That nerve seems like an example of a local optimum. The way to avoid this is either to increase the mutation rate, or modify the fitness function so the interm

That nerve seems like an example of a local optimum. The way to avoid this is either to increase the mutation rate, or modify the fitness function so the intermediate states aren't so heavily penalised. We can't easily do that with real evolution but it's configurable with evolutionary algorithms, as with most machine learning algorithms. The hard part is finding those parameters.