Stability analysis of recurrent neural networks using dissipativity

Abstract

The purpose of this work is to describe how dissipativity theory can be used for the stability analysis of discrete-time recurrent neural networks and to propose a training algorithm for producing stable networks. Using dissipativity theory, we have found conditions for the globally asymptotic stability of equilibrium points of Layered Digital Dynamic Networks (LDDNs), a very general class of recurrent neural networks. The LDDNs are transformed into a standard interconnected system structure, and a fundamental theorem describing the stability of interconnected dissipative systems is applied. The theorem leads to several new sufficient conditions for the stability of equilibrium points for LDDNs. These conditions are demonstrated on several test problems and compared to previously proposed stability conditions. From these novel stability criteria, we propose a new algorithm to train stable recurrent neural networks. The standard mean square error performance index is modified to include stability criteria. This requires computation of the derivative of the maximum eigenvalue of a matrix with respect to neural network weights. The new training algorithm is tested on two examples of neural network-based model reference control systems, including a magnetic levitation system.