I would like to construct two NARX networks, with the output of the first fed in as input to the second (along with other inputs). I only have data for the final target (i.e. the target of the second NARX), hence I am unable to train the first NARX independently and hence again, I am resorting to training the two NARXs simultaneously.
Each of the two NARXs represent a physical process, thus it would be beneficial to keep them as two neural networks in order to get insight into the intermediate output (i.e. the output from the first NARX).
I considered constructing the series of two NARXs in Simulink and then training the two in one go, but unfortunately Simulink is used more for simulating 'trained' neural networks rather than training them with the known learning algorithms commonly used with neural networks.
Any suggestions are appreciated.