FUNCTION MODELING WITH NEURAL NETS.
Abstract
Research concerned replication of the impulsive response (i.e. modeling) of arbitrary functions or plants. The significant results to date are a theory and test of a closed-loop multi-in-put multi-output modeler with parallel internal structure and some theoretical results have been obtained on adaptive selection of an optimal training rate. The cloded-loop modeler is capable of generating real and complex poles, can exactly relicate a stable plant whose order does not exceed that of the modeler, and can determine the impulsive response of more complex plants with least mean square error. Training rules, instrumentation and test results are demonstrated. An adaptive modeler should achieve high accuracy rapidly. But if the plant's output is corrupted, or if perfect modeling cannot be achieved, then fast training makes the adaptive weights oscillate and the modeler does not represent the plant. The training rate should be high in order to achieve modeling rapidly but should be low in order to achieve a least mean square error solution. A training algorithm is demonstrated for adaptive selection of training rates in an open-loop modeler such as to minimize the weighted sum of mean square modeling error plus the mean square rate of change of the adaptive weights. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 14, 1966
- Accession Number
- AD0483991
Entities
People
- Frederic D. Powell
- Johannes G. Goerner
Organizations
- Bell Aircraft Corporation