Tuesday 24th September 2019
Ground Floor Seminar Room
25 Howland Street, London, W1T 4JG
A unified view of entropy-regularized Markov decision processes
Entropy regularization, while a standard technique in the online learning toolbox, has only been recently discovered by the reinforcement learning community: In recent years, numerous new reinforcement learning algorithms have been derived using this principle, largely independently of each other. So far, a general framework for these algorithms has remained elusive. In this work, we propose such a general framework for entropy-regularized average-reward reinforcement learning in Markov decision processes (MDPs). Our approach is based on extending the linear-programming formulation of policy optimization in MDPs to accommodate convex regularization functions. Our key result is showing that using the conditional entropy of the joint state-action distributions as regularization yields a dual optimization problem closely resembling the Bellman optimality equations. This result enables us to formalize a number of state-of-the-art entropy-regularized reinforcement learning algorithms as approximate variants of Mirror Descent or Dual Averaging, and thus to argue about the convergence properties of these methods. In particular, we show that the exact version of the TRPO algorithm of Schulman et al. (2015) actually converges to the optimal policy, while the entropy-regularized policy gradient methods of Mnih et al. (2016) may fail to converge to a fixed point.
Gergely Neu is a research assistant professor at the Pompeu Fabra University, Barcelona, Spain. He has previously worked with the SequeL team of INRIA Lille, France and the RLAI group at the University of Alberta, Edmonton, Canada. He obtained his PhD degree in 2013 from the Technical University of Budapest, where his advisors were Andras Gyorgy, Csaba Szepesvari and Laszlo Gyorfi. His main research interests are in machine learning theory, including reinforcement learning and online learning with limited feedback and/or very large action sets.