Pitfalls of learning a reward function online

In some agent designs like inverse reinforcement learning an agent needs to learn its own reward function. Learning the reward function and optimising for it are typically two different processes, usually performed at different stages. We consider a continual (``one life'') learning approach where the agent both learns the reward function and optimises for it at the same time. We show that this comes with a number of pitfalls, such as deliberately manipulating the learning process in one direction, refusing to learn, ``learning'' facts already known to the agent, and making decisions that are strictly dominated (for all relevant reward functions). We formally introduce two desirable properties: the first is `unriggability', which prevents the agent from steering the learning process in the direction of a reward function that is easier to optimise. The second is `uninfluenceability', whereby the reward-function learning process operates by learning facts about the environment. We show that an uninfluenceable process is automatically unriggable, and if the set of possible environments is sufficiently rich, the converse is true too.

[1]  Laurent Orseau,et al.  AI Safety Gridworlds , 2017, ArXiv.

[2]  Stuart Armstrong,et al.  Good and safe uses of AI Oracles , 2017, ArXiv.

[3]  Robert Riener,et al.  Rehabilitation Robotics , 2013, Found. Trends Robotics.

[4]  John Salvatier,et al.  Agent-Agnostic Human-in-the-Loop Reinforcement Learning , 2017, ArXiv.

[5]  Anca D. Dragan,et al.  Inverse Reward Design , 2017, NIPS.

[6]  Marcus Hutter,et al.  Avoiding Wireheading with Value Reinforcement Learning , 2016, AGI.

[7]  D. Kahneman Thinking, Fast and Slow , 2011 .

[8]  Shane Legg,et al.  Reward learning from human preferences and demonstrations in Atari , 2018, NeurIPS.

[9]  Anca D. Dragan,et al.  Cooperative Inverse Reinforcement Learning , 2016, NIPS.

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

[11]  Tom Everitt,et al.  Towards Safe Artificial General Intelligence , 2018 .

[12]  Illtyd Trethowan Causality , 1938 .

[13]  Angelo C. Loula,et al.  Language Evolution and Robotics: Issues on Symbol Grounding and Language Acquisition , 2006 .

[14]  I-Ping Chen,et al.  Design Aspects of Scoring Systems in Game , 2017 .

[15]  Michael I. Jordan,et al.  Advances in Neural Information Processing Systems 30 , 1995 .

[16]  Kee-Eung Kim,et al.  Inverse Reinforcement Learning in Partially Observable Environments , 2009, IJCAI.

[17]  Wray L. Buntine Operations for Learning with Graphical Models , 1994, J. Artif. Intell. Res..

[18]  Farbod Fahimi,et al.  Online human training of a myoelectric prosthesis controller via actor-critic reinforcement learning , 2011, 2011 IEEE International Conference on Rehabilitation Robotics.

[19]  Guan Wang,et al.  Interactive Learning from Policy-Dependent Human Feedback , 2017, ICML.

[20]  Marcus Hutter,et al.  Universal Artificial Intellegence - Sequential Decisions Based on Algorithmic Probability , 2005, Texts in Theoretical Computer Science. An EATCS Series.

[21]  Eliezer Yudkowsky Artificial Intelligence as a Positive and Negative Factor in Global Risk , 2006 .

[22]  Kareem Amin,et al.  Towards Resolving Unidentifiability in Inverse Reinforcement Learning , 2016, ArXiv.

[23]  Marcus Hutter,et al.  Reward tampering problems and solutions in reinforcement learning: a causal influence diagram perspective , 2019, Synthese.

[24]  Michèle Sebag,et al.  APRIL: Active Preference-learning based Reinforcement Learning , 2012, ECML/PKDD.

[25]  R. Lathe Phd by thesis , 1988, Nature.

[26]  Andrew Y. Ng,et al.  Pharmacokinetics of a novel formulation of ivermectin after administration to goats , 2000, ICML.

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

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