Energy and Inference in Wireless Sensor Networks
Transcript of Energy and Inference in Wireless Sensor Networks
Energy and Inference inEnergy and Inference inWireless Sensor NetworksWireless Sensor Networks
Vince PoorVince Poor((poor@[email protected]))
CTW 2006CTW 2006May 24, 2006May 24, 2006
Energy & Inference in WSNs
• Salient features of WSNs:
– The primary application is inference
– Information at different terminals is often correlated
– Energy is often severely limited
• Research in WSNs:
– Networking Issues: capacity, delay, routing, etc.
– Applications Issues: primarily distributed inference
• Primary goal:
– Optimize performance within constraints of wireless systems
(i.e., “bandwidth & batteries”)
MotivationMotivation
Energy & Inference in WSNs
Sensor FieldSensor Field
Energy & Inference in WSNs
Access PointAccess Point
SensorsSensors
Basic Set UpBasic Set Up
Topics of TodayTopics of Today’’s Talks Talk::¯̄ Energy efficiency in shared-access networksEnergy efficiency in shared-access networks¯̄ Collaborative Collaborative beamformingbeamforming¯̄ Energy issues in distributed inference (briefly)Energy issues in distributed inference (briefly)
Energy & Inference in WSNs
ENERGY EFFICIENCY INENERGY EFFICIENCY INSHARED-ACCESSSHARED-ACCESS
NETWORKSNETWORKS
Energy & Inference in WSNs
Competition in Shared-Competition in Shared-Access NetworksAccess Networks
• Sensors transmit to an access
point via a shared channel.
• Sensors are like players in a
game, competing for resources
to transmit their data to the AP.
• The action of each sensor affects
the others.
• Can model this as a non-cooperative game, with payoff
measured in bits-per-joule.
• First, we digress …
APS
S S
S
Energy Efficiency in Shared Access Networks
Shared-Access ChannelShared-Access Channel
Modulator Channel
010…
110…
SignalProcessing
110…
Modulator
SignalProcessing 010…
... ...
... ...
Multiuser Detection: receiver processing for shared-access systems
Energy Efficiency in Shared Access Networks
€
r2 (t)
€
rP (t)
... ...
... ...
€
r1(t)Sensor 1: 010…
Sensor 2: 110…
Sensor N: 011…
MultipathMultipath, Multi-antenna Case, Multi-antenna Case
Energy Efficiency in Shared Access Networks
Space-Time MUD Structure
•• XISO (XISO (P=P=11) ) requiresrequires no beam-formers no beam-formers•• Flat fading (Flat fading (LL=1=1))requires requires no no RAKEsRAKEs•• Decision logic: Decision logic: Optimal (ML, MAP), linear, iterative, adaptive.Optimal (ML, MAP), linear, iterative, adaptive.
... ...
TemporalMatchedFilters
{k, l, p}
DecisionLogic
110…010…
011…
BeamFormers{k, l}
RAKEs{k}
€
r2 (t)
€
rP (t)
€
r1(t)
... ... ... ... ... ... ... ...
NN××LL××PP NN××LL NN
N Sensors; P Receive Antennas; L Paths/User/Antenna
Energy Efficiency in Shared Access Networks
Linear MUDLinear MUD
DecisionLogic
... ... == QuantizerLinear
Transformation(LT)
... ...
Key ExamplesKey Examples::
•• Matched Filter/RAKEMatched Filter/RAKE Receiver Receiver: : LT = identityLT = identity
•• DecorrelatorDecorrelator: : LT = channel inverterLT = channel inverter (i.e., zero-forcing) (i.e., zero-forcing)
•• MMSE DetectorMMSE Detector: : LT = MMSE estimateLT = MMSE estimate of the transmitted symbols of the transmitted symbols
Energy Efficiency in Shared Access Networks
Game Theoretic FrameworkGame Theoretic Framework
€
uk = utility =throughput
transmit power=Tk
pk bitsJoule
Tk = Rk f(γk), where f(γk) is the frame success rate, and γk is
the received SIR of sensor k.
Game: G = [{1,…,N}, {Ak}, {uk}]
N: total number of sensors
Ak: set of strategies for sensor k
uk: utility function for sensor k
Energy Efficiency in Shared Access Networks
[[MeshkatiMeshkati, Poor, Schwartz, , Poor, Schwartz, MandayamMandayam, , IEEE Trans. COMIEEE Trans. COM, Nov. 2005., Nov. 2005.]]
An Uplink GameAn Uplink Game
• For a fixed linear MUD at the uplink receiver, each sensor
selects its transmit power to maximize its own utility.
• Th’m: f sigmoidal ⇒ Nash equilibrium (i.e., no user can
unilaterally improve its utility) is reached when each sensor
chooses a transmit power that achieves γ*:
• I.e., Nash equilibrium (NE) requires SIR balancing.
f(γ*) = γ* f′(γ*)
Energy Efficiency in Shared Access Networks
RemarksRemarks
• The NE is unique, and can be reached iteratively as
the unique fixed point of a nonlinear map.
• Effects of Detector Choice:
– We can use the NE to examine the effects of uplink receiver
choice on energy efficiency.
– Of interest are the classical matched filter, the (zero-
forcing) decorrelator, and the MMSE detector.
Energy Efficiency in Shared Access Networks
• Random CDMA: N sensors; spreading gain G
• Load: α = N/G (i.e., the number of users per dimension)
• Large-system limit: N, G →∞ , with α fixed.
Nash EquilibriumNash EquilibriumUtilityUtility vs vs. Load (Large-System Limit). Load (Large-System Limit)
m = # receive antennas
Energy Efficiency in Shared Access Networks
Effects of Delay ConstraintsEffects of Delay Constraints
• For some messages (e.g., alarms), delay is important.
• Delay model (ARQ):
– X represents the number of transmissions needed for a given
packet to be received without error, so that:
– We can represent a delay requirement as a pair (D,β):
– Thus, we have a constrained game, with γk ≥ γk’.
P(X=m) = f(γ) [1 - f(γ)]m-1 , m = 0, 1, …
Energy Efficiency in Shared Access Networks
P(X≤D)≥β ⇔ γ ≥ γ’
[[MeshkatiMeshkati, Poor, Schwartz, ISIT05., Poor, Schwartz, ISIT05.]]
NE for Multiple Delay ClassesNE for Multiple Delay Classes
Energy Efficiency in Shared Access Networks
• Traffic is typically heterogeneous
with multiple delay classes.
• A given delay class c will have its
own SIR constraint: γc’
• At NE all sensors in class c will
SIR-balance to max{γ*,γc’}.
• Tight delay constraints on one class can affect the energy
efficiencies of all sensors due to increased interference levels.
SIR
User's Utility
η1
η2
γ1 γ
2 γ* ~ ~
SIR, γ
Utilit
y, u
γ* γ’
----
----
----
----
----
----
----
----
----
----
----
----
----
----
----
---
----
----
----
----
---
2-Class Example: Utility Loss2-Class Example: Utility Loss
Energy Efficiency in Shared Access Networks
• RCDMA in the large-system limit: N, G →∞ , with α = N/G fixed.
• Class A: (DA,βA) = (1, 0.99)
• Class B: (DB,βB) = (3, 0.90)
α = 0.1
α = 0.9
EnhancementsEnhancements
• Nonlinear MUD (ML, MAP, PIC, etc.): SIR-balancing also leads to a
Nash equilibrium for certain nonlinear MUDs for RCDMA in the
large system limit. [w/ D. Guo; Allerton’05]
• Multicarrier CDMA: Actions also include choice of a carrier; at NE
(when it exists) each sensor transmits on its single, best, carrier
+ SIR balancing. [w/ M. Chiang; JSAC’06]
• Delay w/ Finite Backlog: Add queuing. [w/ R. Balan; CITIA Wkshp 06]
• Adaptive Modulation/Coding: Actions also include choice of a
modulation. [w/ A. Goldsmith, et al., GLOBECOM 06, submitted]
Energy Efficiency in Shared Access Networks
COLLABORATIVEBEAMFORMING
[[OchiaiOchiai, , MitranMitran, Poor, , Poor, TarokhTarokh, , IEEE Trans. SPIEEE Trans. SP, Nov. 2005., Nov. 2005.]]
Energy & Inference in WSNs
Collaborative Beamforming
CollaborativeCollaborative Beamforming Beamforming
A
S
S
S
S
S
SensorSensorClusterCluster
AccessAccessPointPoint
S
S
S
What are the properties of a beam formed
collaboratively by randomly placed sensors?
-24
-18
-12
-6
0
-180 -120 -60 0 60 120 180
Observation Angle Observation Angle φφ [deg] [deg]
Aver
age
Powe
r [dB
]Av
erag
e Po
wer [
dB]
N N = 16= 16
NN = 256 = 256
RR//λλ==11RR//λλ==22RR//λλ==88 ff= 2GHz= 2GHz
RR=1.2m=1.2m
-24
-18
-12
-6
0
-180 -120 -60 0 60 120 180
Average Average Beampattern Beampattern ExampleExample
N N = number of sensors= number of sensorsRR = radius of cluster = radius of clusterλλ = wavelength= wavelength
Collaborative Beamforming
Average beam has nice properties. Life is good.
ButBut, average doesn, average doesn’’t represent t represent the the realizationsrealizationsof the sensor array!of the sensor array!
• As cluster radius RR becomes larger relative to wavelength λ,
the main beam becomes sharper.
• Sidelobe level of average beampattern with N sensors is
approximately 1/N.
• Peak ave. sidelobe value does not depend on RR/λ, but the
peak location does.
• There are no grating lobes.
Average Average Beampattern Beampattern PropertiesProperties
Collaborative Beamforming
-12
-6
0
-180 -120 -60 0 60 120 180Observation Angle Observation Angle φφ [deg] [deg]
Powe
r [dB
]Po
wer [
dB]
N N =16=16
RR//λλ==22
AverageAverageRealizationRealization
Maximum Maximum sidelobe sidelobe peak could be peak could be much much higher than its average!higher than its average!
-12
-6
0
-180 -120 -60 0 60 120 180
-18
Ave. Ave. Beampattern vsBeampattern vs. Realization. Realization
Collaborative Beamforming
• Maximum peak of sidelobe corresponds to worst-case
interference.
• We use level-crossing theory to analyze this issue.
• For large N the beam is Rice-Nakagami in the sidelobes.
• Modeling sidelobes as a complex stationary Gaussian
process, approximate upper bound on sidelobe
distribution can be found.
• Simulations show good agreement.
Distribution of Max Distribution of Max Sidelobe Sidelobe PeakPeak
Collaborative Beamforming
9
12
1 10 100 1000
Side
lobe
Si
delo
be L
evel
Abo
ve A
vera
ge [
dB]
Leve
l Abo
ve A
vera
ge [
dB]
Normalized Radius Normalized Radius RR//λλ
Maximum Maximum sidelobe sidelobe peakpeakgrowth if we growth if we allowallow 10% 10% of ofrealizations to realizations to exceed this levelexceed this level
1%1%0.1%0.1%
- 10 log- 10 log1010 NN + 9 + 9
6
9
12
1 10 100 1000
Required Required Sidelobe Sidelobe Level MarginLevel Margin
Collaborative Beamforming
TargetTarget
GPSGPS
φφ11
φφ22
φφkk++ϕϕ11
++ϕϕ22
++ϕϕkk
Each node autonomouslyestimates relative phaseoffset from pilot signalfrom target and absoluteclock.
Closed-Loop Phase AcquisitionClosed-Loop Phase Acquisition(Self-Phasing Arrays)(Self-Phasing Arrays)
Collaborative Beamforming
Average beampattern can be expressed as
-20
-15
-10
-5
0
-5 0 5 10 15 20
Beampattern Beampattern with Phase Jitterwith Phase Jitter
Beam
patte
rn
Beam
patte
rn D
egra
datio
n [d
B]De
grad
atio
n [d
B]
Loop Filter SNR [dB]Loop Filter SNR [dB]
-20
-15
-10
-5
0
-5 0 5 10 15 20
i.i.d.i.i.d.Tikhinov Tikhinov jitterjitterCollaborative Beamforming
ENERGY ISSUESENERGY ISSUESIN DISTRIBUTEDIN DISTRIBUTED
INFERENCEINFERENCE(BRIEFLY)(BRIEFLY)
Energy & Inference in WSNs
HH11: : signal field signal field ++ noise noise HH00: : noise onlynoise only
Energy-Efficient Sensor Scheduling
NeymanNeyman-Pearson performance of an -Pearson performance of an NN-sensor net measured-sensor net measuredvia the via the error exponent,error exponent, KK, of the miss , of the miss probprob.:.:
K ~ - log PM(N)/N
Distributed Inference
[[Sung, Tong, Poor, Sung, Tong, Poor, IEEE Trans. ITIEEE Trans. IT, Apr. 2006, Apr. 2006]]
•• KK can be obtained in can be obtained in closed formclosed form using using state-space modelstate-space model..•• BehaviorBehavior w.r.t. correlation strength w.r.t. correlation strength depends on SNRdepends on SNR..•• This can be used to This can be used to schedule sensorsschedule sensors to transmit to transmit
collaboratively for optimal collaboratively for optimal energy efficiencyenergy efficiency..
Sensor Scheduling Via K
Distributed Inference
Distributed Learning
• Exemplars are distributed
among the sensors in
some way.
• Communications capacity
between the sensors and
fusion center is limited.
• Question: Can we construct algorithms so that
optimal inferential functions can be learnedconsistently (N→∞) consuming little transmit power?
APS
S S
S
[[PreddPredd, , KulkarniKulkarni, Poor, , Poor, IEEE Trans. IT, IEEE Trans. IT, Jan. 2006Jan. 2006]]
Distributed Inference
AccessPoint
S
SS
S
{X1 ,Y1} {X2 ,Y2}
{X4 ,Y4} {X3 ,Y3}
X
X X
X
{0,1}/A
{0,1}/A{0,1}/A
{0,1}/A
Distributed-Data Network
We can construct algorithms for the sensors and AP such that:
• Classification: transmitting 1 bit/sensor/decision is enough.
• Regression: transmitting log2(3) bits/sensor/decision is enough
Distributed Inference
Collaborative Regression
• N sensors at locations
{xi} take measurements:
• Using message-passing-
type algorithms, sensors
can collaborate with their
neighbors to estimate f.
S
S
S
i
S
S
S
S
S
yi = f(xi) + ni
Distributed Inference
E.g., An Algorithm
Note: Converges to a relaxation of the centralized RKHS estimator.
• To initialize, the sensors:• agree on a kernel K(.,.)• localize (i.e., estimate xi)• share positions with neighbors• measure field locally (i.e. observe yi)• set zi = yi
• To estimate the field:for t=1,…,T
for s = 1,…, NQuery: Sensor s queries zi from neighbors
Compute:
Update: Updates neighbors zi = fs,t(xi)
Distributed Inference
[[PreddPredd, , KulkarniKulkarni, Poor, ITW06, Uruguay, Poor, ITW06, Uruguay]]
Energy Efficiency
• Overall error decreases with
size of the neighborhoods.
• But, energy consumed by
message-passing increases
with neighborhood size.
• Question: What are the trade-offs?
MSE
MSE
ConnectivityConnectivity
Distributed Inference
Energy-per-Sensor vs.N
Distributed Inference
Tota
l Ener
gy
Tota
l Ener
gy/N/N
NN (number of sensors) (number of sensors)
rrNN = = NNαα
αα ={.30, .35, .40, .45}
PKP
Centralized
Local Averaging
Mean-Square Error vs.N
Distributed Inference
MSE
MSE
NN (number of sensors) (number of sensors)
rrNN = = NNαα
αα ={.30, .35, .40, .45}
PKP
Centralized
Local Averaging
• We’ve examined issues (primarily signal
processing) affecting the energy efficiency of
wireless networks:
• Energy efficiency in shared access systems
• Collaborative beamforming
• Energy issues in distributed inference, briefly
SummarySummary
Energy & Inference in WSNs
Thank You!