[davelnewton]

This brings back memories... One of the greatest heads-down hacking I did was back in the day, running some variant of Common Lisp, working myself through Koza's _Genetic Programming_ book. The designs arrived at were literally non-human. The story of the FPGA tone recognition circuit (spoiler: lot-specific design) was delightful.

[pmoriarty]

My favorite was using GP to evolve a soccer team that outcompeted the handcrafted competion and won the Robo Cup championship in 1997.[1]

I still wonder why GP never got to be as popular as NN's (aka the currently fashionable "deep learning").

Was it just that GP's didn't perform as well? I find that somewhat hard to believe, as Koza's books are chock full of impressive results, and there are hundreds more papers on them.

GP's also have the virtue of ultimately being analyzable and understandable, at least in some cases (I'm not confident enough to say in all cases). That is a feature that NN's seem to lack, and it's becoming a big problem for some critical systems where life, important decisions, and/or ethics are involved.

[1] - http://www.genetic-programming.com/hc/lukesoccer.html

[andretti1977]

GP/GAs were more popular than NNs before NNs started to use gpu a few years ago and had already some real world applications like antenna design or circuit solvers. That said GP (or GA) didn't gathered the same fame NNs have now since using gpu with NNs let NNs provide results far better than GA/GP using only cpu. If someome will have a smart idea on how to get the advantages of using the gpu with GA/GP then GA/GP will become seriously popular.

[pmoriarty]

I don't think GP's not being able to use GPU's is the issue, because GP's have run on GPU's for a long time:

http://gpgpgpu.com/

[antidesitter]

Uber and OpenAI have written on the subject of evolutionary algorithms in a contemporary context:

https://eng.uber.com/deep-neuroevolution/

https://blog.openai.com/evolution-strategies/