Date: Feb 4, 2013 9:27 AM
Author: Steven Lord
Subject: Re: Specify/change the error used by train command



"AM76 " <wwwalaam2005@hotmail.com> wrote in message
news:kenobb$3jd$1@newscl01ah.mathworks.com...
>> I don't understand.
>>
>> Please explain "output of the NN is multiplied by a system"
>>
>> If you know what the output of the system should be for a given NN input,
>> why don't you just train the net to yield that desired output?
>>
>> Hope this helps.
>>
>> Greg

>
> Thanks for reply, Greg
>
> I have a desired target T and a plant system P. I want to use the NN as a
> controller which will be infron of the plant P. In more details, if the
> target T responds by an input in, then I insert the input to the NN to
> results, Onn, then this, Onn, will be used as input to the plant P. I want
> output of the plant P should track the target T. Therefore, I need the
> error T-P should be back propagated through the Plant Inverse and then
> used to train the NN. I am trying to use the train command and I don't
> figureout how to specify a special error to be used for the gradient in
> train command? Or any other suggestions.


So you're creating a neural network control system? If you're using Simulink
to do so, there's documentation for that.

http://www.mathworks.com/help/nnet/neural-network-control-systems.html

> I have to reprogram the whole feedforward and backpropagation from scratch
> without the facility which available in train command,


Why?

If this is for a school project and the goal of the course for which you're
doing this is NOT to learn neural networks but to control the system, ask
your professor why they're placing this limitation on you.

> I did it but it's slow and need a lot of work to specify the correct
> initials and gradient.


--
Steve Lord
slord@mathworks.com
To contact Technical Support use the Contact Us link on
http://www.mathworks.com