Ch6-Normalized Least Mean-Square Adaptive Filtering

Similar documents
Ch4: Method of Steepest Descent

Adaptive Filters. un [ ] yn [ ] w. yn n wun k. - Adaptive filter (FIR): yn n n w nun k. (1) Identification. Unknown System + (2) Inverse modeling

Ch5: Least Mean-Square Adaptive Filtering

CHAPTER 4 ADAPTIVE FILTERS: LMS, NLMS AND RLS. 4.1 Adaptive Filter

Linear Optimum Filtering: Statement

Convergence Evaluation of a Random Step-Size NLMS Adaptive Algorithm in System Identification and Channel Equalization

2.6 The optimum filtering solution is defined by the Wiener-Hopf equation

Adaptive Filtering Part II

Adaptive Filter Theory

A Derivation of the Steady-State MSE of RLS: Stationary and Nonstationary Cases

CONTENTS. Preface Preliminaries 1

Revision of Lecture 4

Linear Models for Regression

EE482: Digital Signal Processing Applications

2262 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 47, NO. 8, AUGUST A General Class of Nonlinear Normalized Adaptive Filtering Algorithms

V. Adaptive filtering Widrow-Hopf Learning Rule LMS and Adaline

Lecture 6: Block Adaptive Filters and Frequency Domain Adaptive Filters

Lecture Notes in Adaptive Filters

4. Multilayer Perceptrons

Normalized Minimum Error Entropy Algorithm with Recursive Power Estimation

Performance Comparison of Two Implementations of the Leaky. LMS Adaptive Filter. Scott C. Douglas. University of Utah. Salt Lake City, Utah 84112

Sparse Least Mean Square Algorithm for Estimation of Truncated Volterra Kernels

ESANN'2003 proceedings - European Symposium on Artificial Neural Networks Bruges (Belgium), April 2003, d-side publi., ISBN X, pp.

EE482: Digital Signal Processing Applications

A METHOD OF ADAPTATION BETWEEN STEEPEST- DESCENT AND NEWTON S ALGORITHM FOR MULTI- CHANNEL ACTIVE CONTROL OF TONAL NOISE AND VIBRATION

Acoustic MIMO Signal Processing

Adaptive Filtering. Squares. Alexander D. Poularikas. Fundamentals of. Least Mean. with MATLABR. University of Alabama, Huntsville, AL.

26. Filtering. ECE 830, Spring 2014

LMS and eigenvalue spread 2. Lecture 3 1. LMS and eigenvalue spread 3. LMS and eigenvalue spread 4. χ(r) = λ max λ min. » 1 a. » b0 +b. b 0 a+b 1.

RADIO SYSTEMS ETIN15. Lecture no: Equalization. Ove Edfors, Department of Electrical and Information Technology

LMS Algorithm Summary

On-line Support Vector Machine Regression

ELEC E7210: Communication Theory. Lecture 4: Equalization

Reduced-cost combination of adaptive filters for acoustic echo cancellation

A low intricacy variable step-size partial update adaptive algorithm for Acoustic Echo Cancellation USNRao

Numerical optimization

EFFECTS OF ILL-CONDITIONED DATA ON LEAST SQUARES ADAPTIVE FILTERS. Gary A. Ybarra and S.T. Alexander

LEAST-SQUARES parameter estimation techniques have. Underdetermined-Order Recursive Least-Squares Adaptive Filtering: The Concept and Algorithms

Lecture 3: Linear FIR Adaptive Filtering Gradient based adaptation: Steepest Descent Method

Stochastic Subgradient Method

Sample ECE275A Midterm Exam Questions

Stable Adaptive Momentum for Rapid Online Learning in Nonlinear Systems

NSLMS: a Proportional Weight Algorithm for Sparse Adaptive Filters

Image restoration: numerical optimisation

15th European Signal Processing Conference (EUSIPCO 2007), Poznan, Poland, September 3-7, 2007, copyright by EURASIP

Numerical optimization. Numerical optimization. Longest Shortest where Maximal Minimal. Fastest. Largest. Optimization problems

TRACKING and DETECTION in COMPUTER VISION

ADAPTIVE FILTER THEORY

Sliding Window Recursive Quadratic Optimization with Variable Regularization

ECE4270 Fundamentals of DSP Lecture 20. Fixed-Point Arithmetic in FIR and IIR Filters (part I) Overview of Lecture. Overflow. FIR Digital Filter

Lagrange Relaxation and Duality

Variable Learning Rate LMS Based Linear Adaptive Inverse Control *

Constrained Optimization and Lagrangian Duality

SGN Advanced Signal Processing Project bonus: Sparse model estimation

State-Space Methods for Inferring Spike Trains from Calcium Imaging

Parametric Signal Modeling and Linear Prediction Theory 4. The Levinson-Durbin Recursion

ELEG-636: Statistical Signal Processing

Analysis of incremental RLS adaptive networks with noisy links

Data Fusion of Dual Foot-Mounted Zero Velocity Update (ZUPT) Aided Inertial Navigation Systems (INSs) using Centroid Method

Q-Learning and Stochastic Approximation

IS NEGATIVE STEP SIZE LMS ALGORITHM STABLE OPERATION POSSIBLE?

Condensed Table of Contents for Introduction to Stochastic Search and Optimization: Estimation, Simulation, and Control by J. C.

min f(x). (2.1) Objectives consisting of a smooth convex term plus a nonconvex regularization term;

Convex Optimization. Newton s method. ENSAE: Optimisation 1/44

Linear Models for Regression

Linear Regression. CSL603 - Fall 2017 Narayanan C Krishnan

Linear Regression. CSL465/603 - Fall 2016 Narayanan C Krishnan

Chapter 2 Fundamentals of Adaptive Filter Theory

NONLINEAR PLANT IDENTIFICATION BY WAVELETS

The Derivative of a Function Measuring Rates of Change of a function. Secant line. f(x) f(x 0 ) Average rate of change of with respect to over,

Numerical Optimization Professor Horst Cerjak, Horst Bischof, Thomas Pock Mat Vis-Gra SS09

EE 367 / CS 448I Computational Imaging and Display Notes: Image Deconvolution (lecture 6)

Logarithmic Regret Algorithms for Strongly Convex Repeated Games

Lecture 1: January 12

SGN Advanced Signal Processing: Lecture 4 Gradient based adaptation: Steepest Descent Method

Design of Norm-Optimal Iterative Learning Controllers: The Effect of an Iteration-Domain Kalman Filter for Disturbance Estimation

Mobile Robotics 1. A Compact Course on Linear Algebra. Giorgio Grisetti

Gradient Descent. Dr. Xiaowei Huang

Principles of forecasting

Adap>ve Filters Part 2 (LMS variants and analysis) ECE 5/639 Sta>s>cal Signal Processing II: Linear Es>ma>on

SNR lidar signal improovement by adaptive tecniques

CONSTRAINED OPTIMIZATION OVER DISCRETE SETS VIA SPSA WITH APPLICATION TO NON-SEPARABLE RESOURCE ALLOCATION

ADAPTIVE FILTER THEORY

Department of Electrical and Electronic Engineering

Statistical Machine Learning from Data

Ch 5.7: Series Solutions Near a Regular Singular Point, Part II

Optimal Control Theory

ENGR352 Problem Set 02

Performance Analysis and Enhancements of Adaptive Algorithms and Their Applications

Statistical and Adaptive Signal Processing

Least Squares SVM Regression

Support Vector Machines. CSE 6363 Machine Learning Vassilis Athitsos Computer Science and Engineering Department University of Texas at Arlington

Chapter 1.6. Perform Operations with Complex Numbers

A DISSERTATION SUBMITTED TO THE FACULTY OF THE GRADUATE SCHOOL OF THE UNIVERSITY OF MINNESOTA BY. Jie Yang

SOS Boosting of Image Denoising Algorithms

An Adaptive Sensor Array Using an Affine Combination of Two Filters

AdaptiveFilters. GJRE-F Classification : FOR Code:

A Convex Combination of Two Adaptive Filters as Applied to Economic Time Series. Adaptive Signal Processing EEL May Casey T.

Optimal and Adaptive Filtering

Instructor: Dr. Benjamin Thompson Lecture 8: 3 February 2009

Transcription:

Ch6-Normalized Least Mean-Square Adaptive Filtering LMS Filtering The update equation for the LMS algorithm is wˆ wˆ u ( n 1) ( n) ( n) e ( n) Step size Filter input which is derived from SD as an approximation Error signal Step size where the step size is originally considered for a deterministic gradient. LMS suffers from gradient noise due to its random nature. Above update is problematic due to this noise Gradient noise amplification when u(n) is large. 1

Normalized LMS u(n) is random instantaneous samples can assume any value for the norm u(n) which can be very large. Solution: input samples can be forced to have constant norm Normalization wˆ wˆ u u( n) ( n 1) ( n) ( n) e ( n) Update equation for the normalized LMS algorithm. Note the similarity bw. NLMS and LMS update eqn.s NLMS can be considered same as LMS except time-varying step size. ( n) u ( n)

Normalized LMS Block diagram very similar to that of LMS The difference is in the Weight-Control Mechanism block. 3

Normalized LMS We have seen that LMS algorithm optimizes the criterion instead of MSE. Similarly, NLMS optimises another problem: From one iteration to the next, the weight vector of an adaptive filter should be changed in a minimal manner, subject to a constraint imposed on the updated filter s output. Mathematically, minimize the squared Euclidean norm of the change, Subject to the constraint wˆ ( n 1) wˆ ( n 1) wˆ ( n) wˆ ( n 1) u( n ) d ( n ) which can be optimized by the method Lagrange multipliers J( n) wˆ( n 1) Re d( n) ˆ w ( n 1) u( n) 4

ˆ ˆ ˆ ˆ w w w w wˆ u J( n) ( n 1) ( n) ( n 1) ( n) Re d( n) ( n 1) ( n) Jn ( ) w ˆ ( n 1) ˆ ˆ w( n 1) w( n) u( n) Proof detail on slides later: 1 Set equal to zero wˆ( n 1) wˆ( n) u( n). 1 d( n) wˆ ( n 1) u( n) wˆ( n) u( n) u( n) 1 1 wˆ ( n) u( n) u ( n) u( n) wˆ ( n) u( n) u( n). en ( ), u ( n) 1 1 wˆ ( n 1) wˆ ( n 1) wˆ ( n) u( n) u( n) e ( n) u( n) 5

In order to exercise control over the change in the tap-weight vector from one iteration to the next without changing the direction of the vector, we introduce a positive real scaling factor denoted by wˆ wˆ wˆ u u( n) ( n 1) ( n 1) ( n) ( n) e ( n) wˆ wˆ u u( n) ( n 1) ( n) ( n) e ( n) The product vector u(n)e(n) is normalized with respect to the squared Euclidean norm of the tap-input vector u(n). is dimensionless, while dimension of μ is inverse of power. We may view the normalized LMS filter as an LMS filter with a time-varying step-size parameter. 6

Proof detail: k=0, 1,,, M-1 7

8

k=0, 1,,, M-1, λ=λ 1 +jλ we multiply both sides of Eq. by u (n - k) and then sum over all possible integer values of k for 0 to M - 1. We thus get 9

k=0, 1,,, M-1 10

Normalized LMS wˆ ( n 1) 1 wˆ wˆ u 1. Take the first derivative of J(n) wrt and set to zero to find ( n 1) ( n) ( n).. Substitute this result into the constraint to solve for the multiplier en ( ), u ( n) 3. Combining these results and adding a step-size parameter to control the progress gives wˆ wˆ wˆ u u( n) ( n 1) ( n 1) ( n) ( n) e ( n) 4. ence the update eqn. for NLMS becomes wˆ wˆ u u( n) ( n 1) ( n) ( n) e ( n) 11

Normalized LMS Observations: We may view an NLMS filter as an LMS filter with a time-varying step-size parameter ( n) u ( n) Rate of convergence of NLMS is faster than LMS u(n) can be very large, however, likewise it can also be very small Causes problem since it appears in the denominator Solution: include a small correction term to avoid stability problems. اثبات با استفاده از روش نيوتن Ch 4 از کتاب سعيد مطالعه شود. 1

Stability of NLMS What should be the value of step size for convergence? Assume that the desired response is governed by d( n) w u( n) ( n) Substituting the weight-error vector ε( n) w wˆ ( n) Additive disturbance into the NLMS update equation we get wˆ( n 1) wˆ( n) u( n) e ( n) u( n) ε( n 1) ε( n) u( n) e ( n) u( n) which provides the update for the mean-square deviation D( n) E ε( n) Where ξ u (n) is called undisturbed error signal ( ) w wˆ ( ) u( ) ε ( ) u( ) n n n n n u d( n) v( n) y( n) e( n) v( n) d( n) y( n) v( n) 13

Stability of NLMS Find the range for so that Right hand side is a quadratic function of, is satisfied when Differentiate wrt and equate to 0 to find opt This step size yields maximum drop in the MSD! For clarity of notation assume real-valued signals 14

Stability of NLMS Assumption I: The fluctuations in the input signal energy u(n) from one iteration to the next are small enough so that Then Assumption II: Undisturbed error signal u (n) is uncorrelated with the disturbance noise (n) Then e(n): observable, u (n): unobservable 15

Stability of NLMS Assumption III: The spectral content of the input signal u(n) is essentially flat over a frequency band larger than that occupied by each element of the weight-error vector (n), hence Then T ( ) u ε ( ) u( ) E n E n n E ε( n) E u ( n) D( n) E u ( n) 16

Normalized LMS 17

Affine Projection Adaptive Filters Mathematically,minimize the squared Euclidean norm of the change, wˆ ( n 1) wˆ ( n 1) wˆ ( n) subject to the set of N constraints wˆ ( n 1) u( n k) d( n k) for k 0, 1,..., N 1 (6.36) where N is smaller than the dimensionality M of the input data space or, equivalently, the weight space. This constrained optimization criterion includes that of the normalized LMS filter as a special case namely, N = 1. We may view N, the number of constraints, as the order of the affine projection adaptive filter. 18

Following the method of Lagrange multipliers with multiple constraints definitions: N 1 k k0 J( n) wˆ ( n 1) wˆ ( n) Re d( n k) wˆ ( n 1) u( n k). An N-by-M data matrix A(n) An N-by-1 desired response vector An N-by-1 Lagrange vector Compact form of cost function A ( n) u( n), u( n 1),, u( n N 1) d ( n) d( n), d( n 1),, d( n N 1) λ ( n),,, 0 1 N 1 ( ) ˆ ( 1) ˆ ( ) Re ( ) ( ) ˆ ( 1) J n w n w n n n n. d A w λ

The derivative of the cost function is: Jn ( ) w ˆ ( n 1) Set equal zero; Rewrite equation (6.36) in new form Then we have wˆ( n 1) wˆ( n) A ( n) λ. 1 wˆ ( n1) A ( n) λ. d( n) Awˆ ( n1) 1 A( n) wˆ ( n 1) A( n) wˆ ( n 1) wˆ ( n) A( n) A ( n) λ. 1 A( n) wˆ( n 1) A( n) wˆ( n) A( n) A ( n) λ 1 d( n) A( n) wˆ ( n) A( n) A ( n) λ 0

The difference between d( n) and A( n) w( n) based on data available at iteration N is N-by-1 error vector Solving for λ e( n) d( n) A( n) wˆ ( n) 1 λ A( n) A ( n) e( n). Finally, we need to exercise control over the change in the weight vector from one iteration to the next, but keep the same direction. ˆ 1 wˆ ( n 1) A ( n) A( n) A ( n) e( n). 1 wˆ ( n 1) A ( n) A( n) A ( n) e( n). 1 wˆ( n 1) wˆ( n) A ( n) A( n) A ( n) e( n). which is the desired update equation for the affine projection adaptive filter 1

Affine Projection Operator Substituting e(n) in the above eq. 1 wˆ ( n 1) wˆ ( n) A ( n) A( n) A ( n) d( n) A( n) wˆ ( n). 1 1 ( n) ( n) ( n) ( n) ˆ I A A A A w( n) A ( n) A( n) A ( n) d( n) Define the projection operator: 1 P A ( n) A( n) A ( n) A( n) The complement projector I P acts on the old weight vector wˆ ( n) to produce the updated weight vector wˆ ( n 1) Defining pseudo-inverse of the data matrix A ( n) A ( n) A( n) A ( n) 1 w ˆ( n 1) I A ( n) A ( n) w ˆ( n) A ( n) d ( n)

Summary of the Affine Projection Adaptive Filter We may view the affine projection filter as an intermediate adaptive filter between the normalized LMS filter and the recursive least-squares (RLS) filter, in terms of both computational complexity and performance. 3

Stability Analysis of the Affine Projection AF Rewrite 1 ε( n 1) ε( n) A ( n) A( n) A ( n) e( n). where 4

Observations on the Convergence Behavior of Affine Projection Adaptive FiIters 1. The learning curve of an affine projection adaptive filter consists of the sum of exponential terms.. An affine projection adaptive filter converges at a rate faster than that of the corresponding normalized LMS filter. 3. As more delayed versions of the tap-input vector u(n) are used (i.e., the filter order N is increased), the rate of convergence improves, but the rate at which improvement is attained decreases. Practical Considerations: Regularization to take care of noisy data Fast implementation to improve computational efficiency I 1 wˆ( n 1) wˆ( n) A ( n) A( n) A ( n) e( n). W6 ; Ch6: 1, 3, 6, 7 5