Use of Hopfield Networks for System Identification and Failure Detection in Autonomous Underwater Vehicles
Abstract
In the early 1980's John J. Hopfield developed a recurrent network based on a model of biological neurons. In his model, each neuron accepts inputs from all other neurons in the network, modifies each input with a weight and converts their sum to an output via the non-linear sigmoid transfer function. This output is then fed back to each of the input paths where the input signals are updated before the next summation. It has been proposed that this network can be successfully applied to the problem of system parameter identification where the weights are functions of the system states and the network, after being allowed to process a continuous block of system states, is guaranteed to converge to the system parameters. This thesis explores the concepts of network stability and solution existence for a time-invariant system. It is shown that the network will converge as expected provided the steady-state solution falls within the range of values of the sigmoid transfer function.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 1991
- Accession Number
- ADA245875
Entities
People
- Alan M. Marsilio
Organizations
- Naval Postgraduate School