RI: Small: Collaborative Research: Speeding Up Learning through Modeling the Pragmatics of Training

多年来开发能够从奖励信号中学习的算法的努力已经产生了大量可以利用基于性能的数值信号的技术。最近使用这些技术学习人类提供奖励的努力进展缓慢，部分原因是人类给出的反馈是离散的，而不是数字的。该项目提供了新的学习算法，专门用于利用人类做出的选择中包含的信息来提供这种离散的反馈。这些算法的灵感来自于人类与狗的伙伴关系，以及人类仅通过离散的反馈和精心构建的任务序列就能教会狗的令人难以置信的事情。本项目开发的贝叶斯学习框架将利用反馈和任务序列中包含的实用意义，从人类反馈中更快地学习。这项工作的最终目标是为人类提供一个更自然的范例，让人类告诉计算机他们想让计算机做什么。为此，项目的努力将为布朗大学带来一个教学模块。学习交流（LE）。LE包括本科生与缺乏服务的少数民族中学生一起工作，让他们参与STEM。他们是展示这项工作更广泛影响的完美观众。LE参与者学习使用Scratch环境和反馈范例的编程组合来指导计算机，这显示了算法的强大。

Years of effort to develop algorithms capable of learning from reward signals have resulted in a plethora of techniques that can leverage numerical signals that vary in value based on performance. Recent efforts to use these techniques to learn from humans providing rewards have been slower to progress, in part, because humans give feedback discretely rather than numerically. This project contributes new learning algorithms designed specifically to leverage the information contained in the choices humans make to provide such discrete feedbacks. The algorithms are inspired by the human-canine partnership, and the incredible things that humans are able to teach dogs using only discrete feedback and carefully constructed sequences of tasks. The Bayesian learning framework being developed in this project will leverage the pragmatic implicatures contained in the feedbacks and tasks sequences to learn more quickly from human feedback. The ultimate goal of this work is to provide a more natural paradigm for humans to tell computers what they would like for them to do. To that end, project efforts will result in a teaching module for Brown University?s Learning Exchange (LE). The LE involves undergraduates working with underserved minority middle school students to engage them in STEM. They are a perfect audience to demonstrate the broader impacts of this work. LE participants learn to instruct computers using a combination of programming with the Scratch environment and the feedback paradigm, which shows how powerful the algorithms are.