Scan Strategies for Adaptive Meteorological Radars Victoria Manfredi

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 q All radars always scan 360 Limited “lookahead” q 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 Sample atmosphere when and where user need is greatest start angle large scan sectors: low quality, but miss fewer storms end angle� q radar configuration, sr q start and end angles of scan sector q scan action, Sr q set of radar configurations q scan strategy q algorithm to select scan actions storm radius + (x, y) 2 -Step scan strategy achieves higher quality than Sarsa( ), especially when little noise in environment (when 1/ is small) q Up: quality with which storm cell i scanned q Us: quality with which sector j scanned q Transition function q Encodes observed environment dynamics q Obtained from simulator Difference between true and observed # of storm cells q Cost function o Np Nd N p o o C = |dij - dij| + (Np -Np) Pm + I(tk)Pr i=1 j=1 Pm = Penalty for missing a storm Pr = Penalty for not rescanning storm k=1 Difference between observed and true storm attribute Storm scanned within Tr decision epochs? q Sarsa( ) q Linear combination of basis functions to approximate value function q 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 sr Sr % covered ] radar rotation speed how well sector ri scanned Us(ri, sr) = Fw(w(sr )/360)] q cost q 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 Up(p, Sr) = max [ Fc(c(p, sr )) [ Fd(d(r, p)) + (1 - ) Fw(w(sr )/360)] Difference in cost One tiling for each state variable Performance Metrics q inter-scan time q number of decision epochs before phenomenon first observed or rescanned q quality distance to phenomena q how well phenomenon p observed Difference in scan quality q Radar actions q q Convergence of Sarsa( ) q 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 q true state : true = [ x, y, x, ]T q observed state : obs = [ x, y ] T q Assume, truet = A truet-1 + N[0, Q] obst = B truet + N[0, R] q 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 q True state q Storm cells arrive according to spatiotemporal Poisson process q Storm cell attributes chosen from distributions derived from real data q Observed state q 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 q Conclusions Lookahead strategies useful when many phenomena Trade-off: quality and frequency with which phenomena scanned Consider only scan quality q simple lookahead strategy sufficient. q Additionally consider inter-scan time (or optimize over multiple metrics of interest) q reinforcement learning strategy useful q Future work q Semi-MDPs or robotic controllers rather than MDPs q More radar and meteorological information q q q 30 km 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.