Deterministic Learning and Pattern-Based NN Control

Abstract

A deterministic learning theory was recently presented for identification, control and recognition of nonlinear dynamical systems. In this paper, we propose a pattern-based neural network (NN) control approach based on the deterministic learning theory. Firstly in the training phase, the definitions of dynamical patterns normally occurred in closed-loop control are given. The closed-loop system dynamics corresponding to the dynamical patterns are identified via deterministic learning. The representation, similarity definition and rapid recognition of dynamical patterns in closed-loop are also presented. A set of pattern-based NN controllers are constructed using the knowledge obtained from deterministic learning. In the test phase, secondly, a pattern classification system is introduced which can rapidly recognize the dynamical patterns in closed-loop. If the dynamical pattern for a test control task is recognized as very similar to a previous training pattern, then the NN controller corresponding to the training pattern is selected and activated, which can achieve exponential stability and guaranteed performance of the closed-loop control system without readaptation and high control gains. The proposed pattern-based NN control approach may provide insight into human's ability to learn and control and possibly lead to smarter robots.