Fast Reinforcement Learning of Dialog Strategies

Abstract

Dialog management is a critical component of an effective spoken language application. It is also one of the most difficult and time consuming to engineer. This paper examines the application of reinforcement learning and Markov Decision Processes (MDP's) to the problem of learning the dialog strategies. It extends work done at AT&T [1] [2] in two directions. First it examines the ability of RL to learn optimal strategies in the presence of speech recognition errors. Second, it describes a technique for reducing the amount of data required to train these models. This is significant as the difficulty of training MDP-based dialog managers is a serious roadblock to deploying them in realistic applications.