Post on 29-Dec-2015
Multi-Scale Sampling
Outline
• Introduction• Information technology challenges• Example: light patterns in forest
Greg Pottie
Pottie@icsl.ucla.edu
Example: Light Pattern Sampling
• Photosynthesis begins above some (species-specific) incident light intensity threshold, and eventually saturates– Pattern of light levels thus conveys more useful information than simple
average of intensity
• Do not necessarily need to reconstruct the field
• Selected statistics may be sufficient– E.g. histograms of intensity
levels, characterization of light level dynamics
• Natural scenes are very complicated– Shadows from many levels
Some Early NIMS PAR Measurements
• Large variations over short distances– Pure static network will
require unreasonable density for field reconstruction
• Pure mobile network will give misleading results with respect to rapid dynamics (e.g. branches blowing in wind)
• Some type of hybrid strategy suggested
Adaptive Sampling Strategies
• Over-deploy: focus on scheduling which nodes are on at a given time
• Actuate: work with smaller node densities, but allow nodes to move to respond to environmental dynamics
• Our apps (Terrestrial, Aquatic, Contaminant) are at unprecedented scales and highly dynamic: over-deployment is not an option– Always undersampled with respect to some
phenomenon
– Focus on infrastructure supported mobility
– Passive supports (tethers, buoyancy)
– Small number of moving nodes
• Many approaches/experiments explored in past two years
Speedup using Static and Mobile Nodes
t=1
t=2
t=3
• Static nodes act as triggers• Network of static nodes ‘allocates’ tasks to mobile nodes
Sampling
• Lattice/deterministic pattern• Gradient methods (e.g. Newton, LMS)• Simulated annealing (guided walks)• Multi-scale or multi-dwell• Bisection/quad algorithms (decision
trees with multiple depths)• Random walk
• Sampling strategy depends on physical model, objectives, and available resources
• E.g., finding global max or min of a field depends on smoothness:
Smooth
Rough
• Similar stacks can be constructed for other sampling objectives (e.g., reconstruction, gathering statistics)
Multiscale Approach
• Goal is optimization of the hierarchical system– Not merely optimization of devices or any given
layer
– Models, devices, algorithms require co-design
• Context-driven algorithms– No single algorithm/device is best in all
situations
– Context is the state of the next level in the hierarchy; choose resources to apply when drilling down to next level according to this state
– Probabilistic constraints and algorithms lead to more new optimizations
• Examples: adaptive and multi-scale sampling
Optimization Paradigm
Humans
State Detection
Algorithms for each state
Minimize involvement in routine tasks; employ for difficult cognitive tasks
State of next model in hierarchy determines which algorithmic suite to use
Can have multiple algorithms, possibly hierarchically arranged
Play probability game to minimize costs of higher level reasoning; employ hierarchy of algorithmic approaches
Multi-level Processing
• Goal: construction of tree of (re-usable) algorithms and models relating physical structure of forest canopy to light levels on forest floor, and consequences for plant growth.
• Physical Model:– Fixed elements: trunks, major branches,
topography (deterministic)
– Variable elements: branches, leaves, sun position, clouds (statistical)
• Algorithmic Set:– Search algorithms to create maps of canopy and
ground patterns; complexity and choice will vary spatially
– Higher level reasoning to relate data to science question, determine model parameters (e.g., Bayes, rule-based, formal optimizations)
Physical Models
• Models apply at different levels of abstraction– Abstraction level(s) much match query
• Example: information from sources– Attenuation with distance from source
– Statistical description (e.g., Correlations in time/space/transform domain, source and environmental dynamics)
– Disk model
– Statistical aggregations (e.g., flows on graph representing network)
• Model depends also on sensor data– Different statistics at different spatial scales and
for different sensing modes
– Data set affects number of viable hypotheses
• Feasibility depends on algorithmic availability– Need computationally effective suite
Model Uncertainty in Sensor Networks
• How much information is required to trust a model?
• Approach: information theoretic analysis of benefits of redundancy and auditing in model creation
• Will be backed up by simulations and experiments
In how many locations must a field be sampled (by combination of static and mobile nodes) to determine it is caused by one (or more) point sources?
Data Integrity
Multiple nodes observe source, exchangereputation information, and then interactwith mobile audit node
• How can we trust the data produced by a sensor network?
• Approaches– Redundant deployment
– Mobile auditors
– Hybrid schemes
• Components– Reputation systems
– Statistical analysis of information flows
Multi-scale sampling
• A homogeneous screen is placed to create a reflection Er proportional to incident light Ec.
• Camera captures the reflection on its CCD
• The image pixel intensity is transformed to Er using camera’s characteristic curve.
• Sensors with different modes and spatial resolutions– E.g. NIMS PAR sensor and
camera– PAR measures local incident
intensity– Camera measures relative
reflected intensity
• Provides better spatial and temporal resolution, at cost of requiring careful calibration
• Analogous to remote sensing on local scales
Conclusion
• multi-layered systems provide major opportunities– Overall complexity can be significantly reduced, from both
hardware and software perspective– Re-use of components in a variety of settings
• Multi-layered systems present many research challenges– Fundamental research questions in model construction, data
and model uncertainty, information flows among layers, and large-scale systems design/optimization
– Necessary to pursue mix of lab and field experiments to ensure realism in problems and generality of results