When I should stop training a neural network?

I'm working in a neural network with BackPropagation. The network has 6 inputs, 1 hidden layer (6 neurons on that layer) and 1 output. I train the network with algorithms "Levenberg-Marquardt" and "Bayesian Regularization". So, the idea is can "predict" a result but the results are not the right ones according to the table with the historical data.

To stop the training, for the moment, I look the "regression plot", the "Mean squared Error" and "Regression R Values", wich have the "ideal values" but still the results are not accurates and are not even close with datas who "doesn't exist" in the table with the historical data.

What graphic should I look at to know the network is not overfitting or is correctly trained?

Answer 1

[Greg Heath]

The danger is OVERTRAINING an OVERFIT NET. There are several approaches.

1. PREVENT OVERFITTING the I-H-O net by having the number of training equations

Ntrneq = Ntrn* I

be no smaller than the number of unknown weights,

Nw = (I+1)*H+(H+1)*O

i.e., Ntrneq >= Nw

2. PREVENT OVERTRAINING by using a reasonable training goal

mse(target-output) <= 0.01 * mse(target-mean(target')' )

3. PREVENT OVERTRAINING by using a validation subset to implement

EARLY STOPPING

4. PREVENT OVERTRAINING by using TRAIN BR to implemet BAYESIAN

REGULARIZATION LEARNING

EARLY STOPPING (3) is automatic with the default TRAINLM.

Typically I try to implement 1-3 via minimizing H in addition to using EARLY STOPPING with the default training algoithm TRAINLM (2).

On rare equations I will use TRAINBR (4) when 1-3 do not yield satisfactory results.

Searching BOTH NEWSGROUP and ANSWERS using

Greg Ntrneq Nw

should yield zillions of examples

Hope this helps

THANK YOU FOR FORMALLY ACCEPTING MY ANSWER

GREG