When a classifier is generated, either by the GA or by covering, it has an initial average match set size parameter. The value is set to the average match set size over the whole population, a statistic that is updated by the system regularly.

Thus the deletion probability of a classifier that ends up never matching is not zero, but is instead in the range of most other classifiers.

Note that in recent work, where the GA occurs in the action set, the parameter is the average action set size.

A classifier's probability of deletion is proportional to its average match set size. How do you set the probability of those which are never matched and thus, technically, have a deletion probability of 0? Do you set their match set size to 1 and, eventually, multiply by (pop_ave / initial strength) once the unmatched classifier's initial strength falls below the fraction of the population's average (item 2, page 9)?