Remember that elephants (and all other life on this planet) evolved from single-celled organisms (and even from non-life), which are far more different from elephants than even apes are. So drastic evolutionary changes do happen in the real world. But they just take a very long time.
The problem is that not only do we not have millions or even billions of years to evolve our artificial organisms, but that the environments in which our artificial organisms grow, their representation, and the operations performed on them during their "lifetime" and during breeding are incredibly restricted and simplified compared to what happens in the real world.
It's very difficult to model even a tiny fraction of that (see the supercomputer time dedicated to creating a detailed model of even a single neuron, for instance). And, as for the programs that are evolved via GP, it's rare for them to be allowed to create completely arbitrary code. Their instruction sets are usually very limited so as to reduce the search space. They are also usually given quite limited data to work with, are modified very little during their lifetimes (if at all), and the sorts of things that can happen during breeding are similarly curtailed. So it's really no wonder that their existence and evolution is not always so spectacular as what nature has been able to achieve over the millenia.
Now, the problem you mention regarding the disruptive nature of mutation (crossover can also have this effect) is also important. The attempts to solve it have generally focused to finding ways to reduce this disruption (usually by making the crossover or mutation less random). I recommend searching Citeseer for "destructive crossover" to see examples of some proposed solutions.
I meant that changes are not drastic in a single level (at least on phenotype level). You don't expect progeny of an ape-like ancestor to give birth to an elephant like organism. But that is what precisely happens with Genetic Programming.
Remember that elephants (and all other life on this planet) evolved from single-celled organisms (and even from non-life), which are far more different from elephants than even apes are. So drastic evolutionary changes do happen in the real world. But they just take a very long time.
The problem is that not only do we not have millions or even billions of years to evolve our artificial organisms, but that the environments in which our artificial organisms grow, their representation, and the operations performed on them during their "lifetime" and during breeding are incredibly restricted and simplified compared to what happens in the real world.
It's very difficult to model even a tiny fraction of that (see the supercomputer time dedicated to creating a detailed model of even a single neuron, for instance). And, as for the programs that are evolved via GP, it's rare for them to be allowed to create completely arbitrary code. Their instruction sets are usually very limited so as to reduce the search space. They are also usually given quite limited data to work with, are modified very little during their lifetimes (if at all), and the sorts of things that can happen during breeding are similarly curtailed. So it's really no wonder that their existence and evolution is not always so spectacular as what nature has been able to achieve over the millenia.
Now, the problem you mention regarding the disruptive nature of mutation (crossover can also have this effect) is also important. The attempts to solve it have generally focused to finding ways to reduce this disruption (usually by making the crossover or mutation less random). I recommend searching Citeseer for "destructive crossover" to see examples of some proposed solutions.