Evolutionary algorithms are perennially intriguing, but generally other approaches are favored. Odd, considering the incredible success of evolution with extremely limited trials from a computational perspective. I wonder why evolutionary computation does not live up to the expectation?
> I wonder why evolutionary computation does not live up to the expectation?
There's also a social factor here, I think. Neural networks were disregarded due to over-claims and also perhaps due to some bias for machine learning methods that relied on understandable statistical models. In the same way, evolutionary computing is often looked down upon by the rest of the AI community - it's a lot of hacking and hand waving and "try it and see". It doesn't help that there's been a lot of low-quality work in EC. But this bias has perhaps starved EC of talent somewhat; much of the progress in deep learning is probably a result of increased research activity and optimism, i.e. progress is self-reinforcing.
Our top supercomputers can process tens of petaflops. That seems to easily dwarf these biological time scales you mention. There appears to be some magic in evolution that surpasses mere trial limitations.
I'd like to double down and say the top supercomputers today couldn't, in reasonable time, simulate a single generation of the genetic variation in your gut flora alone.
Sure, but that means the magic does not lie in evolution, but in the extreme amounts of information in biological organisms. Evolution is supposed to be able to generate incredibly complex organisms from very simple beginnings, all within a number of trials that can be simulated on today's desktop. Insofar as we displace the problem from evolution to the organism or environment we are saying there is nothing special about evolution. In which case it begs the question what is so special about evolution in the first place?
There are teraflop processors for the desktop now. Our standard 60 GFlop processors will take less than a minute to do a trillion operations. I don't see the computational bottleneck you refer to.
At least from a viewpoint of number of trials, we can easily replicate the number of trials evolution has to work with using an everyday desktop. Why don't we see comparable evolution in the desktop? There is some sort of magic to evolution we have not yet fathomed.