In this paper, we focus on automatic disease diagnosis with reinforcement learning (RL) methods in task-oriented dialogues setting. Different from conventional RL tasks, the action space for disease diagnosis (i.e., symptoms) is inevitably large, especially when the number of diseases increases. However, existing approaches to this problem employ a flat RL policy, which typically works well in simple tasks but has significant challenges in complex scenarios like disease diagnosis. Towards this end, we propose to integrate a hierarchical policy of two levels into the dialogue policy learning. The high level policy consists of a model named master that is responsible for triggering a model in low level, the low level policy consists of several symptom checkers and a disease classifier. Experimental results on both self-constructed real-world and synthetic datasets demonstrate that our hierarchical framework achieves higher accuracy in disease diagnosis compared with existing systems. Besides, the datasets (http://www.sdspeople.fudan.edu.cn/zywei/data/Fudan-Medical-Dialogue2.0) and codes (https://github.com/nnbay/MeicalChatbot-HRL) are all available now.