It's not fancy hmtml, but it's the best genetic simulation I've seen. The "swimmers" exist in 2d space, seek out food and mate. Only the best will reach the food fast enough, and there are colors which a given swimmer will prefer to mate with, and space separating different populations. It's very possible to get several distinct populations which behave in entirely different ways.
Successful parameters could also become more immune to mutation with time, a nice improvement along with "speciation" you cited.
Haha, I've had this same thought before. I wrote a genetic algorithm to work on the sphere packing problem (in a periodic box) and I kept thinking about how I could mutate the settings within the program and optimize them, and then mutate the rate at which those settings themselves changed, and so on. I wonder if there's an optimal point to stop being "meta"?
I say this because this reminded me of a wonderful (and acessible) discussion I once glanced over on the following book: http://www.cs.toronto.edu/~mackay/itprnn/book.pdf (Page 269, pdf 281) -- it offers answers to this sort of question :)
So called, because unfortunately GAs etc. had already taken the name "meta-heuristics".
Perhaps this will help your thinking in the right direction: http://en.wikipedia.org/wiki/No_free_lunch_in_search_and_opt...
I believe this happens in nature. DNA has lots of error correcting stuff to reduce errors.
It's a trade-off between exploration and exploitation. If you speciate then you have little sub-populations exploiting the local maxima, so they might do better. However, now you're not searching such a wide area so you're more likely to spend your time struggling to the top of the molehill that's next to the mountain.
In nature this also happens. Not having crossovers between ludicrously different things is generally a good idea because the landscape is full of tiny maxima. If you change a load of base pairs in your DNA you're less likely to be able to do things like form cells or process glucose than you are to add some awesome new power (unless we're in comic books).
One of the major differences for nature is that it can run all this jazz in parallel. If there are resources, it can happily have 10 trillion bacteria rather than 5 without any cost. When we're simulating, every fitness measurement on a piss-poor crossover is a waste that we should have spent on a more likely solution.
It all depends on the fitness landscape, but both approaches can be better on different real world problems.
But my current instace of the web page is at a point where it's wasting a lot of time on unfit 'mutts' and it seems like I could get to a "best rhinoboat" and "best assdragger" solution faster if it would stop wasting time exploring the mediocre blend.
I didn't dig into this implementation, but it seems like it doesn't carry many (more than one?) individuals between cycles.
A typical work-around for scenarios like this would be to have the various parameters change as the simulation carries on; e.g. later on it might make more sense to increase the number of elites as a way of limiting crossover.
I suspect - would be interested in anyone who's worked it out - that each individual is mutated between each generation.
I tried 10 Elites with high rate of mutation. I'd have expected this to have more "breeding" from the Elites, but it doesn't. It tends to plateau. Occasionally an outlier jumps into the "10", but it's pretty flat.
So.. Is there any concept of "fitness"? If so, it ends up being pretty random, since there is no penalty for a less adaptive mutation.
There is really no advantage in speciation if you only want the single best solution, but it does make sense to try different evolutionary paths as they can end up in totally different places.
If speciation is a better heuristic for finding a better solution faster than the algorithm used here then its advantage is real and useful in any case.
If you divide the population in two, at least half of your resources are wasted on doing tests and evolving traits which will just be discarded. Resources which could have been used to speed up the evolution of the other population. Even if the evolutionary path you choose isn't the optimal one, by putting twice as much resources too it, it still might end up further. And if you divide it into more than 2 species than it's just getting insane.
That's not to say that speciation isn't a useful heuristic, just that it's a costly one. However there are a lot of cases where just throwing more resources at the problem doesn't help, because it's stuck in a local maxima. In that case, going back to an earlier point in the simulation and running it to see if it ends up somewhere else isn't a bad idea, and that's exactly the same as what speciation is (except that you are doing it after the fact, rather than testing both evolutionary paths simultaneously.)
When the hybridization is managed by humans though, it can be useful (e.g.: citrus, mules).
In my instance, there are two predominant "types" of cars that are doing approximately equally well--the "rhinoboat" and the "assdragger". But when these two solutions are combined in the crossover, I think they make an offspring that is terrible.
If the algorithm would mate only the rhinoboats with rhinoboats and assdraggers with assdraggers, I'm convinced we'd end up with some superrhinoboats and superassdraggers, but instead I just get an unhappy compromise.
Maybe make "similarity to self" one of the genotype parameters. A high rating here would mean the phenotype would have a bias against being combined with dissimilar creatures. This allows the propensity towards speciation to develop evolutionarily, too.