Professional Documents
Culture Documents
Block-Based Fuzzy Step Size NLMS Algorithms For Subband Adaptive Channel Equalisation
Block-Based Fuzzy Step Size NLMS Algorithms For Subband Adaptive Channel Equalisation
Block-Based Fuzzy Step Size NLMS Algorithms For Subband Adaptive Channel Equalisation
org
ISSN 1751-9675
Abstract: Judicious selection of the step size parameter is crucial for adaptive algorithms to strike a good balance
between convergence speed and misadjustment. The fuzzy step size (FSS) technique has been shown to improve
the performance of the classical xed step size and variable step size (VSS) normalised least mean square (NLMS)
algorithms. The performance of the FSS technique in the context of subband adaptive equalisation is analysed and
two novel block-based fuzzy step size (BFSS) strategies for the NLMS algorithm, namely xed block fuzzy step size
(FBFSS) and adaptive block fuzzy step size (ABFSS) are proposed. By exploiting the nature of gradient noise
inherent in stochastic gradient algorithms, these strategies are shown to substantially reduce the
computational complexity of the conventional FSS technique without sacricing the convergence speed and
steady-state performance. Instead of updating the step size at every iteration, the proposed techniques adjust
the step size based on the instantaneous squared error once over a block length. Design methodology and
guidelines that lead to good performance for the algorithms are given.
FSS technique that provides reliable implementation can complexity. Clearly, these issues are exacerbated when the
be invoked. techniques are to be implemented into subband adaptive
lters. Other approaches include the NLMS algorithm
To the best of our knowledge, no work on FSS technique based on the Kalman lter [12], non-parametric technique
for subband adaptive lters has been published to date. [13] and neural network method [14].
However, both the VSS and FSS techniques are too costly
to be implemented as step size adaptation is needed for
every subband equaliser. To overcome this problem, we 2 Subband adaptive equaliser
propose two novel block-based fuzzy step size (BFSS)
techniques that exploit the nature of gradient noise in structure and FSS technique
stochastic gradient algorithms to reduce the complexity of In this paper, vectors are represented by boldface lower case
the FSS technique by updating the step size over a block letters and matrices by boldface upper case letters. (.)T and
length. In the xed block fuzzy step size (FBFSS) (.)H symbolise the transpose and Hermitian operations,
technique, it is difcult to obtain the optimum block respectively.
length for practical channels as the channel characteristics
are unknown and time-varying. To circumvent this In this section, we propose a novel subband adaptive
problem, we propose a modied FBFSS approach called equaliser that incorporates FSS as depicted in Fig. 1.
adaptive block fuzzy step size (ABFSS), whereby the block
length of the ABFSS approach is adapted according to the The impulse response of the bth analysis lter fb[n] can be
instantaneous squared error. derived from a real-valued low-pass prototype nite impulse
response (FIR) lter p[n] of even length Lp by using a
It is important to stress that the proposed BFSS strategies generalised discrete Fourier transform (GDFT)
are different from the fuzzy sequential partial update (FSPU)
technique in [10] and the fuzzy logic-controlled block
adaptive ltering (FLC-BAF) in [11]. As compared with fb [n] p[n] e j(2p=B)(bb0 )(nn0 ) n 0, 1, 2, , Lp 1
the conventional FSS technique, the FSPU technique leads b 0, 1, 2, , B 1 (1)
to reduced complexity but suffers from slower convergence,
while the FLC-BAF technique attains faster convergence
speeds at the expense of a higher computational The impulse response of the bth synthesis lters gb[n] can be
Figure 1 Schematic of the subband adaptive equaliser utilising the FSS NLMS algorithm
if the GDFT is selected with a time-offset n0 according to Fig. 2 illustrates the block diagram of a subband adaptive
lter with the step size of the NLMS algorithm updated by
Lp 1 a fuzzy inference system (FIS).
n0 (3)
2
The FIS is based on the principle of fuzzy logic originally
In Fig. 1, the received signal r [n] and the training signal developed by Zadeh [17], which is used to handle linguistic
dtraining [n] are split into B subbands by the analysis lter concepts. In channel equalisation, the FIS provides an
bank and decimated by a factor of X. Oversampled lter adaptation mechanism that converts the linguistic control
banks with an oversampling ratio of B/X . 1 are preferred strategy based on the channel characteristics into an
as any aliased components of this approach would lie in the adaptive parameter control strategy. The FIS consists of
stopband of the respective analysis lter [15, 16]. We three main processors, namely fuzzier, rule inference
dene the input tap-delay line vector for the bth subband as: engine and defuzzier.
r b [m] [rb [m] rb [m 1]; . . . ; rb [m (Lsubband 1)]]T (4) The fuzzier transforms the input variables into their
respective degrees to which they belong to each of the
where m denotes the new sampling instant after the appropriate fuzzy sets by using membership functions
decimation process and Lsubband denotes the length of the (MBFs). The fuzzy sets are used to partition the continuous
subband equaliser. The subband signal rb[m] is obtained by domain of the FIS input and output variables into a small
ltering the input signal r[n] using the analysis lter fb[n]: number of overlapping regions labelled with linguistic terms
to cover the entire universe of discourse. The crisp numerical
Lp 1
X inputs need to be limited to the domain of the input
rb [m] fb [v] r[Xm v] b [ [0, B 1] X , B (5) variables. The output of the fuzzication process
v0
demonstrates a fuzzy degree of membership between 0 and 1.
Let wb[m] denote the weight vector for the bth subband,
In this paper, three fuzzy sets, namely small, medium
wb [m] [wb,0 [m] wb,1 [m] wb,2 [m]; . . . ; wb,Lsubband 1 [m]] T and large are used for the input and output fuzzy
variables, which correspond, respectively, to the squared
(6) error and lter step size. Some examples of membership
functions are depicted in Fig. 3. The input limit for the
The subband equaliser output ^s[m] is given by squared error is very much larger than 0.2, but for clarity, it
is only shown until 0.2 in Fig. 3a. It should be stressed
^sb [m] wH
b [m]r b [m] (7) that the optimum position of the fuzzy sets within their
respective limits for different types of channels is different.
Therefore the error signal within each subband can be
calculated as
Next, the following fuzzy IF-THEN rules are applied:
eb [m] dtrainingb [m] ^sb [m] (8) IF eb2 [m] is large, THEN mb [m] is large.
The equaliser coefcients for different subbands are adjusted IF eb2 [m] is medium, THEN mb [m] is medium.
independently using the individual subband error signal by
means of the NLMS algorithm:
mb [m]
wb [m 1] wb [m] r b [m]eb [m]
a rH
b [m]r b [m]
mb [m]
wb [m] r [m]eb [m] (9)
a jjr b [m]jj2 b
Step 1: Initialisation.
Step 3: Step size and block length adaptation.
1 m1
C1 0 and j IF C1 j, then execute (10) for step size adaptation and
K else
perform the following table lookup for block length
where C1 and K denote the counting index and the block adaptation:
length, respectively. 8
>
> K1 if e2 [n] d1
>
> if d2 e2 [n] , d1
Step 2: Execute (7) and (8) and C1 C1 1. < K2
>
K .. ..
Step 3: Step size adaptation. > . .
>
>
IF C1 j, then execute (10) and reset C1 0. >
> K if d e 2
[n] , dL1
: L1 L
Else set mb[m] mb[m 2 1]. KL if e2 [n] , dL
Step 4: Tap weights adaptation by the NLMS algorithm. where K1 . K2 . . . . . KL21 . KL and d1 . d2 . . . . .
Execute (9). dL21 . dL denote the preset block lengths and switching
thresholds, respectively. The subscript L represents the
Step 5: Repeat steps 2 to 4. number of switching levels.
Reset C1 0.
Else set mb[m] mb[m 2 1].
The problem of block length selection arises when the
channel characteristics are unknown. As a conservative
guide, we suggest setting the block length to be close to
the lter length. This can be justied by the fact that the With this setting, the block length decreases as the number
convergence speed is inversely proportional to the lter of subband iterations increases. It exploits the fact that the
length and in most adaptive ltering applications, FSS at the early training phase is larger and thus a larger
the training duration required for convergence is larger than step size can be utilised for a longer duration and
the lter length. In particular, for subband adaptive unnecessary step size updates can be reduced. A small step
equalisation, the required training duration is found to be size near convergence can also be avoided from being
very much larger than the subband lter length [8, 9]. trapped for a long duration. As the complexity incurred in
Although the suggested K value might not be the optimal block length adaptation using the fullband squared error is
choice from the viewpoint of complexity saving, but in independent of the number of subbands, the complexity
most cases, this suggested block length can effectively can be kept to a minimum. Consequently, the block length
minimise the trapping duration to achieve satisfactory for different subbands at any instant will be identical.
convergence performances while achieving a complexity
lower than that of the FSS NLMS algorithm. As the FSS With regard to the convergence speed and steady-steady
and FBFSS NLMS algorithms are converging, the lter error, the performances of ABFSS NLMS algorithms are
coefcients are approaching their respective steady-state expected to be similar to that of the FSS NLMS algorithm
values. It is understandable that mb [1] for both the FSS if the switching thresholds are properly set. Also, it is
and FBFSS NLMS algorithms will be identical. worth mentioning that the time-varying feature of the
Consequently, this results in the same eb [1] and e[1] block length of the ABFSS NLMS algorithm provides an
values for both algorithms. extra degree of freedom to facilitate algorithm design.
3.2 Adaptive block fuzzy step size (ABFSS) The CFIS is quantied in terms of the number of counts
to perform special instructions that include table lookup,
NLMS algorithm comparison, addition and multiplication. More precisely,
In the ABFSS NLMS algorithm, the block length is adapted the additional costs involved in preparing the input
to the fullband squared error e2 [n] based on the adaptive variable prior to fuzzication are also included in CFIS .
switching thresholds. In this regard, steps 1, 2, 4 and 5 of However, the complexity associated with block length
the FBFSS NLMS algorithm described in Section 3.1 adaptation of the ABFSS NLMS algorithm can be
remain unchanged, with the exception that step 3 being neglected as it is insignicant compared with the overall
modied as in Table 2. complexity. To alleviate the computational load in the
centroid calculation, fewer points (q) should be used as X 3. The SAR is 70.14 dB, which is sufcient to ensure
the main computational burden is incurred by the that the aliasing effect in various subbands is negligible [21].
defuzzication process.
D Lsubband/2 is chosen [22] and Lsubband is set to 53. For
The complexity of a subband adaptive equaliser that the xed step size NLMS algorithm, m 0.4 is utilised. This
requires N training iterations to converge is value is found through computer experiments to give
satisfactory convergence speed and steady-state performance
B 1 over the investigated channel. The MBFs in Fig. 3 have
CT N 8 Lsubband 6 Lp 12 B log2 B
X X been employed for the FSS NLMS and the proposed
algorithms. Block lengths of K 50 and 450 are used for
ABCFIS (12) the FBFSS NLMS algorithm. Table 4 shows the block
length setting of the ABFSS NLMS algorithm.
where A denotes the number of step size adaptations required
for each subband to converge. Likewise, AFSS , AFBFSS and
AABFSS represent the number of step size adaptations needed 5.1 MSE performance comparison
for the FSS, FBFSS and ABFSS NLMS algorithms,
The performance of different algorithms is assessed in terms
respectively. The complexity of the xed step size NLMS
of the MSE for both the transient and steady-state
algorithm can be determined from (12) with A 0.
behaviours. The MSE results were averaged over an
ensemble of 20 runs.
AFSS and AFBFSS can be expressed by (13) and (14),
respectively
In Fig. 4, it is observed that the all FSS-based (FSS,
FBFSS and ABFSS) NLMS algorithms achieve
N
AFSS (13) comparable convergence speeds that are signicantly faster
X
than that of the xed step size NLMS algorithm. It is also
AFSS noteworthy that for the FBFSS NLMS algorithm, a
AFBFSS (14) slightly better convergence performance is achieved with a
K
block length of K 450 over 50. This is because the
Equations (12) (14) are also applicable to a fullband performance of the BFSS techniques is affected by the
counterpart, where B 1 X 1 and Lsubband represents the instantaneous squared error, thus depending on the channel
length of the fullband equaliser. characteristic, a block length value, which gives good block
smoothing results can be found. Also, all algorithms
converge to approximately the same steady-state error.
5 Simulation results However, for applications where only a limited training
We consider a 64-QAM (quadrature amplitude modulation) period is available, better equalisation performances can be
signal transmitted over a channel with a system achieved using the FSS-based NLMS algorithms. For
function h(z) 0.182 0.269 z 21 0.888 z 22 0.269 z 23 instance, it is evident in Fig. 4 that the FSS-based NLMS
0.182 z 24 at signal-to-noise ratio (SNR) 30 dB. To algorithms achieve a performance gain of 4 dB to 5 dB over
maximise both the signal-to-alias ratio (SAR) and power the xed step size NLMS algorithm if only 1500 training
complementarity, an iterative least-squares method is symbols are available.
employed to design the prototype lter [20]. The
oversampled GDFT modulated lter bank with a prototype Subband decomposition of the received signal reduces the
lter length of Lp 60 splits the fullband signal into B 4 eigenvalue spread in each subband. The eigenvalue spread
subband signals, each of which is decimated by a factor of can be approximated by the ratio of the maximum and
From the simulations, it is found that the maximum and 5.3 Computational complexity
minimum step sizes for all FSS-based NLMS algorithms comparison
are identical, which are 1.267 and 0.4, respectively. Because
of space constraint, the step-size transient characteristics of As shown in Section 5.1, the convergence speed of the
the ABFSS algorithm are not shown. As the converged proposed algorithms and the FSS NLMS algorithm are
step size values are identical for all algorithms, similar comparable and they converge approximately after 4000
modelling capability could be achieved by both the overall iterations. AABFSS can be interpreted from Fig. 9 by
and individual subbands, as illustrated in Figs. 4 7. counting the number of step-size adaptations within the
period AFSS . Fig. 10 shows the computational complexities
Fig. 9 depicts the block length transient characteristic of of different NLMS algorithms. The complexities of the
the ABFSS NLMS algorithm. The markers indicate the FSS-based NLMS algorithms are expressed in terms of
Figure 8 Step-size transient characteristic for b 0 and 1 Figure 10 Computational complexity comparison for
using the FSS and FBFSS different NLMS algorithms
percentage with respect to that of the xed step size NLMS [6] SHIN H.-C., SAYED A.H., SONG W.-J.: Variable step-size NLMS
algorithm. and afne projection algorithms, IEEE Signal Process.
Lett., 2004, 11, pp. 132 135
The maximum value of q for a 16-bit processor is around
4000. For a conservative comparison, q 200 is considered. [7] GAN W.S.: Designing a fuzzy step size LMS algorithm,
With respect to the xed step-size technique, the FSS IEE Proc. Vis. Image Signal Process., 1997, 144,
NLMS algorithm requires approximately an extra 37% pp. 261 266
computational complexity whereas only minor increases in
computational complexity (0.0826 0.7435%) are incurred [8] MOHAMAD H., WEISS S., RUPP M., HANZO L.: Fast adaptation of
by the BFSS strategies. fractionally spaced equalizers, Electron. Lett., 2002, 38,
pp. 96 98
6 Conclusion [9] MOHAMAD H., WEISS S., ARIF N.A.M. , ALIAS M.Y.: Subband
decomposition techniques for adaptive channel
We proposed two novel BFSS strategies for the NLMS
equalisation. Proc. IEEE 13th Int. Conf. on Network,
algorithm, namely the FBFSS and ABFSS. Simulation
Malaysia, 2005, vol. 2, pp. 617 621
results show that their convergence speed are similar to that
of the FSS NLMS algorithm but are signicantly faster
[10] LIN H.-Y., HU C.-C., CHEN Y.-F., WEN J.-H.: An adaptive robust
than that of the xed step-size NLMS algorithm.
LMS employing fuzzy step size and partial update, IEEE
Furthermore, all techniques attain almost identical steady-
Signal Process. Lett., 2005, 12, pp. 545 548
state performance. The transient characteristics of the step
size and block length were also analysed. From the
[11] LIN H.-Y., HU C.-C.: Asynchronous multiple-access
complexity analysis, the proposed techniques can greatly
interference suppression using variable block adaptive
reduce the computational burden of the FSS NLMS
ltering in time-varying multipath channels. Proc. IEEE
algorithm. The proposed techniques can also be employed
6th Workshop on Signal Processing Advances in Wireless
for other adaptive ltering problems such as noise
Communications, New York, 2005, pp. 17 21
cancellation and system identication. These intriguing
concepts have implications for all forms of VSS adaptive
[12] LOPES P.A.C. , TAVARES G. , GERALD J.B. : A new type
algorithms, whether they are based on fuzzy technique or
of normalized LMS algorithm based on the kalman lter.
otherwise.
Proc. IEEE 32nd Int. Conf. on Acoustics, Speech,
and Signal Processing, Hawaii, 2007, vol. 3, pp. III-1345
III-1348
7 Acknowledgment
The authors are grateful to the editor and anonymous [13] BENESTY J., REY H., VEGA L.R., TRESSENS S.: A nonparametric
reviewers for their constructive comments that helped in VSS NLMS algorithm, IEEE Signal Process. Lett., 2006, 13,
improving the quality of this paper. pp. 581 584
[14] TUMMARELLO G., NARDINI F., PIAZZA F.: Step size control
8 References in NLMS acoustic echo cancellation using a neural
network approach. Proc. IEEE Int. Symp. on Circuits
[1] HAYKIN S.: Adaptive lter theory (Prentice Hall, New and Systems, Bangkok, Thailand, 2003, vol. 5, pp. V-705
Jersey, 2001) V-708
[2] HARRIS R.W., CHABRIES D.M., BISHOP F.A.: A variable step (VS) [15] MOHAMAD H., WEISS S., RUPP M.: MMSE limitations for
adaptive lter algorithm, IEEE Trans. Acoust. Speech Signal subband adaptive equalisers. Proc. 36th Asilomar Conf.
Process., 1986, 34, pp. 309 316 on Signals, Systems, and Computers, Monterey, California,
2002, vol. 2, pp. 1233 1237
[3] KWONG R.H. , JOHNSTON E.W.: A variable step size LMS
algorithm, IEEE Trans. Signal Process., 1992, 40, [16] WEISS S., STENGER A., STEWART R.W., RABENSTEIN R. :
pp. 1633 1642 Steady-state performance limitations of subband
adaptive lters, IEEE Trans. Signal Process., 2001, 49,
[4] ABOULNASR T., MAYYAS K.: A robust variable step-size LMS- pp. 1982 1991
type algorithm: analysis and simulations, IEEE Trans. Signal
Process., 1997, 45, pp. 631 639 [17] ZADEH L.A.: Fuzzy sets, Inf. Control., 1965, 8,
pp. 338 353
[5] PAZAITIS D.I., CONSTANTINIDES A.G.: A novel kurtosis driven
variable step-size adaptive algorithm, IEEE Trans. Signal [18] ROSS T.J.: Fuzzy logic with engineering applications
Process., 1999, 47, pp. 864 872 (McGraw-Hill, Singapore, 1997)
[19] ALTROCK C.: Fuzzy logic & neurofuzzy applications [21] MOHAMAD H., WEISS S., RUPP M., HANZO L.: Performance
explained (Prentice Hall, 1995) limitation of subband adaptive equalisers, Electron. Lett.,
2006, 42, pp. 1009 1010
[20] WEISS S., STEWART R.W.: On adaptive ltering in
oversampled subbands (Shaker Verlag, Aachen, Germany, [22] WIDROW B. , STEARNS S.D.: Adaptive signal processing
1998) (Prentice-Hall, Englewood Cliffs, NJ, 1985)