随着人工智能（AI）技术的不断发展，大量功能丰富、性能优异的智能导学系统（ITS）在实际教育业务中得到师生广泛欢迎。针对学生习题练习的场景，ITS可以调用相关算法，快速便捷地实现个性化精准评测与学习资源高效推荐，有助于减轻教师日常教学负担，改善学生学习体验。

智能导学系统的核心技术，包括知识追踪、学习资源推荐算法，可以有效为个性化教育赋能，但仍存在改进的空间。知识追踪用于评估与预测学生的动态知识状态，当前存在的主要问题包括：1）知识点之间关系的建模方式较为表层和简单，缺乏对学习内容的深入描述；2）输出结果可解释性弱，较难为下游教育任务服务。学习资源推荐需要根据知识状态在内的用户信息，提供有助于高效、个性化学习的资源或材料，当前存在的主要问题包括：1）注重构建用户的偏好，但忽略了教育场景下复杂多样的学习情况，如：习题多样性、难度变化波动程度，使得推荐目标与实际需求脱节；2）在算法运行过程中缺乏真人教师的调节与参与机制，降低了系统的专业性与灵活性。
针对上述问题，本文以初中物理电学章节为例、以习题练习业务为导向，改进并实现智能导学系统。本研究基于专家绘制的知识逻辑结构图，构建并表征各知识点之间的逻辑关系，作为系统的领域知识模块；然后，利用深度学习基本单元模块的通用性与灵活性，提出新的知识追踪与习题推荐算法，改进知识状态评测、个性化习题推荐两项功能；同时开发了相应的智能导学系统，以在线网页形式对外接受访问，在真实情境下提升用户学习效率。

本文首先研究了知识追踪算法，通过结合图神经网络、正则化项等技术，将知识逻辑结构图中的语义逻辑信息融入做题记录表征、追踪方法、结果输出三部分模块，提出的HKGKT模型可以判别学生的实时知识掌握程度并给出预测结果。在此基础上，进一步提出习题推荐算法，采用深度强化学习算法Double DQN，结合实际教育场景需求设计了状态表征方法、奖励函数与训练策略，使算法能够根据学习者实时状态与多样化学习目标，推荐个性化的习题序列。最后，本文调用完成训练的算法，设计并实现了网页版物理智能导学系统，可以协同教师帮助学生快速提升目标知识点的掌握程度，改善学习效率，同时通过真人使用反馈检验了该技术路径的可行性和实用性。

为验证算法的性能，本文在5个模拟数据集上进行了一系列实验：

KGKT模型在AUC值上对比其他常见知识追踪模型提高了0.10~0.15左右，且可以融入更高质量的知识信息，体现精准性与可靠性；而后续提出的习题推荐算法相比其他6个常规推荐方法在知识状态提升、难度变化等数个可量化结果中均排名第一或第二，说明算法实用效果良好、对复杂多样的推荐目标具备优异的兼容性，能够更好地为实际教学场景需求赋能。

With the continuous development of Artificial Intelligence (AI), a large number of Intelligent Tutoring Systems (ITS) with various functions and excellent performances have become very popular among teachers and students. When students do exercises, ITS can apply some efficient algorithms to realize personalized accurate assessments and efficient recommendations of learning resources quickly and conveniently, which helps to reduce the burden of teachers’ daily teaching and improve students’ learning performance.

The key technologies of ITS, including knowledge tracing and learning resource recommendation algorithms, can effectively promote personalized education, but there is still a large improvement space.  Knowledge tracing is used to evaluate and predict students’ dynamic knowledge state. Its current problems include: 1) The modeling of the relationship between knowledge nodes is relatively superficial and simple, lacking an in-depth description of learning content; 2) The outputs are weakly interpretable, which makes it difficult to serve downstream educational tasks. Learning resource recommendation needs to provide resources or materials that contribute to efficient and personalized learning based on user information, including knowledge status. Its current problems include: 1) Many researches focus on constructing user preferences, but ignore the complex and diverse learning situations in educational scenes, which makes the recommendation goals mismatch the actual needs; 2) These systems are usually lacking in a mechanism for real teachers to regulate and participate in the execution process of the algorithms, which reduces the systems’ professionalism and flexibility.

In response to the above issues, this paper takes the electricity chapter (middle school physics) as an example, guided by the task of exercise practice, to improve and achieve an intelligent tutoring system. This study constructs and characterizes the logical relationships between various knowledge points based on the knowledge logic structure diagram designed by experts, which serves as the domain knowledge module of the system. Then, a new knowledge tracing and exercise recommendation algorithm is proposed by making full use of some basic unit modules in deep learning, which aims to improve the two functions of knowledge state evaluation and personalized exercise recommendation. Also, the access to this ITS is developed in the form of online web pages, promoting users’ learning efficiency in real situations.

This article first studies the knowledge tracing algorithm. By combining various techniques such as graph neural networks and regularization terms, the semantic logic information in the knowledge logic structure diagram is integrated into three modules: problem recording representation, tracing methods, and result output. The proposed HKGKT model can distinguish the real-time knowledge mastery level of students and provide prediction results. Then, an exercise recommendation algorithm is proposed, which adopts the deep reinforcement learning algorithm Double DQN and designs state representation methods, reward functions, and training strategies based on actual educational needs. This enables the algorithm to recommend personalized exercise sequences based on the real-time state of learners and diverse learning objectives. Finally, this article applies the trained algorithm to design and realize a web-based version of the physics intelligent tutoring system, which can collaborate with teachers to help students quickly increase their mastery of target knowledge nodes and improve learning efficiency. What’s more, the feasibility and practicality of this technology path are verified through real person feedback.

To evaluate the performance of the algorithm, two experiments are conducted on 5 simulated datasets: the HKGKT model improves the AUC by about 0.10-0.15 compared to other common methods, and can integrate high-quality knowledge information, which proves its accuracy and reliability. The proposed exercise recommendation algorithm ranks 1st or 2nd in several quantifiable results (e.g., knowledge state improvement, difficulty changes) compared to the other six methods, indicating that it has good practical effects and excellent compatibility with diverse recommendation objectives, and can better empower practical teaching scenarios.