International Journal of Emerging Trends in Signal Processing

Volume 1, Issue 1, November 2012

30

Diversity Added Adaptive-Fuzzy Logic Controlled
Receivers for Direct Sequence Code Division
Multiple Access System
Mangala Ramrao Gajbhiye
#
, Jitendriya Kumar Satapathy
*

#
Electrical Engineering Department, National Institute of Technology
Rourkela-769008, Orissa, India
1
[email protected]
*
Electrical Engineering Department, National Institute of Technology
Rourkela-769008, Orissa, India
2
[email protected]

Abstract A new approach in adjusting a step size and a
forgetting factor of the Least Mean Square (LMS) and
Recursive Least Square (RLS) algorithms respectively using the
fuzzy logic controlled technique is developed for Maximum Ratio
Combining diversity added Multiuser Detection of Direct
Sequence Code Division Multiple Access (DS-CDMA) system in
multipath channels. DS-CDMA is a popular multiple-access
technology for wireless communication. However its
performance is limited by multiple-access interference (MAI)
and multipath distortion. Adaptive techniques for detection of
DS-CDMA signals improves the performance of the CDMA
system by reducing interference among users by employing an
adaptive techniques called LMS and RLS algorithms. The
diversity scheme called maximum ratio combining (MRC) as
pre-processor to the adaptive MMSE detector of DS-CDMA
removes the use of training sequences and the step size of LMS
algorithm and forgetting factor of RLS algorithm which are to be
taken constant can be varied by Fuzzy Inference System. This
receiver employs a fuzzy-logic control mechanism to perform the
nonlinear mapping of the squared error and squared error
variation, denoted by (e
2
, e
2
), into a step size or a forgetting
factor. Simulations have proved its better performance
compared with the conventional Match Filter detector,
performing nearly as well as single user case for multipath
channels.

Keywords Direct-Sequence Code-Division Multiple-Access
(DS-CDMA), Multiuser Detection (MUD), Fuzzy Inference
System (FIS), LMS and RLS algorithms, multipath channel
I. INTRODUCTION
Direct-sequence code-division multiple access (DS-
CDMA), a specific form of spread-spectrum transmission, has
been adopted as the multi-access technology for non-
orthogonal transmission in the third-generation (3G) mobile
cellular systems, such as wideband CDMA (W-CDMA) or
multi-carrier CDMA (MC-CDMA). This sort of the code-
division multi-access techniques requires no time or frequency
coordination among the mobile stations. However, the so-
called near-far problem and the multipath fading are the major
impediments to maintain reliable communication links in
CDMA systems.
It is well known that an adaptive minimum mean-square
error (MMSE) linear receiver [1], [2] & [3] has immunity to
the near far problem and the interference floor in performance
exhibited by the conventional matched filter reception. In
addition, a linear MMSE receiver can be implemented as an
adaptive tapped delay line (TDL), analogous to a linear
equalizer, with a relatively low complexity. However, the
computation of the MMSE solution involves the calculation of
the inverse of the input autocorrelation matrix, which costs a
complexity of O((MN)3). Here M denotes the size of the
MMSE receiving array and N is the processing gain of the
CDMA system so that MN indicates the number of tap
weights of the linear MMSE filter. This cost is even more
expensive when the linear MMSE receiver operates in a non-
stationary multipath environment. In practice, the filter-
coefficient vector of the MMSE-type receiver can be obtained
from the training sequence and the received signal by means
of conventional adaptive filtering techniques, such as the
least-mean square (LMS) and the recursive least-squares
(RLS) [4] approaches. The LMS provides simple
implementation but suffers from slow convergence, while, on
the other hand, the RLS converges much faster as compared
with the LMS, but it possesses more computational
complexity [9]. The drawback of slow convergence of an
LMS-based algorithm, due to its dependence on the
eigenvalue spread, is overcome in an RLS algorithm [4] by
replacing the gradient step-size with a gain matrix, denoted
by R
1
x [n], at the nth iteration.
Maximum Ratio combining (MRC) is optimum way (in the
sense of the least Bit error Rate (BER)) to use information
from different paths to achieve decoding in multipath channel.
It corrects the phase rotation caused by a multipath channel
and then combines the received signals of different paths
proportionally to the strength of each path. Since each path
undergoes different attenuations, combining them with
International Journal of Emerging Trends in Signal Processing
Volume 1, Issue 1, November 2012

31

different weights yield an optimum solution under multipath
channel. [5], [11], [12].
In recent years, the concept of fuzzy logic is used in many
different senses. Fuzzy logic can be treated as a tool for
embedding structured human knowledge into workable
algorithms. In a wider sense, fuzzy logic is a fuzzy set theory
of classes with unsharp or fuzzy boundaries [10]. Systems
designed and developed utilizing fuzzy-logic methods have
been shown to be more efficient than those based on
conventional approaches [6]. In principle, fuzzy-logic
controller (FLC) provides an adaptation mechanism that
converts the linguistic control strategy based on the
characteristics of mobile radio channels into an adaptive
parameter-control strategy. By using the defuzzification, the
fuzzy control decisions are converted to a crisp control
command which is used to adjust properly the level of the
parameter of interest.
Recently, adaptive implementations of the mean-squared
error (MMSE) receivers have received some attention in the
literature, e.g. [2], [3]. The focus of this paper is on improving
the performance of adaptive MMSE receivers by introducing
maximum ratio combiner (MRC) as a preprocessor in
multipath channels. The proposed Adaptive MMSE-MRC
algorithm is applied to the downlink (base station to mobile)
of DS-CDMA system. The Fuzzy Logic Controlled (FLC)
system is employed to perform the nonlinear mapping of the
input variables into a scalar adaptation step-size of the LMS
algorithm and forgetting factor of the RLS algorithm in
response to the channel variation.
In [7], the behavior of different adaptive LMS algorithms
with the fuzzy step size is analyzed. In [8], the authors
proposed the variable forgetting factor linear least-squares
(VFF-LLS) algorithm to improve the tracking capability of
channel estimation. These works motivate the development of
the adaptive linear MMSE CDMA receiver with a fuzzy-logic
controlled two-parameter system of (e
2
, e
2
), where e
2
(n)
indicates the squared error variation at time n. These pair
values of (e
2
, e
2
) are calculated and fed to the FLC system to
assign an exact value of or for the corresponding adaptive
receiver on an iterative basis in order to improve the
convergence characteristic and steady-state MSE
simultaneously. Most of the fuzzy inference rules are derived
by a human expert or extracted from numerical data.
Therefore, it seems natural and reasonable to expect that
wireless communication systems with the use of a two
parameter (e
2
, e
2
)-FLC produce better convergence
characteristics. Furthermore, the pair of (e
2
, e
2
) provides the
FLC system with more precise channel dynamic-tracking and
adaptation capability. This is because the auxiliary
parameter e
2
offers an effective and robust means to monitor
instantaneous fluctuations of a fast-fading multipath channel
and assists the FLC system in selecting an appropriate value
for or . It is remarkable that the proposed FLC-based
approaches produce a faster speed of convergence without
trading off the steady-state performance. Slightly additional
computational load is incurred in the fuzzification (table
lookup), inference (MIN, MAX, and PROD operators), and
defuzzification processes. There is also additional cost which
comes from the preparation of the two input variables, (e
2
,
e
2
), prior to the fuzzification process. Fortunately, these
operations can be done very efficiently in the latest range of
DSPs which provide single-cycle multiply and add, table
lookup, and comparison instructions. The computational load
is compared to other known adaptive MMSE-MRC DS-
CDMA algorithms as well.
This paper continues by introducing the system model in
section II. Linear multi-user detectors and adaptive linear
detectors are described in section III. A Maximal-ratio
combining (MRC) scheme is described in section IV. The
adaptive MMSE-MRC-LMS detector is investigated in
Section V. And in section VI, developed the diversity added
adaptive FLC-LMS and FLC-RLS MMSE Multiuser Detector
of DS-CDMA System in downlink multipath channel.
Simulation results are reported in Section VII, and finally
concluding remarks are summarized in Section VIII.
II. SYSTEM MODEL
The DS-CDMA system modelled in this paper (shown in
Fig.1) [12] is the downlink (from base station to mobile). It is
assumed that the channel is a fully synchronized antipodal
DS-CDMA system with each users data is spread using
an individual quasi orthogonal signature waveform (spreading
code).
The data for all the users is then combined and transmitted
through an additive white Gaussian noise (AWGN) channel.
In this paper the data transmission is assumed to be binary
phase shift keying (BPSK). Assuming a total of K active users
in an M multipath environment and I transmitted symbols
during the observation interval, the received signal can
be written as,
) ( ) ( ) ( ) (
,
) (
1 1
1
0
t n t s t h b A t r
k m k
n
k
N
n
K
k
L
l
k
+ =

= =

=
(1)
where
k k k
E T E A , / = is the energy per symbol.
) (n
k
b
is the n
th
symbol with G chips transmitted in a symbol time
b
T .
In a pass band notation, the transmitted bit b
k
(t) for the k
th
user
can be written as,
) ( ) ( ) (
1
b
N
n
T k k
KT t p n b t b
b
=

=
(2)
where T
b
is the bit duration and (.)
b
T
p is a rectangular pulse
International Journal of Emerging Trends in Signal Processing
Volume 1, Issue 1, November 2012

32



Fig. 1 Multiuser downlink DS-CDMA system
shaping waveform given by

( (
=
otherwise
T t
t p
b
T
b
, 0
0 , 1
) ( (3)
Similarly, the spreading sequence for the k
th
user can be
represented as,
) ( ) ( ) (
1
c
G
g
T k k
KT t p g s t s
c
=

=
(4)
where T
c
is the chip duration and (.)
c
T
p is a pulse-shaping
signal similar to (.)
b
T
p . Since the length of spreading
sequence is G, we have T
b
= GT
c
.
The channel impulse response of the fading channel is
represented by a finite impulse response (FIR) filter
(represented as ) (
,
t h
m k
in (1), where m is number of paths &
it is varying from 1 to M.). Following [12] the three strongest
signals (h
1
, h
2
, h
3
) are used to estimate the signal under
consideration,
2
3
1
2 1
) (

+ + = z h z h h z H (5)
The channel response is stored in a matrix H. The noise n(t)
is a complex AWGN with zero mean and two-sided power
spectral density of No/2 W/Hz.
Equation (1) can be compactly expressed as,
n SHAB r + = (6)
The structures can be defined as follows:
| | S diag S = (7)
| |
k
s s s S ...
2 1
=
(8)
(
(
(
(
(
(

1
1
0 1
0
........ 00
....
0 .......
0 .......
0 ....... 0
M
M
h
h
h h
h
H
(9)
| |
T T T
N b b b B ) ( .... ) 2 ( ) 1 ( = (10)
| | ) ( )..... ( ) ( ) (
2 1
I b I b I b I b
K
= (11)
and r is a matrix which represents all possible received
signals. These signals are then fed to the preprocessor, which
is a bank of matched filters. Based on this compact model (6),
the conventional Matched Filter (MF) detection scheme is
summarized next.
The standard matched filter (MF) is a single user detection
process which utilizes the users own signature sequence only.
At the receiver end, the received signal is passed through a
group of correlators in order to recover the information stream
transmitted by each user. Hence the output of the k
th
correlator
under synchronous condition would be,
( ) r S sign b
H
k MF k
=
) (

(12)
where | | 0 .... .... 0
k
H
k
s S = ,
k
s being the k
th
users signature
code. Equation (12) may be given as,
( )
MF MF k
b sign b =
) (

(13)
where N RHAB b
MF
+ = (14)
In equation (14), R is the normalized cross-
correlation matrix of the spreading sequences and N
is noise component of MF output. If there was no
multipath and mutually orthogonal spreading
sequences were chosen, the conventional (MF)
detector would result in optimal demodulation.
However multipath destroys any orthogonality
present in the spreading sequences and the
demodulated results are unacceptably erroneous.
Hence this b
MF
is then used as an input to multi-user
detectors to improve the bit-error performance [3],
[12], [14] and [15].
III. LINEAR MULTIUSER DETECTORS
There has been great interest in improving DS-CDMA
detection through the use of multi-user detectors. In multi-user
detection, code and timing (and possibly amplitude and phase)
information of multiple users are jointly used to better detect
each individual user. The important assumption is that the
codes of multiple users are known to the receiver a priori [14].
An important group of multi-user detectors are linear and
adaptive linear multi-user detectors. In any detector under this
group, a linear transformation is applied to the single user
matched filter (SUMF) to reduce the MA1 seen by each user
[14]. In this section we briefly review the two most popular of
these, the decorrelating and minimum mean-squared error
linear detectors and one adaptive linear detector.
A. Decorrelating Detector
It achieves perfect demodulation in the absence of noise if
the spreading sequences in S are linearly independent. It is
implemented by premultiplying the matched filter outputs
S
1
(t) A
1

b
1

S
2
(t) A
2

b
2

S
k
(t) A
k

b
k

1
b
2
b
k
b






r(t)
n(t)
Multipath
AWGN
Channel




MF

Receiver
International Journal of Emerging Trends in Signal Processing
Volume 1, Issue 1, November 2012

33

with the inverse of the cross-correlation matrix before
multiplying the sign, i.e.,
1
( )
( )
k MF MF
b sign R b

= (15)
The Decorrelating detector is shown to have many
attractive properties. Foremost among these properties are
[14]:
- Provides substantial performance capacity gains over the
conventional detector under most conditions.
- Does not need to estimate the received amplitudes: In
contrast, detectors that require amplitude estimation are often
quite sensitive to estimation error. (Note that as in the case of
most multi-user detectors, the need to estimate the received
phases can also be avoided through the use of non coherent
detection.
- Has computational complexity significantly lower than
that of the maximum likelihood sequence detector: The per-bit
complexity is linear in the number of users, excluding the
costs of precomputation of the inverse mapping.
Even though the decorrelator achieves perfect demodulation
in the noise-free case, it increases the effect of noise whenever
noise is present. A more significant disadvantage of the
decorrelating detector is that the computations needed to
invert the matrix R are difficult to perform in real time. For
synchronous systems, the problem is somewhat simplified: we
can decorrelate one bit at a time. In other words, we can apply
the inverse of a K x K correlation matrix. For asynchronous
systems, however, R is of order NK, which is quite large for a
typical message length, N.
B. Minimum Mean Square Error (MMSE) Detector
There are many ways one can view the performance of a
linear detector. Perhaps the most widely used is the achieved
Minimum Mean Squared Error (MMSE). A tradeoff can be
made between the increase in performance due to
decorrelation and the loss on performance due to increased
noise level. The linear MMSE detector achieves this optimum
tradeoff by making decisions as [3], [12],
| | ( )
MF LMMSE k
b A No R sign b
1
2
) (
) 2 / (

+ =

(16)
The symbols for section A and B are defined as follows-
A : K x K Users amplitudes diagonal matrix
B : K x 1 Users information bits
S : 1 x K Signature waveforms vector
w : AWGN
b
MF
: K x 1 Output of SUMF vector
R : K x K Cross-correlation matrix
N : K x 1 Noise component of SUMF outputs
An important disadvantage of this detector is that, unlike
the decorrelating detector, it requires estimation of the
received amplitudes. Another disadvantage is that its
performance depends on the powers of the interfering users.
Therefore, there is some loss of resistance to the near-far
problem as compared to the decorrelating detector. Like the
decorrelating detector, the MMSE detector faces the task of
implementing matrix inversion. Thus, most of the suboptimal
techniques for implementing the decorrelating detector are
applicable to this detector as well.
C. Adaptive Linear Detectors
Many adaptive DS-CDMA detectors are based on linear
receivers, especially on MMSE receivers. In this case the goal
is to minimize the MSE in the output of the linear filter. Its
implementation can be done by a simple tapped delay line
filter with an appropriate adaptive algorithm.
1) A Least Mean Square adaption: The Least Mean Square
(LMS) algorithm approximates the steepest descent method.
Following fig.2, coefficient update is as follows

(17)
where n indicates the time and
( ) | |
1 11 12 1
...
T
K
w n w w w = (18)
( )
1 2
, ,...,
T
K
b n b b b ( =

(19)
( ) ( ) ( ) e n d n y n =
(20)

( ) ( ) ( )
1
.
H
y n w n b n =
(21)
The desired signal is d(n). At the beginning of the process,
this signal is set to training sequences. After a while, which
does need to be that long, it will be switched to the output of
the adaptive filter. The step size is bounded to have stable
operation and acceptable convergence speed. For the smaller
value of step size, the convergence speed is lower. The main
advantage of LMS is its simplicity and its robustness to noise.
The complexity is O(NK) [3]
2) Recursive Least Squares Adaption : In this method, a
cost function is based upon the sum of errors up to the current
time. i.e.,
(22)
where n is the current chip time interval and is the forgetting
factor and 0< <1. The recursive least square (RLS) algorithm
provides faster convergence than the LMS method. However,
a major drawback to using RLS is its computational
complexity of O((NK)
2
). In this case, coefficient updates
are as follows
( ) ( ) ( ) ( ) ( )
... , 2 , 1
23 1
*
=
+ =
n where
n n k n w n w
k k





| |
| |
) 24 ( ) 1 ( ) ( ) ( ) ( ) (
) ( ) 1 ( ) ( ) (
~
,...,
~
,
~
) (
) ( ) 1 ( ) ( 1
) ( ) 1 (
) (
,....., , ) (
1 1
2 1
1
1
1 12 11
=
=
=
+

=
=

n p n b n k n p n p
n b n w n d n
b b b n b
n r n p n b
n b n p
n k
w w w n w
H
H
k
T
K
H
T
K k

2
1
) ( ) ( i e n
n
i
i n

= c
( ) ( ) ( ) ( )
*
1 1
1 w n w n b n e n + = +
International Journal of Emerging Trends in Signal Processing
Volume 1, Issue 1, November 2012

34



And the initial conditions are,


where is a small and positive number. The larger results in
slower convergence speed. The convergence speed of RLS
method is higher than the LMS algorithm [3].
IV. MAXIMUM RATIO COMBINING SCHEME
Various Multi-user detection methods have been
investigated in many previous studies [2], [3], [14] and [15] in
AWGN channel which uses a bank of MF as a preprocessor.
But in a multipath fading channel, propagation delay spread
merely provides multiple versions of the transmitted signal at
the receiver. If these multipath components are delayed in
time by more than chip duration, they appear like uncorrelated
noise at a DS-CDMA receiver. Hence CDMA receiver may
combine the time-shifted versions of the original signal by
providing a separate correlation receiver for each of the
multipath signals. So a better technology is the Maximum
Ratio Combining (MRC) (Fig.2) which first identifies few
strong multipath signals and then combines them after
incorporating adjustment for delays. In this scheme, the
spreading code and channel coefficients of the user of interest
are only utilized for the detection process.

Fig. 2 Maximum Ratio Combiner as a preprocessor.
A MF-MRC then utilizes few strongest multipath
components. The outputs of each correlator are weighted to
provide a better estimate of the transmitted signal than is
provided by a single component. Correlator 1 is synchronized
to the strongest multipath component
1
m . Second
multipath component
2
m arrives
1
t later than first component.
The second correlator is synchronized to
2
m . If only a single
correlator is used, once the output of this correlator is
corrupted by fading, the receiver can not correct the value. In
a MF-MRC receiver, if the output from one correlator is
corrupted by fading, the others may not be, and the corrupted
signal may be discounted through the weighting process. The
weighting coefficients
m
o , are normalized to the output
signal power of the correlator in such a way that the
coefficients sum to unity. i.e.

=
=
M
m
m
m
m
Z
Z
1
2
2
o
(26)
where
m
Z is the output from the m
th
correlator [6].
Mathematically the MF-MRC detector can be expressed as:
r H S b
H H
k MRC MF k
=
) (

(27)
where
H
H is the Hermitian of the estimated channel matrix.
In the presence of noise, the optimal LMMSE detector is
considered to be the best linear detector for DS-CDMA
reception [2]. The LMMSE detector for a generic single stage
spreading system is given by
( ) r A No R H S b
H H
k MRC LMMSE k
1
2
) (
) 2 / (

+ =

(28)
where R is the auto-correlation matrix for the signal chip train
received at the mobile set. No/2 is the average power of the
transmitted user signals. A major draw back of LMMSE is the
complexity involved in the computation of the auto-
correlation matrix.
V. ADAPTIVE MMSE-MRC-LMS DETECTOR
From the mathematical expressions given for MF-MRC
detector and LMMSE-MRC detector in the previous section
one can see that except the inverse operation remaining terms
are similar. Hence it is expected to achieve LMMSE
performance by adding a preprocessing stage to the MF-MRC
receiver. The basic principle of MF-MRC receiver is to
provide link improvement through time diversity.
An adaptive least mean squares (LMS) algorithm will be
developed for Adaptive MMSE-MRC detector, which
iteratively produces ( )
1
2
) 2 / (

+ A No R . 'The modified cost

function results in separate LMMSE detectors for each
multipath component. Hence, the adaptive MMSE detector is
actually an adaptive MRC receiver, where each receiver
branch is adapted independently to suppress MAI. The outputs
of adaptive receiver branches are maximal-ratio combined to
reduce decision variable.
Let
) (
,
n
m k
r be the input sample vector during the n
th
symbol
to the m
th
receiver branch. The received signal vectors are fed
to linear filters with impulse response of
) (
,
n
m k
w . The output of
the m
th
receiver branch can be written as,

) (
,
) (
,
) (
,
n
m k
n
m k
n
m k
r w y = . (29)
The error functions,
) (
,
n
m k
e , produced by the difference between
the filter outputs and the reference signals, are used to
update
the filter weights.
) (
,
) (
,
) (
,
n
m k
n
m k
n
m k
y d e = (30)
r(t)
MF
) (
1 , 1
t r
MF
Maximum
Ratio
Combining
) (
, 1
t r
M
) (
1
t r
MRC
MF
Maximum
Ratio
Combining
) (
1 ,
t r
K
) (t r
MRC
K
MF
) (
,
t r
M K



MUD
Or
ALMMSE
Detector
) (

1
t b
) (

t b
K
) 25 ( ,..... 2 , 1
0 ) 0 (
) 0 (
1
=
=
=

n
w
I p
k
k K
o
International Journal of Emerging Trends in Signal Processing
Volume 1, Issue 1, November 2012

35

The product of estimated channel coefficients and data
symbols is the reference signal in the adaptive MMSE-MRC
detector given by,
) ( ) (
,
) (
,

n
k
n
m k
n
m k
b h d = (31)
Either the data decisions or a training sequence can be used
as
) (

n
k
b . The data decisions produced initially by a
conventional MF-MRC receiver are often reliable enough for
adapting the receiver. It is also possible to use the absolute
value of the estimated channel coefficients
) (
,
n
m k
h . Hence, the
proposed adaptive receiver does not necessarily require
separate training sequence. The optimal filter coefficients are
derived using the MSE criterion
|
.
|

\
|
(

2
) (
,
n
m k
e E
, which leads
to the optimal filter coefficients [9]
m k
w
,
.
It is convenient to decompose the filter vector to fixed and
adaptive components:
) (
,
) (
,
n
m k
H
k
n
m k
x S w + = (32)
where | | 0 .... .... 0
k
H
k
s S = is the fixed spreading sequence of
the k
th
user and
) (n
x is the adaptive component. If standard
LMS algorithm is used for adapting the filter, the updates for
the adaptive component can be written as,
) (
,
) (
,
) (
,
) 1 (
,
) (
,
n
m k
n
m k
n
m k
n
m k
n
m k
r e x x
-
+ = (33)
where
) (
,
n
m k
is the time-variant step-size parameter, which
controls the rate of convergence of the algorithm.
VI. FLC-LMS/RLS-BASED MMSE-MRC DETECTOR
A novel approach, which uses the fuzzy inference system
(FIS), is developed here to adjust adaptively the step size for

Fig. 3: Block diagram for the FLC-LMS & FLC-RLS algorithms
the LMS algorithm or the forgetting factor for the RLS
algorithm at each time index, which was taken constant in the
precious section. This proposed fuzzy based MMSE CDMA
receiver provides superior convergence/ tracking
characteristic and smaller steady-state MSE over the
conventional LMS/RLS-MMSE CDMA receivers. The
symbol (n) is employed to stand for both time-dependent
variables (n) and (n) at time n. In this paper, the two-input
one-output FIS, which operates based on the principle of
fuzzy logic proposed originally by Zadeh [10], takes in two
inputs, the squared error (e
2
(n)), and the squared error
variation (e
2
(n)) at the nth iteration.
In general, the basic configuration of the FIS comprises
four essential components, namely, (i) a fuzzification interface,
(ii) a fuzzy rule base, (iii) an inference engine, and (iv) a
defuzzification interface, which map two inputs (e
2
(n),e
2
(n))
into an output (n) for adaptive filtering schemes, as shown
in Figure 4. The fuzzifier converts the real world crisp input to
a fuzzy output described by the membership function. The
inference engine provides the relationship between the fuzzy
input in terms of membership functions and the fuzzy output
of the controller using a set of IF THEN rules derived
from the rule base. The defuzzifier converts the inferences to
provide the crisp output. Generally in a fuzzy system the rule
base is generated in advance with expert knowledge of the
system under consideration.
The general format for the proposed FLC-LMS and FLC-
RLS approaches to assign a suitable (n) at time index n is
formulated as
( ) ( ) ( ) ( ) ( ) ( ) n r n w n d n y n d n e
H
= =
(34)
( ) ( ) ( ) 1
2 2 2
= A n e n e n e (35)
FLC-LMS, FLC-RLS:
( ) ( ) ( ) ( ) n e n e FIS n
2 2
, A = (36)
where e(n), d(n), and (n) represent the error signal, the
desired signal, and the output of the adaptive filter,
respectively, at the time instant n. the desired signal d(n) is
initially produced by the decision directed MMSE-MRC
detector as explained in previous section.
The key concepts of the fuzzy rules are shared and used to
establish a common foundation for both the LMS and RLS
algorithms in order to make the best choice for the .
All the fuzzy inference rules used for the proposed LMS
and RLS algorithms are summarized in tables I and II,
respectively.
The FLC-LMS and FLC-RLS systems have two inputs, one
output and 16 rules respectively.
The term-sets for FLC-LMS algorithm as,
S = 0.01, L = 0.1, M

= 0.05, VL= 0.5
S

= 0.001, L = 0.1, M

= 0.01, VL = 0.3
S

= 0.0003, L = 0.001, M = 0.0006, VL = 0.002
And the term-sets for FLC-RLS algorithm as,
S = 0.01, L = 0.2, M = 0.1, VL= 0.3
S

= 0.005, L= 0.1, M

= 0.03, VL= 0.2
S

= 0.94, L = 0.98, M

= 0.96, VL = 1
TABLE I
RULE MATRIX FOR FLC-LMS
(n)
Fuzzy rule
base
Inference
engine
F
u
z
z
i
f
i
c
a
t
i
o
n

i
n
t
e
r
f
a
c
e

D
e
f
u
z
z
i
f
i
c
a
t
i
o
n

i
n
t
e
r
f
a
c
e

FIR filter
LMS
Fuzzy Inference System (FIS)
Adaptive filter
r(n)
y(n) -
+

e(
n)
D
e
l
a
y

e
2
(n)
d(n)

e(n)
+
e
2
(n) -
e
2
(n-1)
International Journal of Emerging Trends in Signal Processing
Volume 1, Issue 1, November 2012

36

e
2
e
2


S

M

L

VL
S S S M M
M S M M L
L M M L VL
VL S M L VL
TABLE III
RULE MATRIX FOR FLC-RLS


e
2

e
2


S

M

L

VL
S S S M M
M S M M L
L M M L VL
VL S M L VL
VII. SIMULATION RESULTS
Simulation results are presented for synchronous downlink
DS-CDMA system. BPSK modulation was assumed with the
bits 1 being equiprobable and spreading code of length G
with Walsh codes is used. The channel response used for
minimum phase multipath channel is H(z)=1 + 0.5z
-1
+ 0.25z
-2
,
assumed to have three multipath (M=3). For each simulation
10
5
bits were transmitted. Bit error rate (BER) was calculated
by dividing the number of bits classified incorrectly by the
total number of bits transmitted.
Fig. 4 shows the Mean Square Error (MSE) at the output of
the LMS adaptive MMSE multiuser filter for three different
values of step size with G = 32, 3 active users, all have an
equal power of E
b
/N
0
= 8dB. As is seen, for the smaller value
of step size, the convergence speed is lower. Fig. 5 pictures
the MSE of the RLS adaptive filter for three different values
of forgetting factocr with G = 32, 3 active users, all have an
equal power of E
b
/N
0
= 8dB. In this case, the larger results
in slower convergence speed. Comparing figure 5 and 6,
justifies that the convergence speed of the RLS method (even
in worse condition of E
b
/N
0
= 5dB) is still higher than that of
the LMS algorithm. As an example, for = 0.99, after around
2000 iterations the system settles down, while in the LMS for
= 0.001 even after 3000 iterations it still does not give the
final error level.
Fig. 6 and Fig. 10 shows the FLC-LMS and FLC-RLS
systems respectively with two inputs, one output and 16 rules.
The relations between the MBFs and the fuzzy rules in the FIS
of the LMS and RLS algorithms are illustrated in Fig. 7-9 and
Fig. 11-13 respectively. Fig. 14 and Fig.15 shows the FLC-
LMS and FLC-RLS systems output surface plots respectively.
As per the rules made we got the larger range of step size (),
as it converges faster for higher value.
Fig. 16 demonstrates the BER performance of the FLC-
LMS and FLC-RLS adaptive MMSE-MRC detectors as a
function of Eb/No under the MAI for 3 users with spreading
code of length 32 in dispersive multipath channels having 3
paths components. For the sake of comparison the
conventional MF detector, single user bound (SUB) and the
AMMSE-MRC detectors have been considered. Evidently, the
FLC-RLS achieves a much better BER performance than
other schemes, because of the use of a fuzzy variable
forgetting factor in response to the time-varying channels.
Also, results show that the FLC-LMS does not perform as
good as the RLS-based approaches in data demodulation, due
to its slow convergence speed, but the FLC-LMS provides a
much simpler implementation.

Fig. 4: MSE evaluation for the LMS algorithm (training sequence = 100). K =
3 users, G = 32 and Eb/N0 = 8dB.

Fig. 5: MSE evaluation for the RLS algorithm (training sequence = 100 & =
0.001). K = 3 users, G = 32 and Eb/N0 = 8dB.

500 1000 1500 2000 2500 3000
10
-1
10
0
Number of iterations
M
S
E
= 0.99
= 0.9
= 0.7

0 500 1000 1500 2000 2500 3000

10
-1
10
0
Number of iterations
M
S
E
= 0.001
= 0.01
= 0.1

SystemFLClms: 2 inputs, 1 outputs, 16 rules

e
2
(4)
e
2
(4)
. (4)
FLClms
(mamdani)
16 rules

Fig.6:
A
Fig. 6 System FLC-lms: 2inputs, 1 output, 16 rules

International Journal of Emerging Trends in Signal Processing
Volume 1, Issue 1, November 2012

37


Fig. 7: MBF plot of input 1 of FLC-LMS system

Fig. 8: MBF plot of input 2 of FLC-LMS system

Fig. 9: MBF plot of output of FLC-LMS system



Fig. 11 MBF plot of input 1 of FLC-RLS system

Fig. 12 MBF plot of input 2 of FLC-RLS system

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
e
2
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
S M L VL
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
e
2
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
S M L VL
SystemFLCrls: 2 inputs, 1 outputs, 16 rules
e
2
(4)
e
2
(4)
. (4)
FLCrls
(mamdani)
16 rules


Figure
5.35:
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
.
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
SML VL

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0
0.2
0.4
0.6
0.8
1
e
2
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
SM L VL
A
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
e
2
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
S M L VL

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
.
D
e
g
r
e
e

o
f

m
e
m
b
e
r
s
h
i
p
S MLVL

Fig. 10 System FLC-rls: 2inputs, 1 output, 16 rules

International Journal of Emerging Trends in Signal Processing
Volume 1, Issue 1, November 2012

38

Fig. 13 MBF plot of output of FLC-RLS system

Fig. 14: FLC-LMS output surface plot

Fig. 15 FLC-RLS output surface plot

Fig. 16 BER performance comparison of proposed scheme with MF and SUB,
G = 32, K = 3 users and M = 3 paths.
VIII. CONCLUSIONS
An adaptive MMSE-MRC-RLS and FLC-RLS-
based MMSE-MRC were introduced in this paper.
The major advantage of the adaptive MMSE-MRC
approach is its simplicity and it removes the use of
training sequences. The FLC-RLS detector provides
the use of variable forgetting factor. From the
presented results it is found that the adaptive
MMSE-MRC-RLS detector performs better than
conventional MF and the FLC-RLS-based detector
outperforms the conventional and AMMSE-MRC-
LMS detectors and performing nearly as well as
single user case in multipath channel.
REFERENCES
[1] Z. Xie, R. T. Short, and C. K. Rushforth, A family of suboptimum
detectors for coherent multiuser communications, IEEE Journal on
Selected Areas in Communications, vol. 8,no. 4, pp. 683690, 1990.
[2] G. Woodward and B. S. Vucetic, Adaptive detection for
DS_CDMA, Proceedings of the IEEE, Vol. 86, pp. 1413-1434,
Jul.1998.
[3] M. Ghotbi & M.R.Soleymani, MMSE-Based Adaptive Multiuser
Detection of DS-CDMA Signals, CCECE/CCGEL IEEE, May
2005.
[4] C. W. Jim, A comparison of two LMS constrained optimal array
structures, Proceedings of the IEEE, vol. 65, no. 12, pp. 1730
1731, 1977.
[5] J. G. Proakis, Digital Communication, 4th edition, McGraw Hill,
New York, 2001.
[6] V. Kecman, Learning and Soft Computing: Support Vector
Machines, Neural Networks, and Fuzzy Logic Models, MIT Press,
Cambridge, Mass, USA, 2001.
[7] W.-S. Gan, Designing a fuzzy step size LMS algorithm, IEE
Proceedings - Vision, Image and Signal Processing, vol. 144, no. 5,
pp. 261266, 1997.
[8] S. Song, J.-S. Lim, S. J. Back, and K.-M. Sung, Variable forgetting
factor linear least squares algorithm for frequency selective fading
0
0.2
0.4
0.6
0.8
1
0
0.5
1
0
0.1
0.2
0.3
0.4
0.5
e
2 e
2
.
A

0
0.2
0.4
0.6
0.8
1
0
0.5
1
0.485
0.49
0.495
0.5
0.505
e
2 e
2
.
A

International Journal of Emerging Trends in Signal Processing

Volume 1, Issue 1, November 2012

39

channel estimation, IEEE Transactions on Vehicular Technology,
vol. 51, no. 3, pp. 613616, 2002.
[9] S. Haykin, Adaptive Filter Theory, Prentice-Hall, Englewood Cliffs,
NJ, USA, 3rd edition, 1996.
[10] L. A. Zadeh, Fuzzy sets, Information and Control, vol. 8, no. 3,
pp. 338353, 1965.
[11] Wei Li, Hong-Chuan Yang and T. Aaron Gullver, A Practical Rake
Combining Scheme For Synchronous CDMA Systems, IEEE
Communications Society / WCNC 2005.
[12] Chetan Mahendra and Sadasivan Puthusserypady, Multiuser
Receiver for DS-CDMA Signals in Multipath Channels: in
Enhanced Multisurface Method, IEEE Transactions on Neural
Networks, Vol. 17, No. 6, November 2006.
[13] Markku Juntti, Multiuser Demodulation For DS-CDMA Systems
In Fading Channels, Ph.D. Thesis, Department of Electrical
Engineering, University of Oulu, FIN-90570 Oulu, Finland, 1998.
[14] S. Moshavi, Multi-User Detection for DS-CDMA Communication,
IEEE Communication Magazine, Vol. 34, pp. 124-136, Oct. 1996.
[15] A Duel-Hellen, J. Holtzman and Z. Zvonar, Multiuser detection for
CDMA systems, IEEE Personal commun. Mag., pp46-58, Apr.
1995.