Learning Oscillations Using Adaptiv Control

  • We study a model for learning periodic signals in recurrent neural networks proposed by Doya and Yoshizawa [7] that can be considered as a model for temporal pattern memory in animal motoric systems. A network receives an external oscillatory input and adjusts its weights so that this signal can be reproduced approximately as the network output after some time. We use tools from adaptive control theory to derive an algorithm for weight matrices with special structure. If the input is generated by a network of the same structure the algorithm converges globally and does not exhibit the deficiencies of the back-propagation based approach of Doya and Yoshizawa under a persistency of excitation condition. This simple algorithm can also be used for open loop identification under quite restructive assumptions. The persistency of excitation condition cannot be proven even for the matrices with special structure but for a 3d system. For higher dimensional systems we give connections to the theory of linear time-varying systems where this condition is generically true (under assumption which are also needed in the time-invariant case). However, we cannot show that the linearized system related to the nonlinear neural network fulfills these generic assumptions.

Volltext Dateien herunterladen

Metadaten exportieren

Weitere Dienste

Teilen auf Twitter Suche bei Google Scholar
Verfasserangaben:Martin Georg Weiss
URN (Permalink):urn:nbn:de:hbz:386-kluedo-5822
Schriftenreihe (Bandnummer):Berichte der Arbeitsgruppe Technomathematik (AGTM Report) (178)
Sprache der Veröffentlichung:Englisch
Jahr der Fertigstellung:1999
Jahr der Veröffentlichung:1999
Veröffentlichende Institution:Technische Universität Kaiserslautern
Datum der Publikation (Server):03.04.2000
Fachbereiche / Organisatorische Einheiten:Fachbereich Mathematik
DDC-Sachgruppen:5 Naturwissenschaften und Mathematik / 510 Mathematik
Lizenz (Deutsch):Standard gemäß KLUEDO-Leitlinien vor dem 27.05.2011