Least bit error rate adaptive nonlinear equalisers

for binary signalling

S. Chen, B. Mulgrew and L. Hanzo
Abstract The paper considers the problem of constructing adaptive minimum bit error rate
(MBER) neural network equalisers for binary signalling. Motivated from a kernel density
estimation of the bit error rate (BER) as a smooth function of training data, a stochastic gradient
algorithm called the least bit error rate (LBER) is developed for adaptive nonlinear equalisers. This
LBER algorithm is applied to adaptive training of a radial basis function (RRF) equdlialiser in a
channel intersymbol interference (ISI) plus cochannel interfercncc setting. A simulation study
shows that the proposed algorithm has good convergence speed, and a small-size RBF equaliser
trained by the LBER can closely approximate the performance of the optimal Bayesian e q u1' I'iser.
The results also demonstrate that the slandard adaptive algorithm, the least mean square (LMS),
performs poorly for neural network equalisers because the minimum mean square error (MMSE)
is clearly suboptimal in the equalisation setting.

Introduction

Equalisation techniques play a crucial role in coinbating

distortion and interference in modem coniniunication
systems. The topic of equalisation is well researched and a
variety of solutions are available [1-7]. Recently, neural
networks have k e n used as adaptive nonlinear equalisers
[X-171. Typically, adjusting an equaliser's parameters in a
sample-by-sample fashion is required in practical applications to meet real-time computational constraints; and
adaptive training of neural network equalisers is usually
carried out using some stochastic gradient algorithm that
tries to minimise the mean square error (MSE). This kind of
stochastic gradient adaptive algorithm is classically referred
to as the least mean square (LMS). It is a well-known fact
that adaptive training of neural network equalisers often
encounters difficulties. Since the MSE surface is highly
complex in the nonlinear equalisation setting, such difficulties are typically attributed to 'local minima'. However, the
problem is actually more fundamental than this. Surely it is
a strange situation that. on the one hand: the performance
of an equaliser is determined by its bit error rate (BER)
while, on the other hand, a different MSE criterion is used
at the lcarning stage.
For linear equalisers, there is a partial relationship
between the MSE and BER. A small MSE is usually
associated with a small BER. However, even in the linear
case, the minimum mean square error (MMSE) solution in
general is not thc minimum bit error rate (MBER) solution,
and it is well known that the BER &ap between thc linear
MMSE solution and linear MBER solution can he large in
certain situations [18-271. Since the BER is the true
0 IEE, ?MI?
IZZ Procm/hg.$onlinc n o ZllOiOZY4
doi: 10.I049/ip-com:?0030?84

Paper first receivcd 15t11 Jmury ZlH11 and iii rcvlsed form 20th Fehruay 2W2
S . Clien and L. Hatilo are nit11 the Depiinmsnl of Electronics and Computcr
Science. University of Southampton. Hifhfield. Southampton SO17 IBJ. UK
B. Mulgrciv is with the Depanmcnt o l Electronin and Electrical Engineenng,
University of Edinburgh. King's Buildings. Edinhuigh EH9 3JL. UK
I P~oc.-Cor,imus..Vo/ I S 0 A',,. I. F d ~ n r a aXUJZ

performance indicator for equalisation. recent research has

derived some adaptive linear MBER equalisers [20, 24-26],
In the linear case, the properties of the MSE and BER
criteria are also better known. The MSE surface is, of
course, quadratic with a unique global minimum in the
equaliser weight space. The BER surface is much more
complicated and the global minimum solutions fomi a half
line, one end of which approaches infinity and the other end
approaches the origin. The origin of the linear equaliser
weight space is the singular point (discontinuity) ofthe BER
surface (see, for example, [28]).
For nonlinear equalisers, both the MSE and BER
surfaces are highly complex and possess local minima.
Furthermore, unlike the linear case, a small MSE may not
correspond to a small BER. Consider, for cxample, the
inaximutn a posteriori probability or Bayesian symboldecision equaliser [7]. which is optimal in terms of BER.
The MMSE criterion in this case is inappropriate. This is
simply because. multiplying the Bayesian equaliser by a
positive constant, it remains as the optimal Bayesian
equaliser while the resulting MSE value can become very
large. It is worth noting that in theory, adaptive MBER
training can he achieved by adjusting the equaliser's
parameters only when an error occurs. Since the BER is
typically very small in communication systems, such a
strategy would require an extremely long training sequence
and is therefore impractical. Because practical adaptive
algorithms based on the MBER criterion are unavailable to
date for nonlinear equalisers. adaptive training of a neural
network equaliser is typically carried out using the LMS
algorithm, which may sometimes lead to poor BER
performance.
This paper describes the development of a stochastic
gradient adaptive MBER algorithm for nonlinear equalisers
with binary signalling. We adopt an approach similar to the
one used in developing an adaptive MBER linear equaliser
[25, 261- namely that of using a kernel density or Parzen
window estimate to approximate the BER from training
data and deriving a stochastic gradient adaptive algorithm.
The resulting algorithm is called the LBER because of its
links to the MRER criterion, in a manner analogous to the
29

LMS algorithm for the MMSE criterion. This LBER algorithm is tested with an equalisation application in the
presence of channel intersymbol interfarence (ISI), additive
white Gaussian noise and co-channel interference, where a
radial basis function (RBF) network is trained as an
adaptive equaliser. Simulation results obtained show that
the LBER algorithm achieves consistent performance and
'has a reasonable convergence speed. A small-size RBF
network trained by the LBER algorithm can closely approximate the optimal Bayesian performance. The simulation
study also confirms that a neural network equaliser trained
by the LMS algorithm, although converging well in the
MSE, can sometimes produce a poor BER performance.
2

has N I = 2"+"'-' combinations, denoted as bl.l. 1 < / < N I .

From the system model (I), the received signal vector can be
written as
v(k)=F(k)+n(k) =vo(k)+ul(k)+n(k)

It is obvious that F(k), Vo(k) and Fl ( k ) all have finite states,

that is,

r ( k ) =?(k)

(k)

+n(k)

lZ-1

= C a ~ . ; b ~ ( k - i ) + C a l , ; b l ( k - i ) + n ( k () I )
i=U

i=O

(10)

Fl(k)EWI~{vl,,, 1 < [ < N I }

(11)

and
F(k) E W ~ { r o J + v l , l :

I <.j<NUandl </<NI)

+ n ( k ) = ?o(k) +TI

n"-I

%a {Po.,,

I 5 j 5 No}

vo(k) E

Equalisation problem

It is assumed that the channel is modelled as a finite impulse

response filter with an additive noise source [29]. For
notational simplification, it is further assumed that there
exists only one relatively strong co-channel interference.
Thus, the received signal at sample k is given by

where n(k) is a white Gaussian noise with variance

E [ n 2 ( k ) =ut;
]
? ( k ) , ro(k) and r l ( k ) are the noise-free
received signal, desired signal and interfering signal,
respectively; U
,;,,
ti<i<nu-l, are the channel taps and
O<i<nu-l, are the cochannel taps; the desired
and interfering data bu(k) and bl(k) are binary, taking
value from the set { I}, and they are uncorrelated. Let
E[b;(k)]= U; and E[bi(k)] = U:. The signal to interference
ratio of the system is defined by

(9)

(I21

Notice that the set W contains N , = NO

x N I states, that is,
T(k)EWG{v;,1 S i S N , }

(13)

Denote the h h element of bui as b&. Since h e E { f l } ,

the sets WO and W can each be divided into two subsets W!:)
and .%(*I, depending on the value of b e . Thus b e is the
'class' label for y o j t . g o or vnJ-+ q.,t .%. For notational
convenience, define
bjd) = b e >for 1 5 i 5 N,
withj = I

where 1 . 1 denotes the floor function, i.e. 1. J is the largest

integer less than or equal to x.It can be seen that bjd' serves
as the class label for v; E W.
If the channel and co-channel are known, applying the
maximum a posteriori probability principle leads to the
following optimal Bayesian equaliser [30]:
YB(k) =fB(v(k); w )

the signal to noise ratio is defined by

(3)
and the signal to interference plus noise ratio is defined by

The equaliser considered in this study has a sample-decision

finite-memory structure, and it uses the information
contained in the received signal vector

r ( k ) = [r(k)r(k- I ) . . . r(k - m + I)]'

(5)
to produce an estimate of b,,(k-d), where m is the equaliser
order or memory length and d the decision delay [q.
Specifically, the equaliser is defined by
&,(k - d ) = s g n ( y ( k ) ) with
Y ( k ) = f ( r ( k ) ;w )
(6)
where f ( . ; . ) denotes the equaliser map, and w consists of
all the (adjustable) parameters of the equaliser. Such an
equaliser has finite 'states', since

bo(k)=[bo(k)bo(k- I ) . . . b o ( k - n o - m + 2 ) ] '

(7)

has NU= 2m+n0-1 combinations, denoted as buj, 1 <j< Nu,

and
bj(k)=[bl(k)bl(k- I ) - . . b i ( k - n l - m + 2 ) ] '
30

(8)

where equiprobable states in .%U and 91 have been

assumed, and r i E W as is defined in (13). It should he
pointed out that, multiplying y d k ) by a positive constant, it
remains the optimal Bayesian solution even though the
resulting equaliser has a different MSE value. It is also clear
that this Bayesian solution has very high computational
complexity. An often-used equaliser is the linear one:

Y&) = f L ( r ( k ) ; w ) = w'+)
(16)
The most popular solution for the linear equaliser is the
MMSE solution, which can readily be implemented
adaptively using the LMS algorithm.

Derivation of the LBER algorithm

Consider the generic equaliser (6) where the equaliser map1

is realised, for example, by a neural network. Classically.
such a nonlinear equaliser is trained by adjusting w so that
the MSE, E[(bo(k-d)-y(k))2],is minimised. Typically, this
is implemented adaptively using the LMS algorithm. which
IEE

h-oc.-Comnziin..

Vu/ ISV. No 1. Fehmrji 2003

has a very simple form:

and

y(k)=./(r(k); w(k - I))

~ ( k=) w(k - 1 ) p(bo(k - d )
- y ( k ) ) a f ( r ( k ) :+I))

i)v

where is an adaptive gain. However, minimizing the MSE

does not necessarily produce a small BER. A main objective
of this study is to derivc an adaptive algorithm for the
nonlinear equaliser (6) based on the MBER criterion.

3.2 Approximate minimum bit error rate

solution

3.7 Approximate bit error rate expression

The error probability of the equaliser (6) is
f . ( w ) =Prob{sgn(ho(k-d))y(k)<O}

(18)

Define the signed decision variable

y s ( k ) = sgn(bo(k - d ) ) y ( k )

In general the linearisation (21) is valid only for small n(k)

(in some statistical sense). The level of approximation
involved in (21), however, becomes insignificantin coniparison with any stochastic (one-sample) approximation
needed in deriving adaptive training.

If the channel and co-channel are known, an approximate

MBER solution for the equaliser (6) can be obtained by
minimising the approximate BER expression (26). The
gradient of P&{iv)is approximately:

(19)

and denote the probability density function (p.d.f.) of y.Jk)

as p,(v,). Then

l,
0

PE(w)=

P , ( Y ~ )dys

(20)

By linearising the equaliser around i f k ) , the noise-free

received sigma1 vector, it can be approximated as
y(k) = f K k )

+@I;

w)

or
y ( k ) = f ( q k ) ;w )

+e(k) = ?(k) +e(k)

(22)

where ~ ( kis) Gaussian with zero mean and variance

with r, t ,R Basically, the equaliser is approximated as an

additive Gaussian noise model, with the clean simal T ( k )
taking values from the finite set

X d E { n = f(c; w ) ,

1 5i 5W }

(24)

(29)
In the last approximation of (29), we have dropped the term
containing ap@w. That is, p is assumed to be independent
of w . In general, p depends on the value of w, unless the
algorithm has already converged to the (near) optimal
solution wMRBRand p has been fixed to its optimal value.
Theoretical justification for this approximation still needs
to be investigated, but it could be argued on the ground
that this approximation is less significant than the approximation involved in deriving a stochastic one-sample
adaptation.
The following iterative gradient algorithm can be used to
amve at an approximate MBER solution. Given an initial
w(O)> at the Ith iteration. the algorithm computes:

w ( l ) = MI(/ - I ) - ) I V P E ( W ( I ) )

Although the variance p2 could iteratively be calculated by

The p.d.f. of y,(k) can thus be approximated by

and the error probability of the equaliser is approximately

for numerical and convergence considerations, it is preferred

to fix [I to an appropriately chosen constant. Thus, p is
considered as an algorithm parameter that requires tuning.
Notice that if the equaliser (6) is linear, all the approximations done so far can be made exactly, and the variance
of e(k) is p = u:wiw. In this case, we arrive at the exact
MBER solution for the linear equaliser (16) [25, 261.

3.3 Block-data gradient algorithm

In practice, the set is unknown. The key to developing an
effective adaptive algorithm is the p.d.f. p J y J of the signed
where

decision variable y,(k). Kernel density or Parzen window

estimation is known to produce reliable p.d.f. estimates with
short data records and in particular is extremely natural in
dealing with Gaussian mixtures [31, 321. Given a block of K
training samples {(k), bo(&#)}, a kernel density estimate of
31

the p.d.f. is

Simulation study

To test the LBER algorithm for adaptive training of lieurid

network equalisers, the RBF network equaliser of the form
YRBF(k) = . f K B F ( r ( k ) ;

(32)

From the estimated error probability

U

PE(w ) =

/ h(,%)
djA

(33)

-cc

VPE(w)can he calculated:

x sgn(bdk

a/+@); w)

4)

a,v

is used. The equaliser parameter vector w thus contains all

the RBF weights ni, widths Z j and centres L;. The dimension
of IV is therefore
x (in + 2). The derivatives of the
equaliser output with respect to the equaliser's parameters
i n this case are given by

(34)

Thus a block adaptive gradient algorithm can be derived.

At the /th iteration, the algorithm computes:
y$f) = f ( r ( k ) ; w ( l -

I)),

I 5 j 5 iiC

I5 k 5 K

VPc(w(/)) = 1

KJ2?lpk,,

xexp ( - a ) s g n ( b " ( k - d ) )
2p'

w(/)= l V ( 1

1)

8 f ( r ( k )a,;l,w(/- 1))
'

-/IVPE(W(/))
(35)

3.4

Stochastic gradient algorithm

Our real aim is to develop a stochastic gradient adaptive

algorithm with sample-by-sample updating, in a similar
manner to the LMS (17). The LMS algorithm is derived
from its related ensemble gradient algorithm by replacing
the ensemble average of the gradient with a single-datapoint estimate of the gradient. Adopting a similar strategy,
at sample k, a single-data-point estimate of the p.d.f. is

For comparison purposes, both the LMS and LBER

algorithm are used to train the RBF equaliser (39), with the
adaptive gain /I given in the form pk=p,&-''4, where p,, is
an appropriately chosen constant. For the LBER algorithm, the value of p' also needs to be determined.
In all simulations, the first 1iJ2 data points that belong to
the class + I and the first n,/2 data points that belong to the
class - 1 are used as initial centres. The initial weights are set
accordingly. All the widths are initially
to & I /[nC(2no~)""']
set to 80:. Two kinds of BER are used in the simulation
investigation, the true BER that is computed using Monte
Carlo simulation with a sufficiently long test sequence and
the estimated BER calculated usiiig the approximate BER
expression (26) with ql(w) = sgn(bjd')y;/P. The value o f p
is so chosen that the estimated BER agrees with the true
one. This p should not be confused with the algorithm
parameter p used in the LBER. The estimated BER is used
to illustrate the learning rate of an adaptive algorithm, since
computing the true BER of the adaptive RBF equaliser at
each sample k would be computationally too demanding.
Exuiiiple I . The transfer functions of the channel and cowere respectively A&) =0.5+ 1.0~-l and
Al(~)=i.(1.0+0.5z~'). The value of i was set to give an
SIR = 12 dB. The equaliser order was chosen to be i?i = 2
and the decision delay d= I.The sets 9uand 91each had

channel
(36)

Using the instantaneous or stochastic gradient,

a stochastic gradient algorithm is given by

y(k)= f(r(k); w(k - I))
w(k) = w(k

I) + G~ ep x p ( - g )

k):4-l))
Sgn(bo(k d ) ), ? f ( * ( a,,,

assuming that an appropriate p has been chosen. The

influence of p on algorithm performance will be investigated
in a simulation study. Following similar reasoning to the
LMS for the MMSE criterion, the algorithm (38) is called
the LBER algorithm.
32

eight points, and the sct 9 had 64 states. Fig. 1 shows the
together with the decision boundaries of the
sets WOand 9,
linear MMSE and optimal Bayesian equalisers given
SNR=20dB (SINR= 11.36dB). Given this noise and
interference condition, RBF equalisers with four and six
centres were trained by the LMS and LBER algorithms,
respectively. At each sample k? the estimated BER was
calculated for an equaliser with w(/r), and this resulted in the
learning rates plotted in Fig. 2 for the corresponding
adaptive equalisers, where the results were averaged over
100 runs. The LBER algorithm had p' = 200; and
/il,=0.15 for the four-centre RBF equaliser, and p2 = 20;
and /b=O.I for the six-centre RBF equaliser; while the
LMS algorithm had hiu=0.5 for the four-centre RBF and
p,,o=O.4 for the six-centre RBF. For LMS training, the
MSE for an equaliser with w(k) was also calculated using a
block of 100 test samples, and this produced the learning
Ifil? P ~ ~ , ~ . - C " ~ , , , Vol.
, , ~ ~150.
, , . , No. 1. Frbnwri ZWj

6-LMS

1 :,.

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

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

I/)

10-1

600
sample k

200

600

400

1000

CLMS: the 4-centre KBF trained by the LMS with }b=0.5:

6-LMS: the 6sentre RBF trained hy the LMS with }i1,=0.4

'k

4-LBER

b, ( k -3

6-LEER

-3

I"

200

400
600
sample k

800

1000

Fig. 2

Coiiiwgence rates ;I? t e r m of rhe c . s r i t w r d REK, uwruqed

o w 100 runs, Jix c ~ c m p l eI
S I R = I Z d R a n d SNR=2OdB
4-LMS: the 4-ccntrc RBF trained hy the LMS with p,,=O.5:
6-LMS: the bsentre RBF trained by the LMS with ib=0.4
4-LBER: thc 4cenlre RRF trained by the LBER with p,,=O.l5 and

}? = ?Oui:
6-LBER: the 6-cenlre RBF lwinsd by the LREK with p0=O.1
p ? = 202

and

rates in terms ofthe MSE given in Fig. 3. where again the

results were averaged over 100 runs.
It was found that the estimated BERs of an RBF
equaliser with LMS training varied greatly for different
runs. For the four-centre RBF with LMS training, on
average, the estimated BER was hardly reduced, even
though there was around 7dB reduction in the MSE, as can
be seen in Figs. 2 and 3. For the six-centre RBF with the
LMS training, in some runs the BERs were close to that
obtained by the LBER training and for other runs the
BERs were very poor, resulting in, on average, a BER
larger than that of the linear MMSE equaliser. In contrast,
examining the MSE learning rate in Fig. 3 closely shows
that the corresponding MSE reductions were consistent in
different runs. This is not surprising, since the LMS is
IEE P~~,~.-C~~,,,,,,,,,,.,
Vol. IXJ.rV,r I , lihrm,or,, 2NJ3

.*. . .

* x o t

4-LMS

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

1)

= -1

-2

-1

r(iO

Fig. 4 Compor.i.soon ofnro d~i.~io,i

horo!rkirie.s (thin .solid:uduptice
RBF equriliser, thick solid: opiiiiml) /or evo,i?p/c I
S I R = 1 2 d B and SNR=ZOdB
The adaptive RBF cqoaliscr has four centres and is trained by lhe
LBER a l g A h m . The stars indicate the l i i d centre positions

designcd to minimise the MSE not the BER. In comparison, in terms of BER perfomiance_ LBER training was
found to produce consistent results in different runs. and
the six-centre RBF equaliser with LBER training converged
consistently very close to the optimal performance of the 64state Bayesian equaliser.
Typical decision boundaries of the four-centre and sixcentre RBF equalisers trained by the LBER algorithm are
compared with the optimal Bayesian boundary in Figs. 4
and 5. respectively. The (true) BERs of the four-centre RBF
equaliser trained by the LBER together with those of the
linear MMSE and optimal Bayesian equalisers are depicted
in Fig. 6 as functions of SINRs. The BERs of the six-centre
RBF equaliser trained by the LBER are not shown here, as
they are almost indistinguishable rrom the optimal performance. Fig. 7 depicts the same BERs of the three equalisers
as functions of SNRs. The influence of the algorithm
pirameter p2 on the performance of the LBER algorithm
was also investigated. Fig. 8 shows the BERs of the fourcentre RBF eqiialiser trained by the LBER algorithm with a
33

O I

bo(k-l)=l

4-LBER

2 -

-1

1 -

'

. .. .

0 -

-2

4
bo(k-l) =-I

-3

-3

-2

-1

3
0

Fig. 5 Compmison oJ',f,iro fircision bouridaries (tliinsolid mluptioe

RBF equa1i.w thick solid: optimal) for exumpie I

SIR=IZBand SNR=20dB
The adaptive RBF eqwzliser has six centres and is trained by the
LBER algorithm. Thc stan indicale the findl centre positions

, i

-5

r (4

10

20

30

,
40

signal to noise ratio, dB

Fig. 7 Pcr/urmunce mnporison of three equlkcrs in rerms qf

BER qainsr S N R /or Pxrrmpt I
SIR= l2dB
The adaptive REF equaliser has four centres and is trained by the
LBER algorithm

4-LEER

oolimal

-3

4-LEER

10

100

I000

p 2 7'/:

11

13

SINR, dB

Fig. 8 Inflfienceo/pz 011 the peifirmunce ofthe LBER alyorirhm

f i r example I ivilh SIR = 12 dB and S N R =20 dB
The adaptive RBF equaliser has four centres and the algorithm has a
tixcd

@
,

Fig. 6 Performmice C O I I ~ O ~ ~'of

S O rhree
~I
equalisers in terms uf
BER aqaimr SINR .for example I

SIR=lZdB
The adaptive RBF equaliser has four centres and is trained hy the

LBER algorithm

range of p2 in a gwen noise and interference condition,

where it can be seen that the algorithm performance is not
overly sensitive to p2 over a large range of values.
Example 2. The channel transfer function was
Ao(z)=0.3482+0.8704z-1 +0.3482rC2, and the co-channel
transfer function was A,(z)= 7, (0.6+0.8z-') with the value
of 2 chosen to give an SIR = 20 dB. The equaliser order was
set to m = 4 and decision delay d = 1. For this example, the
34

channel state set .go had 64 points and the co-channel state
set .9, had 32 points. Thus, the number of states in W was
2048, and it was computationally too demanding to
implement the optimal Bayesian equaliser. Given
S N R = 18dB (SINR= 15.8SdB), RBF equalisers with 16
and 32 centres were trained by the LMS and LBER
algorithms, respectively. The learning rates in terms of the
estimated BER are plotted in Fig. 9 for the respective
adaptive equalisers, where the results were averaged over
100 runs. The LBER algorithm had p2 = 100; and h= 0.3
for thc 16-centre RBF equaliser, and p2 = 80; and kl=0.3
for the 32-centre RBF equaliser; while the LMS algorithm
had &=0.3 for both the 16-centre and 32-centre RBF
equalisers. For the LMS training, the MSE convergence
IEE Pm.-Cotfimu~.,Vu/. 150. No. I , Fcbrur? 2W3

'I

~------.-.............................

16-LMS

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

-l

s
I

lo-l

.-

2E

10-2

.-I

1
0

+
.-

32-LBER

400

800

1200

1600

2000

sample k

Fig. 9 Conoergence rules in t c m s oftlie estimated BER, uoeruqed

mer 100 rum, .for e.rumple 2
S I R = 2 0 d B and S N R = 18dB
16-LMS: the 16-centre RBF trained by the LMS with &=0.3;
32-LMS: the 32-cenlre RBF trained by the LMS with / k , = 0 . 3 ;
IGLBER Ihe 16-centre RBF trained by the LBER with ~ = 0 . 3and
p2 =

lor;:

10

32-LBER: the 3 2 e n t r e RBF trained by the LBER with ,4=0.3;

02 = 80;

14

18

22

SINR, dB

of three equ1i.ser.i in lerms of

Fig. 11 Performance corq"n
BER uuainsr SINR fur e.xumrJk 2

SIR=ZOdB
Adaptive RBF equalisers have 16 and 32 centres. rcspectively, and are
trained by the LBER algorithm
32-LMS

-l

0.1

I
0

400

800

1200

1600

2000

sample k

Fig. 10 Convergence r a m in terms of the MSE, uwrugd w e r

100 runs, for e.xmpIe 2
SIR=20dB and S N R = 18dB
IbLMS: the I k e n t r e RBF trained by the LMS with &=0.3:
32-LMS: the 32-centre RBF trained by thc LMS with /*=0.3

performance, averaged over 100 runs. are also given in

Fig. 10.
For the 16-centre RBF equaliser with the LMS training,
the estimated BER on average was extremely poor, even
though the corresponding MSE reduction was reasonably
large. For the 32-centre RBF equaliser with the LMS
training, the estimated BER on average was no better than
the performance of the linear MMSE equaliser. Again, it
was found that the estimated BERs varied greatly for
different runs. In some runs, the 32-centre RBF equaliser
with the LMS training were as good as the LBER training,
but for other runs the results were very poor. The LBER
algorithm was seen to produce consistent results. The BERs
of the linear MMSE equaliser as functions of SINR and
SNR are compared with those of the adaptive LBER 16centre and 32centre RBF equalisers in Figs. 11 and 12.
respectively.
IEE Proc-Clinimun.

1'01. 150, No. 1. F e b m q

ZNU

10

15

20

25

30

signal lo noise ralio, dB

Fig. 12 Perfbrmunce conipurison .f three equ1i.cers in lei"

BER uguinrl SNRfor e-ample 2

of

S I R = 2OdB
Adaptive RBF equalisers have I6 and 32 centres. and arc traincd by
the LBER algorithm

Conclusions

A novel adaptive stochastic gradient algorithm called the

LBER has been developed for nonlinear equalisers. This

algorithm has a similar simplicity to the LMS hut is directly
linked to the MBER criterion, the real goal of equalisation.
This LBER algorithm has been applied to train an adaptive
RBF equaliser in the presence of channel ISI, co-channel
35

interference and additive Gaussian noise. Simulation results

have shown that the LBER algorithm achieves consistent
performance and has a good convergence speed. In
particular, a small adaptive RBF equaliser trained by the
LBER can closely approximate the optimal Bayesian
performance. The results also confirm that an adaptive
nonlinear equaliser trained by the LMS algorithm may
produce a poor BER performance, even though it
converges consistently in the MSE. Further research i s
warranted to investigate theoretical convergence analysis of
the LBER algorithm.

References

I OURESHI. S.U.H.: 'Adaotivc eilualisation'. Pruc. IEEE. 1985. 73.

(9). pp. 1349-1387
2 CHANG, R.W., and HANCOCK, J.C.: 'On receiver structures far
channels having memory', IEEE Trans. lnf Theor.", 1966, 12. (4) pp.
463-468
3 ABEND. K., HARLEY, T.J. Jr., FRICHMAN. B.D., and
GUMACOS, C.: 'On optimum receivers for channels having
memory'. IEEE Tronr In$ Theory. 1968. 14. pp. 818-819
4 ABEND, K., and FRITCHMAN, B.D.: 'Statistical detection for
communication channels with intersymbol interfcrenu'. Proc. IEEE
1970. 58,(5). pp. 779-785
5 FORNEY, G.D.: 'Maximum-likelihood squence estimation of digital
sequences in the presence of intersymbol interference'. IEEE Tmnm
lnf Theory. 1972. 18, (3). pp. 363378
6 WILLIAMSON, D., KENNEDY, R.A.. and PULFORD. G.W.:
'Block decision feedback equalization', IEEE Tmnr Cmwmm., 1992,
40. (2). pp. 25s22M
7 CHEN, S.. MULGREW. B., and McLAUGHLIN, S.: 'Adaptive
Bayesian equaliser with decision feedback', IEEE T r m . Signal
P r o c ~ . ~ .1993.
~ . . 41. (9). pp. 2918-2927
8 GIBSON, G.J., SIU. S., CHEN, S., COWAN. C.F.N., and GRANT.
P.M.: 'The application of nonlinear architectures to iidiiptive channel
equalisation'. Praceedings of ICC'YO, Atlanta, USA, lYiM,pp.
312.8.1.-312.8.5
9 SIU. S., GIBSON, G.J., and COWAN. C.F.N.: 'Decision feedback
equalisation using neural network stmctures and performance
comparison with the srandard architmure'. 1EE Proc, I, Gv"un.
Spercli Vis., 1990, 137, (4), pp. 221-225
I O CHEN, S.. GIBSON, G.J.. COWAN, C.F.N., and GRANT, P.M.:
'Adaptive equalization of finite non-linear channels using multilayer
perceptrons', Signal Procesr. 1990, 20, (21, pp. 107-1 19
1I CHEN, S., GIBSON, G.J.. and COWAN. C.F.N.: 'Adaptive channel
equalization using a polynomial-peruptron structure', IEE Proc. I,
C o w " Spmh Vir. 1990, 137, (5). pp. 257-264
12 GIBSON, G.J.. SIU. S., and COWAN. C.F.N.: 'The application of
nonlinear structures 10 the reConstwctiun of binary signals', IEEE
T r m . Sigmd P~ocP.~.~.,
1991. 39, (8). pp. 1877-1884
I3 CHEN, S., GIBSON, G.J., COWAN. C.F.N.. and G R A N , P.M.:
'Reconstruction of binary signals using an adaptive radial-basisfunction equalizer'. Sigrid Procei.~..1991. 22. (I), pp. 77-93
~~

36

14 CHEN. S., McLAUGHLIN, S.. and MULGREW. B.: 'Complexvalucd radial basis function network, Part 11: applicslion 10 digital
communiciitions cliannel equalisation', Sigwl prom^^., 1994, .%, pp.
17ClXX

15

16 CHA. I., and KASSAM. S.A.: 'Channel equalis;sion rising adaptive

complex radial basis function networks'. IEEE .I Sei Arein C O W ~ K ~ , ~ . ~
1995, 13, (I).pp. 122-131
17 CHANG, C.H., SIU. S.. and WEI, C.H.: 'A polynomial-peruptior
based decision feedback equalizer with ii robust learning algorithm',
S i p u i pro re.^.^., 1995. 47, pp. 145-158
18 SHAMASH, E., and YAO. K.: 'On the structure and perform i m e of a linear decision feedback equaliser based on the
ininiinum e n m probability cfiteion'. Proceedings of ICc'74. 1974.
pp. 25FI-25F5
I9 CHEN. S.. CHNG. E.S.. MULGREW. B.. and GIBSON. G.:
'Minimum-BER linciirsombiner DFE'. Proceediiirs of ICC96.
Dallas. Texds, 1996. Vol. 2. pp. 1173-1177
20 YEH, C.C.. and BARRY, J.R.: 'Approximate minimum bit-error rate
equalintion for binary signaling'. Proceedings of ICC97. Montreal.
Canada, 1997. Vol. 2. pp. 1095-1W9
21 CHEN. S.. MULGREW, B., CHNG. E.S.. and GIBSON, G.: 'Space
translation properties and the minimum-BER linear-combiner DFE.
/E P w c . C m " t . , 1998. 14s. (5). pp. 316-322
22 CHEN. S.. and MULGREW. B.: 'The minimum-SER linearcombiner decision feedback equalizer'. IEE Pro<., C r ~ n w w r .1999.
146. (6). pp. 347-353
23 CHEN. S., GUNN. S.. and HARRIS, C.J.: 'Decision feedback
qualiser design using support vector machines'. IEE Proc. Vis. Inuuqi.
2000. 147. (3). pp. 213-219
24 YEH. C.C., and BARRY. J.R.: 'Adaptive minimum bit~errorrate
eqoalizition for binary signaling'. / W E Tr". Cmnmim. 2000. 48.
(71, pp. 1226-1235
25 MULGREW. B., and CHEN, S.: 'Stochastic gradient minimum-BER
decision feedback equaliscn'. P r d i n g s of lEEE Symp. on ;daptive
systems for signal mocessinr. commiiniciitiiin and control. Lakc
iouix, ~ i b e ~Canada.
ta
12act. 2000. pp. 93-98
26 MULGREW. B., and CHEN, S.: 'Adaptive minimummER decision
2001. 81. (7).
feedback equalisers for binary signalling'. Sirjnoi Proce.~.~.,

-pp

Il?OLlAY<I

17,,-11"1

27 ALTEKAR, S.A., VITYAEV, A.E., and WOLF, J.K.: 'Decisionfeedback equalization via scpamting hyperplanes'. IEEE Tr0rz.i
Com,nm.. 2001. 49, (3). pp. 48W86
28 CHEN. S., SAMINGAN, A.K., MULGREW, B.. and HANZO, L.:
'Adaptive minimum-BER linear multiwtr detection for DS-CDMA
2001. 49.
signals in multipath channels', IEEE Tram Siguol Prococu~.~..
(6), pp. 1241247
29 PROAKIS. J.G.: 'DiLllal communications' (McGrawHill. New
York, 1995, 3rd edn)
30 CHEN, S , McLAUGHLIN, S , MULGREW, B , and GRANT,
I

31

32 BOWMAN, A.W.. and AZZALINI. A.: 'Applied smoothing

techniques for data analysis' (Oxford University Press. 1997)