MIT Electrical Engineering and Computer Science EECS Event

Why Robbie Can't Learn: The Difficulty of Learning in Autonomous Agents

Leslie Pack Kaelbling
MIT, EECS, Associate Director, AIL

Monday, November 19, 2001
4:00 PM (refreshments 3:45)
Edgerton Hall, Room 34-101
EECS Colloquium

Abstract

In recent years, machine learning methods have enjoyed great success in a variety of applications. Unfortunately, on-line learning in autonomous agents has not generally been one of them. Reinforcement-learning methods that were developed to address problems of learning agents have been most successful in off-line applications. In this talk, I will briefly review the basic methods of reinforcement learning, point out some of their shortcomings, argue that we are expecting too much from such methods, and speculate about how to build complex, adaptive autonomous agents. I will back up the speculations with recent results demonstrating that a small amount of human-provided input can dramatically speed learning in a real mobile robot.