Truncated backpropagation through time and Kalman filter training for neurocontrol

Abstract

We have recently established the feasibility of training recurrent neural networks by parameter-based decoupled extended Kalman filter (DEKF) algorithms for control of nonlinear dynamical systems. In this paper we investigate the use of truncated backpropagation through time (BPTT) for approximating the required dynamic derivatives that are used by the DEKF training algorithm. The use of this approximation allows the gradient calculations and weight updates by the DEKF algorithm to be performed asynchronously with application of control signals, thereby leading to a scalable, real-time, online training algorithm. We demonstrate in simulation the effectiveness of the BPTT-based DEKF algorithm for the problem of automotive engine idle speed control.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>