We use Reinforcement Learning (RL) to learn question-answering dialogue policies for a real-world application. We analyze a corpus of interactions of museum visitors with two virtual characters that serve as guides at the Museum of Science in Boston, in order to build a realistic model of user behavior when interacting with these characters. A simulated user is built based on this model and used for learning the dialogue policy of the virtual characters using RL. Our learned policy outperforms two baselines (including the original dialogue policy that was used for collecting the corpus) in a simulation setting.
展开▼