How do I change the minimum step size used in training with the TRAINCGF function from the Neural Network Toolbox 4.0.2 (R13)?

I am using some functions from the Neural Network Toolbox in the following code:

net=newff(minmax(training_data),[3,1],{'tansig','purelin'},'traincgf'); 

net.trainParam.show = 1000;

net.trainParam.epochs = 10000;

net.trainParam.goal = 1e-5;

[net,tr]=train(net, training_data,Target);

However, the program stops every time before it reaches the preset 10000 epochs and displays the following message:

 "TRAINCGF, Minimum step size reached, performance goal was not met."

How do I make it run to the 10000 epochs without early stopping?

Answer 1

[MathWorks Support Team]

Tolerance in the Neural Network Toolbox is controlled by several parameters, depending on which SearchFunction is used. When using the TRAINCGF function, the tolerance for the search is calculated using the following formula:

 tol = delta/scale_tol;

Where:

scale_tol = net.trainParam.scale_tol;

delta = net.trainParam.delta;

Making the delta very small will make the tolerance very small. Usually, scale-tol is not changed from the default.

Also, net.trainParam.minstep is used to control the step length used in line search in any iteration.