Martin V. wrote:
> Even if I change the seed to the computer time (instead of default 4357)
> and change the number of generations from 51 to 100 it doesn't help - the
> program still finds only about 70 pieces of food.
> In the literature the evolution should end already in generation 21 where
> it should find an ideal individual which eats all 89 pieces.
>
From my experience the artificial ant has about a 10--20% chance of
success. Steffen Christensen's thesis has a very large number of runs
of the artificial ant. You could check out his results if you're
interested in what performance to expect. But have you tried more than
a few runs? You might like to try perhaps 100 runs---in which case you
could expect 10 to 20 runs to succeed if you use parameters similar to
Koza's.
Another issue that you should consider is the number of timesteps
available. Using Open Beagle I found that 600 (rather than the
specified 400) timesteps were required for Koza's ideal solution to
succeed. Christensen and Oppacher (2002) also noted this issue. I'm
not sure how many ECJ uses by default, but this could certainly be the
root-cause if Koza's "100%-correct" solution isn't working.
Cheers,
Matthew
|