Presentation #8
SuMo-SS: Submodular Optimization Sensor Scattering for
Deploying Sensor Networks by Drones
Komei Sugiura
National Institute of Information and Communications Tech., Japan
Presentation #8
Target task = sensor scattering:
Automatic deployment of disposable wireless sensors by a drone
Motivation
• Deploying sensors by drones instead of humans has advantages in terms of
worker safety and time requirements
Technical challenge
• Optimal plan to maximize information gain from scattered sensors
Human detection
in landslideFlash flood detection
[Abdulaal+, IWRSN14]
Contamination
detection
NP-hard
Presentation #8
We prove that SubModular Sensor Scattering (SuMo-SS) is online, semi-
optimal, and can handle uncertainty
Sensor scattering problem is NP-hard
• Combinatorial explosion
Related work utilized submodular
optimization (e.g. [Krause 08])
• (1-1/e)-approximation = 63% of optimal
score is guaranteed
• However, uncertainty was not handled
Presentation #8
We prove that SubModular Sensor Scattering (SuMo-SS) is online, semi-
optimal, and can handle uncertainty
Sensor scattering problem is NP-hard
• Combinatorial explosion
Related work utilized submodular
optimization (e.g. [Krause 08])
• (1-1/e)-approximation = 63% of optimal
score is guaranteed
• However, uncertainty was not handled
Proposed method (SuMo-SS)
• Handles uncertainty in sensor positions
• Does not suffer from combinatorial
explosion
• (1-1/e)-approximation is guaranteed
Without remote control
x10
Presentation #8
Sensor model
1. Observations from sensor sets follow the
Gaussian distribution
2. Covariance between observations y and y’
can be approximated by the RBF kernel
SuMo-SS does not need to know actual landed positions of sensors
Sensor positions
Input: Covariance between previously scattered
sensors and their target positions
Output: Next target position & info. gain
We give theoretical proof on
submodularity
No information about actual
landed positions is required
Presentation #8
Experimental setup:
We used a simulation environment to make experimental results reproducible
Physical model
• AR.Drone 2.0 with a customized
electromagnetic device for
attaching/detaching a sensor
• Monocular SLAM[Engel+ 14]
Simulation model
• Purpose: To make the results
reproducible
– cf. Estimated lifetime of the
physical drone is <100 h
Presentation #8
Presentation #8
Quantitative results: Proposed method obtained larger mutual information
than baseline and random selection methods
Metric = Cumulative mutual info.
(a) Proposed (SuMo-SS)
(b) Baseline [Krause 08]
(c) Random selection
Sensitivity analysis
• Proposed method outperformed
the baseline[Krause 08] in 43/49
conditions
Deviation in x-axis
Deviationiny-axis
*Average of 10 experiments
Cumulativemutualinfo.
Presentation #8
Target
task
Automatic deployment of disposable wireless sensors
by a drone
Proposed
method
SuMo-SS can deal with uncertainty in sensor positions
Results Proposed method obtained larger mutual information
than baseline and random selection methods
Interactive presentation #8

SuMo-SS: Submodular Optimization Sensor Scattering for Deploying Sensor Networks by Drones

  • 1.
    Presentation #8 SuMo-SS: SubmodularOptimization Sensor Scattering for Deploying Sensor Networks by Drones Komei Sugiura National Institute of Information and Communications Tech., Japan
  • 2.
    Presentation #8 Target task= sensor scattering: Automatic deployment of disposable wireless sensors by a drone Motivation • Deploying sensors by drones instead of humans has advantages in terms of worker safety and time requirements Technical challenge • Optimal plan to maximize information gain from scattered sensors Human detection in landslideFlash flood detection [Abdulaal+, IWRSN14] Contamination detection NP-hard
  • 3.
    Presentation #8 We provethat SubModular Sensor Scattering (SuMo-SS) is online, semi- optimal, and can handle uncertainty Sensor scattering problem is NP-hard • Combinatorial explosion Related work utilized submodular optimization (e.g. [Krause 08]) • (1-1/e)-approximation = 63% of optimal score is guaranteed • However, uncertainty was not handled
  • 4.
    Presentation #8 We provethat SubModular Sensor Scattering (SuMo-SS) is online, semi- optimal, and can handle uncertainty Sensor scattering problem is NP-hard • Combinatorial explosion Related work utilized submodular optimization (e.g. [Krause 08]) • (1-1/e)-approximation = 63% of optimal score is guaranteed • However, uncertainty was not handled Proposed method (SuMo-SS) • Handles uncertainty in sensor positions • Does not suffer from combinatorial explosion • (1-1/e)-approximation is guaranteed Without remote control x10
  • 5.
    Presentation #8 Sensor model 1.Observations from sensor sets follow the Gaussian distribution 2. Covariance between observations y and y’ can be approximated by the RBF kernel SuMo-SS does not need to know actual landed positions of sensors Sensor positions Input: Covariance between previously scattered sensors and their target positions Output: Next target position & info. gain We give theoretical proof on submodularity No information about actual landed positions is required
  • 6.
    Presentation #8 Experimental setup: Weused a simulation environment to make experimental results reproducible Physical model • AR.Drone 2.0 with a customized electromagnetic device for attaching/detaching a sensor • Monocular SLAM[Engel+ 14] Simulation model • Purpose: To make the results reproducible – cf. Estimated lifetime of the physical drone is <100 h
  • 7.
  • 8.
    Presentation #8 Quantitative results:Proposed method obtained larger mutual information than baseline and random selection methods Metric = Cumulative mutual info. (a) Proposed (SuMo-SS) (b) Baseline [Krause 08] (c) Random selection Sensitivity analysis • Proposed method outperformed the baseline[Krause 08] in 43/49 conditions Deviation in x-axis Deviationiny-axis *Average of 10 experiments Cumulativemutualinfo.
  • 9.
    Presentation #8 Target task Automatic deploymentof disposable wireless sensors by a drone Proposed method SuMo-SS can deal with uncertainty in sensor positions Results Proposed method obtained larger mutual information than baseline and random selection methods Interactive presentation #8