Here’s something fun to leave on your screen for a while. It’s an evolutionary car.
Ouches in 3..2..
The problem here is: what’s the the best wheel (and load) size and position to get a little car across a rugged landscape? Genetic algorithms are good at these kinds of problems. Just release a population of slightly differing individuals into the wild, and let the best performers produce offspring that are sort of like they are. Nifty.
12 December 2008 at 12:19 pm
(obligatory comment re your latest post…)
So, you could wait several hours waiting for random mutations to produce a car that makes it 10 seconds into the course, or you could get an engineer or some other intelligent designer (even a 5-year-old made in Someone’s image) to create a car that works in a minute or so.
This is actually a pretty bad use of genetic algorithms. Good cars with the degrees of freedom and physics in the linked demo can be generated with a pretty short grammar (maybe even a quasi-convex set depending on how you parameterise it); and the rules are intuitive.
GA’s have been over-hyped because non-CS people go “ooo… so like, DNA and Darwinism can be used in computers, so they’re kinda like cyborgs amirite?”, when in a lot of cases local search is just as effective. Trying to find a breeding function that retains the mean objective of two feasible solutions is non-trivial for non-trivial problems.
12 December 2008 at 1:24 pm
Well, you’re right, but I have no problem with taking several hours to find a semi-optimal solution. Gotta sleep sometime.
12 December 2008 at 7:44 pm
It’s a lot faster than real genetic evolution, and this was done WITHOUT an ID and that is the point, amirite?
13 December 2008 at 12:31 am
As the time you can wait -> infinity, then an exhaustive search works too. As time -> 0 the likelihood you have a good solution from the GA -> 0. Design is easy and practical here.
On a blog that regularly draws attention to the antics of creationists, I think the point involves an implication that GAs are more than tangentially related to biological evolution. It's poor evidence.
As a side note, one of the corollaries of the simulation argument is that biology was seeded (think Douglas Adams), but I doubt you'll get the orthodox religions using that to support ID.
13 December 2008 at 5:25 am
Why is it bad evidence? It’s incomplete, of course, much like Dawkin’s WEASEL program, but it does show that it’s possible for chance + fitness function to make for better and better organisms. All that is needed is time.
13 December 2008 at 7:23 am
By the same logic, intelligent designer + objective function makes for great ‘organisms’.
You’re extrapolating from 8 degrees of freedom, a strong synthetic selection pressure and forced mutation, to “organisms”. C’mon Daniel, why sink to such a contrived example (other than the hey-kids-science-is-fun factor) when there is better evidence. This is where pseudo-science comes from.
13 December 2008 at 11:30 pm
I find your objections a bit bizarre. Of course it’s a simplification. It’s a toy example. But if you’re going to require complexity on the order of evolving beaks in the Galapagos, you’ll never use any GA’s at all.
Would you have been happier if the little cars could hump each other? and get eaten by predators?