The accurate real-time estimation of
chemical, radioactive, and biological contaminant plumes (and BP oil, Icelandic
ash, …), leveraging targeted deployments of sensor-laden UAVs/UGVs/USVs/UUVs
(unmanned
aerial/ground/surface/underwater vehicles), is a class of interdisciplinary
grand challenge problems of both acute societal impact and intense scientific,
mathematical, and technological interest. This talk will discuss our group's
recent efforts on two algorithms at the core of this class of problems. The
first, which represents a significant advance in estimation theory for
large-scale systems, is the development of the Hybrid Ensemble Smoother (HEnS)
algorithm for data assimilation, which effectively and tractably combines the
key strengths and numerical tractability of the Space/Time Variational (4Dvar,
aka Moving-Horizon Estimation or MHE) approach and the Ensemble Kalman
Filter (EnKF) algorithm used widely in the Numerical Weather Prediction (NWP)
community. The second, which amounts essentially to the dual of this algorithm,
is the development of the Ensemble/Variational Optimization (EnVO) algorithm
for targeting sensor vehicle movements to optimally address the precise identification
and forecasting problems for which these sensor vehicles are deployed, while
complying with the dynamic constraints on the motions of the vehicles
themselves. Experimental testing of these algorithms, leveraging
high-performance computing resources communicating over a low-bandwidth
cellular data link with a dozen UGVs deployed in large smoke plume released in
a parking lot, is also discussed, as well as the potential extension of these
techniques to both wildland and structural firefighting applications.
Prof. Thomas Bewley (BS/MS Caltech 1989, PhD
Stanford 1998) works at the intersection of control theory, fluid mechanics,
numerical methods, and applied math, and has a particular interest in the
analysis, estimation, & forecasting of environmental flow systems. Related
projects include in-situ hurricane and ocean current monitoring leveraging
buoyancy-controlled balloons and drifters, and advanced work in simultaneous
localization and mapping (SLAM) for coordinated deployments of robotic
vehicles. Other recent work includes the use of n-dimensional sphere packing
theory for both derivative-free optimization as well as the design of efficient
new switchless topologies for structured computational interconnects for
massively-parallel computing. Bewley is the sole author of the
forthcoming textbook Numerical Renaissance: simulation, optimization, &
control, as well as its extensive accompanying codebase.
