Learning Dynamic Feedback in Intelligent Tutoring Systems

在线教育资源的迅速增加以及最近对远程教育方法（如大规模在线开放课程（MOOCs））的关注表明，智能辅导系统（ITS）作为一种能够个性化电子学习的自适应教育技术的重要性。经典的信息系统需要学习任务和学习者-系统交互的精确形式化。因此，它们的适用性通常局限于定义良好的领域。此外，它们的劳动密集型准备限制了它们在静态、大规模应用程序中的使用，在这些应用程序中，开发成本不会起重要作用。在FIT项目的第一阶段，我们开发了一个FIT ITS基础设施，该基础设施允许基于机器学习技术在不明确的领域中构建ITS。特别是，我们已经开发了基于原型的机器学习模型，用于结构和结构-度量适应技术，这些模型使ITS解决方案空间在不明确的领域中的自治组织成为可能，在此基础上可以建立反馈提供策略。到目前为止，开发的机器学习技术仅限于单一任务，并且反馈提供不适合个人用户及其进度。DynaFIT的目标是(i)开发可以跨不同任务和用户行为进行泛化的机器学习模型，并在此基础上（ii）通过动态用户自适应反馈和不明确领域的开放学习者模型来增强FIT it。更具体地说，我们将开发跨任务降维（DR）技术，用于在一个共同潜在空间中生成不同解空间的任务无关表示的结构。这支持跨任务的自主信息传输，以及从可能的单一用户行为到相关底层原则的一般化。这种表示构成了获得以下DynaFIT核心组件的关键先决条件：解决方案空间和学习者行为的相关特征的可视化（开放学习者模型）；跨任务的用户行为表示为低维时间序列，其中经典数据分析技术以及FIT项目第一阶段开发的相关学习是可用的；最后，在此基础上，根据该时间序列数据调整动态反馈供应策略。在这个领域，我们将举例详细研究动态对等策略（与大型在线课程高度相关）。在通过跨任务降维处理信息传递时，DynaFIT有助于SPP的一个中心主题：适合学习的表示的自主发展。此外，通过动态反馈提供和开放学习者模型，在定义不清的领域丰富信息技术系统，这对于高动态的大规模教育技术设施（如mooc）具有巨大的潜力。

The rapidly increasing availability of online educational resources and the recent attention gained by distance education methods such as massive open online courses (MOOCs) show the importance of intelligent tutoring systems (ITS) as adaptive educational technologies that can personalize e-Learning. Classical ITSs require an exact formalization of the learning task and learner-system interactions. Hence their applicability is typically limited to well-defined domains. In addition, their labor-intensive preparation restricts their use to static, large-scale applications where development costs do not play a significant role. Within the first period of the FIT project, we have developed a FIT ITS infrastructure which allows the construction of ITSs in ill-defined domains based on machine learning techniques. In particular, we have developed prototype-based machine learning models for structures and structure-metric adaptation techniques which enable an autonomous organization of an ITS solution space in ill-defined domains, based on which feedback provision strategies can be grounded. So far, the developed machine learning techniques are restricted to single tasks, and feedback provision is not tailored to individual users and their progress. The goal of DynaFIT is to (i) develop machine learning models which can generalize across different tasks and user behaviors and, based thereon, (ii) to enhance FIT ITSs via dynamic user-adaptive feedback and open learner models in ill-defined domains. More specifically, we will develop cross-task dimensionality reduction (DR) techniques for structures which generate a task-independent representation of different solution spaces in one common latent space. This enables autonomous information transfer across tasks and a generalization from possibly singular user behavior to relevant underlying principles. This representation constitutes a key prerequisite for obtaining the following central components of DynaFIT: a visualization of relevant characteristics of solution spaces and learner behavior (open learner models); a representation of user behavior across tasks as low dimensional time series, for which classical data analysis techniques as well as relevance learning as developed in the first period of the FIT project are available; finally, based thereon, dynamic feedback provision strategies adjusted to this time series data. In this realm, we will exemplarily investigate dynamic peering strategies (highly relevant for larger online courses) in detail.In addressing information transfer by cross-task dimensionality reduction, DynaFIT contributes to a central topic of the SPP: the autonomous development of suitable representations for learning. Further, the envisioned enrichment of ITSs in ill-defined domains by dynamic feedback provision and open learner models bears great potential for highly dynamic large-scale educational technology facilities such as MOOCs.