexity of solving MDPs" (Littman, Dean, and Kaelbling, 1995).
  • Pashenkova, Rish, and Dechter: Value Iteration and Policy Iteration Algorithms for Markov Decision Problems.
  • Some slides on robot localization that include information on importance sampling.

  • Week 3: TD Learning and n-step Bootstrapping

  • Slides from Thursday: pdf.
  • Harm van Seijen, Hado van Hasselt, Shimon Whiteson, and Marco Wiering, A Theoretical and Empirical Analysis of Expected Sarsa. In ADPRL 2009.
  • A Q-learning video
  • A blog post explaining why double learning helps deal with maximization bias.

  • Week 4: Planning

  • Slides from week 4: The planning ones; The ones on TEXPLORE; Slides on MCTS; Slides on POMDPs.
  • Slides by Alan Fern on Monte Carlo Tree Search and UCT
  • On the Analysis of Complex Backup Strategies in Monte Carlo Tree Search by Khandelwal, Liebman, Stone, and Niekum.
  • A Survey of Monte Carlo Tree Search Methodsby Browne et al.
  • The Dependence of Effective Planning Horizon on Model Accuracy
    by Nan Jiang, Alex Kulesza, Satinder Singh, and Richard Lewis.
    In International Conference on Autonomous Agents and MultiAgent Systems (AAMAS), 2015.
  • A survey on Bayesian RL by Ghavamzadeh et al.
  • A paper that uses MCTS for robot learning: TEXPLORE: Real-Time Sample-Efficient Reinforcement Learning for Robots, by Hester and Stone.

  • Week 5: On-policy Prediction with Approximation

  • Slides from week 5: pdf.
  • Evolutionary Function Approximation by Shimon Whiteson.
  • Dopamine: generalization and Bonuses (2002) Kakade and Dayan.
  • Keepaway Soccer: From Machine Learning Testbed to Benchmark - a paper that compares CMAC, RBF, and NN function approximators on the same task.
  • Boyan, J. A., and A. W. Moore, Generalization in Reinforcement Learning: Safely Approximating the Value Function. In Tesauro, G., D. S. Touretzky, and T. K. Leen (eds.), Advances in Neural Information Processing Systems 7 (NIPS). MIT Press, 1995. Another example of function approximation divergence and a proposed solution.
  • Experiments with Reinforcement Learning in Problems with Continuous State and Action Spaces (1998) Juan Carlos Santamaria, Richard S. Sutton, Ashwin Ram. Comparisons of several types of function approximators (including instance-based like Kanerva).
  • Binary action search for learning continuous-action control policies (2009). Pazis and Lagoudakis. (slides)
  • Least-Squares Temporal Difference Learning Justin Boyan.
  • A Convergent Form of Approximate Policy Iteration (2002) T. J. Perkins and D. Precup. A convergence guarantee with function approximation.
  • Moore and Atkeson: The Parti-game Algorithm for Variable Resolution Reinforcement Learning in Multidimensional State Spaces.
  • Sherstov and Stone: Function Approximation via Tile Coding: Automating Parameter Choice.
  • Chapman and Kaelbling: Input Generalization in Delayed Reinforcement Learning: An Algorithm and Performance Comparisons.
  • Sašo Džeroski, Luc De Raedt and Kurt Driessens: Relational Reinforcement Learning.
  • Sprague and Ballard: Multiple-Goal Reinforcement Learning with Modular Sarsa(0).
  • A post on Deep Q learning. another

  • Week 6: On Policy Control with Approximation and Off Policy Methods with Approximation

  • Slides from week 6, Thursday (Ch 11): pdf.
  • Experiments with Reinforcement Learning in Problems with Continuous State and Action Spaces (1998) Juan Carlos Santamaria, Richard S. Sutton, Ashwin Ram. Comparisons of several types of function approximators (including instance-based like Kanerva).
  • Average Reward Reinforcement Learning: Foundations, Algorithms, and Empirical Results by Mahadaven.
  • Discounted Reinforcement Learning is Not an Optimization Problem by Naik, Shariff, Yasui, and Sutton.
  • Toward Off-Policy Learning Control with Function Approximation
    Maei et al. ICML 2010 - solves Baird's counterexample - Greedy-GQ for linear function approximation control
  • Residual Algorithms: Reinforcement Learning with Function Approximation (1995) Leemon Baird. More on the Baird counterexample as well as an alternative to doing gradient descent on the MSE.

  • Week 7: Eligibility Traces

  • Rich Sutton's slides: pdf
  • The keepaway slides: pdf
  • The corrected version of the eligibility trace example shown in class: pdf
  • The forward and backward views of TD(lambda) are equivalent.
  • Dayan: The Convergence of TD(&lambda) for General &lambda.
  • A paper that addresses relationship between first-visit and every-visit MC (Singh and Sutton, 1996). For some theoretical relationships see section starting at section 3.3 (and referenced appendices). The equivalence of MC and first visit TD(1) is proven starting in Section 2.4.
  • The paper that introduced Dutch traces and off-policy true on-line TD
  • An empirical analysis of true on-line TD: True Online Temporal-Difference Learning by van Seijen et al. (includes comparison to replacing traces)
  • Convergent Temporal-Difference Learning with Arbitrary Smooth Function Approximation
    Maei et al. NIPS 2009 - GTD for nonlinear function approximation policy evaluation
  • Train faster, generalize better: Stability of stochastic gradient descent by Moritz Hardt, Benjamin Recht, and Yoram Singer
  • Keepaway PASS slides, GETOPEN slides and the keepaway main page
  • An extensive empirical study of many different linear TD algorithms by Adam White and Martha White (AAMAS 2016).

  • Week 8: Policy Gradient Methods

  • Slides from class: pdf.
  • Policy Gradient Reinforcement Learning for Fast Quadrupedal Locomotion.
    Nate Kohl and Peter Stone
    In Proceedings of the IEEE International Conference on Robotics and Automation, May 2004.
  • This paper compares the policy gradient RL method with other algorithms on the walk learning: Machine Learning for Fast Quadrupedal Locomotion. Kohl and Stone. AAAI 2004.
  • Overview of Policy Gradient Methods by Jan Peters: http://www.scholarpedia.org/article/Policy_gradient_methods
  • from Jan Peters' group: Policy Search for Motor Primitives in Robotics
  • Szita and Lörincz: Learning Tetris Using the Noisy Cross-Entropy Method.
  • Autonomous helicopter flight via reinforcement learning.
    Andrew Ng, H. Jin Kim, Michael Jordan and Shankar Sastry.
    In S. Thrun, L. Saul, and B. Schoelkopf (Eds.), Advances in Neural Information Processing Systems (NIPS) 17, 2004.
  • PEGASUS: A policy search method for large MDPs and POMDPs.
    Andrew Ng and Michael Jordan
    Some of the helicopter videos learned with PEGASUS.
  • Autonomous Helicopter Control using Reinforcement Learning Policy Search Methods.
    J. Bagnell and J. Schneider
    Proceedings of the International Conference on Robotics and Automation 2001, IEEE, May, 2001.
  • Autonomous reinforcement learning on raw visual input data in a real world application.
    Sascha Lange, Martin Riedmiller, Arne Voigtlander.
    IJCNN 2012.
  • A couple of articles on the details of actor-critic in practice by Tsitsklis and by Williams.
  • Natural Actor Critic.
    Jan Peters and Stefan Schaal
    Neurocomputing 2008. Earlier version in ECML 2005.
  • PILCO: A Model-Based and Data-Efficient Approach to Policy Search.
    Marc Peter Deisenroth and Carl Edward Rasmussen
    ICML 2011
  • The original policy gradient RL paper.
  • Guided policy search
    Sergey Levine and Vladlen Koltun.
    ICML 2013.
    associated videos
  • Learning Neural Network Policies with Guided Policy Search under Unknown Dynamics
    Sergey Levine, Pieter Abbeel. NIPS 2014.
  • Trust Region policy optimization
    John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, Pieter Abbeel. ICML 2015.
  • A post by Karpathy on deep RL including with policy gradients (repeated from week 5)
  • Characterizing Reinforcement Learning Methods through Parameterized Learning Problems
    Shivaram Kalyanakrishnan and Peter Stone.
    Machine Learning (MLJ), 84(1--2):205-47, July 2011.
  • Comparing Evolutionary and Temporal Difference Methods for Reinforcement Learning.
    Matthew Taylor, Shimon Whiteson, and Peter Stone.
    In Proceedings of the Genetic and Evolutionary Computation Conference, pp. 1321-28, July 2006.
  • Evolutionary Function Approximation for RL. Whiteson and Stone, MLJ 2006.
  • Spinning up in Deep Reinforcement Learning Resources from OpenAI on getting up to speed using Deep Reinforcement Learning

  • Week 9: Case Studies and Applications

  • Guided policy search for visual manipulation: pdf video
  • OpenAI hide and seek: blog post
  • OpenAI Rubik's cube: blog post
  • The slides I showed on understanding what Deep RL nodes have learned (in particular LSTM units in a partially observable environment).
  • The slides I showed on AlphaGo
  • The GGP ones.
  • The NEAT+Q ones.
  • Some minimax slides: ppt.
  • Motif backgammon (online player)
  • Tesauro, G., Temporal Difference Learning and TD-Gammon. Communication of the ACM, 1995
  • Practical Issues in Temporal Difference Learning: an earlier paper by Tesauro (with a few more details)
  • Pollack, J.B., & Blair, A.D. Co-evolution in the successful learning of backgammon strategy. Machine Learning, 1998
  • Tesauro, G. Comments on Co-Evolution in the Successful Learning of Backgammon Strategy. Machine Learning, 1998.
  • Modular Neural Networks for Learning Context-Dependent Game Strategies, Justin Boyan, 1992: a partial replication of TD-gammon.
  • A fairly complete overview of one of the first applications of UCT to Go: "Monte-Carlo Tree Search and Rapid Action Value Estimation in Computer Go". Gelly and Silver. AIJ 2011.
  • Some papers from Simon Lucas' group on comparing TD learning and co-evolution in various games: Othello; Go; Treasure hunt.
  • S. Gelly and D. Silver. Achieving Master-Level Play in 9x9 Computer Go. In Proceedings of the 23rd Conference on Artificial Intelligence, Nectar Track (AAAI-08), 2008. Also available from here.
  • Simulation-Based Approach to General Game Playing
    Hilmar Finnsson and Yngvi Bjornsson
    AAAI 2008.
  • Some papers from the UT Learning Agents Research Group on General Game Playing
  • Deep Reinforcement Learning with Double Q-learning.
    Hado van Hasselt, Arthur Guez, David Silver
  • Scaling Reinforcement Learning toward RoboCup Soccer.
    Peter Stone and Richard S. Sutton.
    Proceedings of the Eighteenth International Conference on Machine Learning, pp. 537-544, Morgan Kaufmann, San Francisco, CA, 2001.
  • The UT Austin Villa RoboCup team home page.
  • Greg Kuhlmann's follow-up on progress in 3v2 keepaway
  • Kalyanakishnan et al.: Model-based Reinforcement Learning in a Complex Domain.
  • Making a Robot Learn to Play Soccer Using Reward and Punishment.
    Heiko Müller, Martin Lauer, Roland Hafner, Sascha Lange, Artur Merke and Martin Riedmiller.
    30th Annual German Conference on AI, KI 2007.
  • Reinforcement Learning for Sensing Strategies.
    C. Kwok and D. Fox.
    Proceedings of IROS, 2004.
  • Learning to trade via direct reinforcement
    John Moody and Matthew Saffell
    IEEE Transactions on Neural Networks, 2001.
  • Reinforcement learning for optimized trade execution
    Yuriy Nevmyvaka, Yi Feng, and Michael Kearns
    ICML 2006
  • Adaptive Treatment of Epilepsy via Batch-mode Reinforcement Learning
    Arthur Guez, Robert D. Vincent, Massimo Avoli, Joelle Pineau.
    IAAI 2008
  • PAC Optimal Planning for Invasive Species Management: Improved Exploration for Reinforcement Learning from Simulator-Defined MDPs.
    Thomas G. Dietterich, Majid Taleghan, and Mark Crowley
    AAAI 2013.
  • Design, Analysis, and Learning Control of a Fully Actuated Micro Wind Turbine.
    J. Zico Kolter, Zachary Jackowski, Russ Tedrake
    American Controls Conference 2012.

  • Week 10: Abstraction: Options and Hierarchy

  • Tuesday slides, most courtesy of George Konidaris: pdf.
  • Konidaris and Barto's paper on skill chaining: pdf.
  • Bacon, Harb, and Precup's paper on the Option-Critic architecture: pdf.
  • Vezhnevets el. al paper on feudal networks: pdf.
  • Hausman et. al paper on learning skill embeddings: pdf.
  • Abel et. al paper on value-preserving abstractions: pdf.
  • Tom Dietterich's classic paper on MAXQ: pdf.
  • The Journal version of the MaxQ paper
  • A follow-up paper on eliminating irrelevant variables within a subtask: State Abstraction in MAXQ Hierarchical Reinforcement Learning
  • Automatic Discovery and Transfer of MAXQ Hierarchies (from Dietterich's group - 2008)
  • Hierarchical Model-Based Reinforcement Learning: Rmax + MAXQ.
    Proceedings of the 25th International Conference on Machine Learning, 2008.
    Nicholas K. Jong and Peter Stone
  • Automatic Discovery of Subgoals in RL using Diverse Density by McGovern and Barto.
  • Improved Automatic Discovery of Subgoals for Options in Hierarchical Reinforcement Learning by Kretchmar et al.
  • Nick Jong and Todd Hester's paper on the utility of temporal abstraction. The slides.
  • Lihong Li and Thomas J. Walsh and Michael L. Littman, Towards a Unified Theory of State Abstraction for MDPs , Ninth International Symposium on Artificial Intelligence and Mathematics , 2006.
  • Tom Dietterich's tutorial on abstraction.
  • Nick Jong's paper on state abstraction discovery. The slides.
  • Nick Jong's Thesis code repository and annotated slides

  • Week 11: Learning from Human Input

  • Tuesday slides on imitation learning and IRL
  • TAMER slides
  • Deep TAMER slides
  • behavioral cloning from observation (BCO) slides
  • Some slides on inverse RL from Pieter Abbeel.
  • Maximum Entropy IRL: pdf.
  • Bayesian IRL: pdf.
  • Generative Adversarial Imitation Learning: pdf.
  • Behavioral Cloning from Observations: pdf.
  • Dataset Aggregation (DAgger): pdf.
  • IRL with rankings: pdf.
  • Niekum et al. Learning to assemble IKEA from demonstrations: pdf.
  • Knox et al. Understanding Human Teaching Modalities in Reinforcement LearningEnvironments: A Preliminary Report: pdf.
  • Kaochar et al. Towards Understanding How Humans Teach Robots: pdf.
  • Toris et al. A Practical Comparison of Three Robot Learning fromDemonstration Algorithms: pdf.
  • sequential TAMER+RL and other follow-up papers by Brad Knox.
  • The deep TAMER paper.
  • The BCO paper.
  • Towards Resolving Unidentifiability in Inverse Reinforcement Learning.
    Kareem Amin and Satinder Singh
  • Nonlinear Inverse Reinforcement Learning with Gaussian Processes
    Sergey Levine, Zoran Popovic, Vladlen Koltun.
  • Inverse Reinforcement Learning in Partially Observable Environments
    Jaedeug Choi and Kee-Eung Kim
  • Some papers on IRL and learning by demonstration
  • Deep Apprenticeship Learning for Playing Video Games
  • Maximum Entropy Deep Inverse Reinforcement Learning
  • Generative Adversarial Imitation Learning
  • Generative Adversarial Imitation from Observation
  • Creating Advice-Taking Reinforcement Learners.
    Richard Maclin and Jude Shavlik.
    Machine Learning, 22, pp. 251-281, 1996.
  • Gregory Kuhlmann, Peter Stone, Raymond Mooney, and Jude Shavlik: Guiding a Reinforcement Learner with Natural Language Advice: Initial Results in RoboCup Soccer.
  • Sonia Chernova and Manuela Veloso: Confidence-Based Policy Learning from Demonstration Using Gaussian Mixture Models.

  • Week 12: Multiagent RL and Safe RL

  • Slides from week 12 from Michael Bowling: pdf
  • The ones on grid game pdf
  • Slides on safe RL and IRL pdf
  • Journal version of WoLF
  • Doran Chakraborty and Peter Stone
    Convergence, Targeted Optimality and Safety in Multiagent Learning (CMLeS)
    ICML 2010.
    journal version
    Some slides(ppt)
  • A CMLeS-like algorithm that can be applied
  • Some slides on threats(pdf) - and the relevant paper
  • Busoniu, L. and Babuska, R. and De Schutter, B.
    A comprehensive survey of multiagent reinforcement learning
    IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applicati ons and Reviews, 28(2), 156-172, 2008.
  • Multi-Agent Reinforcement Learning: Independent vs. Coopeative Agents
    by Ming Tan
  • Michael Bowling
    Convergence and No-Regret in Multiagent Learning
    NIPS 2004
  • Kok, J.R. and Vlassis, N., Collaborative multiagent reinforcement learning by payoff propagation, The Journal of Machine Learning Research, 7, 1828, 2006.
  • A brief survey on Multiagent Learning. by Doran Chakraborty
  • gametheory.net
  • Some useful slides (part C) from Michael Bowling on game theory, stochastic games, correlated equilibria; and (Part D) from Michael Littman with more on stochastic games.
  • Scaling up to bigger games with empirical game theory
  • Rob Powers and Yoav Shoham
    New Criteria and a New Algorithm for Learning in Multi-Agent Systems
    NIPS 2004.
    journal version
  • A suite of game generators called GAMUT from Stanford.
  • RoShamBo (rock-paper-scissors) contest
  • U. of Alberta page on automated poker.
  • A paper introducing ad hoc teamwork
  • An article addressing ad hoc teamwork, applied in both predator/prey and RoboCup soccer.
  • Ad hoc teamwork as flocking
  • High confidence policy improvement (Thomas et al.): pdf
  • Safe reinforcement learning via shielding (Alshiekh et al.): pdf
  • Bootstrapping with models: confidence intervals for off-policy evaluation (Hanna et al.): pdf

  • Week 13: Exploration and Intrinsic Motivation

  • Tuesday slides on exploration and IM
  • R-Max - A General Polynomial Time Algorithm for Near-Optimal Reinforcement Learning
    Ronen Brafman and Moshe Tenenholtz
    The Journal of Machine Learning Research (JMLR) 2002
  • Efficient Structure Learning in Factored-state MDPs
    Alexander L. Strehl, Carlos Diuk, and Michael L. Littman
  • The Adaptive k-Meteorologists Problem and Its Application to Structure Learning and Feature Selection in Reinforcement Learning Carlos Diuk, Lihong Li, and Bethany R. Leffler
    ICML 2009
  • Slides and video for the k-meteorologists paper
  • An Analysis of Model-Based Interval Estimation for Markov Decision Processes
    Alexander L. Strehl and Michael L. Littman
    MLJ 2008.
  • A shorter paper on MBIE
  • Model-Based Exploration in Continuous State Spaces
    Nicholas K. Jong and Peter Stone
    The Seventh Symposium on Abstraction, Reformulation, and Approximation, July 2007.
  • TEXPLORE: Real-Time Sample-Efficient Reinforcement Learning for Robots.
    Todd Hester and Peter Stone
    Machine Learning 2012
  • Near-Optimal Reinforcement Learning in Polynomial Time
    Satinder Singh and Michael Kearns
  • Strehl et al.: PAC Model-Free Reinforcement Learning.
  • Safe Exploration in Markov Decision Processes
    Moldovan and Abbeel, ICML 2012
    (safe exploration in non-ergodic domains by favoring policies that maintain the ability to return to the start state)
  • Intrinsically motivated reinforcement learning (Singh et al.)
  • Go-Explore (Ecoffet et al.)
  • Evolved intrinsic rewards for efficient exploration (Niekum et al.)
  • Curiostiy-based exploration for multi-step tasks (Colas et al.)
  • Intrinsically motivated model learning for developing curious robots (Hester et al.)

  • Week 14: Modern Landscape

  • Tuesday slides on distributional RL, metalearning, and multimodal learning
  • Distributional reinforcement learning: pdf
  • Proximal Policy Optimization pdf
  • Open AI's Spinning Up in Deep RL Tutorial: website
  • A follow-up on SAC: Latent Space Policies for Hierarchical Reinforcement Learning
  • Action-Conditional Video Prediction Using Deep Networks in ATARI Games.
    Juhnyuk Oh, Xiaoxiao Guo, Honglak Lee, Richard Lewis, and Satinder Singh.
    Neural Information Processing Systems, 2015.
  • Reinforcement learning with unsupervised auxiliary tasks from Deep Mind includes some action conditional learning.
  • An Introduction to Inter-task Transfer for Reinforcement Learning.
    Matthew E. Taylor and Peter Stone.
    AI Magazine, 32(1):15-34, 2011.
  • Some transfer learning slides; The ones on instance-based transfer; the ones on curriculum learning
  • Improving Action Selection in MDP's via Knowledge Transfer.
    Alexander A. Sherstov and Peter Stone.
    In Proceedings of the Twentieth National Conference on Artificial Intelligence, July 2005.
  • General Game Learning using Knowledge Transfer.
    Bikramjit Banerjee and Peter Stone.
    In The 20th International Joint Conference on Artificial Intelligence, 2007
  • Some other papers on Transfer learning
  • This work addresses the risk of negative transfer and task dissimilarity
    A2T: Attend, Adapt and Transfer Attentive Deep Architecture for Adaptive Transfer from multiple sources
  • This work addresses an improvement to finetuning by adding columns to a deep net and never removing the previously learned weights and avoids catastrophic forgetting.
    Progressive Neural Networks
  • This work explicitly models the differences between two domains to adjust a network trained on one domain and applied to a different one.
    Beyond sharing weights for deep domain adaptation
  • This work trains a network on several task simultaneously and also incorporates expert demonstrations to create general representations that can then be transferred.
    Actor-Mimic Deep Multitaskc and Transfer Reinforcement Learning

  • [Back to Department Homepage]

    Page maintained by Peter Stone
    Questions? Send me mail