A major drawback of evolutionary computation is that only relatively small-scale solutions can be found during the search process—it is currently impossible to evolve a whole chunk of software. We propose a method named evolutionary software improvement, which makes it possible to take some module, or some aspect of an existing software system, and improve just that part using evolutionary computation. We apply evolutionary software improvement to our genetic-programming system Megavac, refining its instruction set for a class of multiinput handling problems. This allows us to develop problemfitting instruction sets in a meta-circular fashion, and leads to surprising solutions evolved by Megavac for the given class of problems.
|Title of host publication||Proceedings of Workshop and Summer School on Evolutionary Computing (WSSEC 2008)|
|Number of pages||4|
|State||Published - 2008|