How to normalize new inputs with mapminmax ?

Hello, I have normalized my data with mapminmax which gives values between -1 and 1. However, when I present new inputs to my network, knowing that the max of these is greater than the maximum of the training data, I will have As the normalization interval [-1,1.5]. what do you think ?

Answer 1

[Greg Heath]

You do not HAVE to normalize anything because the neural net AUTOMATICALLY normalizes input and target to [-1,1].

HOWEVER, I RECOMMEND USING MAPSTD BEFORE TRAINING in order to detect and either modify or remove outliers.

My advice:

 1. Size input and target matrices
 2. Plot x, t, and t vs x 
 3. Apply MAPSTD to deal with outliers
 4. If outliers are modified or removed, 
   replot the affected variables.
 5. Train and obtain e = t-y
 6. plot  y & t and e

Hope this helps

Thank you for formlly accepting my answer.

Greg

[Hamza Ali]

Thank you for your answer. I added these two instructions before training my networks.

net.input.processFcns = {'removeconstantrows','mapminmax','mapstd'};
net.output.processFcns = {'removeconstantrows','mapminmax','mapstd'};

[Greg Heath]

No.!

Use MAPSTD to filter (i.e., delete or modify ) data for which

(z-meanz)./stdz  > yourtreshold

Plotting will help determine what threshold to use for outlier deletion and what threshold to use for outlier modification.

Then use the resulting unnormalized filtered data as inputs to the training algorithm and proceed as usual using defaults.

Although you could

a. Use the normalized filtered data as inputs

and/or

b. Use a nondefault transformation for learning and testing.

Why in the world would you want to juggle two transformations at the same time when it isn't necessary?

Hope this helps,

Greg