Novel algorithms for nonlinear channel equalization using neural vector quantization

Abstract

In this paper we use the Self-Organizing Map (SOM), a well-known neural vector quantization algorithm, to design nonlinear adaptive filters through the Vector-Quantized Temporal Associative Memory (VQTAM) method. In VQTAM, the centroids (codebook vectors) of input clusters found by the SOM are associated with codebook vectors of output clusters, so that the SOM can learn dynamic input-output mappings in a very simple and effective way. In addition, we also propose two VQTAM-based Radial Basis Function (RBF) adaptive filters. Firstly, a global RBF model is built using all the input codebook vectors as centers of M gaussian basis functions, while the hidden-to-output layer weights are given by the output prototypes. Then, a local RBF model is built in a similar fashion, but using only K << M neurons. We evaluate the proposed VQTAM-based adaptive filters in a nonlinear channel equalization task. Performance comparisons with the standard linear FIR/LMS and the nonlinear Multilayer Perceptron (MLP) equalizers are also carried out.