Genetic Programming of Multi-agent System in the RoboCup Domain

by Jonatan Aronsson 

The thesis work was done first at the Department of Electrical and
Computer Engineering, University of Waterloo, Canada and then at our
department, LTH.

The opponents are: Pär Olsson (d99) and Peter Larsson (d99pl).

Wednesday, 26th November, 15.15, glasburen


Abstract

Simulated robotic soccer is a frequently used test method for
contemporary artificial intelligence research. It provides a real-time
environment with complex dynamics and sensor information that is both
noisy and limited. Team coordination between the robots is essential
for success.  
Genetic programming is a method of machine learning that is developed
from the principle of survival of the fittest. A population of
computer programs is generated and each program is tested against a
fitness function. The best programs according to the fitness function
are cloned, mutated, and recombined to create a new generation of
programs. This process continues until the evolved programs satisfy a
user defined criterion. 
In this study, genetic programming is used to teach software robots to
play soccer. The robots quickly learn to chase the ball and kick the
ball towards the goal. Although the strategy is successful, a number
of players develop defensive abilities and recognize that team
coordination is necessary for further development.  
Due to the extremely demanding fitness evaluation several compromises
were made to limit the duration of the 'evolution'. The length was cut
down to 2 weeks per run and the compromises consequently resulted in
weaker robots. 
Conclusions of this study are that more work is needed on the fitness
evaluation, the set of terminals and functions should be extended, and
more computational resources are necessary in order to develop better
performing players.