Neural network generalization and system sensitivity in feedbackcontrol systems

摘要

A new framework for quantifying and analyzing the generalizationability of neural networks in control systems is presented. Rigorousdefinitions to quantify the generalization ability of a neural networkin the context of system control are given. Utilizing these definitions,it is proved that a successfully trained neural network alwaysgeneralize “well” to some extent. This dual property of atrained neural network provides further justification for neuro-controlapproaches, because the added benefit of generalization is nowanalytically assured. In addition, a method for estimating the extent towhich a trained neural network will generalize is presented. The resultsof this work provide new tools for performance analysis ofneural-control systems, and represents a first step towards a rigorousframework for performance-oriented analysis and synthesis of neuralnetworks for control