[Icfp04-discuss] Evolutionary Ants

[Detlef Pleiss]

So, there are some of you who tried the evolutionary approach.
Seen so far: Joseph Gentle, Jin Soo Hwang, Christian Sloma, Jens Hauke.
I myself did so as well.
All in all it looks like that ants implementing human brain generated ideas
perform better than (short time) evolution results.
Then again some of you used evolutionary approaches just to fine tune some
(flip) parameters of hand crafted ant brains.
In my case it was the evolutionary approach that kindled my curiosity, in
how far it would actually generate usable ant brains.
The ant compiler builders got the better ants, but building such a tool
would not have been of so much personal interest to me.
BTW: The evolutionary approach definitely does require the programming of a
working simulator :)
Quoting numbers of some relative strength, without going into detail about
their meaning ...
I found my lightning submissions performing with 12 points
even worse than the example brain counting 14 points
I also tried a completely handwritten ant, that got dismissed with its 42
points
with evolution I got results up to 154
rough checks on others submissions got values so far up to 385

I took the example brain as starting point and was amazed how quickly the
evolution came up with better ants.
I never cared about the detail of how those results work, but I have a
theory why the evolution works fine with the example:
(As I don't have a GUI simulator I can only guess, that) the simple example
strategy has a problem:
once they collected some food, the example ants keep picking it up and
dropping it again in their home.
So, some disruption of just *that* behaviour might easily yield better
results.

I also tried starting the evolution with a random 10000 lines ant brain.
I was amazed to see that even these actually collected some food!
But as during some time they didn't evolve further than a couple points I
dropped that approach.

ciao - Det.