A robust regularization algorithm for polynomial networks for machine learning

Holger M. Jaenisch; James W. Handley

doi:10.1117/12.884284

Translator Disclaimer

19 May 2011 A robust regularization algorithm for polynomial networks for machine learning

Holger M. Jaenisch, James W. Handley

Author Affiliations +

Proceedings Volume 8059, Evolutionary and Bio-Inspired Computation: Theory and Applications V; 80590A (2011) https://doi.org/10.1117/12.884284
Event: SPIE Defense, Security, and Sensing, 2011, Orlando, Florida, United States

Abstract

We present an improvement to the fundamental Group Method of Data Handling (GMDH) Data Modeling algorithm that overcomes the parameter sensitivity to novel cases presented to derived networks. We achieve this result by regularization of the output and using a genetic weighting that selects intermediate models that do not exhibit divergence. The result is the derivation of multi-nested polynomial networks following the Kolmogorov-Gabor polynomial that are robust to mean estimators as well as novel exemplars for input. The full details of the algorithm are presented. We also introduce a new method for approximating GMDH in a single regression model using F, H, and G terms that automatically exports the answers as ordinary differential equations. The MathCAD 15 source code for all algorithms and results are provided.

Citation Download Citation

Holger M. Jaenisch and James W. Handley "A robust regularization algorithm for polynomial networks for machine learning", Proc. SPIE 8059, Evolutionary and Bio-Inspired Computation: Theory and Applications V, 80590A (19 May 2011); https://doi.org/10.1117/12.884284

ACCESS THE FULL ARTICLE