NUC 非均匀性矫校正 红外 图像
Stopping and Restarting Adaptive Updates to Recursive Least-Squares Lattice Adaptive Filtering Algorithms Dept. of Electrical& Computer Eng. Utah State University Logan, Utah 84322-4120 e-mail: jake@ece.usu.edu Phone: (435) 797-7229
Jake Gunther
Qualcomm 5775 Morehouse Drive AQ260J San Diego, CA 92121 e-mail: swang@qualcomm.com Phone: (858) 651-8493
Wang Song
Dept. of Electrical& Computer Eng. Utah State University Logan, Utah 84322-4120 e-mail: tbose@ece.usu.edu Phone: (435) 797-7227
Tamal Bose
Abstract- This paper reports several observations about stopping and restarting adaptive updates to recursive least-squares lattice (LSL) adaptive filtering algorithms. When updates are stopped, the adaptive filter becomes a fixed filter. Simulation examples demonstrate that large output error results from abruptly stopping or restarting adaptive updates. A remedy to the problem is to transition the adaptive updates to an off or on state gradually by driving the unknown system and the adaptive filter simultaneously to the all zero state. This is accomplished by setting the input signal to zero. The length (in number of samples) of the transition period is equal to the length of the adaptive filter. Simulation examples are given to illustrate the problem and the effectiveness of the proposed remedy. I.
INTRODUCTION
two systems.
An adaptive filter is a filter whose parameters, i.e. impulse response coefficients for direct form filter realizations or reflection coefficients for lattice filter realizations, are controlled to make the error e(i)= d(i) -y(i) between a desired signal d(i) and the filter output y(i) as small as possible in the mean-square or least-squares sense. The adaptive filter with impulse response w(n) shown in Figure 1 is configured to identify the unknown system with impulse response h(n). System identification is accomplished by minimizing the error e(i)= d(i) -y(i) between the outputs of the
d( i) x(n) e(n)
Fig. 1. Adaptive filter w(n) configured to identify the unknown system
h((n).
1-4244-0166-6/06/$20.00@2006 IEEE
adaptive filtering algorithm. One motivation for using the LSL algorithm instead of RLS is computational economy. The LSL algorithm has a computational complexity that is linear in the filter order M whereas the standard RLS algorithm has order M2 complexity[2]. We have found that the output error for the LSL algorithm jumps to very high levels when the adaptive updates are turned off or on abruptly. This phenomenon does not seem to have been well documented in the literature and is illustrated below in several examples. This paper reports one way to solve the problem is to insert a transition state between the updates-on and the updates-off states. While in the transition state, the input signal u(i) is be set to zero. The transition state may be exited after M iterations, where M is the length of the adaptive filter. This procedure allows LSL updates to be turned off and on while maintaining a low outp
ut error. The requirement of u(i)= 0 in the transition state may be allowed in some applications
In many applications of adaptive filtering, the impulse response h(i) of the unknown system changes with time. In this case, the adaptive filter tracks the time variation of h(i) after an initial learning transient. In other applications, h(i) may be timeinvariant and updates to the adaptive filter parameters may be halted after an initial learning period. If h(i) does not change, then y(i) continues to be a good estimate of d(i) while the updates are turned off. In some situations, u(i) changes at discrete points in time but is time-invariant between these points. In this case, the adaptive updates may be halted after learning and then restarted when a change is detected. This paper reports several observations concerning stopping and restarting the order recursive adaptive filtering algorithm known as the a priori errorfeedback lattice filter[1],[2],[3],[4],[5],[6],[7] abbreviated here as LSL. It is algebraically equivalent to the well known recursive least-squares (RLS)

Multichannel Recursive-Least-Squares Algorithms and Fast-Transversal-Filter Algorithms...[16], the gradient-adaptive-lattice LMS (GAL) [15] and the DCT-LMS...
Robust recursive least s... 7页 5财富值 Stopping and Restarting ... 6页...There have been a number of studies on adaptive lattice filters. Moro et....
An As-Short-As-Possible Introduction to the Least Squares, Weighted Least Squares and Movin In this introduction to the Least Squares (LS), Weighted Leas...
Robust recursive least squares learning algorithm for...s rule which just updates the synaptic weight ...Similar to the adaptive lattice filter structure [...
Robust recursive least squares learning algorithm for...s rule which just updates the synaptic weight ...Similar to the adaptive lattice filter structure [...
热门文档