How can NEAT “start minimally” in a high-input/high-output domain?
Recall that “starting minimally” means beginning evolution with a population of networks with minimal structure. In the experiments in our papers so far, “minimal structure” means no hidden nodes. However, in domains with a very large number of inputs and/or outputs, things get weird. Let us say you have 50 inputs and 50 outputs, which might be reasonable for a board-game type domain. Then if you start with networks with only direct connections, you would have 2,500 connections. In contrast, a network with 5 hidden nodes would have only 50*5+5*50=500 connections. So at a certain point, when the input/output space becomes quite large, the meaning of a minimal network changes. In general this happens when a reasonable number of hidden nodes for solving the task is significantly lower than the number of inputs and/or outputs. The conclusion is that you need to be creative about the topologies of your initial population in such high-dimensional domains. One option, which has not been explo
