Using the Self-Organizing Map to Design Efficient RBF Models for Nonlinear Channel Equalization

In this paper we show how to build global and local RBF models once the SelfOrganizing Map has been trained using the Vector-Quantized Temporal Associative Memory (VQTAM) method. Through the VQTAM, prototype vectors (centroids) of input clusters are associated with prototype vectors of output clusters, so that the SOM can learn dynamic input-output mappings in a very simple and effective way. Global RBF models are built using all the input prototypes as centers of M gaussian basis functions, while the hidden-to-output layer weights are given by the output prototypes. Local RBF models are build in a similar fashion, but using only K M neurons. We evaluate the proposed RBF models and other global/local neural models in a complex nonlinear channel equalization task.