BackPropagation Through Time (BPTT) is well established tool for training Recursive Neural Networks (RNN). The RNN (from my point of view) is a non-linear dynamic system, as it includes internal states and the activation functions are non-linear.

Therefore, BPTT should be viable option of training/adjusting parameters of any non-linear dynamic system. The motivation to change RNN for dynamic system, is that dynamic system may have desired physical representation (e.g. Equivalent Circuit in electronics) while RNN has hyper-parameters (usually not physically representable).

I have searched my university library and internet, but havent found any literature addressing the use of BPTT for other purpose than RNN training.


1) Is there any problem, why BPTT is not used for fitting dynamic systems to measured data ?

2) If it is suitable, can you share any literature/examples/experience regarding the use of BPTT for training arbitrary (non-linear) dynamic system $\frac{dx}{dt}=f_{(x,y,t)}$ with time-series data?

  • $\begingroup$ Why do you think a dynamic system is backprobagatable----like a neural network? $\endgroup$ – Jiapeng Zhang Jul 3 at 14:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.