Proposição e avaliação de algoritmos de filtragem adaptativa baseados na rede de kohonen

Abstract

The Self-Organizing Network Kohonen (Self-Organizing Map - SOM), by employing an unsupervised learning algorithm, has been traditionally implemented in signal processing area in quantization tasks vector, while MLP (Multi-Layer Perceptron ) and RBF (Radial Basis Function) dominate applications that require the approach of input-output mappings. This type of application is commonly found in adaptive filtering tasks that can be formatted from the perspective of direct and inverse modeling systems such as identification equalization of communication channels. In this dissertation, the range of SOM network applications is extended by proposing neural adaptive filter based on this network, showing that they are viable alternatives to non-linear filters based on MLP and RBF networks. This becomes possible through the use of a newly proposed technique, Quantized Temporal Associative Memory - VQTAM), which basically uses the philosophy called Memory Associative Temporal by Quantization Vector (Vector) network training SOM to perform simultaneous vector quantization of spaces input and output relating to the filtering problem analyzed. From the VQTAM technique are proposed three architectures adaptive filters based on SOM, whose performances were evaluated in identifying tasks and equalization of nonlinear channels. The channel used in the simulations was modeled as an autoregressive process of Gauss-Markov first order, contaminated with Gaussian white noise and provided with nonlinearity of the type saturation (sigmoidal). The results show that adaptive filters based on SOM network have equivalent or superior performance to traditional linear transversal filters and non-linear filters based on MLP.