Convergence in On-line Learning of Static and Dynamic Systems
Abstract
The paper derives new conditions for global convergence of the adaptive moment generation algorithm when applied for on-line, or equivalently, recursive supervised learning of static and dynamic nonlinear systems. The paper also proposes a difference equation based nonlinear dynamic model, that enforces {\em structure} and results in a new type of recurrent neural network. The convergence analysis applies averaging using Ljung's associated differential equation method. It is first proved that the asymptotic update behaviour of the adaptive moment generation algorithm is equivalent to a scaled stochastic gradient update for the standard hyper-parameter setting, or equivalent to a sign-sign update strategy in case the internal filtering of the algorithm is turned off. The analysis is concluded by proving global convergence to the set of parameters that gives a correct input-output description of the true system. The two hyper-parameter settings are evaluated with a Monte-Carlo analysis when the adaptive moment generation algorithm is applied for learning of nonlinear automotive cruise control dynamics. This validates the correct operation of the structured recurrent neural network and confirms the expected reduced convergence speed for the sign-sign update case.
- Publication:
-
arXiv e-prints
- Pub Date:
- January 2025
- DOI:
- arXiv:
- arXiv:2501.03049
- Bibcode:
- 2025arXiv250103049W
- Keywords:
-
- Electrical Engineering and Systems Science - Systems and Control