Performance evaluation of a new variable tap-length learning algorithm for automatic structure adaptation in linear adaptive filters

In this work, a variable-tap length, variable step normalized least mean square algorithm with variable error spacing is proposed. The algorithm finds the optimized tap-length that best balances the complexity and steady state performance in linear adaptive filters. The design provides a systematic procedure with mathematical analysis to select the variable key parameters that affect the structure adaptation. The proposed structure adaptation algorithm maintains a trade-off between the mean square error and convergence speed. A sliding window weight update method is presented along with the tap-length learning algorithm to reduce the structural as well as computational complexity. Guidelines for parameter selection to formulate the optimum tap-length in correspondence with the designed algorithm are shown and assumptions are specified. The proposed algorithm has performed better than the existing fractional tap-length learning methods for both low and high noise conditions. This is achieved because of the unique method adopted in this paper to set dynamic system independent parameters instead of predefined fixed settings.