Automatic pattern recognition and data modeling

[moonman239]

I can't help but think that some day, someone will figure out a way to get a computer to recognize patterns in a given set of data, and fit an equation, if there is any, to that set.

Such a system could be used in areas like forecasting stocks, weather, and sales, or even in air traffic control, which would be by far the coolest use of such a system.

 

[rcgldr]

youtube can already "recognize" some video or audio patterns to search for "matched 3rd party content" in submitted videos, mostly used to include ad's on videos where some of the profit for the ad goes to the "3rd party content" copyright owner.

 

[moonman239]

youtube can already "recognize" some video or audio patterns to search for "matched 3rd party content" in submitted videos, mostly used to include ad's on videos where some of the profit for the ad goes to the "3rd party content" copyright owner.

I see. However, I'm talking about having a computer look at a data set such as:

2,4.2,6,8.5

and figure out that the data can be best modeled by a linear equation.

 

[rcgldr]

I see. However, I'm talking about having a computer look at a data set such as: 2,4.2,6,8.5 and figure out that the data can be best modeled by a linear equation.

Doesn't MATLAB and/or some other mathematical tools have a data fitting / smoothing model that at least suggests what type of equation would be best to do this?

 

[coalquay404]

I see. However, I'm talking about having a computer look at a data set such as:

2,4.2,6,8.5

and figure out that the data can be best modeled by a linear equation.

A computer can't do that for precisely the same reason that a human can't: it's impossible to unambiguously identify that as evidence of a linear relationship unless you supply more information to constrain the model selection. In fact, if you try to fit functions to that particular set of data you'll find that a cubic equation is a better fit than a linear equation.

As to your original question, lots has already been done: artificial neural nets, support vector machines, kernel methods more generally, and many others are all in widespread use in pattern recognition.

 

[gsal]

There is Lowess curve fitting that with cross-validation looks at data and does a curve fit by gluing together whatever fits the best range-wise, in other words, it does not use a single equation, but several for a best fit everywhere.