🚀 go-pugleaf

Satinder Singh, formerly of University of Colorado, formerly of AT&T
Research, presently at Syntek Capital (a VC firm), will be visiting 
November 1.  He will be giving the dept colloquium, but is also giving
an additional talk to folks interested in machine learning.  Here
are abstracts for the two talks.

===============================================================================
Date:  Thursday October 1, 2001
Time:  9:30-10:30
Room:  ECCR 1B08


Cobot in LambdaMOO: A Social Statistics Agent that Learns

Satinder Singh*
Syntek Capital


I will describe our development (at AT&T labs) of Cobot, a
software agent who lives in LambdaMOO, a popular virutal world that is
frequented by hundreds of users. Cobot's goal is to interact with
other members of the community and to become a vital, useful and
accepted part of his social fabric. In the first part of this talk I
will present a detailed discussion of the hand-crafted functionality
that has made him one of the objects most frequently interacted with
in LambdaMOO, human or artificial. In the second part of the talk I
will describe our application of reinforcement learning to allow Cobot
to proactively take actions in its complex social environment, and
adapt his behavior from multiple sources of human reward.

For a Washington Post article on Cobot follow this link:
http://www.washingtonpost.com/wp-dyn/articles/A49940-2000Dec25.html

===============================================================================
Date:  November 1, 2001
Time:  3:30-4:30
Room:  ECCR 265


	Electronic Commerce, Auctions, and Computational Game Theory:
	Some Results that Exploit Computational and Representational
                   Restrictions on the Behavior of Agents

			       Satinder Singh
			       Syntek Capital
	
	Multi-agent games are becoming an increasingly prevalent
formalism for the study of electronic commerce and auctions. In order
to act efficiently in such complex settings agents will adopt both
representational and computational restrictions on their behavior
(much as in the rest of artificial intelligence and machine learning).
Computational game theory studies the impact of such restrictions on
the outcome of games. 

	In this talk, I will present some of my recent work with
colleagues at AT&T labs in this area beginning with a motivating
description of a trading-agent competition in which we
participated. Following that, I will present a result that analyzes
the behavior of agents that incrementally adapt their strategy through
gradient ascent on expected payoff, in the simple setting of
two-player, two-action, iterated general-sum games. I will show that
surprisingly either the agents will converge to a Nash equilibrium, or
if the strategies themselves do not converge, then their average
payoffs will nevertheless converge to the payoffs of a Nash
equilibrium. Finally, I will introduce a compact graph-theoretic
representation for multi-party game theory and provide a provably
correct and efficient algorithm for computing approximate Nash
equilibria in (one-stage) games represented by trees or sparse graphs.