New Methods of Neural Network Training
Abstract
New methods for training computational neural networks for dynamic system identification and control have been created, performance of the training algorithms has been analyzed, and the resulting neural networks have been evaluated. Computational neural networks are shown to have excellent potential for identifying the dynamic models of nonlinear systems and for controlling such systems over their entire operating space. Three topics were addressed: (1) Aerodynamic model identification using sigmoid and radial basis function networks; (2) Control of the preferential oxidizer for a fuel cell power system using a neural network Initializing a neural network (nonlinear) controller so that it replicates the characteristics of a gain scheduled linear controller. This research produced new training approaches that will allow future dynamic systems to work with higher accuracy, greater efficiency, and improved reliability.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 1999
- Accession Number
- ADA370007
Entities
People
- Robert F. Stengel
Organizations
- Princeton University