Thread: Computer Self-Discovers Laws of Physics

Very interesting.

I don't see why this shouldn't be capable of some serious advances, possibly even leading to a unified theory. The human mind may simply turn out to be incapable of noticing the subtle and intricate links between the two major branches of physics, especially as the vast majority of physicists are only able to specialise in one field.

What a misleading article. Firstly, pendulum motion is easy to work out, lastly all the computer did was grunt work. Even worse other people had to do the initial set up and the only thing it did was run through some crappy algorithms.

Anyway, all the computer did was make up a equation from a pre set amount of data, even a idiot can do that if they had something like matlab or mathematica. Computers can't do maths or physics as they don't understand stuff they just crunch numbers.

Originally Posted by Xei

The human mind may simply turn out to be incapable of noticing the subtle and intricate links between the two major branches of physics, especially as the vast majority of physicists are only able to specialise in one field.

As pointed out in the comments about the article a computer told us something we already know by having humans feed in the initial data for them to get something they expected, which is equation of pendulum motion.

Thats why you have people called mathematical physicist. The problem with computers is that they can only prove problems that can be attacked by brute force like proof by cases but not when you have to use different branches of mathematics or even create a new axiom to solve the problem.

The last time I checked Physics and Maths is not like chess. Even then GMs can beat chess programs pretty easy.

At the moment I suppose.

Just wait for computational neuroheuristics.

Originally Posted by Xei

At the moment I suppose.

Just wait for computational neuroheuristics.

Fuzzy logic is the way forward not neuroscience.

Yeah... I'm sorta disappointed with the article too.... It's just data fitting combined with a fitness function...

This is something just about anyone with decent AI knowledge would be able to do (or at least, write a crappy, inefficient solution that sometimes gets stuck on sub-optimal solutions)

I disagree that fuzzy logic is the only way forward though. Neural nets have their place.

It's cool and very useful, but nothing really revolutionary.

I worked with an algorithm that does a similar thing- it comes up with a mathematical model for small genetic or metabolic networks based on measured activities/concetrations of genes/proteins/metabolites.

It doesn't do anything particularly intelligent. It's just a heuristic search through the model space. You have to define possible model structures (equation forms) yourself and the algorithm tries out different combinations of them while simultaneously estimating parameters.