DNN-Rule Hybrid Dyna-Q for Sample-Efficient Task-Oriented Dialog Policy Learning

Reinforcement learning (RL) is a powerful strategy for making a flexible task-oriented dialog agent, but it is weak in learning speed. Deep Dyna-Q augments the agent's experience to improve the learning efficiency by internally simulating the user's behavior. It uses a deep neural network (DNN) based learnable user model to predict user action, reward, and dialog termination from the dialog state and the agent's action. However, it still needs many agent-user interactions to train the user model. We propose a DNN-Rule hybrid user model for Dyna-Q, where the DNN only simulates the user action. Instead, a rule-based function infers the reward and the dialog termination. We also investigate the training with rollout to further enhance the learning efficiency. Experiments on a movie-ticket booking task demonstrate that our approach significantly improves learning efficiency.