Chapter 6 Adaptive Filters
Chapter 6 Adaptive Filters
Chapter 6 Adaptive Filters
Processing
Chapter 6: Adaptive Filters
6.1. Introduction
• The theory of optimum filters was developed
under the assumption that the filter designer
has complete knowledge of the statistical
properties of the SOE.
– Not applicable in real-world applications.
• Adaptive filters can improve their performance,
during normal operation, by learning the
statistical characteristics through processing
current signal observations.
6.2. Applications of Adaptive Filters
• Echo Cancelation
– Some energy on the incoming
branch leaks into the outgoing
branch and returns to the source
as an echo.
• Due to impedance mismatch.
– Echo path is unknown and may
be time changing.
– The main task of the canceller is
to estimate the echo signal with
sufficient accuracy.
• Equalization of Data Communication Channels
– Every pulse propagating through a channel suffers a certain
amount of time dispersion because the frequency response
of the channel deviates from the ideal one of constant
magnitude and linear phase.
– As a result, the tails of adjacent pulses interfere with the
measurement of the current pulse (intersymbol
interference).
– The channel is unknown and possibly time-varying.
– The goal of the equalizer is to restore the received pulse, as
closely as possible, to its original shape.
• Noise Cancelation
– The signal of interest s(n) is corrupted by uncorrelated
additive noise v1(n).
– A second sensor, located at a different point, acquires a
noise v2(n) (reference input) that is uncorrelated with the
signal s(n) but correlated with the noise v1(n).
6.3 Principles of Adaptive Filters
• Every adaptive filter consists of three modules
• Filtering Structure
– Forms the output of the filter using
measurements of the input signal or signals.
– The filtering structure is linear if the output is
obtained as a linear combination of the input
measurements; otherwise it is said to be
nonlinear.
– The structure is fixed by the designer, and its
parameters are adjusted by the adaptive
algorithm.
• Criterion of performance (COP).
– Assess the quality of the output of the adaptive
filter and the desired response (when available)
with respect to the requirements of the particular
application.
– The choice of COP is a compromise between what
is acceptable to the user and what is
mathematically tractable.
– The square error is the most used COP.
• Adaptation algorithm.
– Uses the value of the criterion of performance, or
some function of it, and the measurements of the
input and desired response (when available) to
decide how to modify the parameters of the filter
to improve its performance.
– The complexity and the characteristics of the
adaptive algorithm are functions of the filtering
structure and the criterion of performance.
• A priori information about the signal operating
environment (SOP) is used to choose
– Criterion of performance
– Filtering structure
• The design of an adaptive algorithm from the
criterion performance is the most difficult step
in the design and application of adaptive filters.
Modes of Operation
• Acquisition or convergence mode
– Initial period until it gets reasonably close to its
best performance.
• Tracking mode
– When the SOE change with time, the filter has to
follow the change in SOE.
Acquisition and tracking
• Learning curves of adaptive filters are represented by
plots of MSD, MSE or misadjustment as a function of n.
Optimum vs. Adaptive Filters
• Optimum filters are a theoretical tool and
cannot be used in practical applications
because we do not know the statistical
quantities (e.g., second-order moments) that
are required for their design.
• During normal operations, the filter works with
specific realizations of SOE
– Error formation
– Adaptive algorithm
• The increment is selected to bring C close to
Co with the passage of time.
• Therefore,
Only if
• Since
• Where
– μ is the step-size parameter
• Inserting the value of the gradient
• This iterative optimization can be combined
with filtering
Stability of SDA
• An algorithm is said to be stable if it converges to the
minimum regardless of the starting point.
• Rewriting the SDA algorithm by using principal-
component transformation, the necessary and
sufficient condition for the convergence of SDA is
Rate of convergence of SDA
• The rate (or speed) of convergence depends upon the
algorithm and the nature of the performance surface.
• The most influential effect is inflicted by the condition
number of the Hessian matrix that determines the
shape of the contours of P (c).
• For quadratic surface
• The rate of convergence can be characterized by
the time constant.
• If the time constant of ck,I
• Note that:
– SDA contains deterministic components while LMS
operates on random quantities.
– SDA is not adaptive algorithm (it only depends on
R and d). While LMS depends on the SOE.
• The LMS algorithm can be summarized as
Adaptation in a Stationary SOE
• In theory, the goal of the LMS adaptive filter is to
identify the optimum filter coefficients co.
• If R(n) is invertible
• If we define an adaptation gain vector g(n) as
– Finally estimate ym
• Limitations
– Solution to E{y|wm} requires inversion.
– The one step prediction requires infinite memory.
• If we assume linear data relation model
• then
• If we also assume the following linear signal model