Reinforcement learning-based feature selection for developing pedagogically effective tutorial dialogue tactics

Given the subtlety of tutorial tactics, identifying effective pedagogical tactical rules from human tutoring dialogues and implementing them for dialogue tutoring systems is not trivial. In this work, we used reinforcement learning (RL) to automatically derive pedagogical tutoring dialog tactics. Past research has shown that the choice of the features significantly affects the effectiveness of the learned tactics. We defined a total of 18 features which we classified into four types. First, we compared five feature selection methods and overall upper-bound method seems to be most efficient. Then we compared the four types of features and found that temporal situation and autonomy related features are significantly more relevant and effective to tutorial decisions than either performance or situation related features.