Mauricio Martinez

Efficient Filter Bank Adaptive Filtering

Friday, Apr. 26, 2002

The best known adaptive filtering algorithms are the Least-Mean-Squares (LMS) and the Recursive-Least-Squares (RLS). Several attempts have been made in the past two decades to combine the superior convergence behavior of RLS with the low implementation cost of LMS. The most successful existing approach, known as transform-domain adaptive filtering, uses a signal-independent preprocessing transform (typically FFT) to decorrelate the elements of the signal vector that serves as input to the LMS algorithm. We extend this approach by replacing the fixed transform by a subband splitting filter bank, resulting in a structure we call Filter-Bank LMS (FB-LMS). The added degrees of freedom are used to shorten convergence time, and to reduce the sensitivity of LMS to input signal statistics, with only a modest increase in the overall complexity of the adaptive filter. We present a performance-optimized version of FB-LMS, and compare it in terms of performance and cost with: (i) transform-domain LMS (using FFT), and (ii) a sub-optimal FB-LMS based on classical frequency selective filters, such as Butterworth, Chebyshev, and equiripple FIR (Parks-McClellan). Our examples serve to demonstrate that both our optimal and sub-optimal versions of FB-LMS significantly outperform FFT-LMS, the best available transform-domain alternative.

Committee Members:
D. Brooks
H. Lev-Ari (Thesis advisor)
B. Shafai