2003 IEEE XI11 Workshop on

Neural Networks for Signal Processing

BAYESIAN AND RBF STRUCTURES FOR

W IRELESS CO M M U NICAT IONS DETECTION
L.M. San JosbRevuelta and J. Cid-Sueiro
Depto. de Teoria de la Serial y Comunicaciones e Ingenieria Telemitica
ETSI de Telecomunicaci6n, Universidad de Valladolid
Campus Miguel Delibes, 47011 Valladolid, SPAIN
Phone: +34 983 423660
Fax: +34 983 423667
E-mail: [email protected], [email protected]
Abstract. This paper presents t w o different algorithms for multiuser d e t e c t i o n in wireless D S / C D M A environments. First, a
Bayesian detector which implements merging techniques, based
on N a t u r a l C o m p u t a t i o n selection strategies, for complexity limitation, is analyzed, and, second, a low complexity R a d i a l Basis
Function-based detector is presented. B o t h approaches share in
c o m m o n a low computational load a n d the capability to be implemented even w i t h a high n u m b e r of active users, since their
complexity does not increase exponentially w i t h it. T h e i r perform a n c e a n d characteristics are compared w i t h those of traditional
multiuser detectors, such as the matched filter, the decorrelator
a n d the MMSE detector, as well as w i t h other low complexity det e c t o r s based on Evolutionary Computation methods.

INTRODUCTION
There is current a significant interest in the design of wireless Code-Division
Multiple-Access (CDMA) networks as there is a desire to be able to offer
wireless transmission services to accommodate a variety of diverse sources
[l,21. For instance, the new 3G systems -UMTS in Europe, IkIT2000 in USA
and Japan- include a CDMA scheme, since it allows an efficient use of the
spectrum, frequency reuse, soft hand off, increased capacity and multipath
combating.
In the CDMA access scheme all the users transmit simultaneously using
the same range of frequencies. Each user is distinguished by a unique code
(called spreading code or signature) which is used to modulate its binary
antipodal data. This signature waveform consists in a pseudorandom (PN)
sequence whose bit period (known as chip period) is smaller than the symbol
period of the user information. These user signalling waveforms must be

0-7803-8178-5/03/$17.00 0 2003 IEEE

749

orthogonal or, if not feasible, as uncorrelated as possible, so that the CDMA

detector can extract, free from interferences, the information signal of the
user of interest (UOI) [3].
Since wireless multimedia applications require high data transmission
rates, intersymbol interference (ISI) due to the multipath nature of the wireless channels can not be neglected and, together with the multiaccess interference (MAI) produced by the other cmchannel users, constitutes the major
drawback to the overall system performance [Z].
However, it has been demonstrated that multiuser detection provides very
substantial performance gains over the detection procedures conventionally
used in multiple-access channels. A proper design of the detector allows to
mitigate both IS1 and MA1 (31. The conventional single user receiver is a
filter matched to the spreading code of the user. In the presence of additive
Gaussian noise only (and no additional users), this receiver is optimal with
respect to the probability of detection error. However, this matched filter
(MF) suffers from the near-far problem, i.e., performance is severely degraded
by the effects of large received powers for the interfering users relative to that
of the desired user [4]. The optimal receiver for multiuser detection uses a
maximum-likelihood sequence detection (MLSD). Such a detection scheme
has a complexity which is exponential with the number of users. Due to this
complexity, attention has been devoted to the development of suboptimal
receivers with more moderate complexity. For instance, the decorrelating
detector achieves near-optimum performance when the users spreading codes
form a linearly independent set and retains the near-far resistance of the
optimal detector. While the optimal detector demands the knowledge of the
signatures of all the users as well as their received power, the decorrelating
detector access only to the powers [4].
Neural Networks (NNs) and other Natural Computation-based strategies,
such as Genetic Algorithms (GAS): share some interesting capabilities which
fit very well for the CDMA detection task. NNs are adaptive, near-far resistant and their intrinsic parallelism makes them computationally efficient.
Also, as other nonlinear receivers, NNs outperform linear detectors, since
the optimal decision boundary in DSjCDMA is nonlinear in nature [5] and
nonlinear receiver structures based on NNs or in polynomial series can a p
proximate the decision boundary well, and, hence, achieve superior performance. The performance of different NNs for multiuser detection has been
investigated in [4, 5 , 6, 7 , 81.
In this article we propose two different low complexity multiuser detectors (MUDS). First, a Radial Basis Function (RBF) network-based receiver,
working a t chip level and with reduced complexity and superior performance
over linear structures, is presented. Its efficiency is equivalent to a MLSD in
an AWGN channel. Second, we propose a novel Bayesian approach for MUD,
which was initially (theoretically and with no complexity limitation control
and, hence, impractical) developed in [9]. Now we go a little deeper into
this Bayesian detection scheme, adding some complexity limitation methods to it, and comparing its performance with the RBF detector as well as

750

with some traditional multiuser receivers. Also, in the simulations section,

we will compare the performance of the proposed (Bayesian/RBF) detectors
with that of a GA-based MUD with an extremely low computational load
(i.e., of the type of p-GA), whose internal parameters are on-line adjusted
by making use of fitness dispersion measures. GAS constitute an alternative
method for finding solutions t o highly nonlinear problems, characterized by a
multimodal solutions space, since they efficiently combine the horizontal and
vertical sense of search so as to avoid t o get trapped into these suboptimal
solutions.
The article is organized as follows: first, we describe the DSjCDMA system used. Next, the architectures of both methods proposed are presented.
Afterwards, we show some numerical results from Monte Carlo simulations
and, finally, some conclusions are drawn.

DSICDMA CHANNEL MODEL DESCRIPTION

Consider a binary communications system with K active users, employing
normalized modulation waveforms
and signaling through a dispersive channel with additive white Gaussian noise. We consider this problem in the context of coherent, synchronous Multiple-Access (DS/SSMA)
signaling. User i transmits a sequence of statistically independent symbols,
bi(n)E {*l}, which modulates the PN sequence, s i ( t ) ,so that the spectrum
is spread by a factor N (processing gain) [l]. Thus, the signal transmitted
by the i.th user is

{si(t)}gl,

z,(t) =

b;(n)si(t - nT)

(1)

where the user signaling waveforms are of the form

(2)

where s, = (sZ,o,. . . , S , , N - ~ )is~ the signature vector of user i, T, = T / N

is the chip interval, N the processing gain and + ( t ) the bandlimited chip
pulse signal, which is common to all the users and assumed to be rectangular
in shape. We will also suppose that their energy is normalized, so that

J,'.

ls,(t)l2dt = 1,vz.

Considering a synchronous transmission, the received signal consists in

the sum of the K users signals, in addition to a noise
K

75 1

with g ( t ) being a zero-mean additive Gaussian noise process with power spectral density U:, and E k representing the energy of user k.
At this point we must distinguish between receivers working at symbol or
chip level. Our RBF and Bayesian detectors will work a t the chip level, as we
will show later, so, going a little deeper into this formulation is necessary. In
this case, the receiver has a filter matched to chip pulse $(Tc- t ) ,a sampling
module and a taped linear shift register. When signal $(t) is chosen according
',f I$(t)12dt,
to the interference elimination criterion and with energy &$ = .
each symbol generates a sequence of N samples per symbol, given by
Tj(R)

nT+(j+l)T.

r(t)$(t - nT

- jT,)dt =

nT+jT,

T,

= ~ a , b i ( n ) a i ( j ) & + + g(t+nT+jT&o)dt,
~

j = O , l , ..., N - I

i=l

where we have introduced vector a = (al,. .. ,ax)so as to model a possible

flat fading component. Normalizing with respect to &Q, these N samples can
begathered in a (Nx1)-dimensional vector, rchip(n) = ( r o ( n ).,. . , r ~ - ~ ( n ) ) ~ ,
which can he written
K

rchip(n)=

siaibi(n)

+ g(n) = SAb(n) + g(n)

(4)

i=l

where si is the signature of user i, S = (SI,.. . , S K ) is a N x K matrix

with the transmitted signatures (emh column represents a user signature),
g(n) = (go(n),. . . , g ~ - l ( n )is) ~a N x 1 vector of noise, with components
gi(n) =
.',f g ( t nT jT,)$(t)dt, b(n) = ( b l ( n ) ,...,bK(TZ))T is the
vector containing the symbols transmitted by the K users in the n-th symbol
period, and A is a diagonal matrix holding attenuations ai. Once $ ( t ) has
been properly chosen, g(n) is a multivariant stochastic process of zero-mean,
(U; is
complex-valued, white Gaussian noise, with covariance matrix U ~ I N
the noise variance and IN the identity matrix of order N ) .
Finally, observations rCh'p(n)are processed by the multiuser RBF detector whose outputs are estimates of the transmitted symbols. It is assumed
that all relevant timing and phase information is available.

&

As mentioned before, the optimal single user detector consists of a hank of

filters matched to the set of users signatures. Though conceptually simple, its
main drawbacks are: (i) presence of MA1 when the number of users increases,
and (ii) the near-far effect. One of the possible solutions was proposed by
Verdli and makes use of the knowledge of all the users signatures [3]. Nevertheless, the optimal multiuser detector must evaluate an error metric over
the 2K possible combinations of vector b, and its complexity increases exponentially with the number of users, K , becoming, this way, unfeasible. This
problem has led to the development of many suboptimal detection schemes,
some of them based on Natural Computation techniques.

IS2

DETECTION ALGORITHMS
RBF MUD detector
Radial Basis Function networks have been shown to be efficient solutions
for highly non-linear problems, such as the DSjCDMA environment with
IS1 and MAI. Two different RBF schemes for multiuser detection, which get
completely defined when both the users code waveforms and the channel
impulse response are known, have been proposed in [4, 81. In this case, no
training period is needed and the RBF detector constitutes the symbol MAP
detector when .....................
the noise is Gaussian.

9"

Flat fading
mumuseer channe
............

Figure 1: Structure of the DS/CDMA channel and the RBF-based receiver

Considering multipath effect. which can be modelled with a discrete time
= (ho,.... h v ) , and assuming that the channel response does
not vary within the bit period, the signal associated to chip j t h of bit n is

FIR filter h

i=O

k=1

IS1 can cause the effect of h to last more than one symbol period. To
take this into consideration, whenever ( j -z) < 1, b k ( n ) s r , j - , is substituted
by b k ( n - 1 ) 8 k , N + j - - r , and if ( j - z) > 1, then b k ( n ) s k , j - z is replaced by
bk(n + l ) s k , j - ~ - . .
Figure 1 shows the structure of the DSjCDMA channel
model and the RBF receiver designed to work at chip level.
The shift register is of length N ( M - 1) (N in caSe of no 1%).Next,
we can see the a nodes of the RBF network (working at Tc),each of them
associated to one of the possible symbol combinations of the system users
(ZK). Finally, note that the last block will be a simple sign detector in the

753

case of binary symbols. As kernel function 4 we will use the basis function
based on the Mahalanohis distance measure [lo],

The knowledge of the covariance matrix S allows us to calculate the dispersion of the data in any direction [S, 101. The covariance matrix is derived
as: S = E { ( r ( k )- ci)(r(k)- c * ) ~ }where
,
ci is a noise free state of r ( k ) .
The eigenvectors of S are the principal axes of these hyperellipsoids and the
eigenvalues determine their length. If all the off diagonal elements of S are
zero, Eq. (6) would be reduced to the product of the univariate normal densities and a Gaussian basis function 4, depending on the Euclidean distance
/Ir(k) - cijl and the covariance U* (which determines the radius of influence
for the Gaussian shaped bell), could he used. However, this is not the case,
even with orthogonal spreading codes, due to multipath effects.
The centroids {ci} represent all the possible received signals, a t chip level,
in the absence of noise, associated to all the possible combinations of symbols
transmitted by the K users. In case no IS1 exists, there will be OL = 2K vectors
ci of length N.When IS1 is not negligible, the number of centroids must be
properly increased since adjacent symbols affect the current one.

Bayesian MUD detector

The Bayesian multiuser detector scheme proposed in this section is based on
previous work [9]. In that paper we carried out the mathematical analysis
of a Bayesian detector by means of extending preliminary work on Bayesian
equalization for single user environments [ll]. The algorithm developed in [9]
had a complexity that exponentially increases with the number of user and
with the length of the received symbol sequences. Thus, its interest was only
theoretical. The theoretical analysis was carried out in 191 takes two steps:
(1) data-aided estimation of a, assuming that the symbols transmitted by all
users are known at the receiver. This is the calculus o f p ( a l r ( n l + j ) , B(n))',
with B(n) being the matrix that contains all vectors b(n) from symbol 0 to
the current one: B ( a ) = [b(O) . . . b(n)]; (2) extension to the blind case,
where B(n) is an unknown parameter.
In this paper we use a cmssouer pruning or merging-based sampling selection strategy -taken from selection procedures used in Genetic Algorithmsso as to limit the maximum number of hypothesis which are processed simultaneously. Specifically, we merge together those hypotheses assuming very
similar symbol sequences. This approach has been successfully developed and
proven in single-user environments [ll]. Expressions for the probability density functions corresponding to different hypotheses about the transmitted
'In contrmt to vector rchip(n)defined in (4), which contains the samples corresponding
to the L chips of interval n, from now on, r(nL + j ) will represent the vector with all chip
samples, starting with the first chip of symbol 0 and ending with the j-th chip of the
current, n,symbol interval: (~o(O),rl(O),. . . , r j - ~ ( n ) , r j ( n ) ) .

154

symbol sequences are given in [9]. In order t o minimize the approximation

error, only hypotheses with similar features should be merged.
The proposed Bayesian MUD estimates the sequence of transmitted symbols b(n) once the (n+l ) L received chips (until the end of the n-th symbol
period) have been processed.
Considering the Bayesian MAP decision criterion, the estimate of the
symbol transmitted by user k at time n can be described as follows

The MAP decision of (7) consists in evaluating the posterior probability of

each symbol bl,(n) as the sum of the probabilities corresponding to those
hypothesis associated to the same decision about the symbol n of user k ,

j=O

where Bj,r(n) = airepresents the hypothesis about symbol n of user k in

matrix Bj,and I ( p ) is an indicator function that takes value 1 whenever
condition fl is satisfied, and 0 otherwise. In [9] we developed recursive expressions for the calculus of P{Bi(n)lr(nL j ) } .
This Bayesian multiuser detector considers the users signatures to be a
known parameter, though it does not need any information about the channel noise or fading. The structures which allow the implementation of this
Bayesian equalizer are characterized by a high parallelism and modularity.
In fact, they can be classified as RBF networks where a complezitg limitation
step is introduced between the block with the RBF neurons and the slicer,
which performs the bit sign decision.
A very interesting ability of the Bayesian MUD is its potential t o estimate the BER rates without any reference signal. This is achieved using the
capability of this equalizer to calculate the posterior probabilities associated
to every possible channel state (see [ll,121 for the single user case).

NUMERICAL RESULTS
Simulations were run using Gold sequences as vector signatures s, and considering additive Gaussian noise. Figure 2 shows the bit error rate (BER)
performance of the U 0 1 versus SNR=EI / N O ;with a near-far effect. of 4 dB, for
different multiuser detectors. K = 20 and N = z5. The error probabilities of
those traditional detectors based on MF and decorrelator are also shown. For
the shake of comparison, we have also included a Genetic Algorithm-based
detector developed in [13]. This p-GA works with a very small population
and the horizonal/vertical sense of the search is on-line adjusted by measuring
the diversity of the population by means of its fitness entropy.

755

First, we can see that both the p-GA and the Bayesian MUD have a performance close to the limit of the optimal single user detector (a lower hound
of the multiuser detector can he found by simulating the BER of the single user receiver in the absence of noise), while the traditional MF approach
saturates a t high SNRs. The decorrelating detector offers an intermediate
performance.

I+-

... . .

..j.

..... ...........................

. ..i.......

9
10
SNR (dB)

11

12

Figure 2: Probability of error for different multiuser detectors.

In terms of computational load, the most efficient method is the p G A .
This algorithm needs about 25% of the population needed for the classic GA,
and a number of iterations about 10-20%, to get a similar performance, i.e.,
saving ahout 90 % of time (131. Next, the chip level RBF network offers a nice
trade-off between computational load and efficiency. The Bayesian approach
is the most time consuming method, though it shows very good BER figures
and offers the capability of BER estimation, which can he of much interest
in many applications [11, 121.
In Figure 3, we compare the BERfor different receiver structures (Bayesian,
RBF, MF and MMSE) as a function of the number of users. Results were obtained from a stationary multipath scenario with a channel impulse response
H ( z ) = 0.3482 0.87042- 0.3482z- and E,,,
= 8 dB.
Linear M F and MMSE structures are outperformed by the nonlinear receivers. The MF suffers from MA1 and the MMSE lacks stability to perform
a nonlinear decision boundary. The choice of the Mahalanobis distance in
the RBF basis function allows some advantages over the Euclidean one. The
Euclidean-RBF performance tends towards the MMSE receiver performance
due to the non-spherical shape of the multidimensional clusters.
Finally, we consider a dispersive channel with multipath delay spread of
one symbol interval. The complex channel gains are generated according to
the complex Gaussian distribution with zero mean and unit variances, and
normalized to the zero delay component, i.e., h(0) = 1. The perfect power

756

Figure 3: Probability of Error vs number active of users.

control situation (all users have equal powers) and the near far situation
where each interfering user is 10 dB stronger than the UOI, is simulated.
The performance for K = 6 users is plotted in Figure 4.
The performance of the conventional MF detector, which is matched to
the true composite signal waveform, is also plotted. It can be seen that the
proposed algorithms are near-far resistant in the presence of MA1 and ISI.

CONCLUSIONS

Two different low complexity detectors have been proposed for multiuser detection in DS/CDMA environments with AWG noise, IS1 and MAI. The chip
level-based RBF detector, once trained, does not require to know neither the
energy of the users nor the matrix of correlations. The Bayesian MUD uses
merging techniques in order to keep the receiver complexity practical for real
life use. The performance of both algorithms is close to the limit,of the single
user detector and their complexity is much smaller than that of traditional
approaches. Simulations show that they are near-far resistant and their performance does not saturate at high SNR as happens in the traditional MF
method. Both MUDS can be implemented with a high degree of parallelism
and, in addition, the Bayesian approximation offers the capability of blind
BER estimation.

ACKNOWLEDGEMENTS
The author would like to acknowledge the Spanish ClCYT for research grant
TIC3808-C02-02.

757

loo
, . . ~ ~ + . . . . . ;.. ... ... . . I

- Id'

..

. . . . . . : . . . ..........,.
.. ...........:
:
:
.
.

1
,
I ..........

.-

in4

11

13

15

SNR (dB)

Figure 4: Probability of error for different multiuser detectors.

REFERENCES
[l] S. Glisic and B. Vucetic, Spread Spectrum CDMA Systems for Wireless Communications, UK: Artech House Publishers, 1997.
[2] X. Wang and H.V. Poor, Blind equalization and multiuser detection in dispersive CDMA channels, IEEE Ti-.on Comm., vol. 46, No. 1, Jan. 1998, 91-103.
[3] S. Verdd, Multiuser etection, UK: Cambridge: Cambridge University Press,
1998.
[4] U. hlitra and H.V. Poor, Neural network techniques for adaptive multi-user
demodulation, IEEE J . on Sel. Areas on Comm., vol. 12, No. 9, Dec. 1994.
[5] B. Aazhang, B.P. Paris and G.C. Orsak, Neural networks for multiuser detection in Code Division Multiple Access communications, IEEE Tr. on Comm.,
vol. 40, No. 7, July 1992, 1212-1222.
[6] D.G.M. Cruickshank, RBF receivers for DS-CDMA, Electronics Letters, vol.
32, No. 3, 1996, 188-190.
[7] P.Y. Peng, An efficient multiuser detection algorithm for DS/CDMA by using
neural networks, in Pvoc, IEEE ICC 2001, 2001, 36-40.
[8] R. Tanner and D.G.M. Cruickshank, RBF based receivers for DS-CDMA with
reduced complexity, in Porc. of ISSSTA, South Africa, Sept. 1998.
[9] L.M. San JosB-Revuelta and J. Cid-Sueiro, Extension of blind Bayesian equalization algorithms for multiuser detection, in Proc. IASTED Int'l Conf. on
Wireless and Optical Comm., Banff, Canada, July 2002.
[IO] C.W. Bishop, Neural Networks for Pattern Recognition, UK: Oxford: Oxford
University Press, 1995.
[ l l ] L.M. San JosbRevuelta and J. Cid-Sueiro, A neuro-evolutionary framework
for Bayesian blind equalization in digital communications, Sig. Proc., 83(2),
2003, 325-338.
[12] J . Cid-Sueiro, L. M . San J o s b h v u e l t a and A. R. Figueiras, RRBF based soft
decision Bayesian equalizers for in-service BER Monitoring, Sig. Proc., 80(4),
2000, 741-744.
[13] L.M. San JosB-Revuelta, Reduced complexit,y GA-based multiuser detectors
for DSJCDMA communications, Submitted to: Int'l Conf. on Communication
Systems and Networks, Malaga, Spain, Sept. 2003.

758