Using highly variable data for training neural networks (NN) model may induce "overrating". So for practical reasons normalizing (or standardizing) the inputs in NNs can make training faster and reduce the chances of getting stuck in local optima. If you use Matlab you can use either mapminmax or mapstd as follows: 

[yn, ps] = mapminmax(ymin, ymax)

[yn, ps] = mapstd(ymean, ystd)

where yn is normalized data.

1- MinMax: mapminmax processes matrices by normalizing the minimum and maximum values of each row to [xmin, xmax].  So it process data by mapping their values to minimum and maximum values to [-1 1]. Algorithm: y = (ymax-ymin)*(x-xmin)/(xmax-xmin) + ymin;

2- Z-Score: mapstd processes matrices by transforming the mean and standard deviation of each row to ymean and ystd. So It process data by mapping them having means to 0 and deviations to 1. Algorithm: y = (x-xmean)*(ystd/xstd) + ymean;

Yes, that function can help you to normalize your data into the range you desire.

Regards

Hello My Freinds...

Thank you so much for your answers. Clearly understand.

Best Regards.

Large variability in input data needs to be normalized with respect to (min, max) values and/or with (mean, stddev). This will help in effective training as well as avoiding training to get trapped in local minima. Similarly, the expected output data should be within normalized values, which _may be mapped to actual values.

Dear Soheila Ghambari,

This is an interesting question. In most of supervised neural networks, the normalization is required. This is because in these NNs a transfer function is used (in the forward calculation) which gives outputs in a specific range. And also, the starting point of the optimization process of parameters starts with the calculation of the error as a difference between real and estimated values. And these two sets of values have to have the same range.

I hope this will help you.

Cheers,

José.

Hello Mr.Torrecilla 

Thanks for your answer.Your saying is true.Normalization is required  and I observed it's effect  to reduce the errors .

Best Regards...

its better you take maximum of each individual feature from your training set and divide that feature by max.

Hi! If you are willing to use NN, i sujest first to remove correaltions between the inputs and made them statistically independent. You can use logarithmic returns. After that, you can use logistic function to normalize your data.

A standard approach is to scale the inputs to have mean 0 and a variance of 1. Also linear decorrelation/whitening/pca helps a lot.

If you are interested in the tricks of the trade, I can recommend LeCun's efficient backprop paper.

As you can see from the link, there is no need to normalize data if you standart programs such as FITNET, PATTERNNET, TIMEDELAYNET, NARNET & NARXNET.

All of the normalization and de-normalization is done automatically

http://www.mathworks.com/matlabcentral/answers/25759-normalizing-data-for-neural-networks

Hi everyone

I know that MATLAB NN toolbox has mapminmax function. Do you know how to change min and max values to [0 , 1] instead of defualt value [-1 1] in NN toolbox?