CAREER: Improving Adaptive Decision Making in Interactive Learning Environments

在STEM教育中，交互式学习环境(ILE)为提高学生的表现带来了巨大的希望。虽然传统上，这样的系统侧重于教授学生的主题，但同样重要的方面是教他们如何成为更好的学习者。这个项目的目标是开发一个综合的研究和教育计划，调查如何改善ILE的决策，以及将ILE决策与用户发起的决策相结合的影响。主要的研究目标在于使用最先进的机器学习技术，直接从数据中创建和改进ILE。教育的主要目标在于让学生准备好独立行动，并在他们不知道如何行动的新情况下做出正确的选择。这项工作将有助于通过提高学生在STEM领域的表现，通过教学生做出有效的教学决策，以及通过使这类系统做出的决策对教师和领域专家更加透明，从而使ILES更加有效。该项目将开发并经验性地评估三个ILEE或STEM领域的一般决策框架。该项目自然将研究和教育目标结合在一起。更具体地说，该项目将1)通过调整强化学习来做出类似于人类专家的分层决策，从而促进对强化学习的应用的研究；2)通过引入强调关键决策并能够告知一个人对教育领域的理解的紧凑策略，来促进对ILE和强化学习算法的理解；以及3)通过使用数据驱动的策略来支持学生的决策，并通过体内实验的混合式人机交互决策来最终提高他们的长期问题解决能力。

Interactive Learning Environments (ILEs) hold great promise for improving student performance in STEM education. While, traditionally, such systems have focused on teaching students subject matter, an equally important facet is to teach them how become better learners. The objective of this project is to develop an integrated research and education program that investigates the how to improve decision-making in ILEs and the impact of integrating ILE decisions with user-initiated decisions. The primary research goal lies in creating and improving ILEs directly from data, using state-of-the-art machine-learning techniques. The primary educational goal lies in preparing students to act independently and make good choices in new situations for which they do not immediately know how to act. The work will contribute in enabling ILEs to be more effective by improving student performance in STEM domains, by teaching students to make effective pedagogical decisions, and by making the decisions made by such systems more transparent to both teachers and domain experts.The project will develop and empirically evaluate a general decision-making framework across three ILE or STEM domains. The project naturally integrates both the research and education goals. More specifically, the project will 1) advance research on the application of Reinforcement Learning by adapting it to make hierarchical decisions similar to those of human experts; 2) advance the understanding of ILEs and Reinforcement Learning algorithms by inducing compact policies that highlight key decisions and that can inform one?s understanding of the educational domain; and 3) close the loop by using data-driven policies to support student decision-making and eventually improve their long-term problem-solving abilities through hybrid human-machine interactive decision making in vivo experimentation.