RE: What are your favorite examples of evolution in progress?
November 16, 2017 at 6:37 am
(This post was last modified: November 16, 2017 at 7:10 am by I_am_not_mafia.)
My favourite example of evolution in progress are genetic algorithms. We basically make an abstraction of the evolutionary process and use it for practical purposes. Not only that, but you can reason about what it's doing to improve performance, but also come up with entirely new evolutionary algorithms which are not similar to natural evolution and they work as well.
The religionists think that intelligent design is more plausible but do not understand that it has no explanatory power. I have recently been racking my brain trying to figure out how you could actually build a program that could intelligently design something for you without resorting to loads of blind trials to see if something works in practice. If that was acceptable then we've been using genetic algorithms and searches for many decades now. The best I can imagine for an AI intelligent designer is some heuristic search of a network where the model represents some abstract rules that are very well defined and heavily constrained. So you could use it for Maths, maybe an electronics circuit board perhaps but not for anything that has to interact with a noisy real world.
And this is why I dismiss the arguments about a singularity happening in the near future where it is assumed that robots will start designing themselves. Yeah we know that it is possible to design things intelligently because we do it ourselves, but we're the only species that does this and we have absolutely no idea yet how to do this in AI. The argument for the singularity relies on some techniques that we cannot yet even start trying to solve. It's like asking James Clerk Maxwell to design a television before he has figured out electro-magnetic waves.
If the theory of evolution by natural selection was wrong then my evolutionary algorithms would be useless. I've been using them for twenty years to evolve AI that work in ways that I don't actually understand. It is very easy for me to evolve a neural network for example that is going to be far too complicated for me to figure out. It's actually the main reason that I stopped using neural networks. The last time I tried really understanding an evolved system and being confident that I was correct was back in 2005 and it took me three months. And that's not unusual. Yet if I were to create a program to intelligently design another system, it would have to completely understand every aspect of what it was designing at every stage. This would be a major limitation on what it would be able to achieve compared to using an evolutionary algorithm.
The religionists think that intelligent design is more plausible but do not understand that it has no explanatory power. I have recently been racking my brain trying to figure out how you could actually build a program that could intelligently design something for you without resorting to loads of blind trials to see if something works in practice. If that was acceptable then we've been using genetic algorithms and searches for many decades now. The best I can imagine for an AI intelligent designer is some heuristic search of a network where the model represents some abstract rules that are very well defined and heavily constrained. So you could use it for Maths, maybe an electronics circuit board perhaps but not for anything that has to interact with a noisy real world.
And this is why I dismiss the arguments about a singularity happening in the near future where it is assumed that robots will start designing themselves. Yeah we know that it is possible to design things intelligently because we do it ourselves, but we're the only species that does this and we have absolutely no idea yet how to do this in AI. The argument for the singularity relies on some techniques that we cannot yet even start trying to solve. It's like asking James Clerk Maxwell to design a television before he has figured out electro-magnetic waves.
If the theory of evolution by natural selection was wrong then my evolutionary algorithms would be useless. I've been using them for twenty years to evolve AI that work in ways that I don't actually understand. It is very easy for me to evolve a neural network for example that is going to be far too complicated for me to figure out. It's actually the main reason that I stopped using neural networks. The last time I tried really understanding an evolved system and being confident that I was correct was back in 2005 and it took me three months. And that's not unusual. Yet if I were to create a program to intelligently design another system, it would have to completely understand every aspect of what it was designing at every stage. This would be a major limitation on what it would be able to achieve compared to using an evolutionary algorithm.
