efficient markov chain monte carlo algorithms for …peng/defense.pdf2011/7/16 rong-hui peng...
TRANSCRIPT
2011/7/16
Rong-Hui Peng
Department of Electrical and Computer EngineeringUniversity of Utah
Efficient Markov Chain Monte CarloAlgorithms For MIMO and ISI channels
2011/7/162
Summary of PhD work
• Efficient MCMC algorithms for communication– Application to MIMO channel ([Chen-Peng-Ashikhmin-Farhang], To
appear IEEE Trans. Comm.)– Application to noncoherent channel ([Chen-Peng], to appear IEEE Trans.
Comm.’09)– Application to ISI channel and extend to underwater channel ([Peng-
Chen-Farhang], To appear IEEE Trans. Signal Process.)– Achieve near optimal performance with reduced complexity– Solve the slow convergence problem
• MIMO-HARQ schemes and combining algorithm ([Peng-Chen], to besubmitted to IEEE Trans. Comm.)– Propose new retransmission scheme– Propose new combining algorithm– Increase the throughput significantly– Applicable to Wimax and LTE
2011/7/163
Summary of PhD work
• Nonbinary LDPC codes ([Peng-Chen], IEEE Trans. WirelessComm.’08)– Nonbinary LDPC coded MIMO system, achieve good performance
with reduced complexity– Code design based on EXIT chart– Hardware-friendly construction
• Low encoding complexity• Parallel architecture• Low error floor
• Intern at MERL– Low complexity MIMO detection algorithm for MIMO-OFDMA– Wimax link-level simulation– 1 paper, 1 patent
2011/7/164
Outline
• Background and motivation• Introduction to MCMC technology• MCMC MIMO detection
– Review of MIMO detection– QRD-M and MCMC detector– Hybrid MIMO detector– Complexity analysis
• MCMC ISI equalization– Bit-wise MCMC equalizer– Group-wise MCMC equalizer
• Conclusion
2011/7/165
Background and motivation
ky ˆkuENC MOD CHANNEL DEM/DET DEC
k̂bku kb kdkh
y Hd nMAP detection:
ˆ max ( | , ) ( | , ) ( | , ) ( )
k
k k
k
b P bP b p P
b
y Hy H Y H d d
MAP detector has optimal performance but the complexity isexponential with the dimension of
Our objective: reduce the complexity but still obtain nearoptimal performance
kb
IDD
2011/7/166
Markov Chain Monte Carlo
• MCMC is a class of algorithms for sampling fromprobability distributions based on constructing a Markovchain that has the desired distribution as its stationarydistribution.
• The state of the chain after a large number of steps isthen used as a sample from the desired distribution.
• MCMC is suitable for addressing problems involving high-dimensional summations or integrals
X
f(x)
2011/7/167
Gibbs sampling
• One kind of the MCMC methods.• The point of Gibbs sampling is that given a multivariate distribution it
is simpler to sample from a conditional distribution rather thanintegrating over a joint distribution.
• Generate an initial sample as start state• Using conditional distribution as state transition probability• Each state jumping allows only one variable change• Return best state seen across all iterations (may not be the last one)• Stop after a fixed number of iterations• Solution is sensitive to the starting state, so we typically run the
algorithm several times from different starting points
2011/7/168
State transition in Gibbs sampling
d d1
d2
d3
State d3 d2 d1
S0 -1 -1 -1
S1` -1 -1 +1
S2 -1 +1 -1
S3 -1 +1 +1
S4 +1 -1 -1
S5 +1 -1 +1
S6 +1 +1 -1
S7 +1 +1 +1
y Hd n
),,|(PMFlconditionausing),|( toaccordingSampling
iidpp
dHyHydd
2011/7/169
Gibbs sampler
(0)
( ) ( 1) ( 1) ( 1)0 0 1 2 1( ) ( ) ( 1) ( 1)1 1 0 2 1
Generate initial randomlyfor n= 1 to generate from distribution ( | , , , , )
generate from distribution ( | , , , , )
g
n n n nN
n n n nN
Id p d b d d dd p d b d d d
d
yy
L
LM
( ) ( ) ( ) ( )1 1 0 1 2enerate from distribution ( | , , , , )
end for
n n n nN N Nd p d b d d d yL
2011/7/1610
Why Gibbs sampling works
• Retains elements of the greedy approach– weighing by conditional PDF makes likely to move towards locally
better solutions
• Allows for locally bad moves with a small probability, toescape local maxima (with limitations)
2011/7/1611
ISI channel
• Multipath fading channel
0( ) ( ) ( ) ( ),
for 0,1, , 1
L
ll
y m h m d m l n m
m N L
L
DD
+
)(md
)(0 mh 1 ( )h m 2 ( )h m ( )Lh m
)(mn )(my
…
2011/7/1612
ISI channel
• In matrix form
1
1
1
where : transmitting vector : received vector : channel matrix : dditive noise vector : Equalizer block length :
N
N L
N L N
N L
NL
y Hd n
d Cy CH Cn C
Channel memory
0
1 0
0
1 0
1
0 0 00 0
0 0 0
0 00 00 00 0
L
L L
L L
L
hh h
h h
h h h
h hh
H
L LL
M O O ML L
M O O ML LL O MLL L
2011/7/1613
Detection for ISI channel
• MAP detection is optimal– Efficient implementation of MAP detection is Forward backward
algorithm (BCJR)– Still exponential complexity with channel memory
• Low complexity algorithm– MMSE– Decision-feedback– …
2011/7/1614
Bit-wise MCMC detector
• Transition probability
1
:0
1 1
: : :0 1
:
( ) ( )0 1
( | , ) ( | ) ( )
( | ) ( )
( | ) ( )
( | ) ( | ) ( | ) ( )
( | ) ( )
{ , , , ,
{ }
ak k k
ak
N La
j j L j kj
i N L i La a a
j j L j j j L j j j L j kj j i L j i
i La
j j L j kj i
a n nk
P b a p P b ap P b a
p y P b a
p y p y p y P b a
C p y P b a
b b a
b y y by d
d
d d d
d
b
g
L ( 1) ( 1)1 1
denote the symbol vector corresponding to
,
is mapped
, },
to
b
n nk M
k
N
a a
ib d
b b
d b
L
2011/7/1615
Bit-wise MCMC detector
• Computing the a posteriori LLR– Accurate posteriori LLR involve computations over the whole
block (may be very cumbersome since N can be very large)– Truncated window to approximate
2
1 2,0
1: :1 2 1 2
2
1 2,1
1: :1 2 1 2
1 2
: :
1, 2,
: :
:
1 2
ln
the set which truncate seque
( | ) ( )
( | ) ( )
n
with bits {
ces in
, }
ki i i i
ki i i i
i
i i L i i ll ie e
k ki
i i L i i ll i
i i
k
p P b
p P b
b i k i
b
b
y b
y b
I
I
I I
2011/7/1616
Severe ISI channel
• For channels with severe ISI,bit-wise MCMC suffers fromslow convergence problem
• We found slow convergenceproblem for ISI channel ismainly caused by highposterior correlation
0 0.2 0.4 0.6 0.8 1-60
-50
-40
-30
-20
-10
0
10
Normalized Frequency
PSD
, dB
C1C2
1
2
[ ] 0.227 [ ] 0.46 [ 1] 0.688 [ 2] 0.46 [ 3] 0.227 [ 4]
[ ] (2 04 ) [ ] (15 18 ) [ 1] [ 2] (12 13 ) [ 3] (08 16) [ 4]
h n n n nn n
h n i n i n ni n n
2011/7/1617
Severe ISI channel
• Our solution– Grouping (blocking) the highly correlation variables in Gibbs
sampler
1st iteration:
2nd iteration:
3rd iteration:
4th iteration:
Shift grouping different adjacent symbols over iterationsto speed up the mixing rate of Gibbs sampler
2011/7/1618
Transition probability
• Grouping multiple variables allow a large sample spacecontaining values
0 00
1 0
1 1
0
1 0
1 1
1
1 1
0 0 00 0
0 0 0
0 00 00 00 0
i i
i iL
i P i GL L
i P i G
L L
N L NL
y dhh h
y dy dh h
y dh h hy d
h hy dh
L LM ML
M O O ML L
M MM O O ML LL O M
M MLL L
0
1
1
1
i
i
i P
i P
N L
n
nn
nn
n
M
M
M
2 bM Gr
1 LGP
2011/7/1619
Transition probability
• Can be thought as received signals of a MIMO channelwith G transmit antenna and P+1 receive antenna
• Can apply QRD-M to find samples with large APPs• Gibbs sampler randomly chooses one sample from
samples
: 1 :: : : : 1 : :
::
:
: 1 :
i L i i G i Pi i P i i P i i P i L i i i P i G i P i i P
i i Gi G i Pi i P i i P
y y H d H dH
nd n
%
0r r=
0r
2011/7/1621
Summary of results
• MCMC equalizer for ISI channel is studied• Near optimal performance can be obtained• Slow convergence is mainly caused by high posterior
correlation• QRD-M algorithm is applied to reduce the complexity of
group-wise MCMC equalizer• Parallel implementation is proposed
2011/7/1622
Conclusions
• Efficient MCMC algorithms for MIMO and ISI channel are studied– For MIMO, MCMC works well at low SNR region– For ISI, MCMC works well for moderate ISI
• Solutions for slow convergence are proposed– For MIMO, use QRD-M as good start point– For ISI,
• Use group-wise Gibbs sampler to group high correlated variables• Use QRD-M to reduce the complexity of group-wise Gibbs sampler
• Future work– Study the sensitivity of proposed MIMO detector with more practical
channel model– Study the time-varing ISI channel and adaptive grouping scheme for
group-wise MCMC– Hardware implementation of MCMC
2011/7/1623
Motivation
• Optimal binary code has been designed to approach channel capacity.– Long codes– Irregular
• Nonbinary LDPC code design has been studied for AWGN and showsbetter performance than binary codes.
– Shorter codes– More regular
2011/7/1624
Contribution
• Apply nonbinary LDPC codes to fading channels and MIMO channelsand provide comparison with optimal binary LDPC coded systems
• Propose modified nonbinary LDPC decoding algorithm.• Extend EXIT chart to nonbinary code design• Propose parallel sparse encodable nonbinary code with low encoding
and decoding complexity
2011/7/1625
Introduction of binary LDPC codes
• A subclass of linear block codes• Specified by a parity check matrix (n-k)×n
n: code length k: length of information sequence
+
+
+
VariableNodes
checkNodes
x1
x2
x3
x4
x5
x6
x7
c1
c2
c3
100110101011100010111
3
2
1
721
0Hx
H
T
ccc
xxx
0:0:0:
74313
64322
53211
xxxxcxxxxcxxxxc
2011/7/1626
Definition of nonbinary LDPC codes
• For nonbinary codes, the ones in parity check matrix are replaced bynonzero elements in GF(q)
500720403065200030173
H
2011/7/1627
Application to fading channels
• Channel model
VHSX Mρ
Assume each entry of channel matrix is independent, followsRayleigh fading, and is known by receiver
2011/7/1628
System block diagram
Separate detection and decoding: the detection is performedonly once.
2011/7/1629
Performance comparison
Performance comparison of GF(256) SDD, GF(16) JDD and binary JDDsystem
GF16JDD
GF256SDD
BinaryJDD
802.16e
Performance:GF256>GF16>binary>802.16e
Complexity:GF16<GF256802.16e<binary
2011/7/1630
Construction of Parallel sparse encodable codes
• Motivation– Low encoding complexity– Allow parallel implementation
• Parallel sparse encodable (PSE) codes =Qausi cyclic (QC) cycle codes + tree codes– QC codes: parallel implementation– Cycle code: Low encoding complexity
2011/7/1631
Quasi-cyclic construction
• Quasi-cyclic structure
• Ai,j is a circulant: each row is a right cycle-shift of the row above it and the firstrow is the right cycle-shift of the last row
• The advantage of QC structure– Allow linear-time encoding using shift register– Allow partially parallel decoding– Save memory
nmmm
n
n
,2,1,
,22,21,2
,12,11,1
AAA
AAAAAA
H
2011/7/1632
A new QC structure for GF(q)
000
000
0000
000
2,
2,
,
,
,
'qji
ji
ji
ji
ji
A
• Ai,j is a nonbinary multiplied circulant permutation matrix
)GF(ofelementprimitiveais
},,,{fromchoosenrandomlyis)'GF(ofelementprimitiveais);GF()'GF(
1)1'/(1(10,
q
qqqqq
ji
2011/7/1633
Cycle codes
• With degree 2 variable nodes only• Can be represented by normal graph, every vertex imposes one linear
constraint– Columns => edges; rows => vertices
101100010110100011011001
mH
b0 b1 … b5
c0
c3
c0 c1
c2c3
b0
b2
b3 b1
b4 b5
Parity check matrixNormal graph representation
2011/7/1634
H based encoding of cycle codes
① Find the spanning tree:b0, b3, b4
② Information bits =>Edges outside SPb1=1, b2=1, b5=0
③ Compute coded bitsb3=b2+b5=1b4=b2+b1=0b0=b1+b5=1
c0 c1
c2c3
b0
b2
b3 b1
b4 b5
0 1
1
1 0
1
Not work for nonbinary cycle codes! Why?
For binary code: check c0 is always satisfied.b0+b3+b4=0
2011/7/1635
PSE codes
03c
02c
03b
01c
00c
04b
00b
01b0
2b
05b
13c
12c
13b
11c
10c
14b
10b
11b1
2b
23c
22c
23b
21c
20c
24b
20b
21b2
2b
15b
25b
8b
5c6b
7b
9b
PSE codes = QC cycle codes + tree codes
2011/7/1636
Encoding of PSE codes
• Parity check matrix based encoding• Parallel encoding for QC cycle subcode• Much lower encoding complexity than normal LDPC codes using
generator matrix based encoding because the generator matrix isusually dense
2011/7/1637
Simulation results
Performance comparison of PSE codes over GF(256)with QPP (QC) codes and PEG (non-QC) codes