UTCS FACULTY CANDIDATE: Sevan Ficici/Harvard University When Game Theory Isn't Enough: Engineering Agents for an Open and Imperfectly Rational World ACES 2.402 Thursday March 27 2008

Contact Name: 
Jenna Whitney
Date: 
Mar 27, 2008 11:00am - 12:00pm

There is a signup schedule for this event (UT EID req

uired).

Type of Talk: Faculty Candidate

Speaker/Affiliation:
Sevan Ficici/Harvard University

Date/Time: Thursday March 27 20

08 11:00 a.m.

Host: Risto Miikkulainen

Talk Title: When Ga

me Theory Isn''t Enough: Engineering Agents for an Open and Imperfectly Rat

ional World

Talk Abstract:
Games and game-like scenarios pervade

our lives.
Games transpire in situations that involve strategic
rea

soning arising in myriad ways within economic
social and biological

systems. Many important
engineering problems can also be formulated as

games thereby allowing the mathematics of strategic
reasoning kno

wn as game theory to be applied.
Despite its undeniable success class

ical game
theory makes certain assumptions that place
important li

mitations on its applicability. First
game theory assumes that all age

nts are fully
rational actors. Second game theory makes a
closed-w

orld assumption asserting that every
player knows its gamut of strateg

ic alternatives.
While many problems do fit the assumptions of
game
theory the history of human (and pre-human)
activity reveals a consta

nt stream of strategic
innovation and divergence from rational behavior

;
the real world is in fact open-ended and populated
with imperfec

tly rational actors. How can we build
computer agents to successfully p

articipate in such
a world? This talk presents research aimed towards <

br>building such agents. We first discuss the idea of open-
endedness a

nd provide a formalism with which to relate
the process of open-ended s

trategic innovation to game
theoretic solution concepts. We then presen

t experimental
work on the construction of computer agents whose decisi

on
making is informed by models of human reasoning; these
models a

re learned from observed human behavior. We
conclude with directions f

or future work on improving the
scalability of game theoretic reasoning

.