Max Mu reached problem

Hi everyone,

I have a closed loop NARX model for which I am trying to decide on the architecture so apply many 10 runs for hidden node numbers from 1 to Hmax.

While the code runs what I realized was that for (roughly) 9 runs out of 10 the model performance is quite good but for 1 run it says max mu reached so quickly that it terminates training after a few iterations so the model performance is really bad.

What might be the reason for this? I am sure there is no overfitting in other runs. I tried increasing mu_max and decreasing mu_inc they didn't work.

And is it a problem? Because it uses different initial weights every time so what if it does not learn good enough in 1 run. At the end I will use only one of these architectures, why would it be a problem?

Any ideas??

Thank you very much in advance..