5 ms·
This is brilliantly cool, and utterly hypnotic. I still always wonder about implementations of genetic algorithms, though. In real life, from which inspiratio
by ColinWright 11y ago
This is brilliantly cool, and utterly hypnotic. I still always wonder about implementations of genetic algorithms, though. In real life, from which inspiration has been taken, the phenotype is often radically affected by small changes in the genotype. In computer implementation of the concept this is rarely the case. The computer versions are small, clean, spare, and nowhere near as messy as the real life version.
I wonder if the messiness actually matters. I wonder if "optimization by genetic algorithm" would be significantly improved by using a more faithful algorithm.
========
Anyway, here's a previous discussion of this particular toy:
https://news.ycombinator.com/item?id=5942757 https://news.ycombinator.com/item?id=5942757 (169 comments)
Other submissions:
https://news.ycombinator.com/item?id=10596079 https://news.ycombinator.com/item?id=10596079
https://news.ycombinator.com/item?id=10134390 https://news.ycombinator.com/item?id=10134390
https://news.ycombinator.com/item?id=5952145 https://news.ycombinator.com/item?id=5952145
https://news.ycombinator.com/item?id=2226137 https://news.ycombinator.com/item?id=2226137
Also, genetic development of walkers:
https://news.ycombinator.com/item?id=8911719 https://news.ycombinator.com/item?id=8911719 (74 comments)
- spooningtamarin 11y agoGenetic algorithms aren't really that brilliant. They are simple, and that's what makes them popular among researchers. There's no formal science behind them. I bet that thinking about the problem more would allow you to find an algorithm that works better than genetic algorithms all the time. Just look at travelling salesman problem, or vehicle routing problem. Smart heuristics beat these hybrid-genetic-memetic-algorithms all the time.
- adrianN 11y agoIt's wrong to say that there is no formal science behind them. It's true that practice is ahead of theory in the field of randomized search heuristics like evolutionary and genetic algorithms, but there are researchers working on it. For example it has been proven that crossover can be beneficial http://people.mpi-inf.mpg.de/~doerr/papers/gecco2008_1.pdf http://people.mpi-inf.mpg.de/~doerr/papers/gecco2008_1.pdf There is also a book on the topic that summarizes some of the current research http://www.worldscientific.com/worldscibooks/10.1142/7438 http://www.worldscientific.com/worldscibooks/10.1142/7438
- sevensor 11y ago> I bet that thinking about the problem more would allow you to find an algorithm that works better than genetic algorithms all the time. This is true. It's been formalized as the No Free Lunch theorem by Wolpert and Macready. > Genetic algorithms aren't really that brilliant. They are simple, and that's what makes them popular among researchers. There's no formal science behind them. There is a formal science behind them -- conferences, journals, the whole nine yards. The problem is, there's a dominant notion of genetic algorithms that's 25 years out of date, and nobody's been able (or tried too hard) to break through the noise. (Disclaimer: I work in the field.)
- TelmoMenezes 11y ago> There's no formal science behind them. I'm not sure what that means. Science is empiricism, and genetic algorithms are a wide field of research with a lot of experimentation behind them. There are many hypothesis about how they work, why they work and what their limitations are. These hypothesis are tested through experimentation, just like in other scientific fields. > Smart heuristics beat these hybrid-genetic-memetic-algorithms all the time. Sure, this is almost guaranteed by the "no free-lunch theorem": https://en.wikipedia.org/wiki/No_free_lunch_theorem https://en.wikipedia.org/wiki/No_free_lunch_theorem There might be some exotic cases where genetic algorithms are actually the best solution, but nobody knows that. But that's not the point of genetic algorithms. The point is having a generic solution that can be applied with some success when you don't have the time or the resources to find those "smart heuristics". And there might be cases where we are just not smart enough to find them. Some of these cases might be related to the effort towards developing generic AI. You can also claim that it is always possible to write a more efficient computer program directly in machine code, if you ignore time and resources. Another interesting property of evolutionary computation is enabling some form of artificial creativity. Check out this antenna, created by NASA using evolutionary computation: https://en.wikipedia.org/wiki/Evolved_antenna https://en.wikipedia.org/wiki/Evolved_antenna Also, if you will forgive me a bit of self promotion, check out my own work on the discovery of complex network generators: http://www.nature.com/articles/srep06284 http://www.nature.com/articles/srep06284
- chriswarbo 11y ago>The point is having a generic solution that can be applied with some success when you don't have the time or the resources to find those "smart heuristics". And there might be cases where we are just not smart enough to find them. Indeed, the same could be said of any "smart heuristics" too: > Exact solutions beat these smart heuristics all the time. ;)
- spooningtamarin 11y agoMost of the time, heuristics are faster than operators in genetic algorithms. Same goes for heuristics being faster than exact solutions for harder problems.
- loopbit 11y agoGenetic algorithms are not brilliant, but their results sometimes can be. I remember reading and article[0] about a researcher using genetic algorithms on FPGAs. One paragraph stuck with me through all these years: "Dr. Thompson peered inside his perfect offspring to gain insight into its methods, but what he found inside was baffling. The plucky chip was utilizing only thirty-seven of its one hundred logic gates, and most of them were arranged in a curious collection of feedback loops. Five individual logic cells were functionally disconnected from the rest— with no pathways that would allow them to influence the output— yet when the researcher disabled any one of them the chip lost its ability to discriminate the tones. Furthermore, the final program did not work reliably when it was loaded onto other FPGAs of the same type." Now, I agree with you that, even for a problem as simple as that one (or specially because it's a simple problem, maybe) a "normal" algorithm would perform better. But the fact that the genetic algorithm was able to find the effects of magnetic flux (those five disconnected logic cells) in that specific FPGA is nothing short of brilliant. The problem I've always seen (not an expert, by any means, but I've played around with them a bit) with genetic algorithms is that they require a weird balance. For simple problems, a normal algorithm will probably work better. As the complexity of the problem increases, the scoring and evaluating rules' complexity increases as well (I've got no hard data, of course, but my feeling is that the complexity of the scoring process increases much faster than the complexity of the problem), making a well thought algorithm a better/easier solution. This, and that any mistake in the scoring process might throw away thousands and thousands of generations. [0]http://www.damninteresting.com/on-the-origin-of-circuits/ http://www.damninteresting.com/on-the-origin-of-circuits/
- ionforce 11y agoThis story is like the poster child of success with genetic algorithms.
- segmondy 11y agoWhy would you use a genetic algorithm to solve a traveling salesman problem or vehicle routing problem? Get real. That's a very bad mapping. Genetic algorithm thrives when you have a lot of variables that can be optimized to gain a favorable behavior and those optimizations are unknown. Saying it's not brilliant because it's simple is like saying that binary search or merge sort is not brilliant because they are simple.
- spooningtamarin 11y agoCheckout latest research in Vehicle Routing Problem, its filled with GA/Ants and similar nonsense. On the other hand you have TSP which murdered all those GA algorithms with a proper Lin-Kernighan heuristic implementation. I'm not sure if we are talking about same problems. > Saying it's not brilliant I said they were simple and that's why people like them.
- vidarh 11y agoComputational resources are steadily getting cheaper. Peoples time is not. Which is why GA/GP matters: It's part of a growing toolbox to allow us to find interesting candidate algorithms for certain classes of problems without spending as much human time to figure out cleaner solutions. If you wish you can then have humans figure out why the solution works and use that knowledge to come up with something better.
- verdverm 11y agoI agree! I spent a few years with Genetic Programming for Symbolic Regression. After understanding the nature of the problem, I came up with a dynamic programming based heuristic that beats the pants off of GP. I called it Prioritized Grammar Enumeration. http://github.com/verdverm/pypge http://github.com/verdverm/pypge http://www.math.binghamton.edu/dept/ComboSem/worm-chiu.pge_gecco2013.pdf http://www.math.binghamton.edu/dept/ComboSem/worm-chiu.pge_g...
- sconxu 11y agoDidn't know about the walkers version. Thanks!
- randomsearch 11y ago> the phenotype is often radically affected by small changes in the genotype This is not true. If it were true, then the small mutations caused during reproduction of humans (for example) would result in huge changes to the phenotype. Typically, the genotype is fairly robust, for obvious reasons. OTOH, in genetic algorithms there is usually much less redundancy and neutrality in their genomes, because they are designed to solve problems, rather than simply survive.
- ColinWright 11y agoDo you speak from direct knowledge, or is this a deduction you're making? I can't find any clues in your profile or other contributions, so I don't know whether this is your area of expertise - in which case I would bow to it - or if it's something you've deduced from other reading. The reason I ask is because your assertion is at odds with my understanding. My understanding is that most small mutations lead to no apparent change at all, hence the robustness. Some small mutations lead to catastrophic changes, as do almost all large mutations. But some small mutations lead to noticeable changes in the phenotype, and those are the ones that then get selected against. Some of them are beneficial in the existing circumstances, some are not. But the point is that it's small mutations that lead to a comparatively large change in the phenotype. And that's my point, which you echo: > ... in genetic algorithms there is usually > much less redundancy and neutrality in their > genomes, because they are designed to solve > problems, rather than simply survive. Yes. Current "genetic algorithms" don't actually do anything like the systems they have been inspired by. I wonder if real progress would be better if the implementations were more faithful.
- raverbashing 11y agoHere's the difference between GA and GP In this case it's GA, so your genome directly maps to characteristics. So (let's say) chassis density is N bits. Mutate the MSB of that and the change will be big GP on the other hand would "build a car". And what usually happens during evolution is that some aspects gain resistance to mutation through some form of redundancy. So instead of Chassis density being a number it might be a crazy math expression that is not affected by small changes. Also a lot of "genes" have no effect on the final product. But having a bigger genome (in this way) effectively lowers mutation rates. It is "junk DNA" but actually has a purpose
- sklogic 11y agoI played a bit with the "more faithful" genetic algorithms, got some promising results (getting out of the local minima better). The key is to use hierarchical encoding in your chromosome: instead of directly encoding traits/functions/properties/whatever, encode a program (maybe even in a Turing-complete representation) which in turn generates those traits/functions/properties/whatever, or even generates a next level of code. This is closer to how the biological systems work, and allows for much higher sensitivity to small changes.
- ionforce 11y agoDo you have more information about this approach? I've never heard about it until now.
- sklogic 11y agoI tried to dig out the references from my mess of bibliography (I worked on this topic long before Mendeley and alike appeared), unsuccessfully. I remember that I ran into this technique independently first (in a context of encoding expression trees) and then found a number of other examples in the literature (e.g., encoding the neural network topology, etc.).