create an Evolutionary Neural network ???

I need to create an Evolutionary Neural network and i used the function

net = patternnet(hn); but i tuned the weights manually but what i need is to have the output value of the neural network when the input passes through the layer until the output node and get the result without any back propagation of the errors neither adjutement of the weights .So how can i do that ? Can i define my training function as input of the patternet(hn,@my_fun) or newff(...,@my_fun) is it possible ? the second idea is to create the neural network with the function patternet or any one else , but what i need here is to stop the neural network after his first ieration which means don't let teh neural network adjusting the weights by itself .

[Walter Roberson]

Stopping after one iteration was discussed in your (still-active) question http://www.mathworks.co.uk/matlabcentral/answers/47034-how-can-i-stop-the-neural-network-after-just-one-iteration

Answer 1

[Greg Heath]

Since this is a pattern recognition algorithm, for c classes or categories, your targets should be columns of the c-dimensional unit matrix. For an I-H-O (O=c) node topology, use the layer equations

h = tansig(IW*x+b1); % Output of hidden layer

y = logsig(LW*h+b2); % Output of output layer

e = t-y; % error

NMSE = mse(e)/var(t,1,2)% Normalized mean-square-error

NMSEgoal = 0.005

Genetic Search: Consult a reference because I'm making this up.

[ I N ] = size(x)

[ O N ] = size(t)

Neq = N*O % Total number of scalar training equations

1. Standardize x

2. Initialize the random number generator

3. Choose the number of hidden nodes, H

This will yield Nw = (I+1)*H+ (H+1)*O unknown weights to be estimated from

Neq equations. Require Neq > Nw but desire Neq >> Nw . So choose

H << Hub = (Neq-O)/(I+O+1)

4. Generate M random sets of Nw weights

5. Use the weights in the equations and choose the best B nets.

6. Randomly mutate m% of the B weight sets

7. Generate (M-B) new random sets of Nw weights

8. Repeat 4-7 until the NMSEgoal is reached or the number of repetitions has exceeded a specified limit.

9 . If NMSEgoal is reached, STOP. Otherwise go to 3 and increase H

Again, see a genetic algorithm reference,

Thank you for officially accepting my answer.

Greg

[Mariem Harmassi]

Thank u for answering what i search is the sim function , i set my network configuration and call the sim function it will return me the estimated output of the network and the performance without changing the weights .Is it true ?? i confused the sim and train fuctions .

[Greg Heath]

After you have obtained the weights via a genetic algorithm, you can load them into a net and use sim or net to obtain the output as explained in my 2nd answer.

As explained in that answer, avoid using any of the NNTB functions until the genetic algorithm has converged to a final set of weights.

Answer 2

[Mariem Harmassi]

Can someone help me please ????

Answer 3

[Greg Heath]

My combined equations for y and h are equivalent to the sim and net functions when calculating the output

y = sim(net,x);

or

y = net(x);

If you follow my directions above, there is no reason to use the Toolbox functions patternnet, configure, train or sim.

However, once you have converged to a final set of weights using the genetic code above, you can load them into a net using

net = patternnet(H);

net.IW{1,1} = IW;

net.LW{2,1} = LW;

net.b{1} = b1;

net.b{2} = b2;

Otherwise, you will have to load weights into the net at EACH step for EACH set of candidate weights. Obviously, you might be retired before the design converges.

Hope this helps.

[Mariem Harmassi]

does it have the same computation time when using the defined funtions above and the sim ??

[Greg Heath]

Definitely not. Genetic solutions of ANY problem typically take much much longer.

Greg