800 likes | 951 Views
Victoria Manfredi Thesis Defense August 13, 2009 Advisor: Jim Kurose Committee: Andrew Barto Deepak Ganesan Weibo Gong Don Towsley. Sensor Control and Scheduling Strategies for Sensor Networks. Motivation.
E N D
Victoria Manfredi Thesis Defense August 13, 2009 Advisor: Jim Kurose Committee: Andrew Barto Deepak Ganesan Weibo Gong Don Towsley Sensor Control and Scheduling Strategies for Sensor Networks
Motivation Wireless, Closed-Loop Sensor Network Cameras, radars: cannot collect data simultaneously from all environment locations Where to focus sensing? Data Sensor Controls Multiple users making conflicting sensor requests Changing network topology How to make routing robust to network changes? How to accommodate multiple users? Bursty, high-bandwidth data, many-to-one data routing to sink: congestion How to make sensing robust to delayed and dropped packets?
Contributions • Adaptive sensors • where to focus sensing in adaptive meteorological radar network? • show lookahead strategies useful when multiple small phenomena, trade-off between scan quality and re-scan interval • accommodating multiple users? • identify call admission control problem, give complexity results • How to make sensing robust to delayed, dropped packets? • show good application-level performance possible in closed-loop sensor network when congestion if sensor control prioritized • How to make routing robust to network changes? • propose routing algorithm, show can significantly control overhead while minimally degrading % of packets delivered
Outline • Adaptive sensors • where to focus sensing? • multiple users • Prioritizing sensor control traffic • Robust routing in dynamic networks • Conclusions • Adaptive sensors • where to focus sensing? • multiple users • Prioritizing sensor control traffic • Robust routing in dynamic networks • Conclusions
Where to focus sensing? CASA: Adaptive Meteorological Radar Network small scan sectors high quality, but may miss storm large scan sectors low quality, but miss fewer storms
Sensing Strategies What are ``good” sensing strategies? • sit-and-spin • all radars always scan 360 • myopic • consider only current environmental state • limited lookahead • Kalman filters to predict storm cell attributes k time-steps ahead • full lookahead • formulate as Markov decision process • reinforcement learning to obtain policy: Sarsa()
Storm Tracking Application • Radar network • Storm model • storms arrive according to spatio-temporal Poisson process • storm dynamics from Kalman filters • Radar sensing model • observed attribute value = true attribute value + Gaussian noise 30km 30 km max storm radius: 4km Depends on scan quality
Performance Metrics • Re-scan interval • how long before storm first observed or rescanned • Scan quality • how well storm observed • function of • scan sector size • distance from radar • % of storm scanned • value between 0 and 1 • Cost • function of re-scan interval, quality, penalty for missing storms • 2-step and full lookahead have similar cost for 2 radars 30 km
Optimize over all radars? 1-Step Lookahead Myopic 1-Step Lookahead Average Quality Myopic Sit-and-Spin Sit-and-Spin Max 8 Storms Max 1 Storm Decreasing gains as optimize over more radars No gains in quality as optimize over more radars
Summary Where to focus sensing? Showed lookahead strategies useful when multiple storms, storm radius (much) smaller than radar radius trade-off scan quality and frequency storm scanned may not need to optimize over all radars in network • Related work • track ground targets from airplanes [Kreucher, Hero, 2005] • our focus: track meteorological phenomena using ground radars
Outline • Adaptive sensors • where to focus sensing? • multiple users • Prioritizing sensor control traffic • Robust routing in dynamic networks • Conclusions
Virtualize sensing resources virtualized private sensor network To each user looks like own private network but user only has virtual slice Users request resources possibly conflicting requests which requests to satisfy? How to accommodate multiple users? Mt. Toby MA1 Tower CS Building Call admission control problem
Call Admission Control Problem Scan 360° • Sensor request • use sensor in particular way possibly during particular time • Sensing strategy • sequence of requests over time • Utility of request j • to requesting user i: uij • to each other user i: uij Scan 360° every 2 min Strategy for User 1 Strategy for User 2 Combine into single utility i uij = uij + uij i i i i Select set of non-interfering requests that maximizes utility
Space of Problems Divisible requests? • Utility received if only part of request satisfied? • Yes • scan x of y elevations • No • obtain full scan of storm Shifting permitted? • Utility received if request executed at different time? • Yes • perform surveillance scan • No • sense storm expected at location (x,y) at time t User 1 Request User 2 Request Interleaved Requests User 1 Request Shifted Request Time Time
Complexity Divisible requests? Yes No Shifting permitted? Shifting permitted? Yes No Yes No Polynomial NP-Complete Polynomial Polynomial Same as fractional knapsack problem Interleave sensor requests Interval scheduling [Arkin, Silverberg, 1987]
Indivisible, Shifting • NP-complete • assume utility independent of when request executed • In NP: can check whether solution correct in polynomial time w1 Sensing Strategy for User 1 Knapsack Problem v1 wN Reduction Capacity W Sensing Strategy for User N vN Time T=W w1 v1 wN vN Utility for satisfying user i’s request • to user i: vi • to each other user: 0
Summary How to accommodate multiple users? • Related work • adaptively select set of sensors for task [Jayasumana et al, 2007] • our focus: virtualizing sensing resources within a sensor • Future work • online, decentralized algorithms • trade-off between maximizing utility and user fairness • implement proposed algorithms in deployed network Requests divisible or fixed in time polynomial-time algorithms Requests indivisible but may be shifted NP-complete
Outline • Adaptive sensors • where to focus sensing? • multiple users • Prioritizing sensor control traffic • Robust routing in dynamic networks • Conclusions
Why prioritize sensor control traffic? • Sensor network • Closed-loop sensor network Bursty, high-bandwidth data Many-to-one routing to sink Congestion Wireless links Data Data spatially, temporally redundant Prefer to delay, drop data rather than control Sensor Controls Data >> control How does prioritizing sensor control traffic over data traffic impact application-level performance?
Closed-loop Sensor Networks • Prioritizing sensor control • impact on packet delays? • impact on data collected? • Control loop delay Control, data share queues e.g., wireless links Priority control delay Data delay FIFO control delay Data Data from control k Data from control k-1 Control k k-1 k+1 Update interval Small data delay, large control delay more data collected in time to compute next sensor control
Better Quality Data • More data samples Cramer-Rao bound: SD(W) ≥ 1 / n I • accuracy sub-linearly with n • Effect of data packet drops? • accuracy sub-linearly with n Radars, Sonars, Cameras, … Fisher information Std Dev of W from # of iid samples Lower bound on std dev of unbiased estimator W (sample mean) from parameter (population mean) Sensing accuracy changes slowly with # of samples
Storm Tracking Application • Network model • obtain sensor control and data packet delays • CASA network is closed-loop sensor network • Sensing model • convert packet delays into sensing error • Tracking model • convert sensing error into storm location error • tracking:compute next scan for radar from 99% confidence ellipse control Deterministic arrivals Delays at bottleneck link dominate, assume wireless links data other Bursty arrivals
Tracking Error + + RMSE = # intervals (truet-obst)2 √ t=1 # intervals + + idx = 1 idx = 25 idx = 55 Per-interval performance gains/losses may accumulate across multiple update intervals
Summary How to make sensing robust to delayed and dropped packets? When network congestion, prioritizing sensor control in closed-loop sensor network quantity, quality of data, and gives better application-level performance • Related work • SS7, ATM, [Fredj et al, 2001] [Kyasanur et al, 2005] • our focus: prioritizing sensor control (not network control)
Outline • Adaptive sensors • where to focus sensing? • multiple users • Prioritizing sensor control traffic • Robust routing in dynamic networks • Conclusions
What do we mean by robust? protocols must adapt network structure changing over time • But if frequent changes, adapting is costly: e.g., in MANET may have as much routing control traffic as data Adapt to every change? • yes: potentially perform optimally, but more overhead • no: likely perform sub-optimally, but less overhead Robust:solution performs well over many scenarios, solution is not fragile
Robust Routing Robust routing:routing subgraph has path from src to dest, as links up/down • src-dest reliability • prob instantaneous path in stochastic graph • want max reliability sub-graph for overhead But, reliability #P-complete to compute, can’t search over all sub-graphs Identify structural properties that make graph reliable, efficiently find subgraph with such properties • Effect of graph structure on src-dest reliability? • show reliability (in limits) dominated by shortest paths, smallest cuts Most robust routing subgraph should contain shortest path and have large min cut
Braid (shortest) k-hop braid: most reliable path + all nodes within k-hops 1-Hop Braid 2-Hop Braid Most Reliable Path s s s d d d Given fixed amount of overhead, is braid most reliable sub-graph? reliability simulations + theoretical analysis
Theoretical Analysis How Reliable is Braid? N s d d d s s Add black node rather than blue nodes? Lemma:Suppose sub-graph contains shortest path and 0<n<N 1-hop nodes. Given 1 or 2 extra nodes, to max reliability, use all 1-hop nodes before any 2-hop nodes Note: lemma does not hold when adding links Partial braid less reliable than 2-disjoint paths for 1p√2/3 2-Disjoint Paths Partial Braid
Braid Routing DSR vs Braid • path breaks • use new DSR path or existing 1-hop braid path • primary difference • control overhead incurred to find this new path Use dynamic source routing (DSR) If no path from src to dest: Step 1:Identify shortest path in network Step 2:Build braid around shortest path Step 3:Perform local forwarding within braid • e.g., flooding, opportunistic routing, backpressure Overheard RREQ and RREP contain 1-hop braid info When link breaks, use braid path back to DSR path
Simulation Set-up • QualNet • Gauss-Markov mobility • BonnMotion to generate traces • min 0.5 m/s, max 2 m/s • speed, angle updates every 100s • 20-80 nodes • 400m transmission radius • 2km x 2km area • 1 constant bit-rate flow • 4 pkts/s, 1 million seconds • 10 runs, each lasting life of flow
Control Overhead DSR # of Control Packets Braid # of Nodes As node density increases, braid incurs fewer control packets than DSR
Control Overhead Route Replies Route Errors DSR Braid DSR Braid # of Nodes # of Nodes Route Requests # of Control Packets DSR Braid # of Nodes Up to ~40% fewer replies Up to ~30% fewer requests Up to ~25% fewer errors Braid incurs fewer route requests, replies, errors than DSR
Packets Delivered and Delay DSR Braid Delay (seconds) % of Packets Delivered Braid (4 million packets) DSR # of Nodes # of Nodes Braid delivers slightly fewer packets, incurs higher delay than DSR
Summary How to make routing robust to network changes? Proposed routing algorithm that control overhead by updating routes less frequently performing local forwarding within routing sub-graph • gains depend on network characteristics • Related work • [Shacham et al 1983] [Lee, Gerla, 2000] [Ganesan et al, 2001] [Ghosh et al, 2007] • our work: differs in structure and/or usage of routing subgraph • Future work • which network characteristics most impact performance? • joint rate control and routing • what should be braid width (trade-off with interference)?
Outline • Adaptive sensors • where to focus sensing? • multiple users • Prioritizing sensor control traffic • Robust routing in dynamic networks • Conclusions
Conclusions • Adaptive sensors • where to focus sensing in adaptive meteorological radar network? • show lookahead strategies useful when multiple small phenomena, trade-off between scan quality and re-scan interval • accommodating multiple users? • identify call admission control problem, give complexity results • How to make sensing robust to delayed, dropped packets? • show good application-level performance possible in closed-loop sensor network when congestion if sensor control prioritized • How to make routing robust to network changes? • propose routing algorithm, show can significantly control overhead while minimally degrading % of packets delivered
Thanks! • Jim • Don, Deepak, Andy, Weibo • Networks Lab • Bruno, Mike, Yung-Chih, Daniel2, Majid, Yu, Bo, Patrick, Junning, Giovanni, Guto, Elisha, Suddu, Bing, Sookhyun, Chun … • ALL Lab • Sridhar, Mohammad, George, Sarah, Khash, Ash, Ozgur, Pippin … • Laurie, Tyler, …. Questions?
distance to storm cell Up(p, Sr) = max [ Fc(c(p, sr )) [ Fd(d(r,p)) + (1-) Fw(w(sr ) / 360)] ] srSr % covered radar rotation speed Us(ri, sr) = Fw(w(sr ) / 360)] Performance Metrics sr : radar configuration, start, end angles of scan sector Sr: set of radar configurations • Re-scan interval • # of decision epochs before storm cell first observed or rescanned • Quality • how well storm cell p observed • how well sector ri scanned
Difference between true and observed # of storm cells o Np Nd Np o C = |dij - dij| + (Np -Np) Pm + I(tk)Pr o j=1 i=1 k=1 Difference between observed and true storm attribute Storm scanned within Tr decision epochs? Performance Metrics Pm := penalty for never scanning storm Pr := penalty for not rescanning storm • Cost • Re-scan time and quality + penalty for never scanning storm cell Goal: maximize quality, minimize re-scan time
Assume: A, B, Q, R initialized using prior knowledge truet = Atruet-1 + N[0, Q] obst = B truet + N[0,R] Limited Look-ahead Strategy Use Kalman filters to predict storm cell attributes 1 and 2 decision epochs ahead • State • Actions • select scan action that minimizes cost • additionally scan any sector not scanned in last T=4 decision epochs True state:true= [ x, y, x, y ]T Observed state:obs= [ x, y ]T y x (x,y)
Full Look-ahead Strategy Markov Decision Process Formulation • State • Actions • Transition function • encodes observed environment dynamics, obtained from simulator • Cost function • obtained from performance metrics • Sarsa() • linear combination of basis functions to approximate value function • tile coding to obtain basis functions, one tiling for each state variable Storm radius y # of storm cells, Upquality of storm cells, Us quality of sectors x + (x,y)
Largest positive value of attribute = (1-u) Vmax / Us(ri,sr) quality scaling term Simulation Set-up • True state • storms arrivals: spatio-temporal Poisson process • storm attributes from distributions derived from real data • max storm radius: 4km • max number of storms • Observed state • observed attribute value = true attribute value plus noise ~ N[0, 2] 10 km or 30 km
Scan Quality Max 1 storm 2Step - Sarsa SitandSpin - Sarsa Avg Difference in Quality (250,000 steps) Max 4 storms 1/ 2-Step scans have higher quality than Sarsa(), especially when little noise in environment (when 1/ is small)
# timesteps 2step Full Ct - Ct t=1 # timesteps Cost 2 radars Average Difference in Cost SitandSpin - Full Lookahead 2StepLookahead - FullLookahead 1/ Full lookahead and 2-step look-ahead have similar costs
Re-scan Interval Sit-and-Spin 1-Step Sarsa() more likely than 2-step look-ahead to scan storm within Tr=4 decision epochs 2-Step P[X≤ x] Sarsa() x = # of decision epochs between storm scans
Related Work Large State-space Reinforcement Learning Radar Control • 2005: Stone, Sutton, Kuhlmann • robot soccer • 2004: Ng, Coates, Diel, Ganapathi, Schulte, Tse, Berger, Liang • helicopter control • 2002: Zilberstein, Washington, Bernstein, Mouaddib • planetary rovers • 2005: Kreucher, Hero • look-ahead scheduling of radars on airplanes for detecting and tracking ground targets • information-theoretic reward, Q-learning • 2005: Suvorova, Musicki, Moran, Howard, Scala • target radar beams, select waveform for electronically steered phased array radars • show 2-step lookahead outperforms one-step look-ahead for tracking multiple targets Weconsider tracking meteorological phenomena using ground radars Sensor Networks • 2005: Mainland, Parkes, Welsh • game theory + reinforcement learning to allocate resources • learn profit associated with different actions, rather than profit associated with different state-action pairs Do not consider infinite-horizon case
Divisible, No Shifting • Polynomial-time • assume utility depends on how much of request executed • select max utility sensor request during each conflicting interval Sensing Strategy for User 1 Sensing Strategy for User 2 Interleaved Sensor Requests