Asymptotic Learnability of Reinforcement Problems with Arbitrary Dependence
We address the problem of reinforcement learning in which observations may exhibit an arbitrary form of stochastic dependence on past observations and actions, i.e. environments more general than (PO) MDPs. The task for an agent is to attain the best possible asymptotic reward where the true generating environment is unknown but belongs to a known countable family of environments. We find some sufficient conditions on the class of environments under which an agent exists which attains the best asymptotic reward for any environment in the class. We analyze how tight these conditions are and how they relate to different probabilistic assumptions known in reinforcement learning and related fields, such as Markov Decision Processes and mixing conditions.
KeywordsOptimal Policy Reinforcement Learning Markov Decision Process Countable Family Average Reward
Unable to display preview. Download preview PDF.
- [BT99]Brafman, R.I., Tennenholtz, M.: A general polynomial time algorithm for near-optimal reinforcement learning. In: Proc. 17th International Joint Conference on Artificial Intelligence (IJCAI 2001), pp. 734–739 (1999)Google Scholar
- [CS04]Csiszar, I., Shields, P.C.: Notes on information theory and statistics. In: Foundations and Trends in Communications and Information Theory (2004)Google Scholar
- [EDKM05]Even-Dar, E., Kakade, S.M., Mansour, Y.: Reinforcement learning in POMDPs without resets. In: IJCAI, pp. 690–695 (2005)Google Scholar
- [PH06]Poland, J., Hutter, M.: Universal learning of repeated matrix games. In: Conference Benelearn 2006 and GTDT workshop at AAMAS 2006, Ghent (2006)Google Scholar
- [dFM04]de Farias, D.P., Megiddo, N.: How to combine expert (and novice) advice when actions impact the environment? In: Thrun, S., Saul, L., Schölkopf, B. (eds.) Advances in Neural Information Processing Systems 16, MIT Press, Cambridge, MA (2004)Google Scholar
- [SB98]Sutton, R., Barto, A.: Reinforcement learning: An introduction. MIT Press, Cambridge, MA (1998)Google Scholar