Crash when all individuals are invalid about ponyge2 HOT 11 CLOSED

Hmm, an interesting one. Crossover currently doesn't allow invalids to be created (i.e. crossover will continue until all children in the population are valid), but there's no such catch for mutation. Should we implement something like that for mutation?

from ponyge2.

from ponyge2.

Probably the default should be to allow crossover to create invalids since that is the published algorithm. Maybe add switches --no-crossover-invalids and --no-mutation-invalids.

from ponyge2.

Switch added for --no_crossover_invalids. However, mutation is not quite so easy to put various checks in place (e.g. limitations on max depth/nodes/invalids) as the current mutation is done with a list(map()) function, necessitating a bit more thought to sort out.

from ponyge2.

Mutation now has the same checks as crossover. All checks have been moved to utilities.representation.check_methods.check_inds. Checks can be extended to suit.

from ponyge2.

Setting the default fitness is now the responsibility of each individual fitness function. By default, we have set the default_fitness value of each fitness function to numpy.NaN. This allows us to calculate population statistics over the entire population by filtering out these np.NaN values using numpy's np.nanmax(), np.nanmin(), and np.nanmean() functions. PonyGE will no longer crash if all individuals are invalid.

from ponyge2.

from ponyge2.

All this info should be in the README I presume. RTFM issue...

from ponyge2.

from ponyge2.

Agreed. It's gonna take a fair while to ensure everything is in the README. When the code review is being done it might be an idea to check that every aspect is documented in the README somehow...

from ponyge2.

from ponyge2.