I have similar issues. Not to mention that you only get one run done,
so no statistical significance.
Things you may want to try:
- decrease the population size
- start with a smaller number of evaluations (say 5) to discriminate
among good and bad individuals; increase the number of evaluations
later on during the run to discriminate among really good individuals.
(check out Grajdeanu and DeJong paper in Late Breaking Papers of
- try using an island model with N islands and 500/N individuals in
- try distribute evaluations to several computers (I am not sure what
the status of that code is, but there is some code for this stuff
that's being brushed up by one of Sean Luke's students).
- be patient
On Dec 1, 2004, at 11:55 AM, Steve Butcher wrote:
> Dear all,
> I'm currently using GP to search the solution space of a problem in
> pursuit and evasion.
> Using MASON as a simulator and ECJ as the GP package, I've come up with
> a system that does one simulation in approximately 4 seconds. As there
> are 20 trials per evaluation, 500 individuals and 51 generations. This
> comes out to approximately 23 days of computing time. Well, assuming
> there is to be a write-up of results and a prepared presentation, this
> put me about 18 days over what I had available so I took some
> However, I'm wondering informally if anyone on the list has tackled
> problems with similar run times.