Pessimism About Unknown Unknowns Inspires Conservatism

📖 Description

arXiv:2006.08753v1 [cs.AI] 15 Jun 2020Proceedings of Machine Learning Research vol 125: 1?30, 2020 33rd Annual Conference on Learning Theory Pessimism About Unknown Unknowns Inspires Conservatism Michael K. Cohen MICHAEL -K-COHEN .COM University of Oxford Research School of Engineering Science Future of Humanity Institute Marcus Hutter HUTTER 1.NET Google DeepMind Australian National University Editors: Jacob Abernethy and Shivani Agarwal Abstract If we could de?ne the set of all bad outcomes, we could hard-co de an agent which avoids them; however, in suf?ciently complex environments, this is infe asible. We do not know of any general- purpose approaches in the literature to avoiding novel fail ure modes. Motivated by this, we de?ne an idealized Bayesian reinforcement learner which follows a policy that maximizes the worst-case expected reward over a set of world-models. We call this agen t pessimistic, since it optimizes assuming the worst case. A scalar parameter tunes the agen...

📊 Game Impacts

💭 Reactions

🔗 Sources

• [1] https://arxiv.org/abs/2006.08753