There are multiple approaches to doing that.  Here are two you may want
to try:

1. Create a custom statistics object that waits until the end of the
run, then takes the best individual and tests it on the unseen examples
2. Use a statistics file that outputs the best individual to a file in
a machine-readable format.  Then start another run that has a single
individual in the population (the population is initialized from a file
containing that single individual) and lasts a single generation.  Here
are some parameters you may want to use:

eval.problem =

pop.subpop.0.size =             1
stat.file =                                     unseen-examples.stat
pop.subpop.0.file =             <filename-with-best-individual>
generations =                   1



On Sep 9, 2005, at 5:40 PM, Allan de Brueker wrote:

> Dear Sean,
> I've build some sort of classifier with ECJ,
> the fitness of the classifier was trained by feeding it a number of
> individuals and their classes.
> Is there a way for me to extract the program generated by ECj and to
> run it
> on a number of other individuals, to test if it works well on unseen
> examples?
> Kind Regars,
> Allan