how define constraint on the outputs of ANN fitting?

Hi All

1- how shall I define a contraint on one of all the outputs ? for example if one of the outputs is an angle , and I don't want it to increase more than 90 , since it's out of the logic of my problem

2- when I train a network in a range of data , should it always be given inputs in the same range or also should work for out of the first input ?

[farzad]

Is there any way possible ?

Answer 1

[Greg Heath]

It depends on the reason for the constraint. Is it math/physics or convenience?

 It is all done by a combination of target scaling and choice of output transfer function.
 For example,  

 1. PURELIN  ( -inf , inf )
 2. TANSIG   ( -a   , +b  ), a, b, finite
 3. LOGSIG   (  0   , +b  ), b finite

Hope this helps.

Thank you for formally accepting my answer

Greg

[farzad]

Thank you Dear Professor

well I have 3 outputs , like a,b,c , for example lets say b is an angle, and it's limit SHOULD be 0< b <90

Does the above transfer function work for multiple outputs ? to put contraints only on one of them ?

[Greg Heath]

"THE" above transfer fnction? ... There are 3, and because of scaling you can use any one of them.

In fact I recommend running a fitnet example of your choice 3 times with the 3 different output transfer functions and examine the differences.

Initialize each design with the same random number state, e.g.,

 rng('default')

This will yield the same random initial weights and trn/val/tst random data division

[farzad]

Thank you very much professor

yes I understand each are separate transfer functions , I will search for them , I use rng(0) before the initiation of the loop, so I hope it's the same thing

[farzad]

I feel that using these transfer functions , might lead that ALL of the outputs go in the same cosntraint , I am not sure how to do it in the way I asked , I mean , to put the constraint only on one of the outputs .

how should I do that ?

[Greg Heath]

First, do it for a single output.

Make sure there is a physical or mathematical reason for the constraint; not just a convenience.

The number of outputs is irrelevant. Concentrate on the targets that have upper AND lower bounds. e.g, sin, cos, tanh, etc. If none, then concentrate on those with one bound. Single bounds can be reversed by just using a minus sign.

[farzad]

Dear Professor

it's not for convenience , it's an angle that can not go beyond 90 , but what do you mean by concentrate on the targets that have upper AND lower bounds ? you mean I train the network with 8 inputs and 1 output ?? unfortunately yet I couldn't know how to use those transfer functions to put the limit , how should I indicate that ?

[farzad]

Dear Professor

I still couldn't find the appropriate solution for this problem

[farzad]

I tried :

net.layers{1}.transferFcn='tansig(0 , 90)';

but it doesn't work !

[farzad]

Dear Professor, I still think I should have all the outputs together and I can not focus on one , because the network should know that there is limit on one of the outputs with respcet to the other 2 outputs , I yet could not find a solution for this case

[farzad]

Anyone ?