22 Apr 2003
(updated 22 Apr 2003 at 23:44 UTC) »
I decided to try a traditional 'sort' algorithm.
This involves taking all the individuals, sorting them
into order based on fitness and then looking at the top
few/bottom few for a solution.
A few cute things showed up.
Firstly, when I replaced the 'cull' pass with the results
of this sort algorithm (ie kill individuals until we're
under the desired population) the solution time (in
generations) dropped by about 2/3.
Secondly, no matter what I try I can't seem to
coax the best solution out of the code. Admittedly with
this kind of stuff you're looking for a good
solution but I can't imagine this algorithm performing
that much worse. The population quickly homogonises when
I'm using the sorted array for reproduction selection
and stopped a few bits short of a perfect solution.
One interesting observation with the sorted algorithm -
a higher population count does _NOT_ necessarily result
in a quicker search. In fact with a population of 10,000
vs 1,000 the average fitness scores over time were
significantly smaller and sometimes decreasing.
Gah, lots of work to do.
(Note: yes, I know this kinda stuff has been done before.
Yes, I know it'd be easy to do a google search, grab
someone's evaluation engine and go from there. It just
wouldn't be as fun.)