Regularization in Reinforcement Learning
Amir-massoud Farahmand, Mohammad Ghavamzadeh, Csaba Szepesvari and Shie Mannor
In: Multidisciplinary Symposium on Reinforcement Learning (MSRL-2009), 18 June 2009, Montreal, QC, Canada.

Abstract

We develop regularized counterparts of standard Approximate Value Iteration and Approximate Policy Iteration algorithms. Our statistical analysis show that these methods have an almost optimal finite-sample sample-complexity convergence rate for value function estimation.

[Edit]