Joint Transmit and Receive Antenna

Selection Using a Probabilistic Distribution

Learning Algorithm in MIMO Systems Learning Algorithm in MIMO Systems
Muhammad Naeem and Daniel C. Lee
1
Outline Outline
M ti ti d P bl F l ti Motivation and Problem Formulation
Estimation of Distribution Algorithm (EDA)
Improved EDA
Performance Comparison Performance Comparison
Conclusion
2
MIMO System MIMO System
E
n
c
o
d
e
M
o
b
i
l
e
R
a
d
i
o
C
h
a
n
n
e
l
e
r
Capacity of a MIMO system increases with the number of antennas Capacity of a MIMO system increases with the number of antennas
Larger number of antennas results in a high hardware cost due to the
large number of RF chains
3
M ti ti
Hardware cost can be significantly reduced by
Motivation
Hardware cost can be significantly reduced by
selecting a subset of antennas from the set of
physically available antennas and using the signals physically available antennas and using the signals
from the selected antennas only, without sacrificing
the advantage of multi antenna diversity.
We need to Choose N
t
transmit antennas from
N
T
transmit antennas and similarly N
r
receive
antennas from N receive antennas
4
antennas from N
R
receive antennas
Motivation Motivation
We denote by the collection of all possible joint transmit
and receive antenna selections. Then, the number of
possible ways of selecting antennas is
N N
| | | |
T R
t r
N N
N N
| | | |
=
| |
\ . \ .
The computational complexity of finding an optimal
Joint Transmit and Receive Antenna selection by
exhaustive search grows exponentially with the number
of transmit and receive antennas.
5
o t a s t a d ece e a te as
Joint Antenna Selection Problem

We denote by in a selection of transmit and receive

antennas
Joint Antenna Selection Problem
antennas.
We denote by the channels formed between
r t
N N
H

We denote by the channels formed between
selected N
t
transmit antennas and N
r
receive antennas. The
channel capacity associated with selected transmit and
i t i
H
H

| |
|
receive antennas is
2
( ) log det ( )( )
where is the average SNR per channel use
r
H
N
t
C H I H H
N

| |
= +
|
\ .
6
where is the average SNR per channel use.
Joint Antenna Selection Problem Joint Antenna Selection Problem
We can model joint transmit and receive antenna selection
problem as a combinatorial optimization problem
max ( ) or C H

2
max log det ( )( )
r
H
N
I H H
N

| |
+
|
\ . t
N
\ .
7
Conventional EDA Conventional EDA
Generate random population
Evaluate Individuals
Sort
X
1
X
2
X
3
X
n
Function
Values
1 1 1 0 0 F
1
2 1 0 0 1 F
2
3 0 1 1 1 F
l

C
Yes
3 0 1 1 1 F
3
... ... ... ... ... ...
1 0 0 1
1 l
F

1 l

l
Convergence
Criterion satisfied
No
Terminate
Generate New
Individuals with Conditional
Prob. Vector
l l

Update Counter l=l+1
X
1
X
2
X
3
X
n
1 1 1 0 0
2 1 0 0 1
Select
Best
Individuals
1 l

1 2 1
( , , , | )
n l
P

=
8
... ... ... ... ...
0 1 0 1
1 l

EDA EDA
EDA can be characterized by parameters and Notations
1. I
s
is the space of all potential solutions
2. F ( ) denotes a fitness function.

3.
l
is the set of individuals (population) at the l
th
iteration.
4.
l
is the set of best candidate solutions selected from set
l
at the
l
th
iteration.
th
5. We denote
l

l

l

l

C l
.where
C l
is the complement of
l
.
6. p
s
is the selection probability. The EDA algorithm selects p
s
|
l
|
individuals from set
l
to make up set
l
.
l
p
l
7. We denote by the distribution estimated from
l
(the set of
selected candidate solutions) at each iteration
8. I
Ter
are the maximum number of iteration
9
Ter
Modified EDA
l

1 l
F

1 l
l l

l l

1 l

1 l

1 l

1 2 1
( , , , | )
n l
P

=
10
Generating the initial population Generating the initial population
11
Weighted EDA Weighted EDA
we propose an idea of adding some skew in estimating the probability
distribution in EDA distribution in EDA.
The skew can be added by giving more weights to the individuals
in
l-1
that have better fitness in estimating the joint probability
l-1
g j p y
distribution
An example of Weight values is
( ) ( )
log log
1 2
l
j
j

= =
( ) ( )
1
, 1, 2,...,
log log
l
j l
l
i
j
i

=
= =
(

12
Simulation Results
13
25
30
Optimal
Random
Norm
Gorokhov Algo
20
s
/
s
/
H
z
)
Decoupled
EDA
10
15
C
a
p
a
c
i
t
y

(
b
i
t
5
0 5 10 15 20 25
0
SNR (dB)

10% O t it SNR With NT 6 Nt 3 NR 30 N 2
14
10% Outage capacity versus SNR. With NT=6, Nt=3, NR=30, Nr=2.
25
O ti l
20
Optimal
Random
Norm
Gorokhov Algo
Decoupled
15
b
i
t
s
/
s
/
H
z
)
Decoupled
EDA
10
C
a
p
a
c
i
t
y

(
b
5
0 5 10 15 20 25
0
SNR (dB)

10% O t it SNR With NT 6 Nt 4 NR 18 N 3
15
10% Outage capacity versus SNR. With NT=6, Nt=4, NR=18, Nr=3
19.6
19.8
EDA-A
EDA-R
EDA-BN
EDA WBN
19.4
/
s
/
H
z
)
EDA-WBN
19
19.2
C
a
p
a
c
i
t
y

(
b
i
t
s
18.8
4 6 8 10 12 14 16 18 20
18.6
EDA Interations

16
10% Outage capacity versus EDA iterations
20.5
19
19.5
20
(
b
i
t
s
/
s
/
H
z
)
18
18.5
C
a
p
a
c
i
t
y

15
20
25
30
20
25
30
17.5
P l i i
5
10
15
15
Number of iterations
Population size
17
Tradeoff between population size and the number of iterations
The number of complex multiplications and additions The number of complex multiplications and additions
ESA
3
R
T R
N
N
N
N
N

|
.
|

\
|
|
.
|

\
|
Decoupled Algorithm
3
T R
R
N N
N
(
| | | |
+
( | |
t r
N N

. \ . \
Decoupled Algorithm
R
t r
N N
( | |
\ . \ .

3 3
N N N N N N +
Gorokhov
3 3
r t T t r R
N N N N N N +
EDA
( )
3

R Ter l
N I
18
The number of complex multiplications and additions The number of complex multiplications and additions
,
, , , ,
R r T t Ter
N N N N I
(

ESA
Decoupled
Algorithm
Gorokhov
Algorithm
EDA
[30 2 6 3 30 20] 23200 1213 1794 1600 [30 , 2, 6, 3,30,20] 23200 1213 1794 1600
[16, 4, 6, 4, 16, 8] 582400 39147 5632 2730
[20 4 6 4 20 8] 1 510
6
103680 6656 3413 [20, 4, 6, 4, 20,8] 1.510
6
103680 6656 3413
[20,8,10,6,20,8] 1.910
9
9.0810
6
65280 11520
19
Conclusions Conclusions
Existing antenna selection schemes are computationally
expensive expensive.
The performance of EDA algorithm is close to the optimal. The performance of EDA algorithm is close to the optimal.
EDA with Cyclic shifted initial population reduces the number
of iterations to reach the optimal solution.
Th f f i ht d EDA i b tt th ll i t f The performance of weighted EDA is better than all variants of
EDA.
20
Thank You
21