You are on page 1of 2

Second Problem Assignment

EE608 - Adaptive Signal Processing


Assigned on: January 15, 2015

Due on: January 22, 2015

Notes:
(1) The computer assignments can be solved using Scilab, GNU Octave, SciPy or any other
free/open source software kit. Solutions that work only on Matlab will not be accepted.
(2) Copying will be dealt with strictly. Institute disciplinary procedures will be invoked if any
form of cheating is detected.
Problem 1
(10 points) This is an example similar to the noise cancellation example we looked at in class. Let
x[n] = d[n] + g[n]
where d[n] = sin(n0 +), 0 = /20 and is a random variable that is distributed uniformly
between and . Assume that g[n] is unit variance white noise. Suppose that a noise process
v2 [n] that is correclated with g[n] is measured by a secondary sensor. The noise v2 [n] is related
to g[n] by a ltering operation:
v2 [n] = 0.8v2 [n 1] + g[n]
(a) Generate N = 500 samples of the processes x[n] and v2 [n].
(b) Derive the Wiener-Hopf equations that dene the optimum p-th order FIR lter for estimating
g[n] from v2 [n].
(c) Using lters of order p = 2, 4 and 6, design and implement the Wiener noise cancellation
[n] and compare the average squared errors for each lter.
lters. Plot the estimated process g
You can use the sample averaged autocorrelation and cross-correlations from the samples
generated in (a), i.e. use
1 N1
rv2 = v2 [n]v2 [n k]
N n=0
rxv2

Assigned: January 15, 2015

1 N1
= x[n]v2 [n k]
N n=0

Due: January 22, 2015

Second Problem Assignment


(d) In some situations, the desired signal may leak into the secondary sensor. In this case, the
performance of the Wiener lter may be severely compromised. To see what eect this has,
suppose the input to the Wiener lter is
v0 [n] = v2 [n] + d[n]
where v2 [n] is the ltered noise dened above. Evaluate the performance of the Wiener noise
canceller for several dierent values of for lter orders of p = 2, 4 and 6. Comment on your
observations.
(e) Analytically derive the required rv2 [k] and rgv2 [k] values and observe whether their statistically obtained average values match the analytical ones. You can assume that the process is
two-sided and stationary at the instances of interest (assume that n is a large positive number
and k < n, so that the process looks stationary. Recall homework 1 problem 3(b)).
You dont need to characterize rv2 [k] and rgv2 [k] for all k. Just provide the values that are
required to obtain the p-th order Wiener lter for (c).
Problem 2
(10 points) In this exercise, we see how gradient descent is aected by the step size .
We want to design an order 1 (p = 1) Wiener lter to obtain the desired signal d[n] from x[n],
with both the processes jointly wide-sense stationary. The following details are given:
rd [0] = 6
rx [0] = 6, rx [1] = 3
rdx [0] = 3, rdx [1] = 2

(a) Assume that the lter (not necessarily the optimal lter) used to lter x[n] to obtain d[n]
is
T
T
= [w[0] w[1]] = [x y] . Plot the error surface (x, y) as a 3D plot.
(b) Implement gradient descent to nd the optimal for various step sizes. That is, implement
the following iteration:
n+1 = n [n]
where [n] is the gradient of evaluated at the current n . Set your initial guess 0 to
[0 0]T , and run the iterations for = 0.001, = 0.01 and = 0.1. For each value, plot
n [0] versus the iteration counter n, and the error [n] versus n. What do you observe for
various values, and what is your conclusion?

Assigned: January 15, 2015

Due: January 22, 2015

You might also like