how to get the transfer function

i have data in the form of 6 inputs and 2 outputs (80*6 and 80*2 tables)

i want to optimize the transfer function (f(x1,x2,x3,x4,x5,x6,y1,y2)) with genetic algorithms, but i dont know how to get it first, any ideas?

im also using neural networks on the data,can it be done using these?

if it helps,i can do with the transfer function of each of the outputs separately (f(x1,x2,x3,x4,x5,x6,y1),f(x1,x2,x3,x4,x5,x6,y2))

[chris rz]

or else, if this cannot be done, is it possible to optimize a black box model using genetic algorithms after identifying it somehow (model identification toolbox or sth)?

Answer 1

[Greg Heath]

The terminology "Transfer Function" is only applicable to the spectral analysis of linear systems. Clearly it is inappropriate for this problem.

A model for the nonlinear I/O relationship y = f(x) can be constructed using a Feedforward Multilayer Perceptron Neural Network with a single hidden layer. AFAIK the NN Toolbox does not offer a genetic learning algorithm. In addition, the Global Optimization Toolbox containing genetic algorithm functions does not consider Neural Network Design.

You can search the internet including the comp.soft-sys.matlab and comp.ai.neural-nets archives.

The I/O relationship for a single hidden layer FFMLP can be characterized by the matrix equations

h = tanh(W1*x+b1); y = W2*h + b2;

(x,h,y) = (input, hidden variable, output)

(W1,b1,W2,b2) = unknown weight matrices to be estimated

y = W2*tanh(W1*x+b1)+ b2 ;

If the training target is t, the error is

e = (t-y);

and the objective function to be minimized is

MSE = mean(e(:).^2);

Hope this helps.

Greg