Scan Strategies for Adaptive Meteorological Radars
Victoria Manfredi, Jim Kurose, U Massachusetts Amherst
3. Scan Strategies
Abstract
We address the problem of adaptive sensor control in dynamic resource-constrained sensor networks. We focus on
a meteorological sensing network comprising radars that can perform sector scanning rather than always scanning
360. Such a network is currently being developed and deployed by the Collaborative Adaptive Sensing of the
Atmosphere (CASA) NSF Engineering Research Center. We compare three sector scanning strategies. The sitand-spin strategy always scans 360. The limited lookahead strategy additionally uses the expected environmental
state K decision epochs in the future, as predicted from Kalman filters, in its decision-making. The full lookahead
strategy uses all expected future states by casting the problem as a Markov decision process and using
reinforcement learning to estimate the optimal scan strategy. We show that the main benefits of using a lookahead
strategy are when there are multiple meteorological phenomena in the environment, and when the maximum radius
of any phenomenon is sufficiently smaller than the radius of the radars. We also show that there is a trade-off
between the average quality with which a phenomenon is scanned and the number of decision epochs before
which a phenomenon is rescanned.
Sit-and-spin
 All radars always scan 360
Limited “lookahead”
 Use Kalman filters to predict future attributes of storm cells 1 and 2-steps ahead

y
1. Meteorological Radar Control
Full “lookahead”
Information Extraction
Distributed
radars
Cyril
Rush
Springs
Hazard
detection and
prediction
Lawto
n
large scan sectors:
low quality, but
miss fewer storms
Sample atmosphere when and
where user need is greatest
start angle
 radar configuration, sr
 start and end angles of scan sector
 scan action, Sr
 set of radar configurations
 scan strategy
 algorithm to select scan actions
storm
radius
 Us: quality with which sector j scanned
 Transition function
 Encodes observed environment dynamics
 Obtained from simulator
Difference between true and
observed # of storm cells
 Cost function
o
Pm = Penalty for missing a storm
Pr = Penalty for not rescanning storm
Np Nd
N
p
o
o
C =   |dij - dij| + (Np -Np) Pm + I(tk)Pr
k=1
Difference between observed
and true storm attribute
Storm scanned within Tr
decision epochs?
 Sarsa()
 Linear combination of basis functions to approximate value function
 Tile coding to obtain basis functions
Interscan time
Sarsa() is more likely than 2-step
lookahead to scan a storm within
Tr=4 decision epochs
radar rotation speed
how well sector ri scanned
Us(ri, sr) = Fw(w(sr )/360)]
 cost
 inter-scan time and quality plus penalty for never scanning a phenomenon
Goal: maximize quality while minimizing inter-scan time
Sarsa() scans have slightly lower cost
than 2-step lookahead scan strategy
6. Conclusions and Future Work
4. Simulator
]
Difference in cost
One tiling for each state variable
Performance Metrics

2-Step scan strategy achieves higher quality
than Sarsa(), especially when little noise in
environment (when 1/ is small)
 Up: quality with which storm cell i scanned
i=1 j=1
 inter-scan time
 number of decision epochs before phenomenon first observed or
rescanned
 quality
distance to phenomena
 how well phenomenon p observed
Up(p, Sr) = max [ Fc(c(p, sr ))  [  Fd(d(r,p)) + (1-) Fw(w(sr )/360)]
Difference in scan quality
 Radar actions

% covered
+
(x,y)
end angle
sr Sr
Convergence of Sarsa()
 Number of storm cells

x
2. CASA Radar Network
small scan sectors:
high quality, but
may miss storm
Matrices A, B, Q, R
initialized using prior
knowledge
Markov Decision Process

y
End users: warning,
emergency management,
commercial, research
… and control
Where, what, when, how to sense?
y
 true state
: true = [ x, y, x,
]T
 observed state : obs = [ x, y ] T
 Assume,
truet = A truet-1 + N[0, Q]
obst = B truet + N[0,R]
 Observed state of environment
Distributed, adaptive
computation
Chickasha

x
(x,y)
Challenge: adaptive sensor control in dynamic resource-constrained
meteorological sensor (radar) nets
5. Results
 True state
 Storm cells arrive according to spatiotemporal Poisson process
 Storm cell attributes chosen from
distributions derived from real data
 Observed state
 observed attribute value equals true
value of plus some noise ~ N[0, 2]
Largest positive value of attribute
 = (1-u) Vmax / 
Us(ri,sr) quality
scaling term
 Conclusions



30 km
Lookahead strategies useful when many phenomena
Trade-off: quality and frequency with which phenomena scanned
Consider only scan quality
 simple lookahead strategy sufficient.
 Additionally consider inter-scan time (or optimize over multiple metrics of interest)
 reinforcement learning strategy useful
 Future work
 Semi-MDPs or robotic controllers rather than MDPs
 More radar and meteorological information
Acknowledgments: The authors thank Don Towsley for his input. This work was supported in part by the
National Science Foundation under the Engineering Research Centers Program, award number EEC-0313747.
Any opinions, findings and conclusions or recommendations expressed in this material are those of the
author(s) and do not necessarily reflect those of the National Science Foundation.