Back propagation learning of MLP doesnt convergate, why?

Hi, I am working on image pattern recognition in Matlab but for some certain reasons I cannot use nn toolbox (I have to rewrite it into labview for deployment). So I developed learning algorithm for 2 layers MLP but it always stucks in local minima. I tried Back Propagation Algorithm and genetic algorithm as well but both don't calculate weights good enough I tested it for simple data and few targets = both work perfect but for more parameters it's not good.

When I use nnstart IF with same data, I get perfect results but I am not able to rewrite it or use the calculated weights to get the network only represented by matrices. I use the same weights, transfer function, topology, data pre-procesing, still have different results.

Any suggestions that could be helpful in my suffering?

Thanks in advance.

Jakub

My code for learning:

% x - one sample % t - corresponding target % w1,w2 - weights % n - learning rate

% w_new_1,2 - new weights

function [w_new_1,w_new_2] = back_prop(x,t,w1,w2,n)

    y_1 = logsig(x'*w1);                            %output for 1th layer
    y_2 = logsig(y_1*w2);                           %output for 2nd layer

    sigma_2=y_2.*(1-y_2).*(t(:,1)'-y_2);            %error for 2nd layer

    for i=1:length(sigma_2) 
        w_new_2(:,i)=w2(:,i)+(sigma_2(i).*y_1)'*n;    %update weights between hidden layer and output
    end

    sigma_1=y_1.*(1-y_1).*(sigma_2*w_new_2');       %error for 1th layer
    X_act=x;

    for i=1:length(X_act)
        w_new_1(i,:)=w1(i,:)+(sigma_1.*X_act(i))*n;   %update weights between hidden layer and output
    end
end