MM-NEAT versus EShyperNEAT

[evolvingfridge]

Is there reason why no one compared performance of MM-NEAT versus EShyperNEAT or other variations (compared similarly to as are classification algorithms in UCR Archive) ?

[drallensmith]

Huh. This question inspired me to do a search - thanks! - and to find material on work by Risi and others to do a sort of MM-HyperNEAT - see here.

[evolvingfridge]

@drallensmith, MB-HyperNeat is very interesting, but why use MB-HyperNeat and not multiple MM-NEAT or RBF-NEAT combining multiple NEAT can be achieved in many ways not only by using HyperNeat. What is primary issue with just simply training multiple smaller networks versus using MB-HyperNeat. I personally feel like most publications about neuroevolution struggle a bit to compare against other NEAT variations, as example UCR Archive is great for classification algorithms since you can see for what problem domain what algorithm is most likely will be best for a job.

P.S. Thanks for suggesting opening new issue for this question.

[drallensmith]

I think part of the problem with comparisons is that, with large alterations, it's hard to do an ablation study in which some portions of the changes are disabled to see what effect it has; the changes can be rather interdependent. (This is for comparisons against a more classical NEAT, while having the inputs/outputs/etc as comparable as possible.)