Combinação afim de algoritmos adaptativos (Affine combination of adaptive algorithms). Escola Politécnica, University of São Paulo, São Paulo, Brazil, 2009 (available only in portuguese).
In order to improve the performance of adaptive filters, the combination of algorithms is receiving much attention in the literature. This method combines linearly the outputs of two filters operating in parallel with different step-sizes to obtain an adaptive filter with fast convergence and reduced excess mean squared error (EMSE). In this context, it was proposed an affine combination of two least-mean square (LMS) filters, whose mixing parameter is not restricted to the interval [0, 1]. Hence, the affine combination is a generalization of the convex combination. In this work, the affine combination of two LMS algorithms is extended to the supervised algorithms NLMS (normalized LMS) and RLS (recursive least squares), and also to blind equalization, using the constant modulus algorithm (CMA). A steady-state analysis of the affine combination of the considered algorithms is presented in a unified manner, assuming white or colored inputs, and stationary or nonstationary environments. Through the analysis, it was observed that the affine combination of two algorithms of the same family can provide a 3 dB EMSE gain in relation to its best component filter and consequently in relation to the convex combination. To ensure that the combined estimate is at least as good as the best of the component filters, three new algorithms to adapt the mixing parameter were proposed and analyzed. Using the analysis results of these algorithms in conjunction with the results of the transient analysis of adaptive filters, the transient behavior of the affine combination was analyzed. Through simulations, a good agreement between analytical and experimental results was always observed. In the blind equalization case, a combination of two CMA equalizers with different initializations was also proposed. The simulation results suggest that the affine combination can avoid local minima of the constant modulus cost function.
Adaptive filters, Affine combination, Steady-state analysis, Transient analysis.