Actually, the fact that mutation testing is useful is a sign of one of the major difficulties with automatic algorithm generation: a tiny change in a program ty

Actually, the fact that mutation testing is useful is a sign of one of the major difficulties with automatic algorithm generation: a tiny change in a program typically destroys a piece of functionality completely, which means it's hard to follow any sort of fitness landscape to incrementally create a working program that passes all your tests. When writing an algorithm you're looking for a needle in a haystack, not tweaking parameters to find the peak of a broad mountain, and that takes a lot of the punch out of automated methods. What this means, in part, is that brute-force search will often be just as effective as evolutionary or hill climbing methods. Where evolutionary methods start to shine is when you allow them to reuse bits of code that have already been found to be effective, esp. in other problems - that corresponds a lot more closely to the way that humans write code, we lean on our memories of the types of things that have worked before, for everything from syntax to algorithm design to system architecture.