Training Neural Networks with Weight Constraints
Abstract
Hardware implementation of artificial neural networks imposes a variety of constraints. Finite weight magnitudes exist in both digital and analog devices. Additional limitations are encountered due to the imprecise nature of hardware components. These constraints can be overcome with a stochastic global optimization strategy which effectively searches the range of the weight space and is robust to quantization and modeling errors. Evolutionary programming is proposed as a solution to training networks with these constraints. This work investigates the use of evolutionary programming in optimizing a network with weight constraints. Comparisons are made to the backpropagation training algorithm for networks with both unconstrained and hard-limited weight magnitudes. Neural networks, Analog, Digital, Stochastic
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 1993
- Accession Number
- ADA264665
Entities
People
- John R. Mcdonnell
Organizations
- Naval Command, Control and Ocean Surveillance Center