Vous êtes sur la page 1sur 5

Multi-User MIMO Precoder Design via Genetic Search

Mouncef Benmimoune and Daniel Massicotte, Senior Member, IEEE


Universit du Qubec Trois-Rivires, Electrical and Computer Engineering Department
Boul. des Forges, Trois-Rivires, Qubec, CANADA
Laboratory of Signal and System Integrations
Email:{mouncef.benmimoune, daniel.massicotte}@uqtr.ca
AbstractThis paper presents a new precoding approach based
on a genetic algorithm to search for coded codewords applied to a
multi-user MIMO system. The main idea hinges on the definition
of coded codewords in which no codebook is previously stored at
transmitter and receiver. In this manner, the precoder aims to
maximize the sum capacity in order to solve the optimization
problem by searching for the best coded codewords for all users.
The performance evaluation and complexity analysis show that
our approach offers outstanding performance and low
complexity compared with the multi-user MIMO precoding
scheme based on the Grassmannian codebook. The proposed
method uses a multiplier-free precoding suitable for VLSI
implementation.
Keywords-multi-user MIMO; precoding; perfect CSIT; genetic
algorithm

I.

INTRODUCTION

In wireless communication, the use of multiple antennas at


either end, also called the Multi-Input Multi-Output (MIMO)
technique, allows system capacity to increase linearly with the
minimum number of antennas at both the transmitter and
receiver sides. MIMO systems also enhance the link reliability
of the transmission system [1]. Nevertheless, these benefits are
achievable only when the receiver knows the channel state
information (CSIR). Channel knowledge at the transmitter
(CSIT), however, offers further performance enhancement,
since it helps to increase the transmission rate, removes
interferences in the multi-user case and therefore reduces the
complexity of the MIMO receiver [2]. In the literature, the
precoding technique targets these advantages by exploiting
CSIT in order to customize the signal before transmission.
Furthermore, precoder design depends on the performance
criterion and the kind of CSI available at the transmitter. The
latter can often be obtained by two methods. In a time division
duplexing (TDD) communication system, the CSIT can be
inferred from the receiver channel through reciprocity, since
the channel is the same viewed from both sides of the wireless
link. In the frequency division duplexing (FDD) mode, on the
other hand, the receiver is required to transmit channel
knowledge at the transmitter by feedback. However, it is
impractical to send the full CSIT, especially in the case of a
rapidly time-varying wireless channel.
The use of a codebook-based limited feedback strategy to
provide partial CSI to the transmitter was proposed to solve this
problem [3], [4]. In this case, a codebook known to both sides
of the wireless link is considered, and for each channel
realization only the index of the codeword is sent to the
transmitter. In [3-7], Grassmannian subspace packing was used
to design the codebook. By cons, in [8], a Kerdock codebook is
used. In all codebook designs, the construction of the codebook

is completed by optimizing a cost function. This requires an


exhaustive search of all possible solutions, especially when an
increase is observed in codebook size and the parameters that
characterize the system. Accordingly, the codebook principle is
based on the idea that all channel realizations are divided into
several groups where each group corresponds to different
codewords. In other words, the precoding vector does not
correspond perfectly to each channel realization, which limits
this approach. Several new approaches have recently been
proposed aimed at reducing the exhaustive search. For instance,
in [9] the authors suggest an optimal multi-user MIMO linear
precoding scheme with LMMSE detection based on particle
swarm optimization (PSO), but this requires huge complexity.
In [10] the Grassmannian codebook is considered, and the
choice of precoding vectors that maximize the sum capacity is
accomplished with a genetic algorithm (GA) where a
favourable comparison was done with [6]. To reduce the
complexity of the codeword search, some works propose a
different approach that is better suited for practical VLSI
implementation (e.g. [7], [8]).
In this article, we propose a new near-optimal multi-user
MIMO precoder based on a complex value coded by a binary
symbol. The genetic algorithm is used to solve the optimization
problem at issue while providing a low-complexity search
vector precoding suitable for VLSI implementation. Notice that
in our approach a codebook does not need to be stored at the
receiver and transmitter.
For notation, we use lower-case bold letters for vectors and
upper case bold letters for matrices; denotes the Euclidean
norm of the vector; | | denotes the absolute value of ; and
denote Hermitian and matrix inversion, respectively. The
identity matrix is denoted by .
The remainder of this paper is organized as follows:
Section II presents the multi-user MIMO system model.
Section III gives the precoding principle based on a genetic
algorithm. Simulation results and complexity analysis are
presented in Section IV, and Section V draws the conclusion.
II.

MULTI-USER MIMO SYSTEM MODEL

We consider a MIMO broadcast channel in which a base


station (BS) equipped with
antennas communicates with
mobile stations (MS) and where each MS is equipped with
antennas, as depicted in Fig. 1. After precoding, the equivalent
MS can be
baseband input-output relationship of the
written as
(1)
where is the data symbol to be communicated to the MS,
denotes the
precoding matrix,
is the
flat

This paper is sponsored by the Natural Sciences and Engineering Research


Council of Canada.

978-1-4244-8327-3/11/$26.00 2011 IEEE

2
#

User 2

2
#

Precoder

User 1

Nr

User 1

Fig. 2-a. All genes making up a chromosome are grouped into


portions formed by two adjacent bits. It is interesting to note
that all alternating portions belong to the same user.
Chromosome size ( ) depends, therefore, on the number of
users and transmitter antennas; hence,
2
. The number
of different possible chromosome is defined as 2
.

User 2

Nr

User K

w1

Nt

#
2
#

Nr

fading Rayleigh channel matrix and


is the noise vector with
. At transmitter, the total
distribution according to
0,
transmit power must satisfy the following constraint
H

(2)

Since the MMSE receiver is used to estimate the


user is
transmitted symbols, the decoding vector of the
given by [5]
(3)
(4)
(5)
represents the

column of

Value

00

01

-1

10

11

-i

Wx

User K

Figure 1. Downlink Multi-User MIMO system.

where

Code

where
III.

User 1

User 2

-1

-i

Chromosome x

a)

b)

Figure 2. Representation of a) coded codeword based on 4 values and b)


2 and
2.
chromosome construction for

The idea is to find the best coded codewords to enable each


user to maximize a fitness function. A cumbersome search
process is thereby avoided to ensure enhanced performance.
Accordingly, we apply the GA to optimize the codeword, ,
used at the transmitter. The GA is naturally iterative, with each
iteration
0, 1, 2, ,
1 consisting of several
mechanisms (steps) that encompass the entire evolutionary
process as illustrated in Fig. 3.

.
Initialize the
Population, Pop(0)

In this case, the received signal-to-interference and noise


user is [5]
ratio (SINR) for the
,

w2

(6)

denotes the transmit power of the

Evaluate the fitness


function for a whole
population, Pop(g)

user.

MULTI-USER MIMO PRECODING BASED ON CODED


CODEWORDS

Our aim is to find precoding vectors that ensure maximum


system performance while reducing the exhaustive search
called for by the methods in the literature. We know that
precoding performance improves with the codebook dimension
used (e.g. [3]). The purpose of this paper is to propose a
precoding vector based on a complex value coded by a binary
symbol (gene). In our approach a codebook does not need to be
stored at the receiver and transmitter, and as a result, no
memory is used.
The precoding matrix is represented by a chromosome in
which each chromosome, denoted by , is formed by a binary
string consisting of
genes (binary symbol
0,1 ; as
illustrated in Fig. 2-a, and each codeword value is coded with
two bits as 1, 1, ,
00,01,10,11 . The chromosome
construction, , based on the coded codeword idea is shown in
Fig. 2-b as an example of 2 transmit antennas and 2 users. In
this example, we have a precoding matrix formed by 2
and
. Each value of the precoding
codewords vectors
matrix elements is coded with two bits using the code table in

Create a new
generation Popnew(g)

Generation

Apply selection

g=0,1,2,,G-1

Apply crossover and


mutation operators

Figure 3. Flowchart for GA processing.

Step 1 Initialization: This step consists of initializing the


population,
0 , composed of chromosomes, where each
chromosome is initialized by a random process.
Step 2 Fitness Evaluation: This step operates based on
selective function
and governs the action of retaining
each individual according to its ability to satisfy the
requirements of this function. Because the precoding matrix
(chromosome) is built to provide better system performance,
we chose a criterion that maximizes the sum capacity. Thus, the
selective function represents MU-MIMO sum capacity.

log 1

(7)

It should be noted that the selective function is calculated


for each chromosome of the population.
Step 3 Selection: The goal of this step is to retain the most
efficient chromosomes and progressively eliminate the least
efficient. Prior to selection, however, we must calculate the
normalized version of
, also called the relative selective
, such as
function

(8)

steps, we found that a complexity analysis is needed only in the


latter. Therefore, we first evaluated the complexity of our
approach in terms of numbers of SINR to calculate for all users
provided this forms the most resource-intensive material in the
selective function. Table I shows the number of SINR required
for each method to accomplish precoding (Note that
!
). In our approach, no codebook is stored at the receiver
!
!
and transmitter.
TABLE I. COMPARISON OF THE NUMBER OF SINR TO CALCULATE.

(9)
The decision concerning selection will depend on its
integer part
. Thus, if the integer part of
is nonzero,
the corresponding chromosome will be retained, whereas if the
integer part of
is zero, the associated chromosome will be
eliminated and replaced by the chromosome with a maximum
value of
.
Step 4 Crossover and Mutation: Crossover is based on the
exchange of information between two chromosomes. To
guarantee sum capacity growth, we took a group of
chromosomes (those most successful and which represented
the parents) and engendered
remaining
chromosomes (children) by crossing the
chromosomes
between them.
The mutation operation aims to introduce errors of copies
and consists of randomly modifying one or several
chromosome genes, thereby bringing about a dilation of the
search range. In our case, only one gene (bit) of each
chromosome is changed by the bits inversion. It should be
noted that the mutation is applied to the whole population,
except for a group formed of
chromosomes, which
represents those that are most successful.
Step 5 New Population: From the previous steps, we
conserve chromosomes to form the new population
for the next generation.
Steps 2 to 5 are repeated during generations. When these
last are completed, the choice of the precoding matrix will
correspond to the chromosome with a maximal sum capacity
value.

A. Complexity Analysis
Since crossover and mutation do not require considerable
hardware resources compared with evaluation and selection

GA

56
7280
2 541 504

56
800
4000

!
O

(10)

TABLE II. COMPUTATIONAL COMPLEXITY REQUIREMENT FOR A REFERENCE


.
METHOD AND PROPOSED GA METHOD FOR
Operation
12

[5]
8

L 13

11
4
4K

GA
3

13
6K

EVALUATION RESULTS

In this section, we present the simulation results to evaluate


the proposed GA precoder. To this end, we first perform a
complexity analysis and, based on the low complexity
presented, compare our method with a precoder using a
Grassmannian codebook of size [5]; hereafter, this called the
reference method. Next, we compare performances in terms of
error probability and sum capacity with regard to the precoding
presented in [5], where [6] has shown that is better than the
Block Diagonalization (BD).

[5]

Next, we evaluated the complexity of our approach in terms


of the arithmetic operations needed for the precoding process.
However, it is important to note that in our proposal, equations
(3) to (9) are considered when computing complexity. By cons,
for reference method, [5], we consider the equations (3) to (7).
Table II shows the number of arithmetic operations required for
the precoding in each method. To simplify the comparison of
the VLSI implementation complexity point of view, we
. According to Tables I
therefore assume that
and II, we can state that our method reduces the complexity of
the precoder, particularly if the number of users and the
Grassmannian codebook size are significant. Moreover, the
precoding matrices produced by our approach consist of
1, 1, ,
, which saves 4
2
1
1 as
trivial multipliers where a multiplier-free implementation of the
precoding matrix is applied in VLSI. Thus, considering the
practical limitations on the number of users
, we
can see that the complexity ratio is

Special

IV.

Number of SINR calculation


General case
2,
8,
4
7
4,
16,
20
10
4,
64,
20
50

4
2

1
1

Fig. 4 shows the evolution of the complexity ratio versus


the codebook size of the reference method. The complexity
evaluation was carried out following a change in the number
of users and for a product
that is worth 100. According to
the different shapes of Fig. 4, we observe that for the reference
method using a codebook size of 16, our approach can
dramatically reduce complexity, since we record a reduction
factor of 2, 30 and 200 for 2, 4 and 8 users, respectively.

10

-1

[5] with P=8


GA with G=4 and m=7
GA with G=20 and m=10

K=2
K=4
K=8

10

10

-2

10

BER

Complexity Ratio Grassmannian/GA

10

10

10

-3

-1

10

-2

10

10

11
12
13
Codebook Size (P)

14

15

10

16

-4

10

SNR (dB)

a)

Figure 4. Complexity ratio between the reference method [5] and our GA
search using
100 versus the codebook size ( ) used in reference
method for 2, 4 and 8 simultaneous users.

1
0.9
0.8
0.7

[5] with P=8

0.6
CDF

Fig. 5 still shows the complexity ratio, but this time it is


versus the number of users that communicate simultaneously
with the BS. The present evaluation was obtained for a
variation of the codebook size used in the reference method
and for a product
100 for GA. The shapes of Fig. 5
demonstrate first, that our approach is useful for
12, and
second, that the maximum ratio complexity is seen in a
number of users, which are worth almost half the codebook
size. As a result, we observe maximum reduction factors of 12
and 300 for codebook sizes of 12 and 16 respectively. Note
that the complexity ratio, shown in Fig. 4 and 5, decreases
linearly with the product
.

0.5
0.4
0.3

GA with G=4 and m=7

0.2
GA with G=20 and m=10

0.1
0
4

10
12
14
16
Sum capacity (bps/Hz)

18

20

22

b)
17

10

P=8
P=12
P=16
P=32

10

16

Average sum capacity (bps/Hz)

Complexity Ratio Grassmannian/GA

10

10

10

10

15

14

13

GA with G=20 and m=10


GA with G=4 and m=7
[5] with P=8

12

10

-1

10

11
1

10

SNR(dB)

-2

10

4
5
6
Simultaneous User (K)

Figure 5. Complexity ratio between the reference method [5] and our GA
search using
100 versus the number of users ( ) for different
codebook size ( ) used by the reference method.

B. Performance Evaluation
To study performance, we considered the MIMO broadcast
transmitter antennas
channel where the BS is equipped with
by means of which it transmits a QPSK symbol to
simultaneous users with equal power. Each of the latter is
equipped with
receiver antennas and a MMSE detector. We

c)

Figure 6. Performance results for a 2 2 MU-MIMO system with 2


users evaluated in equivalent complexity scenario for the reference
method [5] (
8) and the proposed GA method with (
7,
4
and
10,
20): a) BER versus SNR, b) CDF of the sum capacity,
and c) average sum capacity.

also assumed that the BS transmitted symbols to each MS over


quasi-static Rayleigh fading channels with the frame length .
It should be noted that all results presented in this section were
obtained by Monte Carlo simulations, which implies that the
transmission simulations were repeated
times.

sizes. The same observations can be made about the sum


capacity, since our method still prevails. Thus, we observe the
ergodic sum capacity of 20 bps/Hz and 14 bps/Hz for our
proposal (case 50 chromosomes) and the reference scheme
respectively, making for a capacity gain of about 6 bps/Hz.

1
0.9
0.8

[5] with P=16

0.7

V.

CDF

0.6
0.5
GA with G=20 and m=10

0.4
0.3
0.2

GA with G=20 and m=50

0.1
0
10

12

14

16
18
20
22
Sum capacity (bps/Hz)

24

26

28

Figure 7. CDF of the sum capacity for a 44 MU-MIMO system with 4


users evaluated in MUI scenario for the reference method [5] (
16)
and the proposed GA method (
10 and 50,
20).

We considered that
1024 ,
10000 . In all
simulations, is fixed at 40% of the population, since this is
the optimal percentage for best performance, and population
size ( ) and number of generations ( ) are varied to compare
the performance.

In this paper, we presented a new precoding method based


on research of the coded codewords using GA applied to a
MU-MIMO transmission context. Our proposal demonstrates
evolutionary behaviour in terms of channel realization while
maintaining low complexity arithmetic operations. Compared
with a reference method requiring a codebook [5], the
performance evaluation in terms of error probability and sum
capacity along with a complete complexity analysis have
shown that our method offers outstanding performance and low
complexity regardless of transmission scenario and the fact
there is no need for a previously-stored codebook.
ACKNOWLEDGMENT
The authors wish to thank the Natural Sciences and
Engineering Research Council of Canada for its financial
support.
REFERENCES

In the first scenario studied, we considered the same order


of complexity for both methods with a transmission
configuration using
2,
2 and
2 . From
Table I, we observe the same complexity for
8,
7
and
4.

[1]

A comparison of BER versus SNR for both methods is


given in Fig. 6-a. This figure shows clearly that for the same
order of complexity, our method outperforms a reference
method of 2 dB at 1% of BER.

[3]

Fig. 6-b describes the cumulative distribution function


(CDF) of the sum capacity for both methods. Note that in this
simulation, the SNR is set at 5 dB. Fig. 6-b demonstrates that
our method performs in terms of sum capacity and,
consequently, transmission rate, since it produces a sum
capacity gain of 2.5 bps/Hz for
200. We also observe
that our method leads to a gain of 2 bps/Hz even for an
equivalent complexity (
28). The plot of the ergodic
MU-MIMO sum capacity versus the SNR is given in Fig. 6-c.
Compared with the reference method, we observe a gain of
6.5 dB and 8 dB for 14 bps/Hz of sum capacity regardless of
the SNR level for
200 and 28, respectively.
Next, in the second scenario studied, we changed the
transmission parameters to observe the influence of multi-user
interference (MUI) in a lower computational complexity of our
method compared with the reference method. Hence, we
consider:
4,
4,
4.
As observed in Fig. 7, we consider SNR = 5dB for this
simulation, the reference method used a codebook size of 16,
and our GA search used
20 for two different population

CONCLUSION

[2]

[4]

[5]

[6]

[7]

[8]
[9]

[10]

M. Jiang and L. Hanzo, Multiuser MIMO-OFDM for next


generation wireless systems, Proceedings of the IEEE, vol. 95,
no. 7, pp. 1430- 1469, July 2007.
R. Calderbank, A. Goldsmith, A. Paulraj, H. V. Poor, and E.
Biglieri, MIMO Wireless Communication, Cambridge Univ.
Press, 2007.
D. J. Love, R. W. Heath, Jr., and T. Strohmer, Grassmannian
beamforming for multiple-input multiple-output wireless
systems, IEEE Trans. Inf. Theory, vol. 49, no. 10, pp. 27352747, 2003.
D. J. Love and R. W. Heath, Jr., Limited feedback unitary
precoding for spatial multiplexing systems, IEEE Trans. Inf.
Theory, vol. 51, no. 8, pp. 2967-2976, 2005.
S. Fang, G. Wu, and L. H. Li, Multi-User MIMO Linear
Precoding with Grassmannian Codebook, International
Conference on Communications and Mobile Computing, pp.
250-255, 2009.
S. Fang, L. H. Li, and P. Zhang, Grassmannian precoding MUMIMO scheme, Journal of China Universities of Posts and
Telecommunications, Elsevier, vol. 15, no. 3, pp. 105-109, Sept.
2008.
M. Benmimoune and D. Massicotte, Grassmannian precoding
for multiuser MIMO system based on the maximal SJNR
criterions, Int. Symp. Wireless Comm. Syst., pp. 471-475, UK,
Sept. 2010.
, Multi-user MIMO precoding with Kerdock codebook,
Int. Symp. Wireless Comm. Syst., pp.71-75, UK, Sept. 2010.
S. Fang, G. Wu, and L. H. Li, Optimal multiuser MIMO linear
precoding with LMMSE receiver, EURASIP Journal on
Wireless Communications and Networking, Article ID 197682,
pp. 1-10, 2009.
S. C. Huang et al., Hybrid genetic algorithm for joint precoding
and transmit antenna selection in multiuser MIMO systems with
limited feedback, IEEE Veh. Technol. Conf., pp. 1-5, 2010.

Vous aimerez peut-être aussi