Figure 1.

Feedback loop of the reinforcement learning environment for training the MCI diagnosis agent. The user simulator trained from the original dialogue corpus is used to generate simulated user response to new questions from the MCI diagnosis agent (i.e., the Reinforcement Learning Agent). At each conversational turn, the “user state” of the simulated patient is updated based on the questions asked by the MCI diagnosis agent. We designed a Dialogue Manager which produces a reward signal to the MCI diagnosis agent based on the quality of questions asked.