Using Deep Reinforcement Learning to Decide Test Length

This study explored the application of deep reinforcement learning (DRL) as an innovative approach to optimize test length. The primary focus was to evaluate whether the current length of the National Board of Chiropractic Examiners Part I Exam is justified. By modeling the problem as a combinatorial optimization task within a Markov Decision Process framework, an algorithm capable of constructing test forms from a finite set of items while adhering to critical structural constraints, such as content representation and item difficulty distribution, was used. The findings reveal that although the DRL algorithm was successful in identifying shorter test forms that maintained comparable ability estimation accuracy, the existing test length of 240 items remains advisable as we found shorter test forms did not maintain structural constraints. Furthermore, the study highlighted the inherent adaptability of DRL to continuously learn about a test-taker's latent abilities and dynamically adjust to their response patterns, making it well-suited for personalized testing environments. This dynamic capability supports real-time decision-making in item selection, improving both efficiency and precision in ability estimation. Future research is encouraged to focus on expanding the item bank and leveraging advanced computational resources to enhance the algorithm's search capacity for shorter, structurally compliant test forms.