Building Cognitive Tutors with Programming by Demonstration: When Simulated Students help Cognitive Modeling and Educational Studies

本计画将开发一个新颖的认知型教师智能写作系统。该系统包括一个机器学习设施-称为模拟学生-生成一个“认知模型”，可以像学生一样执行主题任务。模拟学生学习认知模型中的产生式规则，就像人类学生一样：通过观察解决问题的例子，阅读相关的说明，并获得解决方案尝试的反馈。作者提供了示例、英语说明和解决方案反馈，因此更像是一名教师而不是程序员。通过这种方式，模拟学生允许作者绕过目前构建认知导师的认知模型组件所需的耗时且难以学习的编程。模拟学生背后的关键技术是通过演示进行编程，其中机器学习紧密集成到用户界面中。我们计划使用基于归纳逻辑编程和统计关系学习的学习方法。利用所提出的智能创作系统，作者将执行以下步骤来构建导师：（1）为导师构建图形用户界面（GUI）;（2）演示解决方案，通过用GUI解决问题并用指导消息注释步骤，从而为模拟学生提供训练数据以学习复制演示的产生式规则;以及（3）测试产生式规则，并向模拟学生反馈其对另一组问题的解决方案。这项工作建立在一套认知导师创作工具（CTAT）之上，该工具包括一个内置的GUI构建器、一个演示解决方案步骤并对其重新编码的工具，以及编写和测试产生式规则的工具。拟议的智能创作系统将通过将模拟学生集成到CTAT中来实现，因此所有三个创作步骤都将是现有创作工具的自然扩展。该项目在推进智能系统技术和通过允许教师成为作者来提供更好的认知导师方面具有更广泛的影响。

This project will develop a novel design for an Intelligent Authoring System for Cognitive Tutors. The system includes a machine learning facility - called the Simulated Student - that generates a "cognitive model" which can perform the subject task just as students do. The Simulated Student learns the production rules in the cognitive model much like a human student: by observing examples of worked-out problem solutions, reading associated instructions, and getting feedback on solution attempts. The author provides the examples, English instructions, and solution feedback, and thus acts more like a teacher than a programmer. In this way, the Simulated Student allows authors to bypass the time-consuming and difficult-to-learn programming that is currently required to build the Cognitive Model component of a Cognitive Tutor. The key technology behind Simulated Students is programming by demonstration, in which machine learning is integrated tightly into a user interface. We plan to use learning methods based on Inductive Logic Programming and Statistical Relational Learning. With the proposed Intelligent Authoring System, an author will perform the following steps to build a tutor: (1) build a graphic user interface (GUI) for the tutor; (2) demonstrate solutions, by solving problems with the GUI and annotating steps with instructional messages, thus providing training data for the Simulated Student to learn production rules that replicate the demonstrations; and (3) test the production rules and giving the Simulated Student feedback on its solutions for another set of problems.This effort builds upon a suite of Cognitive Tutor Authoring Tools (CTAT), which includes a built-in GUI builder, a facility to demonstrate solution steps and recode them, and facilities to write and test production rules. The proposed Intelligent Authoring System will be implemented by integrating the Simulated Student into CTAT, so all three authoring steps will be a natural extension of existing authoring tools. This project has broader impacts in advancing intelligent systems technology and in providing better cognitive tutors by allowing teachers to be authors.