Problem training Neural network

Hi,

I am trying to train a simple multilayer feed forward neural network using a genetic algorithm. I have been trying to train the network to learn a sine wave from 0:6pi. However, the search stops abruptly without finding the global optimal solution. I usually get a MSE of 1e-2 or greater. I tried with 4 neurons to 20 neurons in the hidden layer, but the search still ends with non-optimal solutions. I also tried to train a ANN using the the neural network toolbox's GUI nftool, which gives very good results, even with 4 neurons in the hidden layer. Isn't genetic search supposed to better than BP based training? Or is something wrong with my code?

The function I am using for optimization is as defined below:

function [err]=ANN(weights)

global len; %len is the number of neurons in the hidden layer

w=weights(1:len); %weights from input layer to hidden layer 

u=weights(len+1:2*len); %bias at hidden layer

v=weights(2*len+1:3*len); %weights from hidden layer to output layer
b=weights(3*len+1); %bias at the output layer

outh=sigmtf(w,u,input);%Sigmoid transfer function

outpma=outh.*repmat(v',[1 length(input)]);

outp=ones(1,length(w))*outpma+b; %final output of the network

err=mean(((outp-target)).^2); error to be minimized

%Function used to train the error is from the GA toolbox.

ga(@ANN,len*3+1)