From: Tyler Robison (trobison_at_cs.washington.edu)
Date: Sun Oct 19 2003 - 23:32:30 PDT
Evolving Robot Tank Controllers
Jacob Eisenstein
This paper discusses using genetic algorithms to create a
competitive robot in the game/simulation of RoboCode.
The most important idea is of course the use of genetic algorithms
to create a robot in a game where the robots are almost exclusively
hand-coded. Although RoboCode is a simple and simulated environment, it
provides an interesting experiment in which genetic algorithms can result
in a product that can out perform many robots designed by people; a very
encouraging demonstration of the usefulness of genetic programming.
Another important idea is that although robots are evolved, the success of
the method doesn't depend entirely on the evolution. In the paper we see
repeatedly that the author's choices in implementation are critical to the
success of the robot (this is especially evident in the descriptions of
different fitness functions on pages 8 & 9).
Despite encouraging results, there are numerous flaws. First and
foremost, impressive results were mentioned when robots were evolved to
fight the showcase robot "SquigBot", these results are not very
meaningful. The evolved bot could only win 50% of the time, when multiple
starting positions were used, when evolved specifically to fight SquigBot,
whereas SquigBot could presumably hold it's own against numerous types of
bots. The results are impressive, but it cannot be said that the evolved
robot is better than SquigBot in a general sense.
Another flaw in my mind is the author's treatment of the topic of
the gun. It is certainly understandable, as the author explains it, that
evolved robots would shun the use of the gun; it is costly to use (as it
drains energy), and would require significant sophistication before it
could even be remotely useful; something that is unlikely to happen in an
evolutionary system, in which improvements need to be somewhat gradual.
The author suggests the explanation that perhaps using the gun is a poor
strategy, and that perhaps the fact that most hand-coded bots use the gun
is just preconceived notions of effectiveness and irrationality of their
creators. While this could certainly be true, it is untested, and comes
off as rationalizing the evolved gun-less bots. That gun-using bots can
be effective is evidenced by SquigBot, which has 400 lines of code for
it's targeting system. It is conceivable that genetic algorithms could be
used to develop effective gun-using bots (the author mentions that adding
trigonometric functions might work), and that whole area is left
unexplored.
A significant research question in my mind involves exploring
the usage of the gun, and as a result, targeting, as I think it would open
up an entirely new side to the evolution.
Another question is what would happen if a similar experiment were
carried out given more computing power and time; it is mentioned
repeatedly that the author was constrained in certain situations by the
time it takes to evolve each generation (particularly when testing against
multiple adversaries in multiple starting positions), and so being able to
carry out the evolution faster and for longer might develop more effective
robots (and it might also expose inherent plateaus and limitations of the
technique).
This archive was generated by hypermail 2.1.6 : Sun Oct 19 2003 - 23:32:31 PDT