Evolving populations of expert neural networks

In standard neuroevolution, the goal is to evolve one neural network that would compute the right answer most often. However, it often turns out that the population as a whole could perform even better, if we could only choose the right network for each input. One way to do this is to evolve networks that output not only the answer, but also an estimate of that answer's correctness. Experiments in the handwritten character recognition domain show that such an evolutionary process, combined with an effective technique for speciation, can create a population of networks that collectively performs better than any individual network.