Scholarlink

<p>Practical teaching in ideological and political theory courses faces persistent challenges in adapting instructional strategies to diverse learner profiles and multidimensional educational objectives. This paper proposes a knowledge graph-enhanced multi-agent deep reinforcement learning framework that addresses these challenges through three integrated components. To begin with, a domain-specific knowledge graph is constructed via a semi-automatic extraction pipeline that combines a Chinese-BERT-BiLSTM-CRF named entity recognition model with verification by three subject-matter experts (inter-annotator Kappa = 0.83), encoding knowledge concepts, teaching activities, learner competencies, and value dimensions as structured triples embedded through the TransR model. Next, the teaching process is formalized as a partially observable Markov game involving three cooperative agents—a Teaching Strategy Agent, a Learning Path Planning Agent, and a Teaching Evaluation Agent—trained through the QMIX value decomposition architecture with graph attention network-based state representations and gated fusion mechanisms. Third, a potential-based reward shaping method grounded in knowledge graph topology is introduced to accelerate policy convergence while preserving optimal policy guarantees. Across seven independent runs on interaction data from 4826 students at three universities, the proposed framework attains a 37.2 ± 0.9% knowledge mastery improvement rate and 26.9 ± 0.8% value identification growth rate, outperforming both classical and recent state-of-the-art multi-agent baselines (MAPPO, QPLEX). Ablation studies confirm that TransR embeddings, graph attention aggregation, and reward shaping contribute complementary benefits, while the graph-guided shaping reduces convergence time by over 50%. A monolithic single-agent ablation further demonstrates the empirical necessity of the three-agent decomposition. The framework offers a replicable paradigm for applying advanced AI techniques to value-oriented educational contexts.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Knowledge graph-enhanced multi-agent deep reinforcement learning for intelligent teaching decision-making in ideological and political theory practical courses

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Ruijun Yin

摘要

Practical teaching in ideological and political theory courses faces persistent challenges in adapting instructional strategies to diverse learner profiles and multidimensional educational objectives. This paper proposes a knowledge graph-enhanced multi-agent deep reinforcement learning framework that addresses these challenges through three integrated components. To begin with, a domain-specific knowledge graph is constructed via a semi-automatic extraction pipeline that combines a Chinese-BERT-BiLSTM-CRF named entity recognition model with verification by three subject-matter experts (inter-annotator Kappa = 0.83), encoding knowledge concepts, teaching activities, learner competencies, and value dimensions as structured triples embedded through the TransR model. Next, the teaching process is formalized as a partially observable Markov game involving three cooperative agents—a Teaching Strategy Agent, a Learning Path Planning Agent, and a Teaching Evaluation Agent—trained through the QMIX value decomposition architecture with graph attention network-based state representations and gated fusion mechanisms. Third, a potential-based reward shaping method grounded in knowledge graph topology is introduced to accelerate policy convergence while preserving optimal policy guarantees. Across seven independent runs on interaction data from 4826 students at three universities, the proposed framework attains a 37.2 ± 0.9% knowledge mastery improvement rate and 26.9 ± 0.8% value identification growth rate, outperforming both classical and recent state-of-the-art multi-agent baselines (MAPPO, QPLEX). Ablation studies confirm that TransR embeddings, graph attention aggregation, and reward shaping contribute complementary benefits, while the graph-guided shaping reduces convergence time by over 50%. A monolithic single-agent ablation further demonstrates the empirical necessity of the three-agent decomposition. The framework offers a replicable paradigm for applying advanced AI techniques to value-oriented educational contexts.