As noted, in a traditional classifier system, classifiers get discovered and "converged upon" (meaning they are the only ones in the population, or the only strong ones) that solve the problem--more or less. If the environment should change, they may no longer constitute a good solution.

However, since there are no other significant classifiers in the population, the only way the system can test better responses to the new environment is to actually discover (via the GA or perhaps covering) new classifiers to test. This will be difficult, since the GA has only the old, no-longer-so-good classifiers to work from.

XCS, in contrast, aims to evolve a population that forms a relatively complete mapping of the payoff consequences of each action in each situation (X x A => P). So XCS knows not only which action is most remunerative in a given situation, but what the other actions are worth. I.e., it has the other classifiers.

Consequently, if the environment changes in the sense of changing the payoffs, XCS can adapt by simply modifying the predictions of existing classifiers--it does not have to discover new ones.

I have not yet experimented with XCS in changing environments, though!

So does this mean that your system will work better than traditional classifier systems in changing environments as well, where new classifiers might suddenly be called for?