Scholarlink

<p>To address the lack of personalization and dynamic adaptability of learning paths in virtual reality music education, this study applies a deep reinforcement learning framework based on PPO to generate adaptive learning paths, overcoming the limitations of traditional fixed paths in long-term learning effectiveness and personalized support. A Markov decision process is constructed with an 18-dimensional state space (rhythm comprehension level, hand movement stability, cognitive load index, etc.), and a ternary reward function is designed, centered on skill gain, cognitive load balance, and engagement maintenance. The system is deployed in a VR piano teaching environment developed in Unity, integrating Leap Motion gesture tracking, a Tobii eye tracker, and an Empatica E4 physiological sensor. Real-time interaction between the PPO agent and the VR front-end is implemented using PyTorch. The combination of deep reinforcement learning and VR technology not only enables real-time state perception with a high level of immersion but also significantly enhances the adaptability of the teaching system through intelligent decision-making, highlighting the “1 + 1 &gt; 2” advantage of technological integration. Experimental results show that the PPO method significantly outperforms baseline methods (such as DQN and A2C) in path fitness (0.936, the average cosine similarity between the learning path and the ideal teaching state sequence defined by both real data and flow theory), skill gain rate (0.042 min⁻<sup>1</sup>), and convergence speed (89 rounds). For four types of virtual students (gifted, Diligent, struggling, and interest-driven), the PPO-generated pathways are tailored to each student’s status. A study with real-life users shows that the post-test skill levels for the expert and PPO pathways are 0.59 and 0.66, respectively (<i>p</i> = 0.012*). The flow experience (FSS-2: 42.6 ± 2.8) is significantly better than that of Group A (expert pathway), and cognitive load (NASA-TLX: 52.1 ± 6.7) is significantly reduced. This study validates the effectiveness and superiority of deep reinforcement learning for dynamic optimization of personalized pathways in VR education.</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

Adaptive learning path generation in virtual reality music education via deep reinforcement learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月28日

查看全文
  • Xiaolu Wu,
  • Yeping Wang

摘要

To address the lack of personalization and dynamic adaptability of learning paths in virtual reality music education, this study applies a deep reinforcement learning framework based on PPO to generate adaptive learning paths, overcoming the limitations of traditional fixed paths in long-term learning effectiveness and personalized support. A Markov decision process is constructed with an 18-dimensional state space (rhythm comprehension level, hand movement stability, cognitive load index, etc.), and a ternary reward function is designed, centered on skill gain, cognitive load balance, and engagement maintenance. The system is deployed in a VR piano teaching environment developed in Unity, integrating Leap Motion gesture tracking, a Tobii eye tracker, and an Empatica E4 physiological sensor. Real-time interaction between the PPO agent and the VR front-end is implemented using PyTorch. The combination of deep reinforcement learning and VR technology not only enables real-time state perception with a high level of immersion but also significantly enhances the adaptability of the teaching system through intelligent decision-making, highlighting the “1 + 1 > 2” advantage of technological integration. Experimental results show that the PPO method significantly outperforms baseline methods (such as DQN and A2C) in path fitness (0.936, the average cosine similarity between the learning path and the ideal teaching state sequence defined by both real data and flow theory), skill gain rate (0.042 min⁻1), and convergence speed (89 rounds). For four types of virtual students (gifted, Diligent, struggling, and interest-driven), the PPO-generated pathways are tailored to each student’s status. A study with real-life users shows that the post-test skill levels for the expert and PPO pathways are 0.59 and 0.66, respectively (p = 0.012*). The flow experience (FSS-2: 42.6 ± 2.8) is significantly better than that of Group A (expert pathway), and cognitive load (NASA-TLX: 52.1 ± 6.7) is significantly reduced. This study validates the effectiveness and superiority of deep reinforcement learning for dynamic optimization of personalized pathways in VR education.