Fixed-weight controller for multiple systems

We demonstrate here a perhaps unexpected result: the ability of a single fixed-weight time-lagged recurrent network, properly trained, to act as a stabilizing controller for multiple (here 3) distinct and unrelated systems, without explicit knowledge of system identity. This capability, which may be regarded as a challenge to the usual understanding of what constitutes an adaptive system, seemed plausible to us on the basis of our earlier results on both multiple time-series prediction and robust controller training. We describe our training method, which has been enhanced toward enforcing stability of the closed-loop system and dealing with process noise, and provide some results.

[1]  Paul J. Werbos,et al.  Backpropagation Through Time: What It Does and How to Do It , 1990, Proc. IEEE.

[2]  Kumpati S. Narendra,et al.  Gradient methods for the optimization of dynamical systems containing neural networks , 1991, IEEE Trans. Neural Networks.

[3]  G. V. Puskorius,et al.  Training of robust neural controllers , 1994, Proceedings of 1994 33rd IEEE Conference on Decision and Control.

[4]  G. V. Puskorius,et al.  Training controllers for robustness: multi-stream DEKF , 1994, Proceedings of 1994 IEEE International Conference on Neural Networks (ICNN'94).

[5]  Johan A. K. Suykens,et al.  Artificial neural networks for modelling and control of non-linear systems , 1995 .

[6]  Lee A. Feldkamp,et al.  Dynamic neural network methods applied to on-vehicle idle speed control , 1996, Proc. IEEE.

[7]  Lee A. Feldkamp,et al.  Adaptation from fixed weight dynamic networks , 1996, Proceedings of International Conference on Neural Networks (ICNN'96).