An emotional student model for game-play adaptation

Game-based learning offers key advantages for learning through experience in conjunction with offering multi-sensorial and engaging communication. However, ensuring that learning has taken place is the ultimate challenge. Intelligent Tutoring Systems (ITSs) have been incorporated into game-based learning environments to guide learners’ exploration. Emotions have proven to be deeply intertwined with cognitive and motivational factors. ITSs attempt to recognise and convey emotion in order to enhance students’ learning and engagement. The ITS student model is responsible for attainment of adaptability and understanding of learners’ needs. It is not clear which emotions are relevant to the teaching-learning experience, or what antecedents and interpersonal differences are involved in determining an emotion. Therefore, student modelling involves uncertainty. Creating an emotional student model that can reason about students’ observable behaviour during online game-play is the main goal of our research. The analysis, design and implementation for this model are our central focus here. The model uses as a basis the Control-Value theory of achievement emotions and employs motivational and cognitive variables to determine an emotion. A Probabilistic Relational Model (PRM) approach was applied to facilitate the derivation of three Dynamic Bayesian Networks (DBNs) corresponding to three types of achievement emotions. Results from a prototyping exercise conducted along with the outcome-prospective emotions DBN are presented and discussed. In future work a larger population of students will be employed to develop an accurate DBN model to incorporate into PlayPhysics, an emotional game-based learning environment for teaching Physics.

Research highlights
► We aim to reason about students’ emotions using a game-based learning environment.
► We use Control-Value theory to derive our emotional computational model.
► Results of Multinomial Logistic Regression indicated the most relevant predictors.
► The final Dynamic Bayesian Network obtained an overall accuracy of 70.49%.
► Negative and neutral emotions were identified more accurately (21 cases out of 28).