How to use a trained neural network as objective function in fminsearch?

I have trained a neural network, with input x that is a matrix 12x22000 and a target t 1x22000, and i have got an output y 1x22000, now i want to optimaize one element of my output with the tool fminsearch but i don't know how to write the objective function 'fun'.

What should i put as 'fun'?

x = input;
t = output;
trainFcn = ['trainlm'];  
hiddenLayerSize = 50;
net = feedforwardnet(hiddenLayerSize,trainFcn);
net.divideFcn = ['divideblock'];
net.performFcn = 'mse';

[net,tr] = train(net,x,t); 

y = net(x);

%now i want use fminsearch(fun,x0,options)

[Matt J]

Optimize with respect to what unknowns? If you are trying to refine the weights, the output surely depends on many, many network weight parameters, which is not suitable for fminsearch.

[Leonardo Anzellotti]

optimize with respect to the 12 input variables for each output

Answer 1

[Matt J]

fminsearch.is unlikely to be able to handle 12 unknowns well. You should probably use fminunc if you have the Optimization Toolbox. Either way, the 'fun' input would be,

fun=@(x) net(x);

[Saurabh Sharma]

How can we get a trained Gaussian process regression machine learning model in a mathematical equation form? How to write 'fun' if we want to minimise three or four models at same time?

Thankyou