Upload
annis-shepherd
View
218
Download
0
Embed Size (px)
Citation preview
Aris Moustakas, University of Athens
CROWN KickoffNKUA
Power Control in Random Networks
with N. Bambos, P. Mertikopoulos, L. Lampiris
Aris Moustakas, University of Athens 2
• Power – Frequency allocation
• Random Networks
• min “P” subject to “SINR”– “P”: total power – power per user – power per user <Pmax
– “SINR”: SINR contraints on all (some) connections = connectivity
• Impediments in the analysis of Power Control– Randomness in
• Distance between Tx-Rx
• Fading coefficients
• Interference location/strength
– Interference (interaction – domino effect)
– Constraints (max power makes problem non-convex)
Resource Allocation
ii
ii
PSINRP
i
)(min P
Aris Moustakas, University of Athens 3
• Simplify (enough) problem so that can obtain analytic solution
• Take – Minimize (total) power subject to power constraints
– Linear SINR constraints result to conical section
or
• Good news: If solution exists, can be reached using distributed algorithms (e.g. Foschini-Miljanich)
• Simplify neglect random fading,
• Problem still non-trivial due to randomness in positions and interference– Two specific examples of randomness
Power control
11
jijijiii PgPg
2/22
1
arrg
ji
ij
1iMP
ii
ii
PSINRP
i
)(min P
Aris Moustakas, University of Athens 4
• Start with ordered (square) lattice of transmitter-receiver pairs.
• (a) With probability p erase transmitter– Intermittency of transmission
– Randomness of network
• (b) With probability p/(1-p) locate users at distance a1/a2
– Models randomness of location of users
Models of Randomness: The Femto-Cell Paradigm
�
� 1
� 2
�
Aris Moustakas, University of Athens 5
• Both represent simple models with all important ingredients: – PC, randomness, interference
• After some algebra:
– Ei = 0,1 with probability p/(1-p) (erasures)
–
• Assume M circulant : eigenvalues
• Using Random Matrix Theory:
where
• β plays role of shift (β=0, when p=0)
Solution Approach
EuIEMEEu 1
0lim
1
1
Ttotave pN
Pp
jig
jigM
ij
iiij
1
)0(
1
1
1
ppave
j
jiiqijii eggq ||1)(
)(qdqp
Aris Moustakas, University of Athens 6
• When p=0 blowup at a given γ
• p>0 moves singularity to the right.
• Pave does not diverge
• Var diverges
Hint: a finite number
(1?) of nodes diverges
Metastable state?
Analysis
Aris Moustakas, University of Athens 7
• In reality system is unstable (max/ave)
• One – two dimensional systems very accurate
• Questions:– Probability of instability as a function of γ?
– Fluctuations btw samples?
Analysis
Aris Moustakas, University of Athens 8
• Introduce max-power constraint
• Distributed version:
• λ=1/Pmax
• Use 3 methods to find optimum:– Foschini-Miljanic
– Best-Response
– Nash
Resource Allocation
iiP
i PSINRui
)(maxmax P
Aris Moustakas, University of Athens 9
Type of problem
No Pure Nash Equilibrium
Players Best Respond
3 General Categories
Payoff:
Introduction
max
)()(P
PSINRPU i
i
γ_))-SINR(1log( rThroughput:
Aris Moustakas, University of Athens 10
One Mixed Nash Equilibrium
1,1
1,
max
111
2
2max
111
iP
PP
iP
PP
ni
iii
niP
P
niP
PP
n
iii
,1
,
max
2
1
1max
21
2
niP
P
niP
PP
ni
iii
,1
,
max
2
1
1max
21
2
niP
P
niP
PP
ni
iii
,1
,
max
1
2
2max
11
1
Aris Moustakas, University of Athens 11
3 mixed Nash equilibria
niP
P
niP
PP
ni
iii
,1
,
max
2
1
1max
21
2
niP
P
niP
PP
ni
iii
,1
,
max
1
2
2max
11
1
niP
PP
niP
PP
ki
ni
iii
i
,)(
1
,
2,0
max
122
1
1max
12
12
1
niP
PP
niP
PP
ki
ni
iii
i
,)(
1
,
2,0
max
11
11
2
2max
11
11
2
niP
PP
niP
PP
ki
ni
iii
i
,)(
1
,
12,0
max
122
1
1max
12
12
1
niP
P
niP
PP
ki
ni
iii
i
,1
,
2,0
max
11
2
2max
11
11
2
Aris Moustakas, University of Athens
Single & Double Pure Nash Equilibria
12
Single Equilibrium
Two Equilibria
Aris Moustakas, University of Athens 13
Average Payoffs & Throughput Comparison
1 Nash:
Throughput:
)0,1( 2nP
FM: Best Response:
max
2,0P
Pk
n
i
in
i
i
P
P
P
P
1 max
2
1 max
1 13
1,1
3
1
3 Nash:FM:
BR
)0,0(
max
2,0P
Pk
n
i
in
i
i
P
P
P
P
1 max
2
1 max
1 12
1,1
2
1
Pure Nash:
Throughput:
FM:
Best Response:
)0,0( )1,0( 11
nP )0,1( 11
nP
2/
1 max
22
2/
1 max
121 1
3
1,1
3
1 n
i
in
i
i
P
P
P
P
2/
1 max
122
2/
1 max
21 1
3
1,1
3
1 n
i
in
i
i
P
P
P
P),(
max
2,0P
Pk
Aris Moustakas, University of Athens 14
• Max-power constraint brings new features
• Nash – game on restricted power feasible and better than other cases
• BR not bad
• Generalisable to more users?
• Analytic estimates?
Questions
Aris Moustakas, University of Athens 15
• Optimize network connectivity using collaborative methods inspired by statistical mechanics (Task 2.1)
– Power – connectivity fundamental trade-off:– Tradeoff between connectivity and number of frequency bands.– Design and validation of distributed message passing algorithms
• Develop distributed message passing methods to achieve fundamental limits of detection and localization of a network of primary sources through a network of secondary sensors (Task 2.2)
– Detection of sources using compressed sensing on random graphs and Cayley trees
– Effect of additive and multiplicative noise on detection– Application of compressed sensing on two-dimensional graphs with realistic
channel statistics
• Develop decentralized coordinated optimization approaches (Task 2.3).– design self-coordinated, fast-convergent wireless resource management techniques– convergence, stability and the impact of operation on different time scales on the
performance.
Goals
Aris Moustakas, University of Athens 16
• Minimize power subject to constraints
• Interactions due to interference
• Simplifications:– Random graphs (1d-2d-inft d)
– gij = 0,1
– Power levels
• Use replica theory
Power Control – Connectivity tradeoff
1..
min
1
jijiii
ii
PgPgts
P
Aris Moustakas, University of Athens 17
• Minimum number of colors needed to color network with interference constraints– E.g. no adjacent nodes in same color
• Simplifications:– Random graphs (Bethe lattice / Erdos-Renyi)
– gij = 0,1
• Use replica theory and – Graph coloring
• Message passing algorithms
Connectivity – Frequency Bandwitdth
Aris Moustakas, University of Athens 18
Cooperative Sensing
Sensor Node
Transmitter Node
Signal
Sensor Communication
Aris Moustakas, University of Athens 19
• Minimize power subject to constraints
• Models:– H known and P discrete (on-off)
– Random graphs
– H random valued
– H in a given geometry • possible locations of sources
– With/without noise
• Use replica theory
• Compressed sensing (sparsity)
• Message passing
Collaborative Sensing and Localization
yP
zPHty ai
iaia
|Prmax
)(
Aris Moustakas, University of Athens 20
• Minimize power subject to constraints
• Models:– H known and P discrete (on-off)
– Random graphs
– H random valued
– H in a given geometry • possible locations of sources
– With/without noise
• Use replica theory
• Compressed sensing (sparsity)
• Message passing
Collaborative Sensing and Localization
yP
zPHty ai
iaia
|Prmax
)(