ModelDriven Data Acquisition in Sensor Networks Part II

1
Model-Driven Data Acquisition in Sensor
Networks- Part II -

• Paper by Amol Deshpande, Carlos Guestrin, Samuel
  R. Madden, Joseph M. Hellerstein and Wei Hong
• In Proceedings of VLDB 2004
• Presented by Shantha Ramachandran
• Oct. 4, 2004

2
Last Time

• BBQ probabilistic model used to create
  observation plan for sensor network
• Probability density function
• Value, range and average queries

3
Outline

• Introduction
• Overview of approach
• Model-based querying
• Choosing an observation plan
• Experimental results
• Extensions and future directions
• Related work
• Conclusions

4
Choosing an Observation Plan

• Pdfs can be conditioned on the value of observed
  attributes o
• Gives us a more confident answer to a query
• Important which attribute to observe?
• Focus select attributes that are expected to
  increase the confidences in the answer to the
  query, at a minimum cost

5
Cost of Observations

• Let O 1,..,n be a set of observations
• Expected cost C(O) of observing attributes O
• C(O) Ca(O) Ct(O)

6

• Ca(O) data acquisition cost
• Sum of energy required to observe attributes O
• Ca(O) SieOCa(i)
• where Ca(i) is the cost of observing attribute i

7

• Ct(O) expected data transmission cost
• Depends on data collection mechanism used to
  collect observations from network
• Depends on network topology
• If topology is unknown or changing, cost function
  is basically random
• Therefore, assume networks with known topologies

8

• Network graph
• Set of edges ?
• Each edge eij
• 2 link quality estimates pij, pji
• Probability that packet from i will reach j
• Assume pij, pji are independent
• Expected of transmission acknowledgement
  messages required to guarantee a successful
  transmission is 1 / pijpji
• Use these values to estimate transmission cost

9

• Choose a simple path through network that visits
  all sensors, observes O, returns
• Ct(O) is defined to be expected cost of this path
• C(O) Ca(O) Ct(O)

10
Improvement in Confidence

• Observing attributes O should improve the
  confidence of posterior density
• Should be able to answer query with higher level
  of confidence

11

• Suppose we have a range query Xi e ai, bi
• We can compute benefit Ri(o) of o
• Ri(o) maxP(Xi e ai, bio), 1- P(Xi e ai,
  bio)
• Ri(o) measures our confidence after observing o

12

• For value and average queries
• Ri(o) P(Xi e xi-e, xieo)
• xi is the posterior mean of Xi given o

13

• Specific value o of O is not known a priori
• Must compute expected benefit Ri(O)
• Ri(O) ?p(o)Ri(o)do

14

• We may have range or value queries over multiple
  attributes
• Trying to achieve a particular marginal
  confidence over each attribute
• Must decide how to trade off confidences between
  different attributes

15

• For a query over attributes Q 1,,n
• Define total benefit R(o) as either
• R(o) minieQRi(o)
• R(o) 1/Q SieQRi(o)
• Focus minimize number of mistakes made by query
  processor
• Use average benefit to decide when to stop
  observing new attributes

16
Optimization

• We have so far defined R(O) and C(O) as expected
  benefit and cost
• Different sets of observed attributes lead to
  different benefits and costs
• If user wants confidence level 1-d, we want to
  pick a set of attributes O that meet the
  confidence at minimum cost
• minimizeO C(O),such that R(O)gt 1- d
• This is generally NP-hard

17

• Two algorithms for solving optimization problem
• Exhaustive search
• Greedy algorithm

18
Exhaustive Search

• Exhaustively search over all possible subsets of
  possible observations, O
• Finds the optimal subset
• Exponential running time

19
Greedy Algorithm

• Uses a greedy incremental heuristic
• Initialize with empty set of attributes, O ø
• For each attribute Xi not in set
• Compute new expected benefit R(OUi) and cost
  C(OUi)
• If some set G reaches desired confidence
• Then pick from G the one with lowest total cost
• And terminate
• Else if G ø, we have not reached our desired
  confidence
• Then add attribute with highest benefit over cost
  ratio
• Repeat until desired confidence is reached

20
Experimental Results

• Measure performance of BBQ on real world data
  sets
• Goal demonstrate that BBQ provides ability to
  efficiently execute approximate queries with user
  specified confidences

21
Data Sets

• Results based on running experiments on two real
  world data sets
• Collected using TinyDB

22
Garden

• One month trace of 83,000 readings
• 11 sensors in a redwood tree at UC Botanical
  Garden in Berkeley
• Sensors placed at four different altitudes
• Collected light, humidity, temperature and
  voltage readings once every five minutes
• Data split into training and test data sets
• Model built on training set

23
Lab

• 54 sensors in Intel Research, Berkeley lab
• Collected light, humidity, temperature and
  voltage readings
• Also collected network connectivity information
• 8 days of readings
• 6 days training
• 2 days test

24
Query Workload

• Two sets of query workloads
• Value queries
• Predicate queries

25
Value Queries

• Main type of queries anticipated
• Ask to report sensor readings at all sensors
• Within error bound e
• With specified confidence d

26
Predicate Queries

• Selection queries over sensor readings
• Ask for all sensors that satisfy a certain
  predicate
• With specified confidence d
• Also looked at average queries
• Do not present these results

27
Comparison Systems

• Compare BBQ against
• TinyDB-style Querying
• Approximate-Caching

28
TinyDB-style Querying

• Query disseminated into sensor network using tree
  structure
• At each mote, sensor reading is observed
• Results reported back along same tree to base
  station
• Combine results on the way back to minimize
  communication costs

29
Approximate-Caching

• Base station maintains view of readings at all
  motes
• View is guaranteed to be within a certain
  interval of the actual sensor readings
• If value of sensor falls outside this interval,
  motes are required to report new reading

30
Methodology

• BBQ used to build model of training data
• Includes transition model for each hour of the
  day
• Generate traces from test data by taking one
  reading randomly per hour
• Issue one query per hour

31

• Model computes a priori probability for each
  predicate
• Choose one or more sensor readings to observe if
  confidence bounds not met
• Execute generated observation plan
• Updates model with observed values from test data
• Compares predicted values for non-observed
  readings to test data from that hour

32

• Measure accuracy
• Compute average mistakes per hour
• How many reported values are further away from
  the actual values than the specified error bound
• Number of predicates whose truth value was
  incorrectly approximated

33

• TinyDB
• All queries answer correctly
• Approximate-Caching
• Values reported upon deviation
• No mistakes!

34

• Compute cost and accuracy for each observation
  plan
• Both acquisition cost and communication cost

35
Garden datasetValue Based Queries

• Want to analyze performance of value queries on
  garden in detail
• Show effectiveness of BBQ
• Query requires system to report temperatures at
  all motes to within specified e
• Confidence 97, e varied

36

• Vary e from 0 to 1degrees Celsius
• Cost of BBQ falls rapidly as e increases
• Percentage of errors stays below 5 confidence
  threshold

Figure 4 Relative Costs 1
37

• For reasonable values of e
• BBQ uses significantly less communication
• Approximate-Caching always reports values to
  within e
• Does not make mistakes
• Average observation error is close to BBQ

38

• Percentage of sensors that BBQ observes by hour
• Varying e

Figure 5 Number of sensors 1
39

• As e gets small (lt0.1), must observe all nodes on
  every query
• Variance between nodes high enough that it cannot
  infer value of one sensor from anothers with any
  accuracy
• As e gets large (gt1), few observations are needed
• Changes in one sensor predict values of others
• Intermediate e
• More observations are needed, especially during
  times when readings change drastically

40
Garden DatasetCost vs. Confidence

• Compare cost of plan execution
• confidence 80 to 99
• epsilon varying between 0.1 and 1.0

Figure 6 Energy and errors vs. confidence
interval and epsilon 1
41

• Decreasing confidence intervals or epsilon
  reduces energy per query

• Confidence 95
• Errors 0.5
• Reduce expected energy cost from 5.4 J to 150 mJ
  per query
• Factor of 40 reduction

Figure 6A 1
42

• Meet or exceed confidence interval in almost all
  cases

• Except 99 confidence

Figure 6B 1
43
Additional Experiments

• Performance of greedy algorithm vs. optimal
  algorithm
• Performance of dynamic filter vs. static model

44
Garden DatasetRange Queries

• Ran a number of experiments with range queries
• Average number of observations required for 95
  confidence

Figure 7 BBQs performace 1
45

• 3 different range queries
• Temp. in 17,18
• Temp. in 19,20
• Temp. in 21,22
• In all 3 cases, error rates were at or below 5
• Different range queries required observations at
  different times

46
Lab Dataset

• Similar experiments run on lab dataset
• Higher number of attributes
• Temperatures harder to predict than outdoors
• Human intervention -gt randomness

47

• Cost incurred answering value query
• Confidence varied
• As required confidence drops, BBQ is more
  efficient

Figure 8A Energy vs. confidence interval and
epsilon 1
48

• BBQ achieved specified confidence bounds in
  almost all cases

Figure 8B Errors vs. confidence interval and
epsilon 1
49

• Example traversal executing a value query
• 99 confidence, e0.5 degrees C
• Initial set, 8 am

Figure 9 Traversals of the lab network 1
50
Extensions and Future Directions

• So far, authors have focused on core architecture
  of BBQ
• Goal unifying probabilistic models with
  declarative queries
• Several possible extensions

51
Conditional Plans

• Generate plans that include early stopping
  conditions
• Generate plans that explore different parts of
  the network, depending on the values of observed
  attributes

52
More Complex Models

• In particular, models that detect faulty sensors
• Answer fault detection queries
• Give correct answers to queries in the presence
  of faults

53
Outliers

• Does not work well in current implementation
• Only way to detect outliers is to continuously
  sample sensors
• Outlier scheme would likely have high sensing
  cost
• Probabilistic techniques are expected to work to
  avoid excessive communication

54
Support for Dynamic Networks

• Current approach works best in static network
• Systematic study how network topologies change
  over time
• New sensors added, existing sensors move
• Topology change recovery strategies
• Find alternate routes through network

55
Continuous Queries

• Currently, exploration plan re-executes at root
  node
• May be possible to install code that causes
  devices to periodically push readings during
  times of high change

56
Related Work

• Substantial work has been done on approximate
  query processing in the database community
• Using model-like synopses for query answering
• Instead of probabilistic models
• Most do not use correlations

57
AQUA Project

• Proposes sampling-based synopses that can provide
  approximate answers to a variety of queries using
  a fraction of the total data in the database
• Includes tight bounds on the correctness of the
  answers
• Designed to work in and environment where it is
  possible to generate independent random samples
  of data

58

• Does not exploit correlations
• Lacks predictive power of probabilistic models
• Others propose exploiting correlations through
  graphical model techniques for approximate query
  processing

59
Approximate Caching

• Olsten et al.
• Provides bounded approximation of values of a
  number of cached objects at some server
• Server stores cached values along with absolute
  bounds for deviation
• When objects notice values outside bounds, they
  send an update

60

• This requires cached objects to continuously
  monitor values
• High energy overhead
• Can detect outliers
• BBQ cannot

61
IDSQInformation Driven Sensor Querying

• Probabilistic models
• Estimation of target position in tracking
  applications
• Sensors tasked to maximally reduce positional
  uncertainty of target

62
ACQPAcquisitional Query Processing

• Query processing in an environment like sensor
  networks
• Must be sensitive to costs of acquiring data
• Main goal avoid unnecessary data acquisition

63
CONTROL

• Provide interface that allows users to see
  partially complete answers with confidence bounds
  for long running aggregate queries
• No correlations

64
Conclusions

• Proposed architecture for integrating database
  systems with correlation-aware probabilistic
  model
• Do not directly query the network
• Rather, build model from stored and current
  readings
• Answer SQL queries by consulting model

65

• Advantages to using model in sensor network
• Shield users from faulty sensors
• Reduce number of expensive sensor readings and
  radio transmissions

66

• BBQ shows encouraging order of magnitude
  reductions in sampling and communication costs
• BBQ general architecture is seen as proper
  platform for answer queries and interpreting data
  from real-world environments like sensornets
• Conventional database technology is not equipped
  to deal with lossiness, noise and non-uniformity
  which is inherent in such environments

67
Thank YouQuestions?
1 A. Deshpande, C. Guestrin, S. Madden, J.
Hellerstein, W. Hong. Model-Driven Data
Acquisition in Sensor Networks. In Proc. Of the
30th VLDB Conference, 2004.