Computer Program Deduces Laws of Physics

Terry Olaes

I Used to be the [H] News Guy
2FA
Joined
Nov 27, 2006
Messages
4,646
A computer program developed by researchers at Cornell accomplished in a little over a day what it took physicists centuries to complete: extrapolating the laws of motion from a swinging pendulum. I’m trying to resist the obvious tie-ins to Skynet and HAL 9000 regarding artificial intelligence and eventual sentience. ;)

Lipson and Schmidt designed their program to identify linked factors within a dataset fed to the program, then generate equations to describe their relationship. The dataset described the movements of simple mechanical systems like spring-loaded oscillators, single pendulums and double pendulums — mechanisms used by professors to illustrate physical laws.
 
Someone needs to tell these guys that MS Excel can already fit an equation to given data :D

Ok funny stuff aside, this is really nothing surprising, you go a brute force attack at a set of data, have it look at the "most promising" results, then re-attack that, and repeat until you find a solution. It's called evolutionary computations, been around for a while, and yeah it does show some promise... but like I said it's a brute-force type attack.

I am curious exactly what was done, I mean did they need to program gravity into the computer? First guys who came up with laws of physics didn't have a million points of data from a single experiment and then extrapolate to that. Also curious if they stopped it when they saw some common laws of physics pop up, or if they pushed it in that direction.

It also mentions some other laws that were unknown that popped up, I'm curious to see what they are.
 
I wonder when computers will be able to find out what we don't know instead of just confirming what we already now...
 
that's stupid. "AI" will always evolve around what we give it or what we know. no mere program will go beyond that unless u introduce biological elements in the equation whether animal or human brains. then it'll become sentient. otherwise, no sentient program AI will be truly 'intelligent' without the artificial side.

otherwise, NASA would've been installing AI programs to satellites traveling to outer space because that'll be useful to extrapolate NEW theories no humans have thought of.

not going to happen.
 
They got it all wrong, we don't need computers to extrapolate the laws of physics for us, we got geeks and nerds to do that already. What we really need is a computer that can extrapolate the laws which govern the behavior of women, something which is currently beyond the capability of every male on the planet.
 
Maybe you're missing the point.

I can't read the article from work, however the code to do such a thing is simple.

Feed in 5 starting forms to a least squares

e
log
x
x^2
x^1/2

drop the form with the highest residual, reinput the remaining and repeat. You can solve 90% of the problems in the world doing that.
 
I can't read the article from work, however the code to do such a thing is simple.

Feed in 5 starting forms to a least squares

e
log
x
x^2
x^1/2

drop the form with the highest residual, reinput the remaining and repeat. You can solve 90% of the problems in the world doing that.

I rest my case. And you don't need to read the article to understand the point.
 
I rest my case. And you don't need to read the article to understand the point.

Ok, I just broke out my iphone and read the article. It's nothing but exactly what I described. Least squares determining linear systems. They aren't describing complex nonlinear (continious or discontinious) systems.
 
Ok, I just broke out my iphone and read the article. It's nothing but exactly what I described. Least squares determining linear systems. They aren't describing complex nonlinear (continious or discontinious) systems.

Jesus christ, man. The point is not to make robots solve things for us.
 
I mean, the point is more about developing AI and filling in what we might be missing. But purely synthetic robots will always be limited by our ability to program them.
 
I mean, the point is more about developing AI and filling in what we might be missing. But purely synthetic robots will always be limited by our ability to program them.

My entire point is it is nothing special. This is still limited by the basic starting equations they are feeding into thier least squares. They are just feeding a lot of equations into a least squares and filtering out the ones that end up with crazy residuals. While normally you would look at the graph for your initial guess at a starting equation. This is brute force at it's worst.
 
Back
Top