Thursday, November 14, Volen 101, 2:00-3.00 pm
We describe our development of Cobot, a software agent who lives in LambdaMOO, a popular virtual world frequented by hundreds of users. We present a detailed discussion of the functionality that made him one of the objects most frequently interacted with, human or artificial.
Our initial work allowed Cobot to interact reactively; namely, he could collect social statistics and report them to users, and engage in simple conversations. Eventually we used a novel application of reinforcement learning to allow Cobot to take actions proactively, and adapt his behavior from multiple sources of human reward. Over a period of five months of training, Cobot received over 3100 reward and punishment events from over 250 different LambdaMOO denizens, and learned nontrivial preferences for a number of specific users. Finally, we allowed Cobot the ability to move beyond his immediate text-based environment to provide real-time, two-way, natural language communication between a phone user and the multiple users in LambdaMOO.
For each of the applications developed for Cobot, we describe a number of the challenging design issues we faced, both techincal and social, and the methods we used to address them. We report a number of empirical findings from several user studies.
Bio: Dr. Charles L. Isbell, Jr. received his bachelor of science degree in Computer Science in 1990 from the Georgia Institute of Technology. Awarded a fellowship from AT&T Bell Labs, he continued his education at the Artificial Intelligence Laboratory at the Massachusetts Institute of Technology. After earning his PhD from MIT in 1998, Charles joined AT&T Labs/Research. In the fall of 2002, he returned to Georgia Tech to join the faculty of the College of Computing.
Charles' research interests are varied, but the unifying theme of his
work in recent years has been using statisitical machine learning
techniques to build autonomous agents that engage in life-long
learning of individual preferences. These agents build models of the
usage patterns of individuals, rather than discovering trends in large
datasets. His work with agents who interact in social communities has
been featured in The New York Times, the Washington Post and Time
magazine's inaugural edition of Time Digital magazine, as well as in
several technical collections.
Host: Tim Hickey