Sep 5, 2018 · 1 min read
This is a very well-written article; however, as you pointed out in your description of NEAT, its key discriminating capability is in its ability to principally “complexify” the network topology. The example is only evolving the network weights, which makes sense as this makes the implementation is far more straightforward. I’m curious, however, if you have any thoughts on how topological evolution would be implemented in a layer-oriented framework like Keras.