Self-Organizing Search in the Web of Things Kay Rmer, University of - PowerPoint PPT Presentation

Self-Organizing Search in the Web of Things Kay Römer, University of Lübeck

Networked Embedded Sensing Research  Protocols  Adaptation to interference  Programming Models  Role assignment  Services  Content-based Sensor Search  Minimally-Invasive Management  Systems  Body sensor networks

Motivation Camera Accelerometer  Mobile phones equipped GPS Ambient Light Bluetooth Microphone with sensors and connected GSM Magnetometer to the Internet WLAN Proximity  Sensors published on the Web: state of the real world available in real-time  Search the real world by  Search the real world by its current state ! its current state ! 2

3 Mensa empty quiet A Web of Real-World Places tation hectic full S upermarket quiet full S

Web of Things  Web presence of things, people, and places with real-time state information  Web of real-world entities, not Web of sensors  High-level states, not raw sensor data  Searching the Web of Things  Search for real-world entities: places, people, things, …  by their current state: empty, hot, broken, …  in real-time 4

Searching the Real World: Examples  Quiet picnic places at waterfront?  Route through city avoiding traffic jams?  Which rental station has bicycles available?  Where are many people who share my interests?  Which trains from A to B are not crowded?  Where to enter train to get free seat?  Supermarkets with short waiting queues? 5

Problem: Content-based Sensor Search  Find sensors reading given state in real time  Potentially huge, distributed set of candidate sensors  More state updates than queries, push not a good idea!  Sensor output is highly dynamic  Indexing sensor output not a good idea!  We need only a limited number of results at a time  Heuristics to select good candidates! 6

Approach: Sensor Ranking  Sensors create prediction model using past readings Indexing Time  Prediction models are published on the Web  Search engine periodically indexes prediction models  Prediction models are used to rank candidate sensors Query Time  Highest ranking sensors are read first  Goal: Minimize the number of read sensors 7

System Model  Sensor maps discrete time to a finite discrete set of states: s : T  V   Sensor output time series: s( t i ) = v i Sensor Output ... V 1 V 2 V 3 V 4 ... V present ... t 1 t 2 t 3 t 4 t present ... Time 8

System Model (Continued)  Prediction model maps query time and query value to a probability estimate:    P T V  : 0 , 1  P( t , v ): Probability that s( t ) = v search for v model construction Sensor ... V 1 V 2 V 3 V 4 ... V c ... ? Output ... t 1 t 2 t 3 t 4 ... t C ... t Time time window TW forecasting horizon 9

Query Resolution Example: Quiet places at waterfront 1. Filter static (waterfront, occupancy) 2. Predict (quiet) 3. Rank 4. Read 5. Return .7 .1 .9 .5 .6 .2 .2

Ranking Metrics  Normalized overhead for reading non-matching sensors  Ranking error e(t,v) Number of non-matching sensors above last matching sensor Rank of last matching sensor  Top-m ranking error e top (t,v,m)  Dito, but only first m sensors considered 15

Ranking Metrics: Examples Suboptimal ranking: Optimal ranking: S1 S1 • e = 4/8 • e = 0/6 S2 S2 • e top = 2/5 • e top = 0/5 S3 S3 S4 S4 m m S5 S5 Last match S6 S6 Sensors that S7 S7 match the query Last match S8 S8 Sensors that do not match the S9 S9 query 16

Prediction Models  Focus on people-centric sensors  Tend to show periodic behaviour  Requirements  Accurate predictions for forecasting horizons that match indexing frequencies (days - weeks)  Deal with imperfect periodic behavior 17

Considered Prediction Models  Single-period prediction model (SPPM)  Assumes single dominant period of known length (e.g., 1 week)  Multi-period prediction model (MPPM)  Assumes multiple periodic processes of unknown length (e.g., 1 week, 4 weeks)  Select appropriate models at runtime 18

Single-Period Prediction Model (SPPM)  Assumption: Single dominant period with length p p p p p Sensor V 1 V 1 V 2 V 4 V 4 ... ... ... ? V 3 Output t C t 1 t 3 t 4 ... t t 2 ... Time time window TW P(t,green) = 2/4 P(t,red) = 1/4 19

Number of consecutive appearances of symbol α in period p at offset l Max. possible occurences Multi-Period Prediction Model (MPPM)   p l  Periodic symbol ( , , , )  α : symbol ; p : period ; l : offset ;  : support  Example: α =blue, p =2, l =1 p =2 p =2 p =2 l =1 Sensor Output   ps  blue ,2,1,2     3 21

Inferring Prediction Estimates Query for value v=b at time t=6 1. Filter periodic symbols Same value:  = v  Same phase: l  t mod p  2. P(v,t) = max  v = b ?   p l b 2 0 0.7 g 3 1 0.1 b 4 2 0.9 0 1 2 3 4 5 6 23

Adjustment Process  Faulty/malicious sensors, inaccurate S1 predictions may result in persistent S2 misranking S3  Individual ranking error for each sensor S4  S8 ranked to low: increase prediction value S5  S7 ranked to high: decrease prediction value S6  Idea: adjustment term for each sensor S7  Updated after each query using ranking error S8 E7 = -2/9 E8 = +6/9 S9 24

Adjustment Process: Feedback Loop C3 A1 + S1 S1 S3 S1 P1 P3 + + E3 E1 C3 A2 + S2 S2 P2 P3 S3 S2 + + E2 E3 C3 A3 + S3 S3 S3 S3 P3 P3 + + E3 E3 C3 A4 + S4 S4 P4 P3 + + E4 E3 S4 S3 C3 A5 + S5 S5 S3 S5 P3 P5 + + E3 E5 25 25

Evaluation  Simulation of a realistic search engine  Periodic rebuild and indexing of models (1 week)  Periodic queries for possible values  Measure average ranking error  Prediction models: Random, SPPM, MPPM  With / without adjustment 26

Evaluation: Data Sets  MERL motion detector dataset  50 PIR sensors in office building  PIR output mapped to “ free” and “ occupied”  With and without a “faulty” sensor  ETH room reservation system  7 “sensors”  Room occupancy: “ free ” or “ occupied ”  With and without “synthethic” multiperiod sensor  Bicing data set (in progress)  350 bicycle rental stations in Barcelona  Number of available bicycles: “no”, “few”, “many”

Average Ranking Error: MERL Average Ranking Error P P D D MPPM SPPM+AP SPPM+AP MPPM M SPPM A A N N P + A + A P M M R R S P P P P M M 30

Average Ranking Error vs. Top m Average Ranking Error x5.5 x10 #Top Entries 31

Summary  Ubiquitous sensors connected to Internet  Search for real-world entities by current state  Sensor Ranking, a primitive for content-based sensor search utilizing prediction models  Adjustment process to alleviate persistent inaccurate rankings  Promising results on real-world data sets  Ongoing work  Improved ranking based on correlations  Building a search engine 32

Ads  Act-Control-Move: Beyond networked Sensors  Summer School, Schloss Dagstuhl, August 15-21, 2010  www.cooperating-objects.eu/school  IEEE SUTC (Sensor Networks, Ubiquitous & Trustworthy Computing)  Conference, Newport Beach, California, June 7-9, 2010  sutc2010.eecs.uci.eu  SESENA (Software Engineering for Sensor Nets)  ICSE Workshop, CapeTown, South Africa, May 3, 2010  www.sesena.info