Hands-On Neuroevolution with Python

Information patterns in the hypercube

Why should such limitations on the location of nodes be imposed in the first place? Wouldn't it be nice if the implicit clues that had been drawn from the connectivity patterns became the guidelines of where to place the next node to represent the natural regularities of the physical world better?

The areas with uniform connection weights encode a small amount of information and hence have little functional value. At the same time, the areas with vast gradients of weight values are extremely information-intensive. Such areas can benefit from additional network nodes being placed to represent a much finer encoding of the natural process. As you may recall from our discussion of the HyperNEAT algorithm, ...