Login
Computer Science @ Rochester
Tuesday, May 03, 2005
1:00 PM
CSB 703
Ph.D. Thesis Proposal
Phillip Michalak
University of Rochester
Natural Language Dialogue in Procedure Learning
There are multiple reasons for wanting computers to be able to perform some of the tasks that humans do, varying from practical concerns of human safety and task efficiency to the goal of developing a complete intelligent agent. In order for a computer program to perform these tasks, it must somehow acquire knowledge of the steps involved in performing the task. Traditional knowledge engineering approaches to acquiring this knowledge are manpower intensive, requiring a computer programmer to acquire proficiency in a domain in order to code the knowledge in a form suitable for computer reasoning. Recent work in procedure learning has shown that this burden can be ameliorated by having the computer system learn from expert demonstration. In my work, I will focus on a procedure learning method that uses natural language descriptions of a task as it is being demonstrated in order to learn the structure of the task (sub-task decomposition and causal relationships among sub-tasks) as well as a procedure (sequence of primitive actions) for accomplishing the task. This approach reduces task annotation requirements without sacrificing the capability to learn richly structured task models. In this talk I motivate and describe this approach to procedure learning.