Evolving QWOP gaits
Venue
GECCO '14 Proceedings of the 2014 conference on Genetic and evolutionary computation, ACM, Vancouver, pp. 823-830
Publication Year
2014
Authors
Steven Ray, Vahl Scott Gordon, Laurent Vaucher
BibTeX
Abstract
QWOP is a popular Flash game in which a human player controls a sprinter in a
simulated 100-meter dash. The game is notoriously difficult owing to its ragdoll
physics engine, and the simultaneous movements that must be carefully coordinated
to achieve forward progress. While previous researchers have evolved gaits using
simulations similar to QWOP, we describe a software interface that connects
directly to QWOP itself, incorporating a genetic algorithm to evolve actual QWOP
gaits. Since QWOP has no API, ours detects graphical screen elements and uses them
to build a fitness function. Two variable-length encoding schemes, that codify
sequences of QWOP control commands that loop to form gaits, are tested. We then
compare the performance of SGA, Genitor, and a Cellular Genetic Algorithm on this
task. Using only the end score as the basis for fitness, the cellular algorithm is
consistently able to evolve a successful scooting strategy similar to one most
humans employ. The results confirm that steady-state GAs are preferred when the
task is sensitive to small input variations. Although the limited feedback does not
yet produce performance competitive with QWOP champions, it is the first autonomous
software evolution of successful QWOP gaits.
