Algorithmic Challenges in Sensor Networks

1
Algorithmic Challenges inSensor Networks
Leonidas Guibas Computer Science Dept. Stanford
University
2
Some Obvious Desiderata

• Naturally, we want algorithms that are
• efficient
• distributed
• energy-aware
• robust

CACM, 04
3
Use Data Redundancy to FightImperfect Calibration

• Topological data analysis
• Large-scale features of a signal landscape should
  be visible even without perfect
• localization
• time-synch
• sensor calibration

4
Keeping In-Network Processing Simple

• Compressed sensing
• Exploit network coding to produce universal data
  aggregators
• Data read-out proportional to signal landscape
  complexity

Base station
Data aggregation
5
Global Advice for Local Methods

• Local, greedy, reactive methods are natural for
  sensor networks
• Such methods can get stuck is local minima
• How much global knowledge is needed to bypass
  most of these
• How can this knowledge be obtained robustly?

6
Resource-Adaptive Algorithms

• The resources (e.g., computation, storage)
  available to a sensor network algorithm may vary
  dynamically, even as the algorithm runs
• How can we design algorithms that move smoothly
  over a quality vs. resources trade-off curve,
  even as they run?

q-digest update Shrivastava et. al., SenSys04
7
The Impact of AlternativeInformation
Representations

• The identity management problem
• Operations
• evidence about identity
• target mixing
• Marginal probabilities require global
  normalization at each evidence event
• Accumulated log-likelihoods do not ...

8
Parting Thoughts

• Successful algorithmic design depends upon
  certain stable abstractions that can describe the
  functionalities of a sensor network, as well as
  their associated costs
• Good data models are essential as well the
  special structure of the data in each application
  must be exploited for the best algorithms

9
The End