AForge.NET

  :: AForge.NET Framework :: Articles :: Forums ::

Input scaling in the TimeSeries sample

The forum is to discuss topics from different artificial intelligence areas, like neural networks, genetic algorithms, machine learning, etc.

Input scaling in the TimeSeries sample

Postby Motig » Sat Dec 01, 2012 9:30 am

I have a question about input scaling.

In the TimeSeries sample of this library, you multiply the input with a "data transformation factor" which is defined as "1.7 / chart height" and then subtract "0.85" from it.

Code: Select all
double factor = 1.7 / chart.RangeY.Length;


Code: Select all
input[i][j] = ( data[i + j] - yMin ) * factor - 0.85;


What are the numbers 1.7 and 0.85 ? I guess there's a just a mathematical concept I'm missing, but I've been unable to comprehend it.
Motig
 
Posts: 3
Joined: Sat Dec 01, 2012 8:13 am

Re: Input scaling in the TimeSeries sample

Postby Motig » Sat Dec 01, 2012 9:42 am

I just realized 0.85 = 1.7 / 2

That led me to think what is the meaning of the number 1.7 .... and I came to a conclusion that it is somehow related to the fact that the bipolar sigmoid range's length is 2 (from -1 to +1) ...

And the reason of subtracting 0.3 from 2 has something to do with scaling the input not to the exact range of -1 to 1 but to -0.7 to +0.7 and I know I have already read why is it good to do so but I don't remember...but I'm going to experiment and see what's going to happen.

Well, at least, writting it down has made me answer it myself, sorry for spamming your forums :D
Motig
 
Posts: 3
Joined: Sat Dec 01, 2012 8:13 am

Re: Input scaling in the TimeSeries sample

Postby andrew.kirillov » Sat Dec 01, 2012 2:03 pm

The [-0.85, 0.85] range was used in that sample for mapping training data. The idea was (not sure if that proved to work well) to give network a chance to predict beyond that range. If you would map all training data to [-1, 1], then the network would never predict anything which is smaller/greater then the smallest/greatest number in the training set.
With best regards,
Andrew


Interested in supporting AForge.NET Framework?
User avatar
andrew.kirillov
Site Admin, AForge.NET Developer
 
Posts: 3453
Joined: Fri Jan 23, 2009 9:12 am
Location: UK

Re: Input scaling in the TimeSeries sample

Postby ganjawar » Thu Nov 13, 2014 8:14 pm

for compute y after learn neural net using yfactor, but how compute y if i using train net and i have only one x.then y factor i dont know, because i have only x, and xfactor = 2.0 / 1; and xmin = null ,because i input only one number. finally, how use net after learning, if i have only one input and i want comput one output?
ganjawar
 
Posts: 6
Joined: Tue Nov 11, 2014 1:48 pm




Return to Artificial Intelligence