作者: G.V. Puskorius , L.A. Feldkamp
关键词:
摘要: We have recently established the feasibility of training recurrent neural networks by parameter-based decoupled extended Kalman filter (DEKF) algorithms for control nonlinear dynamical systems. In this paper we investigate use truncated backpropagation through time (BPTT) approximating required dynamic derivatives that are used DEKF algorithm. The approximation allows gradient calculations and weight updates algorithm to be performed asynchronously with application signals, thereby leading a scalable, real-time, online demonstrate in simulation effectiveness BPTT-based problem automotive engine idle speed control. >