Effects of localized inputs on neural network training

Kristina A. Richardson, Robert F. Stengel

Research output: Contribution to conferencePaper

2 Scopus citations

Abstract

Procedures for the design and training of a radial basis function network to represent dynamic aerodynamic data are outlined. The local support of the radial basis function network approximation allows new information from a dynamic maneuver to be considered, while retaining knowledge in regions not visited. The design procedure includes both the fc-means clustering and extended Kalman filter algorithms for nodal parameter determination. The training procedure uses an input-space distance criterion to incorporate new information into the training set. The optimal number of network nodes is investigated through generalization ability. Numerical experimentation indicates that networks designed and trained with these algorithms learn locally and generalize to new data.

Original languageEnglish (US)
Pages135-143
Number of pages9
DOIs
StatePublished - 1998
Event23rd Atmospheric Flight Mechanics Conference, 1998 - Boston, United States
Duration: Aug 10 1998Aug 12 1998

Other

Other23rd Atmospheric Flight Mechanics Conference, 1998
CountryUnited States
CityBoston
Period8/10/988/12/98

All Science Journal Classification (ASJC) codes

  • Energy(all)
  • Aerospace Engineering

Fingerprint Dive into the research topics of 'Effects of localized inputs on neural network training'. Together they form a unique fingerprint.

  • Cite this

    Richardson, K. A., & Stengel, R. F. (1998). Effects of localized inputs on neural network training. 135-143. Paper presented at 23rd Atmospheric Flight Mechanics Conference, 1998, Boston, United States. https://doi.org/10.2514/6.1998-4158