OGPT: A Guidance-Oriented Periodic Testing Framework with Online Learning, Online Testing, and Online Feedback

Yimeng Ren, Yuhu Shang, Kun Liang, Xiankun Zhang, Yiying Zhang

Abstract: Periodic testing (PT) is part and parcel of instructional process, which targets at measuring student proficiency level on specific stage. In general, most previous PTs follow an inflexible offline-policy method, which can hardly adjust testing procedure using the online feedback instantly. In this paper, we develop a dynamic and executed online periodic testing framework called O\(^3\)GPT, which selects the most suitable questions step by step, depending on student’s previous timestep’s real-time feedback. To begin with, we employ a stacked GRU to update student’s state representation instantly, which could well capture the long-term dynamic nature from their past learning trajectories, leading to the testing agent perform effective periodic testing. Subsequently, in Stage2, O\(^3\)GPT incorporates a flexible testing-specific reward function into the soft actor-critic algorithm (SAC) to guarantee the rationality of all selected questions. Finally, to set up the online feedback, we test O\(^3\)GPT on an on-line simulated environment which can model qualitative development of knowledge proficiency. The results of our experiment conducted on two well-established student response datasets indicate that O\(^3\)GPT outperforms state-of-the-art baselines in PT task.