Shimmei, M.*, & Noboru, M. (2019). Evidence-Based Recommendation for Content Improvement UsingReinforcementLearning. In S. Isotani, A. Ogan, B. McLaren, E. Millán, P. Hastings & R. Luckin (Eds.), Proceedings of the International Conference on Artificial Intelligence in Education (pp. 369-373). Cham, Switzerland: Springer.

Abstract: One of the most challenging issues for online-courseware engi- neering is to maintain the quality of instructional elements. However, it is hard to know how each instructional element on the courseware contributes to stu- dents’ learning. To address this challenge, an evidence-based learning- engineering method for validating the quality of instructional elements on online courseware is proposed. Students’ learning trajectories on particular online courseware and their final learning outcomes are consolidated into a state transition graph. The value iteration technique is applied to compute the worst actions taken (a converse policy) to yield the least successful learning. We hypothesize that the converse policy reflects the quality of instructional ele- ments. As a proof of concept, this paper describes an evaluation study where we simulated online learning data on three hypothetical pieces of online course- ware. The result showed that our method can detect more than a half of the ineffective instructional elements on three types of courseware containing var- ious ratios of ineffective instructional elements.