Uyarlanır yayınım LMS stratejileri

In this thesis new distributed adaptive algorithms for the in-networking parameter estimation problem are proposed. They are cooperative and resistant to link failures. The individual nodes run local least-mean squares (LMS) algorithm to estimate the common parameter of interest and then share these estimates with nodes in vicinity. Neighbor nodes use these data to update their own estimates by combining received estimates and processing the resulting aggregate estimates in the local adaptive LMS filters. This strategy is known as the diffusion LMS algorithm.In the first chapter of the thesis stability and convergence of the diffusion LMS algorithm is introduced. Theoretical statement of the evolution of the mean-square deviation (MSD) and excess mean-square error (EMSE) are given in short as stated in literature. Simulations show perfect match between experimental and theoretical evolution of these error measures for diffusion algorithm. Also experiments show that this algorithm has a faster convergence and better performance (tens of dB difference in MSD and EMSE) compared to noncooperative LMS.The second chapter of the thesis contains main contributions of the research. In the diffusion LMS algorithm, aggregation step comprises of combining neighbor estimates by weighing them with constant coefficients. Contrary to this approach, in proposed adaptive diffusion algorithms another adaptation layer is introduced to update these weighing coefficients at every iteration. The weights are constrained to produce a) convex, b) affine combination or c) may not have any constraints. For adaptation purpose gradient-descent algorithm is used. Simulations show that in some cases adaptive diffusion LMS algorithms have faster convergence than classical diffusion algorithm with penalty in larger MSD and EMSE values in the steady-state.