From a practical point of view the machine resources consumed by any system which suffers from bloat will prevent extended operation of that system. However in practice we may not wish to operate the system continually. For example it may quickly find a satisfactory solution or better performance may be achieved by cutting short its operation and running it repeatedly with different starting configurations [16, page 758,].
In some data fitting problems growth in solution size may be indicative of ``over fitting'', i.e. better matching on the test data but at the expense of general performance. For example [1, page 309,] suggests ``parsimony may be an important factor not for `aesthetic' reasons or ease of analysis, but because of a more direct relationship to fitness: there is a bound on the `appropriate size' of solution tree for a given problem''.
By providing a ``defence against crossover'' [12, page 118,] bloat causes the production of many programs of identical performance. These can consume the bulk of the available machine resources and by ``clogging up'' the population may prevent GP from effectively searching for better programs.
On the other hand [3, page 84,] quotes results from fixed length GAs in favour of representations which include introns, to argue we should ``not ... impede this emergent property [i.e. introns] as it may be crucial to the successful development of genetic programs''. Introns may be important as ``hiding places'' where genetic material can be protected from the current effects of selection and so retained in the population. This may be especially important where fitness criteria are dynamic. A change in circumstance may make it advantageous to execute genetic material which had previously been hidden in an intron.
In complex problems it may not be possible to test every solution on every aspect of the problem and some form of dynamic selection of test cases may be required [22]. For example in some cases co-evolution has been claimed to be beneficial to GP. If the fitness function is sufficiently dynamic, will there still be an advantage for a child in performing identically to its parents? If not, will we still see such explosive bloat?