Deep Reinforcement Learning and the Deadly Triad

We know from reinforcement learning theory that temporal difference learning can fail in certain cases. Sutton and Barto (2018) identify a deadly triad of function approximation, bootstrapping, and off-policy learning. When these three properties are combined, learning can diverge with the value estimates becoming unbounded. However, several algorithms successfully combine these three properties, which indicates that there is at least a partial gap in our understanding. In this work, we investigate the impact of the deadly triad in practice, in the context of a family of popular deep reinforcement learning models - deep Q-networks trained with experience replay - analysing how the components of this system play a role in the emergence of the deadly triad, and in the agent's performance

[1]  Longxin Lin.  Self-Improving Reactive Agents Based on Reinforcement Learning, Planning and Teaching , 2004, Machine Learning.

[2]  R. Sutton,et al.  Gradient temporal-difference learning algorithms , 2011 .

[3]  Alex Graves,et al.  Asynchronous Methods for Deep Reinforcement Learning , 2016, ICML.

[4]  Richard S. Sutton,et al.  Introduction to Reinforcement Learning , 1998 .

[5]  Hado van Hasselt,et al.  Double Q-learning , 2010, NIPS.

[6]  Ian H. Witten,et al.  An Adaptive Optimal Controller for Discrete-Time Markov Environments , 1977, Inf. Control..

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

[8]  Alex Graves,et al.  Playing Atari with Deep Reinforcement Learning , 2013, ArXiv.

[9]  David Budden,et al.  Distributed Prioritized Experience Replay , 2018, ICLR.

[10]  David Silver,et al.  Deep Reinforcement Learning with Double Q-Learning , 2015, AAAI.

[11]  Richard S. Sutton,et al.  Neuronlike adaptive elements that can solve difficult learning control problems , 1983, IEEE Transactions on Systems, Man, and Cybernetics.

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

[13]  Martha White,et al.  An Emphatic Approach to the Problem of Off-policy Temporal-Difference Learning , 2015, J. Mach. Learn. Res..

[14]  Tom Schaul,et al.  Dueling Network Architectures for Deep Reinforcement Learning , 2015, ICML.

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

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

[17]  John N. Tsitsiklis,et al.  Analysis of Temporal-Diffference Learning with Function Approximation , 1996, NIPS.

[18]  R. Bellman,et al.  Dynamic Programming and Markov Processes , 1960 .

[19]  Shalabh Bhatnagar,et al.  Natural actor-critic algorithms , 2009, Autom..

[20]  Hado van Hasselt,et al.  Reinforcement Learning in Continuous State and Action Spaces , 2012, Reinforcement Learning.

[21]  Sean R Eddy,et al.  What is dynamic programming? , 2004, Nature Biotechnology.

[22]  Leemon C. Baird,et al.  Residual Algorithms: Reinforcement Learning with Function Approximation , 1995, ICML.

[23]  R. Sutton.  On The Virtues of Linear Learning and Trajectory Distributions , 2007 .

[24]  Martha White,et al.  Linear Off-Policy Actor-Critic , 2012, ICML.

[25]  Sanjoy Dasgupta,et al.  Off-Policy Temporal Difference Learning with Function Approximation , 2001, ICML.

[26]  Richard S. Sutton,et al.  A Convergent O(n) Temporal-difference Algorithm for Off-policy Learning with Linear Function Approximation , 2008, NIPS.

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

[28]  Chris Watkins,et al.  Learning from delayed rewards , 1989 .

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

[30]  Patrick M. Pilarski,et al.  Horde: a scalable real-time architecture for learning knowledge from unsupervised sensorimotor interaction , 2011, AAMAS.

[31]  Martin L. Puterman,et al.  Markov Decision Processes: Discrete Stochastic Dynamic Programming , 1994 .

[32]  Shalabh Bhatnagar,et al.  Convergent Temporal-Difference Learning with Arbitrary Smooth Function Approximation , 2009, NIPS.

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