From: Lillie Kittredge (kittredl_at_u.washington.edu)
Date: Mon Oct 20 2003 - 12:28:44 PDT
Jacob Eisenstein's "Evolving Robot Controllers" describes his attempts to use genetic algorithms to create controllers for RoboCode tanks. His objective was to evolve controllers for the tanks that would rival hand-coded controllers.
One of the central issues was the representation chose for the genome of each tank. Though the first instinct might be to try to evolve the java code itself, it's unlikely one would get compilable code out of such an effort. Instead, he created his own architecture for robot behaviour, which had more discrete sections, and parameters to each. This was helpful because it meant that tanks would be able to produce viable offspring, but it also put some limitations on the kinds of thinks that tanks could do.
I rather liked that the evolved tanks rarely fired, and instead avoided other tanks' fire and occasionally rammed. I'd like to extrapolate some sort of lesson about pacifism from this fact, but I don't think it'd really apply.
I would be rather interested to see more about coevolution in this environment. I've seen evolved agents in other environments which are competing to get to a goal, and will evenetually learn to fight eachother. Perhaps if there was a bonus in the fitness function which reinforced tanks for damaging competetors, there would be a lower incidence of the "catatonic" tanks Eisenstein mentions in his coevolved populations. It'd also be interesting to see how co-evolving tanks would deal with a hand-coded tank in their midst that fired upon them all. Would they defeat the agressor and then go back to pacifism, or would they keep figting eachother once the aggressor was defeated? (Assuming that they could defeat it.)
The only major complaint I have about this paper is how much the results hinge on the situations where the environment is stacked in favor of the evolved tanks- making only one adversary, or letting the tank start at the same positon every time. I'd be more impressed if the tanks evolved in the "multiple adversaries, multiple starting positions" situtation had done better than the hand-coded tanks.
What I wonder about this whole endeavor is whether intelligent fighting behavior really can be evolved, or whether intelligence iteself must first be evolved (which can't be acheived in as limited as environment as this) and then applied to the situtaion.
This archive was generated by hypermail 2.1.6 : Mon Oct 20 2003 - 12:28:45 PDT