Re: Choosing neurons to express in HyperNEAT
- Hi Oliver, I have not heard of anyone doing that, so I believe it's an original idea. It is indeed similar to the LEO, as you point out, but perhaps a more blunt instrument. As a practical tool, it may indeed work well, so it could be worth some experimentation, but of course unlike ES-HyperNEAT it still leaves some questions about how to lay out the initial substrate neurons (from which the NEO might prune).
One aspect of ES-HyperNEAT that I think is particularly interesting theoretically is that it has in principle no maximum density of neurons it can encode. As long as the pattern encoded by the CPPN has information at a higher resolution than checked so far by the quadtree information extraction algorithm, the algorithm can probe deeper at finer and finer resolutions. That way, ES-HyperNEAT really can (at least theoretically) discover brains with billions of neurons or more. The NEO would not have this interesting theoretical property because it would be picking from a pre-decided density. Nevertheless it could still be practically useful.
--- In email@example.com, Oliver Coleman <oliver.coleman@...> wrote:
> Has anybody tried using a Neuron Expression Output (akin to link expression
> output) with HyperNEAT to choose which neurons to express?
> Of course there is ES-HyperNEAT, but NEO would have the advantage of being
> simpler (KISS) and faster to run (a relatively large amount of time can be
> spent decoding in ES-HyperNEAT if the evaluation function is relatively