Reinforcement Learning Applied to Adaptive Classification Testing

Abstract

This study investigates how computerized adaptive classification testing task can be considered as a sequential decision process and made accessible to Reinforcement Learning. The proposed method performs a sequential item selection that learns which items are most informative, choosing the next item depending on the already administered items and the internal belief of the classifier. A simulation study shows its efficiency for tests which require to make a confident classification decision with as few items as possible. Solutions for a variety of practical problems using the proposed method are considered in this study.