You are on page 1of 2

INSTITUTO MILITAR DE ENGENHARIA (IME), 19 DE AGOSTO DE 2015, RIO DE JANEIRO, RJ

Design of Experiments in Adaptive Filtering
Rigel P. Fernandes

Abstract— This report presents an overview in the Design
of Experiments applied to adaptive filtering. The sources of
this work were classes annotation, text-book, and literature
review. This work also presents a study of the influence of
algorithm parameter L on convergence speed, misadjustment,
and computational effort.

0.1
0.08
0.06
0.04
0.02
0

I. I NTRODUCTION

II. E XAMPLE
According to Seneca (5 BC - 65 AD), the Roman philosopher: “The path of precept is long, that of example short
and effectual”. As a result, the examples that follows are used
to help make the concepts of DOE presented in this study
easier to understand.
Figure 1 presents the input of an Adaptive Filter implemented using IQRD–RLS. This sequence is White Noise with zero
mean and σ 2 = 10−3
Figure 2 presents the output of the system, learning curves
of the IQRD–RLS, that is a way to assess the algorithm implemented. The performance of the algorithm in this controlled
experiment is affected by parameter L, λ, and the filter order
N.

−0.02
−0.04
−0.06
−0.08
−0.1

Fig. 1.

0

100

200

300

400

500

White noise with zero mean and σ 2 = 10−3 .
Learning curve (a priori error) of the Inverse QRD−RLS

−10
L=1
L=2
L=3
L=4
L=5

−20

−30
E[|e(k)|2]dB

Design of Experiments (DOE) is often used for planning
and conducting experiments, as well as to analyze the results
obtained. The main purpose of this tool is to reduce time
and costs related to experiment analysis, it also allows users
to evaluate and objectively make conclusions about results
obtained. This method is found in literature as a tool to validate
results and is supported by the experience of researchers [1].
DOE supports researchers for a better identification of the
inputs, the factors involved (assesses if these are controlled
or not), and the outputs.
It is likely that extensive experiments will allow users to find
optimal solution for a problem. However, in real world it is
likely that experimenters face problems due to time or budget
constraints. These constraints may not allow extensive tests to
be done. Therefore, an statistical approach offers results that is
very close to optimal solution with less effort (or resources).
For achieve this, the tool provides guidelines for experimenters use their experience for assessing factors that may
influence experiments. In fact, DOE is strongly influenced by
the state of knowledge in the field [2].
This method allows users to analyze interaction between
factors. This analysis may provide information about what
values for the factors that will impact in the output. The
simplest way to analyze factor is the one-factor-at-a-time.
This strategy consists in vary one factor with the others held
constant. However, DOE has many other tools to plan and
conduct experiments, e.g. factorial design, fractional factorial
designs [1].

−40

−50

−60

−70

0

100

200

300

400

500

k

Fig. 2.

MSE with L varying from 1 to 5.

It should be noted that as L increases, the convergence speed
decreases. For instance, L = 1 converges before 20th k time
instant. On the other hand, L = 5 converges close to the 500th
k time instant. Misadjustment after the convergence seems to
be the same for every value of L. Parameter L indicates the
number of samples for the next update, i.e. if L = 1, w is
updated every sample, if L = 2, w is updated every 2 samples.
This approach presents a compromise between convergence
speed and computational effort1 . This means that one may
trade off convergence speed for less mathematical operations during the execution of this IQRD–RLS, i.e. one could
adjust parameter L in order to save DSP devices energy
1 Computational effort in this work is the number of mathematical operations
that one algorithm calculates. If one can avoid one or more operations the
effort is smaller. On the other hand, computational complexity is the maximum
number of mathematical operations that one algorithm demands to run

2004. The parameter L in this algorithm does not trade off convergence speed and misadjustment. . the misadjustment of L = 2. “Design of experiments. i. [2] J. the less amount of energy spent the better. methods. vol. the classic IQRD–RLS. Diniz. “Affine projection and recursive least squares adaptive filters employing partial updates. 950–954. pp. Springer. Conference Record of the Thirty-Eighth Asilomar Conference on. 3. RIO DE JANEIRO.” Journal of the Franklin Institute. 19 DE AGOSTO DE 2015. [3] P. and resources in academic research. Khong et al. However. where updates is an integer number. A. The number of updates is updates = Lk . Jacquez. R EFERENCES [1] D.. and concepts that may allow engineers to perform experiments accurately spending less time. 2. S.INSTITUTO MILITAR DE ENGENHARIA (IME). 2000.” in Signals. 5 are very similar to L = 1. Systems and Computers. C. but in SMF coefficients are updated only when the output estimation error is higher than the pre-determined upper bound [3]. Depending on the system. RJ running IQRD–RLS. 2004. Naylor. vol. Montgomery. Naylor. It should be noted that less time is spent during the execution of this approach regarding conventional IQRD–RLS because less updates are performed when L > 1. 4 and. IQRD–RLS with parameter L is a good approach to systems that pay more attention at saving energy than convergence rate. 1997. Adaptive filtering. 335. investigates how to minimize computational complexity by employing partial updates of w and guarantee good convergence speed. IEEE.e. W. John Wiley & Sons. 1998. 1. Design and analysis of experiments. no. Set–membership filtering (SMF) are similar to this simple approach. [4] P. pp. III. computational complexity and misadjustment remain the same. A. The trade off in this paper regards computational complexity and convergence rate [4]. 259–279. C ONCLUSION DOE offers a variety of tools. Therefore. After the convergence. This method validates conclusions of researches by employing statistical concepts to the works being assessed.