neeraj jaggi assistant professor dept of electrical engineering and computer science wichita state...
Post on 19-Dec-2015
214 views
TRANSCRIPT
Neeraj Jaggi
A SS I STA N T P R O F E SSO R
D E P T O F E L E C T R I C A L E N G I N E E R I N G A N D C O M P U T E R SC I E N C E
WICHITA STATE UNIVERSITY
1
Rechargeable Sensor Activation under Temporally Correlated
Events
Outline
Sensor Networks
Rechargeable Sensor System Design of energy-efficient algorithms Activation question – Single sensor scenario
Temporally correlated event occurrence Perfect state information
Structure of optimal policy Imperfect state information
Practical algorithm with performance guarantees
2
Neeraj Jaggi Dept of EECS Wichita State University
3
Sensor Nodes Tiny, low cost Devices Prone to Failures Redundant Deployment Rechargeable Sensor Nodes
Range of Applications
Important Issues Energy Management Quality of Coverage
Sensor Networks
Neeraj Jaggi Dept of EECS Wichita State University
4
Rechargeable Sensor System
Quality of Coverage
Discharge
Recharge
Event Phenomena
Renewable Energy
Randomness
Spatio-temporal Correlations
Activation Policy
Control
Rechargeable Sensors
Neeraj Jaggi Dept of EECS Wichita State University
Research Question5
How should a sensor be activated (“switched on”) dynamically so that the quality of coverage is maximized over time ?
A sensor became ready. What should it do ? Activate itself now :
Gain some utility in the short-term
Activate itself later : No utility in the short term Activate when the system “needs it more”
Neeraj Jaggi Dept of EECS Wichita State University
Temporal Correlations6
Event Process (e.g. Forest fire) On period (HOT) Off period (COLD) Correlation probabilities
0.5 < ( , ) < 1
( = = 0.8)
Performance Criteria – Single Sensor Node Fraction of Events Detected over time
oncp
offcp
oncp
offcp
Neeraj Jaggi Dept of EECS Wichita State University
Sensor Energy Consumption Model7
Discrete Time Energy Model Operational Cost (1) Detection Cost (2) Recharge Rate (qc)
Probability (q) Amount (c)
recharge
sensor not activated(no discharge)
qc
activation policy
K
δ1
discharge - Off period
discharge - On period
δ1+δ2sensor activated
Neeraj Jaggi Dept of EECS Wichita State University
System Observability8
Perfect State Information Sensor can always observe state of event process
(even while inactive)
Imperfect State Information Inactive sensor can not observe event process
Neeraj Jaggi Dept of EECS Wichita State University
Approach/Methodology9
Perfect State Information Formulate Markov Decision Problem (MDP) Structure of Optimal Policy
Imperfect State Information Formulate Partially Observable MDP (POMDP) Transform POMDP to equivalent MDP (Known techniques) Structure of Optimal Policy Near-optimal practical Algorithms
Neeraj Jaggi Dept of EECS Wichita State University
Perfect State Information10
Markov Decision Process State Space = {(L, E); 0 ≤ L ≤ K, E є [0, 1]}
L – Current Energy Level, E – On/Off period Reward r– one if event detected; zero otherwise Action u є [0, 1]; Transition probabilities p
Optimality equation (average reward criteria)
h* – state variables λ* – optimal reward
Neeraj Jaggi Dept of EECS Wichita State University
Perfect State Information (contd.)11
Approximate Solution
Closed form solution for h* does not seem to exist
Value Iteration
Activation Algorithm L << K
Sensitive to system parameters when L ~ K
Optimality equation (average reward criteria)
H* – variables Lambda* – optimal reward
Neeraj Jaggi Dept of EECS Wichita State University
Perfect State Information (contd.)12
Optimal Policy Structure Randomized algorithm P* is directly proportional to the recharge rate Energy balance
Average recharge rate equals average discharge rate in steady state
On Period ? ActivateYes
No
Sufficient Energy ?
No
Do Not Activate
Yes
Prob. ≤ P* ?YesNo
Neeraj Jaggi Dept of EECS Wichita State University
Imperfect State Information13
Partially Observable Markov Decision Process State Space Observation Space Optimal actions depend on current and past observations
(y) and on past actions (u)
Transformation to equivalent MDP 1
State – Information vector Zt of length |X|
Zt+1 is recursively computable given Zt, ut and yt+1
Zt forms a completely observable MDP
Equivalent rewards and actions 1 Neeraj Jaggi Dept of EECS Wichita State University
Equivalent MDP Structure14
Active Sensor – Observation = (L, 1) or (L, 0) State is the same as observation Zt has only one non-zero component
Inactive Sensor – Observation = (L, Φ) Let state last observed = E, number of time slots inactive =
i Zt has only two non-zero components Let pi= prob. that event process changed state from E to 1-
E in i time slots State = (L, E) with prob. 1 - pi
State = (L, 1 – E) with prob. pi
Zt is a function of (L, E, i)Neeraj Jaggi Dept of EECS Wichita State University
Transformed MDP State Space – (L, E, t) L – Current Energy Level E – State of Event process last observed t – Number of time slots spent in inactive state
Optimal Policy Structure f0 – (L, 0, t), f1
– (L, 1, t)
[1=c=1, 2= 2, = 0.6, = 0.9, q = 0.1]
Imperfect State Information (contd.)
15
Off Period – Reluctant Wakeup
On Period – Aggressive Wakeup
oncp
offcp
Neeraj Jaggi Dept of EECS Wichita State University
Practical Algorithm16
Correlation dependent Wakeup (CW) Activate during On Periods; Deactivate during Off
Sleep Interval (SI*) Derived using energy balance during a renewal
interval
-optimal ( ~ O(1/β)); β = 2/1
A – ActiveI – Inactive
Y – On, N – OffSI – sleep duration
t1, t2 – renewal instances
Y Y Y Y Y N
A A A A A A I
Y Y Y N
A A A A I
SI
I
tt2t1
Neeraj Jaggi Dept of EECS Wichita State University
Simulation Results17
Energy balancing Sleep Interval SI*
[ = 0.6, = 0.9, SI* = 7] [ = 0.7, = 0.8, SI* = 18]
[1 = c = 1, 1 = 6, q = 0.5, K = 2400]
oncp
offcp
oncp
offcp
Neeraj Jaggi Dept of EECS Wichita State University
Contributions18
Structure of Optimal Policy
EB Policy is Optimal for Perfect State Information
EB Policy is near Optimal for Imperfect State Information
Coauthors Prof. Koushik Kar , Rensselaer Polytechnic Institute Prof. Ananth Krishnamurthy, Univ. of Wisconsin
Madison 5th International Symposium on Modeling and
Optimization in Mobile Ad hoc and Wireless Networks (WIOPT) April 2007
ACM/KLUWER Wireless Networks 2008 (Accepted )
Neeraj Jaggi Dept of EECS Wichita State University
Q & A19
THANK YOU !!
Neeraj Jaggi Dept of EECS Wichita State University
Policies – AW, CW20
AW (Aggressive Wakeup) Policy Activate whenever L ≥ 2 + 1
Ignores temporal correlations Optimal if no temporal correlations
CW (Correlation dependent Wakeup) Policies Activate during On periods; deactivate during Off Upper Bound (U*
CW) State unobservable during inactive state Performance depends upon sleep duration
Neeraj Jaggi Dept of EECS Wichita State University
How long should sensor sleep ?
MDP – State Transitions21
State (L, 1): L ≥ 2 + 1Action u = 1 (activate)
Next state : (L + qc – δ1 – δ2, 1) with probability q.pc
on
(L + qc – δ1, 0) with probability q.(1 – pcon)
(L – δ1 – δ2, 1) with probability (1 – q ).pcon
(L – δ1, 0) with probability (1 – q ).(1 – pcon)
Reward r = 1 with probability pcon; 0 otherwise.
Action u = 0 (deactivate) Next state :
(L + qc, 1) with probability q.pcon
(L + qc, 0) with probability q.(1 – pcon)
(L, 1) with probability (1 – q).pcon
(L, 0) with probability (1 – q).(1 – pcon)
Reward r = 0
Neeraj Jaggi Dept of EECS Wichita State University