MLP artificial neural networks can be used for prediction. In order to avoid making suggestions we standarize input data (with linear or non-linear method) providing data with values from near the same range e.g. .
Do we need to standardise the output? Or should we leave output data with original values?
Many thanks for hints.
Hubert
It depends on the activation function at the output layer. For instance, if you use a tansigmoid you should normalize targets to fall in the range [−1, 1]. If you use a linear activation function in the output layer then you could work with the original values.
usually, you can obtain a standard results through ANN. If you want to give more interpretation on your data, I think you can repeat the simulation and collect the results and do some statistic analysis.
Thank you guys for the answers. Special thanks to Jose. Activation function in a output layer it is a key to the answer.
Thank you for the help.
Hubert
Fatima,
I try to prognose delay in contract completion date. Input data have different values (e.g. contract sum in milion and unemployment rate in %), so I standardize them. Output is oryginally in days: from zero to hundreds. I've standardized thre output with the same method as all data. Can you say more about influence of application on standarization of output? Or where to read about it?
Best regards
Hubert
Just to elaborate slightly on Jose's answer, for a prediction problem your ANN is learning an approximation of the function between your input and output data.
Commonly this is done through gradient descent which relies on calculation of the error between predictions and true values for each instance. Now obviously your gradient descent wont work if your ANN output is constrained to a range of values by your activation function (such as Sigmoid with range [0, 1]) which your real output values fall outside of, though a quite a few (inc. Matlab) libraries accommodate for this by standardizing your output.
To answer your question, when training ANNs for regression problems the normal way, you will need your output in the same form as your real data. This can be choosing a sensible activation function (Linear allows continuous values for example) or by standardizing your output data to match the range of your activation function.
Thanks Tim,
So I will try to find out in matlab how to switch off its internal standarization not to double this action.
Best regards
Hubert
Don't take my word for it that Matlab does so, it's been a while since I used it! Glad I could help though.
Kind regards,
Tim
VERY IMPORTANT !!!
Hi everybody,
In my present case I predict the time with ANN basing (among other factors) on planned time. There are real Times (as an output) too. Analysing results these three values are compared
I realized that, planned time (one of input) and real time (output, for teching) shold be divided by the same numer during standardization.
OTHERWISE
If we find max for planned Times and divide all planned Times by this max to standardize and ...
... when we find another max for real Times and repeate procedurę in this column...
then, when we will come to comparing planned and predicted Times (based on real teached), we will be comparing not mached values...
UNLESS
we come back from standardized numbers to real ones, separately for each type of data.
I revised it already. Thanks for attention :-) I hope it will help you
Regards
Hubert
- Badges
- Science topic