next up previous contents index
Next: Non-gradient methods Up: Gradient-based algorithms Previous: Real-time recurrent learning   Contents   Index


Other derivative-based methods

It is also possible to train DTRNN using the extended Kalman filter (EKF, see e.g. (Haykin, 1998, 762ff)) of which RTRL may be shown to be a special case (Williams, 1992); the EKF has been successfully used in many applications, such as neurocontrol (Puskorius and Feldkamp, 1994). The EKF is also related to RLS (recursive-least squares) algorithms.



Debian User 2002-01-21