Comments on: Strandbeesten http://kevin.saff.net/2007/06/12/strandbeesten/ some adventure Wed, 03 Dec 2008 23:32:58 +0000 http://wordpress.org/?v=2.3.1 By: John the Statistician http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1195 John the Statistician Thu, 30 Nov 2000 00:00:00 +0000 http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1195 Yeah, I could see that would be a nice bit of research: can you learn domain specific transformation rules for an underconstrained genetic programming problem? Of course, right now the approach is to start with a limited operator set, which is generally much easier. Yeah, I could see that would be a nice bit of research: can you learn domain specific transformation rules for an underconstrained genetic programming problem? Of course, right now the approach is to start with a limited operator set, which is generally much easier.

]]> By: Kevin http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1194 Kevin Mon, 30 Nov 2009 00:00:00 +0000 http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1194 I don't know anything about genetic programming, but I would like to see some models based on the genetic switches that DNA actually uses for regulation. This can be thought of as a kind of programming language - I think some folks at MIT were using this to program bacteria to flash in interesting patterns and such, but I can't find the link. It's pretty clear that genetic switches are well-behaved under gene duplication events, so it would be interesting to see what you could do with that in a computer model. I don't know anything about genetic programming, but I would like to see some models based on the genetic switches that DNA actually uses for regulation. This can be thought of as a kind of programming language - I think some folks at MIT were using this to program bacteria to flash in interesting patterns and such, but I can't find the link. It's pretty clear that genetic switches are well-behaved under gene duplication events, so it would be interesting to see what you could do with that in a computer model.

]]> By: John the Statistician http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1193 John the Statistician Sun, 30 Nov 2008 00:00:00 +0000 http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1193 Ah, I see what you mean. You do see more complex operators of the kind you are talking about in genetic programming than in vanilla genentic algorithms. It is definitely true that genetic programming with a large operator set is so unconstrained as to be poor at search for many problems. Ah, I see what you mean. You do see more complex operators of the kind you are talking about in genetic programming than in vanilla genentic algorithms. It is definitely true that genetic programming with a large operator set is so unconstrained as to be poor at search for many problems.

]]> By: Kevin http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1192 Kevin Sun, 30 Nov 2008 00:00:00 +0000 http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1192 Hmm, I don't think I got across what I was trying to. I do think genetic algorithms are a clever way of searching a parameter space, but I would never call them evolution. All genetic algorithms I have seen involve constant genome sizes and only point mutations, which is not enough to generate the complexity we see in biology. The evolution of complex biological characters really requires <a href="http://en.wikipedia.org/wiki/Gene_duplication" rel="nofollow">gene duplication</a> events, including for example <a href="http://en.wikipedia.org/wiki/Transposon" rel="nofollow">transposable elements</a> which are known to be a driving force behind bacterial resistance to antibiotics. You don't get any of that stuff if your model uses a fixed genome size and only point mutations, which is why GA's are restricted to search problems. Hmm, I don't think I got across what I was trying to. I do think genetic algorithms are a clever way of searching a parameter space, but I would never call them evolution. All genetic algorithms I have seen involve constant genome sizes and only point mutations, which is not enough to generate the complexity we see in biology. The evolution of complex biological characters really requires gene duplication events, including for example transposable elements which are known to be a driving force behind bacterial resistance to antibiotics. You don't get any of that stuff if your model uses a fixed genome size and only point mutations, which is why GA's are restricted to search problems.

]]> By: John the Statistician http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1191 John the Statistician Fri, 30 Nov 2007 00:00:00 +0000 http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1191 I do agree there is a gap between the different ways he wants to use evolutionary approaches, but I think you can't blame him for having aspirations. It's very clearly driving some singular work. I'm willing to grant him some nonsense given the willingness he's demonstrated to take on this kind of project. I agree that evolution is the wrong search technique to use, BUT ONLY if you can make very stringent assumptions about the lack of noise. As for the efficiency of evolution, sexual reproduction can support a rate of improvement O(1/sqrt(G)) without losing good traits, where G is the number of bits in the genome, which is very impressive for a local search algorithm (see Information Theory, Inference, and Learning Algorithms, chapter 19, http://www.inference.phy.cam.ac.uk/mackay/itprnn/ps/265.280.pdf). So, altogether, if you're doing a parameter search, the general purpose processor you are using is probably wasting a lot of energy on error correcting memory. I do agree there is a gap between the different ways he wants to use evolutionary approaches, but I think you can't blame him for having aspirations. It's very clearly driving some singular work. I'm willing to grant him some nonsense given the willingness he's demonstrated to take on this kind of project.

I agree that evolution is the wrong search technique to use, BUT ONLY if you can make very stringent assumptions about the lack of noise. As for the efficiency of evolution, sexual reproduction can support a rate of improvement O(1/sqrt(G)) without losing good traits, where G is the number of bits in the genome, which is very impressive for a local search algorithm (see Information Theory, Inference, and Learning Algorithms, chapter 19, http://www.inference.phy.cam.ac.uk/mackay/itprnn/ps/265.280.pdf). So, altogether, if you're doing a parameter search, the general purpose processor you are using is probably wasting a lot of energy on error correcting memory.

]]> By: John the Statistician http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1189 John the Statistician Fri, 30 Nov 2007 00:00:00 +0000 http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1189 Did I forget to tell you about him? http://www.itconversations.com/shows/detail762.html He actually does genetic algorithms to design these creatures. His ultimate hope is to be able to build his own reproducing mechanical species. Did I forget to tell you about him? http://www.itconversations.com/shows/detail762.html

He actually does genetic algorithms to design these creatures. His ultimate hope is to be able to build his own reproducing mechanical species.

]]> By: Becky http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1188 Becky Wed, 30 Nov 2005 00:00:00 +0000 http://kevin.saff.net/2007/06/12/strandbeesten/#comment-1188 Wow! That's really neat! That has to take a long time to figure out how to get the motion to work as you envision it. Wow! That's really neat! That has to take a long time to figure out how to get the motion to work as you envision it.

]]>