Methods for Competitive Co-evolution: Finding Opponents Worth Beating
Chris Rosin
University of California at San Diego
2:30pm Friday July 14 in Room 2310
Co-evolution refers to the simultaneous evolution of two or more
genetically distinct populations with coupled fitness landscapes. In
this talk I consider ``competitive co-evolution,'' in which the
fitness of an individual in a ``host'' population is based on direct
competition with individual(s) from a ``parasite'' population.
Measured fitness is used to direct the evolution of these populations
with a genetic algorithm. I will present results of the application
of competitive co-evolution to three game-learning problems:
Tic-Tac-Toe (TTT), Nim and a small version of Go. Two new techniques
in competitive co-evolution will be described. ``Competitive fitness
sharing'' changes the way fitness is measured, and ``shared sampling''
alters the way parasites are chosen for testing hosts. Experiments
using TTT and Nim show a substantial improvement in performance when
these methods are used. Preliminary results using co-evolution for
the discovery of cellular automata rules for playing Go will be
presented.
(Work done in collaboration with Rik Belew, CSE-UCSD, now visiting
Computer Science, Univ. Wisc.)