CAREER: Socially Guided Machine Learning

职业：社会引导的机器学习

• 批准号: 0953181
• 负责人: Andrea Thomaz
• 金额: $ 54.11万
• 依托单位: Georgia Tech Research Corporation
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2010
• 资助国家: 美国
• 起止时间: 2010-01-01 至 2015-12-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0953181&HistoricalAwards=false
• 关键词: CAREER; Socially; Guided; Machine; Learning

There is currently a surge of interest in service robotics, a desire to have robots leave the labs and factory floors to help solve issues facing society. But if robots are to play a useful role in domains ranging from eldercare to education, they will need the ability to interact with ordinary people and to acquire new relevant skills after they are deployed; we cannot pre-program these robots with every skill they might conceivably need. The PI's approach to solving this critical issue is Socially Guided Machine Learning (SG-ML). In this project she will explore ways in which machine learning agents can exploit principles of human social learning. An important question for SG-ML is "What is the right level of human involvement?" Previous efforts in machine learning systems that use human input have tended to hold this level constant (e.g., guidance oriented approaches that are completely dependent on human instruction in order to learn, and exploration oriented approaches with limited input from a human partner). The PI, taking her inspiration from human learning and from her prior work in robot learning, posits that a robot should be able to explore and learn on its own, while also taking full advantage of a human partner's guidance when available. The PI's goal in this work is to successfully incorporate self and social learning within a single SG-ML framework, enabling a robot learner to dynamically adjust to varying levels of human involvement in the learning process. To this end, the PI will seek to make progress toward four main objectives:1) Motivations for learning: Typically machines learn because they are programmed to do so, unlike children and animals who learn because they are motivated to master their environment. A key component of this work is computational motivations that drive a robot to good learning opportunities.2) Multiple learning strategies: As mentioned above, an SG-ML framework should have a repertoire of self and social learning mechanisms working together. A central issue of this research is how the robot should best arbitrate and manage the use of multiple strategies.3) Transparency devices: Learning is a collaborative activity. The learner's behavior has to be understandable, and has to express a certain level of internal state to help the teacher guide the learning process. Transparency is a fundamental issue of this work, developing robot behaviors that successfully communicate the progress of the learning process.4) Engagement mechanisms: In human social learning, teaching is a rewarding process for both the learner and the teacher. This is a positive feedback loop from which a machine learner could benefit. A primary component of this work is to develop mechanisms that make the teaching process rewarding.Broader Impacts: The long-term promise of this research is robots in society able to adapt and learn from everyday people. The core principles developed in this work will one day enable robots to adapt and learn about the changing needs of people in their homes, or staff in a hospital. The lessons learned about social learning with robots will be relevant both to computational devices and to human-computer interaction in general. The PI will exploit the fact that social robot projects like this one generate particular interest in the community to conduct outreach programs in local area high schools, to raise awareness about the work of women in science, and to stimulate the American public's interest in science.

目前，人们对服务机器人的兴趣激增，希望机器人离开实验室和工厂车间，帮助解决社会面临的问题。 但是，如果机器人要在从老年人护理到教育等领域发挥有用的作用，它们将需要与普通人互动的能力，并在部署后获得新的相关技能;我们无法为这些机器人预先编程，使其具备可能需要的所有技能。 PI解决这个关键问题的方法是社会引导机器学习（SG-ML）。 在这个项目中，她将探索机器学习代理可以利用人类社会学习原理的方法。 SG-ML面临的一个重要问题是“人类参与的正确水平是多少？“以前在使用人类输入的机器学习系统中的努力往往保持这个水平不变（例如，完全依赖于人类指令以便学习的导向指导的方法，以及具有来自人类伙伴的有限输入的导向探索的方法）。 PI从人类学习和她之前在机器人学习方面的工作中获得灵感，她认为机器人应该能够自己探索和学习，同时也充分利用人类伙伴的指导。 PI在这项工作中的目标是成功地将自我学习和社会学习整合到一个SG-ML框架中，使机器人学习者能够动态地适应人类参与学习过程的不同程度。 为此，PI将寻求朝着四个主要目标取得进展：1）学习动机：通常机器学习是因为它们被编程这样做，而不像儿童和动物那样学习是因为它们有动力去掌握它们的环境。 这项工作的一个关键组成部分是推动机器人获得良好学习机会的计算动机。2）多种学习策略：如上所述，SG-ML框架应该具有一系列共同工作的自我和社会学习机制。本研究的一个中心问题是机器人应该如何最好地仲裁和管理多种策略的使用。3）透明设备：学习是一种协作活动。学习者的行为必须是可理解的，并且必须表达一定程度的内在状态，以帮助教师指导学习过程。 透明度是这项工作的一个基本问题，开发成功沟通学习过程进展的机器人行为。4）参与机制：在人类社会学习中，教学对学习者和教师来说都是一个有益的过程。 这是一个积极的反馈循环，机器学习者可以从中受益。 这项工作的一个主要组成部分是开发使教学过程获得回报的机制。更广泛的影响：这项研究的长期前景是社会中的机器人能够适应日常生活并向人们学习。 在这项工作中开发的核心原则有一天将使机器人能够适应和了解人们在家中或医院工作人员不断变化的需求。 关于机器人社会学习的经验教训将与计算设备和一般的人机交互相关。 PI将利用这样的社会机器人项目在社区中产生特别兴趣的事实，在当地高中开展外展计划，提高人们对女性科学工作的认识，并激发美国公众对科学的兴趣。