Abstract:

Intelligent Tutoring Systems (ITS) and Affective Tutoring Systems (ATS) can successfully teach courses to users at all levels of education. These systems however, are typically built around a specific subject, making reusability in different domains difficult. In contrast, Adaptive Learning Systems are domain-agnostic, being able to teach more than one course and focus on its presentation by adapting it to the user’s learning preferences. These adaptations are usually based on Learning Styles theories. Two disadvantages of Adaptive Learning systems are that they do not offer any type of feedback to the user, and that there are doubts in the scientific community concerning the validity of Learning Styles due to lack of relevant scientific evidence.

In order to combine the advantages of ITS, ATS and Adaptive Learning Systems, we built the AfflogRL tutor, an ATS with the following specifications: The domain, tutoring, and student models are designed using the Predicate Calculus and the Event Calculus and implemented following the Answer Set Programming (ASP) formalism. In order to formulate a teaching strategy for the tutor to follow, the system makes use of a Reinforcement learning (RL) agent that utilizes the experience it accumulates by interacting with users.

To evaluate the system, a simple course on how to play the “Settlers of Catan” board game was used. After training the RL agent, evaluation showed that users had high learning gains from their interaction with the system. We also compared AfflogRL with a previous version of the system that utilizes the Learning Styles of the user to guide the actions of the tutor rather than a RL policy. Results showed that AffLogRL performed marginally better than its predecessor, indicating that with more training samples, AfflogRL will perform significantly better.

Keywords:

Affective Tutoring systems, Event Calculus, ASP, Reinforcement Learning.