questions about "plotperform" and performance graph

Hi,

when i searched Matlab Help and typed "plotperform", it explains this by an example with a picture MSE vs the epochs, Having training, test and validation curve. why it is written "best validation performance is 0.4778 at epoch 4" ? In other words i want you to explain me how can i analyse a MSE vs epochs graph! How one graph shows a good or bad network? my another question is whether a good network is one, in which all curves for training, test and validation converges to the same point or for example the picture depicted in the help - plotperform page can be considered as a good trained network?

Thanks

Answer 1

[José-Luis]

MSE or mean squared error, gives you the difference between observation and simulation. The lower the better. In this case it means that the best results were attained after four epochs.

A good/bad network. That question is difficult to answer. What might be an acceptable error for some applications might be disastrous for others. However, MSE vs epoch will tell you if your network has done all it can for the data. In the example, you can see that the MSE stabilizes. If it didn't, if it was jumping all over the place, it would mean that your network has a hard time finding an optimal solution and that your configuration might not be ideal.

All curves converging to the same point would mean that your network performs equally well (or equally bad) in both training an validation. Good or bad depends on your application.

[Respaco]

Thank you so much for your perfect explanations.

In that example can we say that at epoch 7 that all train, test and validation curves becomes flat without jumping, is where the best result is expected to find?

As far as i understood from what you said, it is not necessary that the curves for train, test and validation becomes so close to each other that we could say the network is trained well. Am i right? But when each curves converges to a specific value (no matter how much the value for MSE reagrding each train,test and validation differ from each other) we can say the network works good?

Another question! When it is asked to show a convergence chart showing the MSE vs the iteration run for your network, Is it essential that the graph has all types (train,test,validation), or it can have only 'train' or 'train and test', ...?

Many thanks

[chaudhry]

waiting for the answer