Sequential Adaptive Sensor Management Alfred O Hero III

Sequential Adaptive Sensor Management Alfred O. Hero III† Dept. of Electrical Engineering and Computer Science, The University of Michigan 4 th Year ARO MURI Review June, 2006

Sequential Sensor Resource Allocation Predict performance for each possible sensing action Progress (since June 05) • Theory of information gain (IG) scheduling – Result: IG bounds risk (Kreucher CDC 05). – Implication: IG is a universal surrogate • Classification reduction for RL Time update information state under each available action model Compute expected improvement for each sensing action Deploy action with best predicted performance improvement – Result: Generalization error bounds (Blatt, Thesis-06). – Implication: Minimum # samples and model/measurement complexity • Adaptive energy allocation and waveform selection – Result: LARS reduction of optimal adaptive waveform selection policy (Rangarajan: ICASSP 06) – Implication: linear-complexity solution to exponential-complexity problem • IRIS: sensor management for STW Measurement update info state – Result: IRIS adaptive illuminator placement strategy w/ confidence maps – Implication: Information-directed path planning for STW (Marble: 06) 4 th Year ARO MURI Review June, 2006

Progress Highlighted Today 1. Adaptive energy allocation and waveform selection [Rangarajan&Raich&Hero] 2. Iterative Redeployment of Illumination and Sensing (IRIS) for STW [Marble&Raich&Hero] 4 th Year ARO MURI Review June, 2006

Progress 1: Adaptive Waveforms • Sequentially illuminate a medium and measure backscatter using an array of sensors. • Applications to mine detection, ultrasonic medical imaging, foliage penetrating radar, nondestructive testing, communications, and active audio. • GOAL: Optimally design a sequence of waveforms using an array of transducers – To image a scatter medium (Estimation). – To track targets (Tracking) – To discover strong scatterers (Detection). 4 th Year ARO MURI Review June, 2006

Progress 1 a: Energy allocation for D&E • Let past observations be 1. Adaptive Energy Management 2. Active Waveform Design under average energy constraint • Energy allocation question: Given transmission of certain waveforms , how much can we gain through optimal energy allocation between various time steps (Rangarajan: 2005)? 4 th Year ARO problem, " ICASSP-2005. R. Rangarajan, R. Raich and A. O. Hero, "Optimal experimental design for an inverse scattering. MURI Review June, 2006

Gains more than 5 d. B!!!! • RESULT : We prove through optimal energy designs, we can achieve at least 5 d. B gain (compared to one-step strategy) for estimation problems (imaging). • How much can we gain for target detection? ? th Year ARO MURI Review June, 2006 4

Results for target detection • Two-step energy design procedure – ~ 2 d. B gain or 20% decrease in average error for same SNR. • How much improvement can be achieved asymptotically with time? (Work in progress) 4 th Year ARO MURI Review June, 2006

Progress 1 b: Active waveform selection • • M possible waveforms Can only send p out M, p < M+1 Design criterion: Optimal solution: subset selection, is intractable 4 th Year ARO MURI Review June, 2006

Simplification via rule ensembles • We approximate the decision statistic at receiver (detector, estimator, classifier) by a weighted sum of non -linear functions (rule ensembles (Friedman: 2005)) of subsets of q measurements at time t • Special case (GAM) for estimating state variable s • Reduced GAM waveform selection criterion Friedman, J. H. and Popescu, B. E. "Predictive Learning via Rule Ensembles. " ARO MURI Review June, 2006 4 th Year (Feb. 2005)

Solution via convex relaxation • Convex relaxation (Tibshirani: 1994) of waveform design criterion (Rangarajan: 2006) Tibshirani, R. "Regression selection and shrinkage via the lasso" Technical Report (June. 1994). R. Raghuram, R. Raich and A. O. Hero, "Single-stage waveform selection for adaptive resource constrained state estimation, " 4 th Year IEEE Intl. Conf. on Acoustics, Speech, and Signal Processing, Toulouse France, 2006. ARO MURI Review June, 2006

Summary comparisons • HMM diffusion with bi-level variance • Diffusion measured in Gaussian additive noise with one of possible subsets of n=5 waveforms 4 th Year ARO MURI Review June, 2006

Numerical results Future Directions • Sensor network localization/tracking problem. • Combine optimal energy allocation with waveform selection. (Work in progress) 4 th Year ARO MURI Review June, 2006

Progress 2: Iterative Redeployment of Illumination and Sensing (IRIS) Elements of IRIS strategy l Initial illumination with physical antenna array l Antenna array is deployed at an initial location and illuminates the region of interest. l Sparse reconstruction image reconstruction (Ting: 2006) is performed l Form Confidence Map of Image l Confidence map (Raich: 2005) is computed using initial image and side information l Select a region of low confidence from confidence map l Simulate external energy/resolution field induced by virtual transmitter l Place virtual transmitter in low confidence region and apply FEM, Mo. M, PO to estimate electric field distribution outside the building l Compute induced energy or gradient field (wrt perturbation of virtual transmitter location) l Re-illuminate with physical antenna array at maximum of simulated field M. Ting, R. Raich and A. O. Hero, "Sparse image reconstruction using a sparse prior, " ICIP 2006 R. Raich and A. O. Hero, "Sparse image reconstruction for partially unknown blur functions, " ICIP 2006 4 th Year ARO MURI Review June, 2006

IRIS Illustration: Sensor Illumination Table Initial Sensor Position/Configuration Transmitter Sink Point Scatterer Chair Weapons Cache Wall 4 th Year ARO MURI Review June, 2006

IRIS Illustration: Confidence Map Iterative image reconstruction (Fessler&Hero: TIP 95) Sparsity constrained deconvolution (Nowak&etal: TSP 03) Image confidence maps (Raich&Hero: ICASSP 06) Sink Weapons Cache Wall Low confidence region 4 th Year ARO MURI Review June, 2006

IRIS Illustration: Virtual backillumination Possible sensor locations Sink Weapons Cache Wall 4 th Year ARO MURI Review June, 2006

IRIS Illustration: Predict Energy/Resolution MAP Sink Weapons Cache Wall 4 th Year ARO MURI Review June, 2006

Sparse image reconstruction and confidence mapping • MAP-EM Formulation • Separates deconvolution from denoising: • EM-MAP iterations for image x and confidence map • • Properties • Iterates monotonically increase likelihood • Deconvolution (E) only involves adjoint of forward operator • Fast implementation with wavenumber migration approx for H Ting, M, Hero, A. O. , “Sparse Image Reconstruction Using a Sparse Prior, ” ICIP 2006. 4 th Year ARO MURI Review June, 2006

Sparse Reconstruction Example • 4 th Year ARO MURI Review June, 2006

IRIS illustration for STW Initial illuminator location Accessible Region Simulated Scene Chair Weapons Cache Table Inaccessible Region Empty Space interior Sink 1 m aperture Accessible Region External Wall – Permittivity: 10 Thickness: 0. 2 m Length: 10 m 4 th Year ARO MURI Review June, 2006

IRIS for STW – Iteration 1: Sparse reconstruction Standard Wavenumber Migration 1 m Aperture Sparse iterative Reconstruction (10 iterations) (Marble&Raich&Hero: 06) 4 th 1 m Aperture Review June, 2006 Year ARO MURI

IRIS for STW – Iteration 1: Confidence Mapping Sparse Prior w = 0. 25 a = 0. 5 l Confidence Map shows pixels that have high confidence of being “empty space”. l Quantitative map: Ambiguous pixels 4 th Year ARO MURI Review June, 2006

IRIS for STW – Iteration 1: Insert virtual transmitter and simulate field Energy-optimal Energy Mapping 1 m sensor placement Virtual Transmitter IG-optimal 1 m sensor placement KL Mapping Virtual Transmitter 4 th Year ARO MURI Review June, 2006

Spectral Information Gain KL Divergence – Information Gain Reference Field Observation Location -Electric Field From Transmitter k. Horizontal Perturbation Field 1 2 3 Virtual Transmitters Div Map= Div(E 1, E 3)+ Div(E 1, E 2) KL Divergence is a measure of Discrimination Error Probability Vertical Perturbation Field 4 th Year ARO MURI Review June, 2006

IRIS for STW – Iteration 2: Insert virtual transmitter and simulate field Energy-optimal 1 m sensor placement Virtual Transmitter Info-optimal 1 m sensor placement Virtual Transmitter 4 th Year ARO MURI Review June, 2006

IRIS for STW – Iteration 3: Insert virtual transmitter and simulate field Energy Mapping Virtual Transmitter IG-optimal 1 m sensor placement Cross Range [m] Energy-optimal 1 m sensor placement 4 th Year ARO MURI Review June, 2006

IRIS for STW – Comparisons to fixed aperture 3 1 m 2 1 10 m Aperture 1 m 4 1 m 1 m Aperture 4 th Year ARO MURI Review June, 2006

Personnel on A. Hero’s sub-Project (20052006) • Raviv Raich, post-doctoral researcher – BS Tel Aviv University – Ph. D Georgia Tech, May 2004 • Neal Patwari, post-doctoral researcher – BS Virginia tech – Ph. D, Univ of Michigan, Sept. 2005 • Doron Blatt, 4 th year doctoral student – BS Univ. Tel Aviv – Ph. D Univ of Michigan, May 2006 • Raghuram Rangarajan, 5 th year doctoral student – BS IIT Madras – Dept. Fellowship/MURI GSRA • Jay Marble, 5 th year doctoral student – BS UIUC – MURI GSRA – Presently employed at NVRL 4 th Year ARO MURI Review June, 2006

Pubs Since June 2005 – Theses of students funded on MURI • "Performance Evaluation and Optimization for Inference Systems: Model Uncertainty, Distributed Implementation, and Active Sensing, " Ph. D Thesis, The University of Michigan, May 2006. – Journal • “Adaptive Multi-modality Sensor Scheduling for Detection and Tracking of Smart Targets”, C. Kreucher, D. Blatt, A. Hero, and K. Kastella, Digital Signal Processing, vol. 15, no. 4, July 2005. • "Multitarget Tracking using the Joint Multitarget Probability Density, " C. Kreucher, K. Kastella, and A. Hero, IEEE Transactions on Aerospace and Electronic Systems, 39(4): 1396 -1414, October 2005 (GD Medal winner 2005). • "Convergent incremental optimization transfer algorithms: application to tomography", S. Ahn, J. A. Fessler, D. Blatt, and A. Hero, IEEE Trans. on Medical Imaging, vol. 25, no. 3, pp. 283 -296, March 2006 4 th Year ARO MURI Review June, 2006

Pubs Since June 2005 – Conference • "Sequential Design of Experiments for a Rayleigh Inverse Scattering Problem, " R. Rangarajan, R. Raich, and A. O. Hero, Proc. Of IEEE Workshop on Statistical Signal Processing (SSP), Bordeaux, July 2005. • "APOCS: a convergent source localization algorithm for sensor networks, " D. Blatt and A. O. Hero, IEEE Workshop on Statistical Signal Processing (SSP), Bordeaux, July 2006 • "Incremental optimization transfer algorithms: application to transmission tomography", S. Ahn, J. A. Fessler, D. Blatt, and A. Hero, IEEE Conf on Medical Imaging, Oct. 2005. • "A Comparison of Task Driven and Information Driven Sensor Management for Target Tracking, " C. Kreucher, A. Hero, and K. Kastella, 44 th IEEE Conference on Decision and Control (CDC) Special Session on Information Theoretic Methods for Target Tracking, December 12 -15 (Invited) • "Single-stage waveform selection for adaptive resource constrained state estimation, " R. Raghuram, R. Raich and A. O. Hero, IEEE Intl. Conf. on Acoustics, Speech, and Signal Processing, Toulouse France, June 2006. • "Optimal sensor scheduling via classification reduction of policy search (CROPS), " D. Blatt and A. O. Hero, 2006 Workshop on POMDP's, Classification and Regression (Intl Conf on Automated Planning and Scheduling (ICAPS)), Cumbria UK, June 2006. (Invited) 4 th Year ARO MURI Review June, 2006

Synergistic Activities and Awards since June 2005 • • • Sensip Nov 2005, A. Hero plenary speaker Member of ARO MURI (John Sidles PI) awarded in 2005 for MRFM sensing and image reconstruction Member of AFOSR MURI (Randy Moses PI) awarded in 2006 for multi-platform radar sensing Member of ISP team (Harry Schmit PI) General Dynamics, Inc – – – • K. Kastella: collaboration with A. Hero in sensor management, July 2002 C. Kreucher: former doctoral student of A. Hero, continued collaboration M. Moreland Melbourne: collaborator in area of sensor management Ben Shapo: MS student collaborator in area of sensor management Mike Davis: MS student collaborator in area of satellite MIMO ARL – NRC ARLTAB: A Hero is member of NAS oversight/review committee – ARLTAB SEDD: A. Hero participated in yearly review • • • Night Vision Lab: Jay Marble spent two weeks of Aug 2005 with Steve Bishop EIC of Foundations and Applications of Sensor Management (Springer - 2006) Contributor, IEEE Proceeedings Special Issue on Large Scale Complex Systems, Editor S. Haykin. 4 th Year ARO MURI Review June, 2006

Synergistic Activities (ctd) • In May 2005 UM Student Jay Marble was at Georgia Tech (working with Waymond Scott) • In Aug 2005 Jay Marble was at Night Vision Lab (working with Steve Bishop) : • • Indirectly support the Autonomous Mine Detection System (AMDS) Identify new data sets for algorithm validation: “Check Test 1” (April 2005) Apply multi-stage reinforcement learning algorithms to Army problems. Further develop demonstration software for illustrating algorithm performance. • A. Hero visited AFRL Rome (B. Bonneau) in Nov. and gave invited presentation at Sensip on sensor management and at the “Old Crows Conference” at AFRL. • Collaboration with Eric Michielsson on IRIS started in fall 2005 – led to several proposals to DARPA, ARO. 4 th Year ARO MURI Review June, 2006

Transitions • Transition of SM methods to control of sensor swarms (GD) – resulted in GD sensor net demo. • Marble visited NVRL for 1 month in summer 2005 to demo UM mine detection software • June 2006 Marble is now full-time employee at Night Vision Research Laboratory 4 th Year ARO MURI Review June, 2006