neural network training terminated prematurely

Hi, i am trying to learn the neural network toolbox from the help section. I have a problem with the folowing code given in the tutorial.

p = [0 1 2 3 4 5 6 7 8];
t = [0 0.84 0.91 0.14 -0.77 -0.96 -0.28 0.66 0.99];
net = newff(p,t,10);
y1 = sim(net,p)
net.trainParam.epochs = 50;
net.trainParam.goal = 0.01;
net = train(net,p,t);
y2 = sim(net,p)
plot(p,t,'o',p,y1,'x',p,y2,'*')

The problem is that the output doesn't match the target when the same input is given. I have decreased the goal to 0.0001 to the end that now training stops before any of the stopping conditions are met (none of the bars in the nntraintool window is filled) while significant error remains. Isn't it supposed to give results identical to the target if the training data is given as the input after training is complete?? and why does training stop before any of the stopping conditions have been met?? thanks.

Answer 1

[tathagata chakraborti]

i think the fact that only 60% are used for training led to the mismatch, changing the default dividerand settings has resulted in 100% match. :)

Answer 2

[Greg Heath]

1. y2 ~= y1 because of network training.

2. In a repeat y3~= y1 and y4~=y2 because newff net creation creates random initial weights. You have to initialize the RNG to the same initial state in order to duplicate a design.

3. Unfortunate choices of random initial weights are not uncommon. That is why I always design 10 nets for each candidate value of H, the number of hidden nodes.

4. Search on

 greg Ntrials

for examples of my double-loop designs.

Hope this helps.

Greg

[Greg Heath]

The cause is probably one of unfortunate random initial weights