How To Change MSE Performance function Input Arguments ?

i would like to implement a recurrent neural network to control a dynamic system , i need to use 2 of my net input arguments as variables of error function instead of network targets and outputs ,so system will dynamically gets to an optimum point. but i cant change mse inputs. in mse code we have mse(net,varargin) so i cant change varargin in my mse code (i have made a copy of mse.m and it works fine without any modifications) because i dont know where varargins are taken and any change will lead to errors. so is there any solution ?

[Greg Heath]

Why do you "need" to do this? What is the analytic form of your error function?

[Amin K]

thanks for your response greg, i have 4 inputs which 2 of them are identically the same thing but one of them is optimum point and other is real time one. The idea, here is that for a particular value of other inputs the error between the optimum point and real time inputs will be minimum when the system operates at the optimum point. If for any reason the system operates at any other operating point other than the optimum point then, the neural network’s weights will be adjusted accordingly in order to produce the output a reference value which will reduce this error and bring the system to maximum possible point. system is dynamic and any changes in network output will affect the real time input of network.

[Greg Heath]

I don't understand what you are trying to do. I doubt if anyone else does either. Including analytics would help ( As I asked for previously!!!)

[Amin K]

this is inputs and output of my recurrent network here e = pm - po and error function E = 1/2 * e^2. this causes weight update rule for hidden layer be like

 dw(N)= b* -dE/d(layerinputs) * layeroutput + a* dw(N-1)

so weight updates will continue till pm - po gets zero and then whole system will work at that situation. as you see there is no targets in this network (unsupervised learning) but the problem is standard performance functions in matlab use targets and outputs of network to reduce the size of error function( e = targs - outs ).

i hope this has been clarified the problem, thanks.

[Greg Heath]

Not yet.

Why does the e arrow go thru the box with the arrowhead sticking out?

Why does the w* line have no arrowhead?

What are vw, w, Po, Pm and w* ?

What is causing the system to be recurrent?

[Amin K]

e arrowhead is symbolically sticked out , its not really an output of neural network. i just cropped this part of system to show schematic of neural network, w* goes to pid controller and causes Po increase to get closer to Pm .

vw is mechanical speed makes shaft of a asynchronous generator rotate, w* and w are rotational speed of rotor reference value and realtime value respectively, po is power generated of machine and pm is maximum power.

w* causes w and po to increase and because of po and pm being variables of error function, network's weights get updated continuously till e gets near zero then weights stay constant and system works at optimum point untill vw changes and the whole procedure starts over.