System identification techniques for adaptive recursive filtering.

Many problems in adaptive filtering can be approached from the point of view of system identification. For an adaptive filter with N adjustable coefficients or weights, the mean-squared value of the output error is a plot, in N + 1 dimensions, of the mean-squared error versus the N coefficient values. If the adaptive filter is non-recursive, the mean-squared value of the output error is a quadratic function of the coefficients. With recursive adaptive filters, the output error is not quadratic and may even have local minima. This paper introduces a new class of adaptive recursive algorithm, which offer a much reduced computational load for basically the same performance, and has a non-vanishing gain which enables the algorithm to remain active and to track time-varying systems. A simplified version of the algorithm, called the steepest descent recursive algorithm, which retains provable convergence at low convergence rates, is well suited to system identification. [ABSTRACT FROM AUTHOR]