تصميم وتقييم منصّة تعليميّة ذكيّة قائمة على تقنيات التعلّم المعزّز

Abstract

Adaptive online learning can enhance learning rewards and reduce the effort required from students, teachers, and course designers. Reinforcement learning is a promising tool for developing educational strategies, as reinforcement learning models can learn the complex relationships between course components, learner interactions, and achieved outcomes.

This research presents the first reinforcement learning model using real-time scheduling of educational activities for a large online course on electrical circuits and electronic laws, enabling students to understand and analyze circuits, close appropriate interrupters, and place electrical components such as resistors and capacitors in the correct locations to perform calculations of current, voltage, and power through effective learning. Our model learns to identify a series of course activities that maximize learning rewards while minimizing the number of required actions. Using Q-Learning algorithm applied to over 1800 learners, we investigate how this scheduling technique affects learning rewards, dropout rates, and qualitative learner responses. The task of reinforcement learning is to train an agent (the virtual teacher) that interacts with its environment (the electronic circuit interface). The agent appears in different scenarios defined by states by performing actions (correct placement of components, incorrect placement of components, correct clicks, incorrect clicks, external clicks). We demonstrate that our model produces better learning rewards using fewer educational activities compared to the linear assignment case. It also yields similar learning gains to self-directed navigation while utilizing fewer educational activities and achieving lower dropout rates. The results indicate that the proposed approach achieved high accuracy with a correct click rate for interrupters and proper placement of electronic components by the agent at 92.5%, demonstrating that the system has been learned correctly and accurately. Meanwhile, the error rate for incorrect clicks on interrupters and improper placement of electronic components was only 7.5%, along with clicks in other areas.