Whenever i run my neural network I get different result in nftool

I built my neural network using only nftool without writing code.Whenever i run my neural network I get different result. What can I do to have the same result?

Answer 1

[Walter Roberson]

This is expected. The default initialization function for weights is random.

You can use rng() before every run that initializes the weights, or you can use a custom weight initialisation that is not random.

[Kangujam]

Thank you @Walter Roberson. I do agree with your answer but I am dealing only with nftool which does not involve any writing of code. Can you please elaborate on how to initialise weight in nftool?

[Walter Roberson]

You will need to give an rng() command in the command window before each time you click to "train".

nftool is a graphical interface to simplify use of a number of routines. As is common with graphical interfaces, it only gives access to a portion of the functionality of the routines. It was not designed to provide an interface to control the random number stream or initializations of weights. It is expected that if your needs outgrow nftool that you will implement your own routines.

[Kangujam]

Thank you @Walter Robertson. But I am still getting different result even after the implementation of the command rng() before I clicked "train" in nftool. Can you please suggest me how to rectify this problem?

[Walter Roberson]

edit nftool

At approximately line 85, you will see

        STATE = trainNetwork(STATE,percentValidate,percentTest,trainingFcnIndex);

Just before that line, insert your call to rng() to set a constant seed.

Remember to save the result under a different name.

[Kangujam]

Thank you for your response @Walter Robertson. It still doesn't work.For same number of neuron, it gives different result.

[Walter Roberson]

What was the line of code that you added?

[Kangujam]

STATE = trainNetwork(STATE,percentValidate,percentTest,trainingFcnIndex); The above line of code is at line 77 and I have added rng() just above it.I have run it for same number of neuron but still different answers are obtained.

[Walter Roberson]

You need to pass a seed to rng() . Like

rng(12345);