Deep Reinforcement Learning from Policy-Dependent Human Feedback

To widen their accessibility and increase their utility, intelligent agents must be able to learn complex behaviors as specified by (non-expert) human users. Moreover, they will need to learn these behaviors within a reasonable amount of time while efficiently leveraging the sparse feedback a human trainer is capable of providing. Recent work has shown that human feedback can be characterized as a critique of an agent's current behavior rather than as an alternative reward signal to be maximized, culminating in the COnvergent Actor-Critic by Humans (COACH) algorithm for making direct policy updates based on human feedback. Our work builds on COACH, moving to a setting where the agent's policy is represented by a deep neural network. We employ a series of modifications on top of the original COACH algorithm that are critical for successfully learning behaviors from high-dimensional observations, while also satisfying the constraint of obtaining reduced sample complexity. We demonstrate the effectiveness of our Deep COACH algorithm in the rich 3D world of Minecraft with an agent that learns to complete tasks by mapping from raw pixels to actions using only real-time human feedback in 10-15 minutes of interaction.

[1]  Andrea Lockerd Thomaz,et al.  Policy Shaping: Integrating Human Feedback with Reinforcement Learning , 2013, NIPS.

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

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

[4]  Sergey Levine,et al.  Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization , 2016, ICML.

[5]  Pieter Abbeel,et al.  Apprenticeship learning via inverse reinforcement learning , 2004, ICML.

[6]  Geoffrey E. Hinton.  Reducing the Dimensionality of Data with Neural , 2008 .

[7]  Wolfgang Ertel,et al.  Reinforcement learning combined with human feedback in continuous state and action spaces , 2012, 2012 IEEE International Conference on Development and Learning and Epigenetic Robotics (ICDL).

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

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

[10]  Jürgen Schmidhuber,et al.  Stacked Convolutional Auto-Encoders for Hierarchical Feature Extraction , 2011, ICANN.

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

[12]  Brett Browning,et al.  A survey of robot learning from demonstration , 2009, Robotics Auton. Syst..

[13]  Peter Stone,et al.  Reinforcement learning from simultaneous human and MDP reward , 2012, AAMAS.

[14]  Ufuk Topcu,et al.  Environment-Independent Task Specifications via GLTL , 2017, ArXiv.

[15]  Geoffrey J. Gordon,et al.  A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning , 2010, AISTATS.

[16]  Craig Boutilier,et al.  Rewarding Behaviors , 1996, AAAI/IAAI, Vol. 2.

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

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

[19]  Maya Cakmak,et al.  Power to the People: The Role of Humans in Interactive Machine Learning , 2014, AI Mag..

[20]  Peter Stone,et al.  Combining manual feedback with subsequent MDP reward signals for reinforcement learning , 2010, AAMAS.

[21]  Peter Stone,et al.  Deep TAMER: Interactive Agent Shaping in High-Dimensional State Spaces , 2017, AAAI.

[22]  Honglak Lee,et al.  Control of Memory, Active Perception, and Action in Minecraft , 2016, ICML.

[23]  R. French.  Catastrophic forgetting in connectionist networks , 1999, Trends in Cognitive Sciences.

[24]  Shie Mannor,et al.  A Deep Hierarchical Approach to Lifelong Learning in Minecraft , 2016, AAAI.

[25]  Katja Hofmann,et al.  The Malmo Platform for Artificial Intelligence Experimentation , 2016, IJCAI.

[26]  Yishay Mansour,et al.  Policy Gradient Methods for Reinforcement Learning with Function Approximation , 1999, NIPS.

[27]  Shane Legg,et al.  Deep Reinforcement Learning from Human Preferences , 2017, NIPS.

[28]  Jing Peng,et al.  Function Optimization using Connectionist Reinforcement Learning Algorithms , 1991 .

[29]  David L. Roberts,et al.  Learning behaviors via human-delivered discrete feedback: modeling implicit feedback strategies to speed up learning , 2015, Autonomous Agents and Multi-Agent Systems.

[30]  Geoffrey E. Hinton,et al.  Learning representations by back-propagating errors , 1986, Nature.

[31]  Andrea Lockerd Thomaz,et al.  Reinforcement Learning with Human Teachers: Evidence of Feedback and Guidance with Implications for Learning Performance , 2006, AAAI.

[32]  Stefanie Tellex,et al.  Minecraft as an Experimental World for AI in Robotics , 2015, AAAI Fall Symposia.

[33]  Geoffrey E. Hinton,et al.  Reducing the Dimensionality of Data with Neural Networks , 2006, Science.

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

[35]  P. Stone,et al.  TAMER: Training an Agent Manually via Evaluative Reinforcement , 2008, 2008 7th IEEE International Conference on Development and Learning.

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

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