Abstract
Reinforcement learning has been used in dialogue systems for several years. However, most of research focuses on how to improve the efficiency and accuracy performance by developing new learning algorithms. Some papers discuss the learning performance with respect to a reward schema, but do not address how rewards can be designed automatically in a more general way. In fact, designing a reward schema in reinforcement learning is not a trivial task. Some problem domains do not have specific cost and reward for reinforcement learning, the designers must conceive a reward schema that can provide with proper reward and cost during problem solving process in order to guide the learning system to acquire the optimal policy for actions. In our information retrieval dialogue system for restaurants, we wish to train the information retrieval system to generate a proper dialogue in order to identify users’ true intention quickly and response with the correct information that users want. We propose a prediction model for user’s preference to help us. The idea of prediction is based on the data dependency as to how much a dialogue action may reduce quantity of information under retrieval and how likely it may be effective to find the information in terms of probability. In this paper, we demonstrate a new method designing a reward schema which is based on this prediction model that can automatically provide rewards for reinforcement learning.