Neural Networks for Complex Valued Signals: A Preliminary Study

Abstract

This article presents the work related to the design and architecture of a special neural network capable of dealing effectively with Complex numbers. The proposed architecture employs parameter space partitioning and a novel partition mapping scheme. An empirical design based partially on the concepts of Rough Sets has been described. The applied signal in the form of Complex numbers is divided into a set (containing both the imaginary and real coefficients) and, a subset (containing of only the real coefficient). These set-subsets are processed by specialized neurons. The proposed architecture displays superior learning speeds and similar accuracy when compared to other established complex-valued-neural-networks.