Prediction using created NARX neural network

Hi everyone

I've created a NARX NN for DC motor in open loop mode (2 inputs and 1 target , delays equal to 2 and hidden layers equal to 25) and I want to test it ( or predict it) on new data input , so here i passed to the closed loop mode and i do the test and notice that my network performance is decreasing (when i compare the estimated output predicted from the new data input with the real mesurement), so i try to train the closed loop to enhance performance but it takes 2 hours and still waiting the end of the training, so please how we can keep the same performance of the network (passing from OL to CL ) ?

1/ Is there any other method instead of doing the training of the closed loop ?

2/ I want to know if it is possible to predict future output using open loop mode with removedelay function ?

Your advice or suggestions will be much appreciated and welcomed.

Best.

[Fernando Quevedo]

Hi Salma, hope i help but I'm not an expert.

Your network could be overfitted with the training set, hence the decrees in performance in other sets. How big are the sets you are using? How much does the system data change every step?

The decrees in performance from openloop to closeloop is very normal, since errors from past times accumulate over time and creates a butterfly effect.

1/ Training in closeloop sounds a little stupid since using the prediction from the system while training will give false data. I belive if you train a closedloop network it will still do the same as an openloop one.

It is normal for long hours of training depending on the hardware you use. Reduce the hidden layer size could speed the results without sacrificing results.

2/ The change of open to close loop makes it easier to use the network, you could prepare the data and use the openloop but results should be the same as closeloop if you use the NN feedback.