I really like the idea of the weekly lightning-fast corewars
tournaments on IRC. Not much chance to use a GA in a half hour window
though.  Or could you.
Last week's rules were to submit warriors that only used values of 
"-1", "0", and "+1". That's certainly GA-friendly. Seized by the
spirit of scientific inquiry, I did some tests and yes, when I ran my
evolver program overnight, it found plenty of warriors that would have
bested those actually submitted. (Naturally, I didn't use the warriors
submitted to the contest to train the evolved ones.) But could it have
been done in 30 minutes?
After tireless investigation (before I got tired of it, I quit) I
believe I have the answer. If I'd started with a clean version of the
evolver program, the value restriction could have been accomplished by
the addition of one line of code. Drawing on my many years of proposal
writing, I'll budget one minute for software development and testing.
Similarly, I'll budget zero time for delivery, leaving 29 minutes to
let the GA run.

;redcode verbose
;assert 1
;name Convolution Chromatograph
;author Dave Hillis
;strategy -   Created using RedRace.c.
;strategy -   An evolving population playing KOTH.
;strategy -   For a contest where only "-1", "0", and "+1" are allowed
mod.ba {    -1, #    -1
sub.i  $     1, $     0
mod.i  #     1, *     0
nop.i  }     0, #     0
mod.b  #    -1, }    -1
spl.i  @    -1, #     0
slt.i  @    -1, #     0
spl.i  }     0, {    -1
jmp.f  }     0, #     0
jmp.i  }     0, $     0
sne.ab >     1, $     1
add.ab #    -1, {     1
sub.i  <     0, #    -1
spl.f  #     1, >    -1
mov.ab #    -1, #    -1
mov.i  }    -1, }     0
end   4

Sadly, five independent runs with different parameters failed to
produce competitive results. I needed an order of magnitude more
processing power or, better yet, a contest on a tiny, 800-line core.
Much later, I had another idea: start evolving on a tiny core, then
switch to a standard size right at the end. I've used this technique
before. The 6th run produced the warrior above, which would have won
comfortably.

benchmark for rdrc: Convolution Chromatograph. 200 battles
1: ;Program "2 Craze 2" (length 83) by "Christian Schmidt"
 390 210
2: ;Program "Bad Stone" (length 6) by "Paul Khuong"
 491 80
3: ;Program "BinFool" (length 20) by "Philip Thorne"
 436 82
4: ;Program "BinFool #3" (length 21) by "Philip Thorne"
 597 3
5: ;Program "Desperate" (length 3) by "Miz"
 348 126
6: ;Program "ere we go" (length 4) by "Miz"
 345 219
7: ;Program "noname" (length 4) by "Lukasz Adamowski"
 593 5
8: ;Program "2365*2^2936+1 is prime!" (length 7) by "John Metcalf"
 193 235
9: ;Program "Repli" (length 4) by "Christian Schmidt"
 506 50
10: ;Program "Safe Clear" (length 2) by "Lukasz Adamowski"
 475 82
11: ;Program "Semi-Skilled Redcoder" (length 8) by "John Metcalf"
 186 414
12: ;Program "tmp" (length 9) by "Joonas"
 591 6
score (scaled to 100 battles) = 214.625000


Conclusion: it's just barely possible but you'd have to be lucky and
think reeeeally fast. In my case, the GA wouldn't be the weakest link.

Dave Hillis