HYBRID STOCHASTIC GRADIENT FOR CONVERGENCE OF ADAPTIVE FILTERS

摘要

ABSTRACTThe present invention provides a technique forupdating the coefficients a(n) of an adaptive filterimpulse response A(n), whereA'(n) = (a0(n) a1(n) . . . am(n))is the adaptive filter impulse response at time n. Theimpulse response is utilized for tracking the outputy(n) of a communications channel, wherey(n) = where H is an unknown impulse response of the channel,H' = (h0 h1 . . . hm),the ' character denoting transposition. The channel isdriven with a known input signal S(n), whereS'(n) = (s(n) s(n-1) . . . s(n-m)).According to the technique, the output x(n) of anadaptive filter, wherex(n) = S'(n) A(n)is compared with the channel output y(n) according tothe Least Mean Squares (LMS) algorithm to develop an LMSerror prediction signal. At the same time, the adaptivefilter output x(n) is compared with the channel outputy(n) according to the Sign Algorithm (SA) to develop anSA error prediction signal. If the LMS error signal isgreater than or equal to a preselected value, then it isused to update the coefficients a (n) of the adaptivefilter impulse response. If the LMS error signal isless than the preselected value, then the SA errorsignal is used to update the coefficients a(n).