Composite Task-Completion Dialogue Policy Learning via Hierarchical Deep Reinforcement Learning

In a composite-domain task-completion dialogue system, a conversation agent often switches among multiple sub-domains before it successfully completes the task. Given such a scenario, a standard deep reinforcement learning based dialogue agent may suffer to find a good policy due to the issues such as: increased state and action spaces, high sample complexity demands, sparse reward and long horizon etc.. In this paper, we propose to use hierarchical deep reinforcement learning approach which can operate at different temporal scales and is intrinsically motivated to attack these problems. Our hierarchical policy network consists of two levels: the top-level meta-controller for subgoal selection and the low-level controller for dialogue policy learning. Subgoals selected by meta-controller and intrinsic rewards can guide the controller to effectively explore in the state-action space and mitigate the spare reward and long horizon problems. Experiments on both simulations and human evaluation show that our model significantly outperforms flat deep reinforcement learning agents in terms of success rate, rewards and user rating.