Neural Net Algorithms for Dynamical Systems

PhD Thesis, 1992

Y. Ming Cheung

Abstract

A neural net based algorithm is devised as an alternative to traditional analogue/numerical integration. The new algorithm consists of a multi-layered neural net integrator model inspired by the neuron organization of the vertebrate retina. A mixture of implicit weight setting, supervised and unsupervised learning is employed. The convergence of this approach proves to be fast when compared to existing models producing comparable results.

When the model is operating in a closed loop system it yields a consistent estimate of the derivatives of pictorial input profiles.

The mapping of the resulting neural net models onto single and multiprocessor systems is examined. A general framework is formulated to permit arbitrary network definition and easy alterations of network parameters.

A parallel processing technique for distributed memory multiprocessor systems is devised. The parallel algorithm yields a large reduction in processing time.

I