CAREER: A Multiagent Teacher/Student Framework for Sequential Decision Making Tasks

点击翻译按钮获取中文摘要

Physical (robotic) agents and virtual (software) agents are becoming increasingly common in industry, education, and domestic environments. Although recent research advances have enabled agents to learn how to complete tasks without human intervention, little is known about how best to have humans teach agents or agents teach other agents or even how agents might teach humans. Considering the full matrix of agent/human learning, in which either an agent or a human can play the role of teacher or student, would increase the potential benefits of leveraging human and agent expertise and knowledge. This project aims to study agent/human learning in the context of sequential decision-making problems, a class of central importance for real-world agent systems. This project aims to develop a novel teacher/student framework that integrates autonomous learning with teaching by another agent or a human. The project plans to develop and evaluate a set of core algorithms to allow: (1) agents to teach agents, thus enabling robust knowledge sharing among agents; (2) humans to teach agents, thus allowing humans to share or transfer common sense or domain-specific knowledge with agents; and (3) agents to teach humans, thus helping humans better understand how to perform or recast sequential decision-making tasks already understood or performed by autonomous agents. In all cases, the goal is to develop methods that significantly improve learning performance relative to learning without guidance from a teacher. Issues to be explored include mismatch between teacher/student abilities, learning from multiple teachers, and shared knowledge representation between teacher/student. The PI plans to focus on several scenarios, each with different sets of assumptions about the knowledge or skill of the student or teacher and the kind of interaction possible between them (e.g., whether the teacher can tell the student what action to take). The techniques developed in the project will be evaluated in a variety of tests domains and will involve simulations as well as actual robots.The teacher/student framework will enable agents to teach other agents and humans, as well as integrate autonomous learning with agent and human teaching. Understanding how to best teach agents is of key importance in developing deployable agent systems. The platform- and domain-independent approach incorporates ideas from multiagent systems, machine learning, human-computer interaction, and human-robot interaction communities, and has the potential to impact each of these areas. This work takes a step towards transitioning agents from specialized systems usable only by experts into useful tools and teammates for people without programming expertise. This project has a strong educational component. The PI teaches at an undergraduate college and undergraduate students will play a crucial role throughout the project. Furthermore, the research produced by this project will be incorporated into five of the PI's courses, providing exciting new material to attract and retain computer science majors. The PI will also continue outreach to secondary school students as well as to underrepresented groups via Lafayette College's S-STEM and Higher Achievement programs.