No, deletion occurs from the population as a whole, not from the match set currently undergoing a GA. Of course the deletion could "hit", by chance, in the current match set.

The basic philosophy is to keep all the match sets approximately the same size. Thus a given classifier's deletion probability is proportional to its estimate of the size of the match set it occurs in. The result is that match sets that get substantially bigger than others suffer deletions bringing them back to the typical size. Similarly for atypically small match sets. They tend to be protected from deletion until the GA gets a chance to bring them up to average size.

This deletion philosophy would certainly be subject to modification if for some reason it is not desired to equalize the match sets.

Does this imply that the only classifiers eligible for deletion are those in the match set?