If a newly created classifier is "merged" into an existing (macro)classifier, the only effect on the latter is to increment its numerosity by one and its "msize" estimate by one. (Msize is the classifier's estimate of the size of the match sets in which it occurs--used for deletion (see point 1. in Section 3.3)). No other parameters are changed.

Notice that if a new classifier *can* be merged, we side-step the problem of choosing its initial parameters--it simply takes on, in effect, the parameters of the classifier into which it is merged, parameters that have already been tested over time. This is an advantage of macroclassifiers not mentioned in the paper.

"Macro classifiers diminish the diversity of the population". I'm not sure what you mean. They do not diminish the condition/action diversity. Maybe you mean that the micro-classifiers represented by a macro-classifier would have a "spread" of predictions, errors, fitnesses, etc. This is true, at least to some extent, although the updating process would gradually bring them to the same values. I don't do micro/macro comparisons any more, because I earlier found no important performance difference. You could be right that it matters, but I think the effect would be slight.

Your mentioning the Time-stamp led me to examine the program in detail. What happens is this. When a classifier is created, it is time-stamped. But also, whenever the GA occurs in a match set, all the classifiers of the set are time-stamped. The paper does not mention the latter time-stamping. I think you were worried that if a new classifier got merged into an existing classifier, its time-stamp would in effect be lost. This is correct. But if the "absorbing" classifier were from the same match set (i.e., the set in which the GA occurred) it would not matter. If not, I just don't know!

Suppose a new (micro) classifier is created. If there already exists a classifier with same condition/action string then the numerosity counter of the macro classifier is incremented by one. What happens to the other parameters such as the Time-Stamp used for the GA?

Macro classifiers diminish the diversity in the population. Don't you think this can have a negative effect in other applications?