探討強化學習演算法之素材推薦機制與AI學習履歷之學習者感知 Learner Perceptions of AI-Powered Learning Portfolios and Personalized Material Recommendation Mechanisms in Reinforcement Learning Algorithms

Abstract

疫情底下的線上自學課程，為學習歷程檔案提供另一種做法，但是，線上自學衍生的學習動機缺乏、無法適時提供對應的學習素材等問題，都對學習成效帶來極大的挑戰。近年來，在人工智慧發展的浪潮下，教育大數據的機器學習技術成為提升線上自學課程與個人化學習的方法之一。本研究依線上學習課程累積的大量數據，定義多向度的學習者特徵與建構機器學習模型，結合學習診斷與自我調整學習，發展一套線上自學課程的練習題推薦機制。本研究透過北部某大學所開設的微積分線上課程進行實驗，實驗組透過強化學習演算法推薦練習題，控制組則是隨機推薦練習題，結果發現實驗組的前測與後測分數達到顯著，同時實驗組學生的學習黏著度也高於控制組。整體推薦後，全部學生的PR值（百分等級）穩定上升，自我調整學習問卷則顯示學習者對於素材推薦機制與所產生的AI學習履歷持正向的感知態度。 The COVID-19 pandemic necessitated alternative pedagogical approaches, with online autonomous learning courses emerging as a viable method for compiling learning portfolios. Consequently, online autonomous learning has garnered increasing scholarly attention. Embodying principles of openness and transcending temporal and spatial constraints, online courses afforded global learners opportunities for continued education during the pandemic. Online courses facilitate enhanced online interaction among students and teachers and allow students to control their learning experience (learner autonomy) and pace. Nevertheless, online autonomous learning presents fundamental challenges. Notably, in the absence of direct teacher and teaching assistant supervision, online autonomous learning tends to lead to lower completion rates and higher dropout rates, concerns currently under investigation by numerous researchers. In contrast to traditional teacher-centered models, online autonomous learning courses prioritize self-directed learning. Learners independently establish learning objectives and strategies commensurate with their personal learning levels to master course content. Through a series of instructional videos, in-class exercises, discussion forums, and other interactive features, an appropriate self-regulated learning mechanism was developed to guide learners toward effective autonomous learning. The exponential growth of big data in recent years has positioned artificial intelligence as a focal point of inquiry across various fields. Machine learning has catalyzed substantial advancements in the field of data science. The accumulation of extensive learning generates substantial volumes of structured and unstructured data, including the personal information of learners and various learning metrics. A growing body of research advocates for the use of data analytics as a viable method to optimize online and adaptive learning processes. Learning diagnosis entail learners’ self-assessment of requisite capabilities for learning tasks and comparative analyses of capabilities against domain expert-established concept structures by employing relevant question parameters, such as difficulty and discrimination. To facilitate this, an automated artificial intelligence material recommendation mechanism was developed, underpinned by several machine learning models. By observing online user learning behavior patterns, learning data and indicators were formulated, enabling the analysis of various online learning behaviors (e.g., watching videos and answering practice questions) and the generation of learning processes that can be viewed by learners. A practice question recommendation mechanism combined with an instant messaging application (LINE) was designed, leveraging teacher-created knowledge maps to assess students’ mastery of concepts. Zimmerman’s cyclical model of self-regulation served as the foundational framework for the recommendation mechanism. A quasiexperimental research design was employed. Participants were recruited from a calculus course taught at a university in northern Taiwan. An experimental group used reinforcement learning–recommended practice questions for self-evaluation, and a control group received randomly assigned questions. Significant improvements in scores were observed in the experimental group, and greater learning stickiness was observed compared with the control group. Consistent percentile rank increases following practice question completion suggest the system’s capacity to deliver personalized recommendations on the basis of individual differences, thereby facilitating concept-specific feedback and adaptive learning. This, in turn, fostered increased teacher–student interaction, mitigated learner isolation, and increased learning motivation, thereby strengthening self-regulated learning abilities. Upon course completion, the participants could autonomously generate artificial intelligence learning portfolios through the system on the basis of diagnostic results, creating a comprehensive record of their learning performance. These portfolios facilitated the elucidation of learner mastery levels through the accumulation of extensive learning data (big data) on the platform. A postcourse self-regulated learning questionnaire survey revealed a positive participant perception of the material recommendation mechanism and generated artificial intelligence learning portfolio. The participants demonstrated strong positive attitudes toward system reliability, learning attitudes, and metacognition but low perceptions of system utility, and low overall usage rates. Enhancing usage incentive, continuously refining the accuracy of the recommendation system’s algorithms, and conducting comparative analyses with existing systems are essential to improve the recommendation system’s perceived utility.

Keywords

• ai學習履歷
• 自我調整學習
• 強化學習
• 教育大數據
• 學習診斷
• ai learning portfolio
• self-regulated learning
• reinforcement learning
• educational big data
• learning diagnosis