Stochastic approximation for learning rate optimization for generalized relevance learning vector quantization

Abstract

Herein the authors apply the stochastic approximation method of Kiefer and Wolfowitz to optimize learning rate selection for Generalized Relevance Learning Vector Quantization — Improved (GRLVQI) neural networks with application to Z-Wave cyber-physical device identification. Recent work on full factorial models for GRLVQI optimal settings has shown promise, but is computationally costly and not feasible for large datasets. Results using stochastic optimization illustrate show fast convergence to high classification rates.