1. MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation 1 Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Overview MURI Annual Review Meeting Randy Moses November 3, 2008

2. 2 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Research Goal Develop an integrated systems theory that jointly treats information fusion, control, and adaptation for automatic target exploitation (ATE). Multiple, dynamic sensors Multiple sensing modes Resource-constrained environments

3. 3 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Research Framework ATE Objectives Sensor Resources ATR/ATE Inferences and Confidences Optimal, Robust Information Fusion Graphical Models Performance Measures Information State Propagation Contextual Models Learning and Adaptation Dynamic Sensor Resource Management Dynamic sensor allocation Efficient sensor management algorithms Multi-level planning Performance uncertainty Adaptive Front-End Signal Processing Decision-directed Imaging and Reconstruction Modeling and Feature Extraction Statistical Shape Estimation Features and Uncertainties Priors and Learned Statistics Measurement Constraints Value of Information

4. 4 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Research Framework Inference-rooted metrics provide the common language across the system  likelihoods  posterior probabilities  information measures

5. 5 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation MURI Program Statement of Work 1. Innovative Front-End Processing. We will develop robust, adaptive algorithms for decision directed front- end waveform processing. We will provide analytical performance assessments and performance bounds for these approaches. Specifically, we will develop 1a) decision-directed imaging and reconstruction techniques. 1b) physics-based parametric feature representations of targets and confusers, and corresponding feature extraction algorithms and uncertainty metrics. 1c) statistical shape theory that integrates with regularized inversion techniques. 2. Optimal, Robust Information Fusion in Uncertain Environments. We will develop fundamental theory and associated algorithms for inference and performance prediction in ATE systems that employ distributed, multi-modal sensing. We will develop hierarchical structures that integrate both signal-level features and feature uncertainties with high-level context. Specifically, we will develop 2a) structured inference methods for graphical models that manage complexity in high-dimensional ATE inference problems. 2b) information-theoretic performance measures for ATE inference assessment and for multimodal data fusion. 2c) methods for approximating and propagating statistical state information under communications and computational resource constraints. 2d) machine learning methods for on-the-fly adaptation of inference structures to evolving target and context scenarios. 2e) contextual models for inference in complex ATE clutter environments. 3. Dynamic Sensor Resource Management for ATE. We will develop a fundamental theory and associated algorithms for on-line, adaptive, active sensor management to support ATE using distributed, multi-modal sensing. Specifically, we will develop 3a) sensor management algorithms that incorporate ATE inference performance metrics. 3b) theory, models and associated algorithms for dynamic sensor management of ATE operations using multiple multi-modal, passive, and active sensing platforms. 3c) joint trajectory optimization and sensor control algorithms that optimize ATE inference metrics. 3d) sensor management algorithms that are robust against unknown or inaccurate models of sensor performance. Remove this slide – but it serves as a useful reminder for planning.

6. 6 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Information Fusion: Key Research Questions Optimal, Robust Information Fusion Graphical Models Performance Measures Information State Propagation Contextual Models Learning and Adaptation Inference on Graphical Models: Structures and algorithms for fusion, tracking, identification Scalable algorithms Learning and adaptation Contextual Information What are the ‘right’ performance measures and bounds for FE and Sensor Mgmt? State propagation in graphical models. How to effectively direct front-end signal processing?

7. 7 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Signal Processing: Key Research Questions Adaptive Front-End Signal Processing (R. Moses, lead) Decision-directed Imaging and Reconstruction Modeling and Feature Extraction Statistical Shape Estimation Physics-based feature representations Decision-directed imaging and reconstruction Feature uncertainty characterization Statistical shape estimation Adaptation and Learning Feature sets and feature uncertainties that permit fusion across modalities Problem formulations that admit context, priors and directed queries

8. 8 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Sensor Management: Key Research Questions Dynamic Sensor Resource Management (D. Castañón, lead) Dynamic sensor allocation Efficient sensor management algorithms Multi-level planning Performance uncertainty Active sensor control that incorporate ATE performance metrics Multi-modal, heterogeneous platforms Scalable real-time algorithms for theater-level missions Robust to inaccurate performance models Integrate ATE performance based on graphical models Manage evolution of “information state” in support of ATE missions

9. 9 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation MURI Payoff Goal: Develop an integrated theory for ATE systems that combines information fusion, platform control, signal processing, and adaptation. Payoff: Systematic design tools for end-to- end design of multi-modal, multi- platform ATE systems. Active platform control to meet ATE objectives. System-level ATE performance assessment methods. Adaptive, dynamic ATE systems. Research Outcomes: An integrated theoretical framework for dynamic information exploitation systems. Theoretical foundations for adaptivity and learning in complex inference systems. New algorithms and performance metrics for coupled signal processing, fusion, and platform control.

10. 10 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation MURI Team UNIVERSITY TEAM: Ohio State University (lead) Randy Moses (PI) Lee Potter Emre Ertin Massachusetts Institute of Technology Alan Willsky John Fisher Mujdat Çetin (also Sabanici U.) Boston University David Castañón Clem Karl University of Michigan Al Hero Florida State University Anuj Srivastava AFOSR: David Luginbuhl  Doug Cochran AFRL POC: Greg Arnold

11. 11 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Year 1 Meeting Feedback Strongly positive on team expertise and interactions. Strongly supportive of research plan Maintain emphasis on fundamental research. Maintain research continuity Interactions among the MURI PIs Interactions with government labs and industry RLM: Update these!

12. 12 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Year 2 Advances Regularized Tomography for Sparse reconstruction Sparse apertures – monostatic and multistatic Sparse ‘objects’ (targets or scenes) Anisotropy characterization 3D Reconstruction for wide angle and circular SAR Decision-directed reconstruction Lots of cross-pollination, especially OSU, BU, MIT Shape Statistics for Curves and Surfaces Shape Analysis Shape distribution; not just point estimates Bayesian Classification from Shapes

13. 13 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Year 2 Advances II Sensor Management: Multistatic radar sensor management using scattering physics Radar resource management with guaranteed uncertainty metrics. Real-time SM algorithms and performance bounds Scalable, flexible inference Lagrangian relaxation and multiresolution methods for tractable inference in graphical models New graphical model-based algorithms for multi-target, multi- sensor tracking Learning Graphical Model structures directly for discrimination GM-based distributed PCA and hyperspectral image discrimination

14. 14 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation MURI Students 10 graduate students and 1 postdoc fully supported by the MURI. 14 graduate students and 1 postdoc working on the MURI team with outside support (e.g. fellowships) or partial funding. 3 PhD and 3 MS degrees awarded in Year 1 RLM Update for Year 2

15. 15 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation ATE MURI Web Page People Publications On-line research collaboration space Code Data resources

16. 16 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Intra-team Interactions MURI research meeting in Boston Dec07 30 participants, 25+ posters, Sensor Management Book Collaboration Foundations and Applications of Sensor Management; Springer 2007; Editors: A. Hero, D. Castañón, D. Cochran, K. Kastella Student Committees D. Castañón (BU) on Jason Williams’ (MIT) dissertation committee J. Fisher (MIT) on Shantanu Joshi’s (FSU) dissertation committee Numerous Research Interaction Visits A. Hero → MIT sabbatical Au06 Fisher → FSU; Joshi Ph.D. committee Moses → 3 visits to MIT Summer 2006 Fisher → UMich July07 Çetin → OSU Aug07 (video-conference)

17. 17 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Publications and Student Degrees 129 Publications 2 book chapters 45 Journal publications 82 Conference proceedings papers Student Graduate Degrees 6 Ph.D. graduates Williams (MIT), Rangarajan (UMich), Joshi (FSU), Malioutov (MIT), Johnson (MIT), Bashan (UMich) 6 M.S. graduates Austin (OSU), Som (OSU), Varhney (MIT), Chen (MIT), Choi (MIT), Chandrasekaran (MIT)

18. 18 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation PHD E. Bashan, Efficient Resource Allocation Schemes for Search. PhD thesis, UMich, 2008. D. Malioutov, Approximate inference in Gaussian graphical models. PhD thesis, MIT, Cambridge, MA, 2008. R. Ranagarajan, Resource constrained adaptive sensing, PhD Thesis, UMich, 2007. J. Johnson, Convex Relaxation Methods for Graphical Models: Lagrangian and Maximum-Entropy Approaches. PhD thesis, MIT, Cambridge, MA, 2008. J. L. Williams, Information Theoretic Sensor Management. PhD thesis, MIT, 2007. S. Joshi, Inference in Shape Spaces with Applications to Computer Vision, Ph.D. thesis, FSU, 2007 MS Z. Chen, Efficient multi-target tracking using graphical models," Master's thesis, MIT, Cambridge, MA, 2008. M. Choi, “Multiscale gaussian graphical models and algorithms for large-scale inference,“ Master's thesis, MIT, Cambridge, MA, 2007. N. Ramakrishnan, “Joint enhancement of multichannel synthetic aperture radar data," Master's thesis, The Ohio State University, Mar. 2008. V. Chandrasekaran, “Modeling and estimation in gaussian graphical models: Maximum-entropy relaxation and walk-sum analysis," Master's thesis, MIT, Cambridge, MA, 2007. K. Varshney, “Joint image formation and anisotropy characterization in wide-angle SAR,” Master’s thesis, MIT, Cambridge, MA, May 2006. C. D. Austin, “Interferometric synthetic aperture radar height estimation with multiple scattering centers in a resolution cell,” Master’s thesis, The Ohio State University, 2006. Student Theses Probably delete this slide

19. 19 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Research Leadership Journal Special Issue Statistical Shape Modeling, IEEE PAMI, 2008 (Srivastava) Themed Workshops Workshop on Challenges and Opportunities in Image Understanding, Jan 2007 (Srivastava) Workshop on Signal and Information Processing, July 2008 (Hero) Conference Special Sessions Sparse Reconstruction methods in Radar SPIE Defense Symposium, April 2008 (Moses) Signal Processing for Inference IEEE 5 th DSP Workshop, January 2009 (Moses) Signal Processing and Learning for Sensor Signal Exploitation 2008 Asilomar Conference on Signal, Systems, and Computers, Oct 2008 (Ertin) Fisher: Statistical Signal Processing Workshop 2007??

20. 20 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Interactions and Transitions I Presentations at 15 conferences Scientific Advisory Boards Castañón: Air Force SAB Hero: ARL TAB Willsky: DARPA POSSE; Willsky: Chief Scientific Consultant to BAE-AIT Hero: National Research Council Research Formulation Workshops Srivastava organized ARO Workshop on Challenges and Opportunities in Image Understanding Moses, Fisher, Hero, Arnold presented Hero organized ARO Workshop on Signal and Information Processing Moses, Fisher, Castañón presented Ertin: AFRL ATR Modeling Workshop

21. 21 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Interactions and Transitions II Transitions: Sparse aperture research -> AFRL GREP program Srivastava statistical shape analysis to Northrup-Grumman via Innovation Alliance Award A. Hero directed Mike Davis (GD-AIS) research on MIMO radar networks MURI LADAR model identification and radar scheduling provided to Lincoln Laboratory Summer internships Julie Jackson: AFRL Dayton, Summers 06 and 07 Kerry Dungan: AFRL Dayton, Summer 07 Ahmed Fasih, SET Corp. Dayton, Summers 06 and 07 Christian Austin: SET Corp. Dayton, Summer 06 Kyle Herrity: FAAC, Inc Ann Arbor MI, Summer 07

22. 22 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Today’s Presentations Algorithms and Bounds for Networked Sensor Resource Management (Castañón, Fisher, Hero) Optimal, Robust Information Fusion in Uncertain Environments (Willsky) Sparse Reconstruction and Feature Extraction (Çetin, Ertin, Karl, Moses, and Potter) Graphical Models for Resource- Constrained Hypothesis Testing and Multi-Modal Data Fusion (Castañón, Fisher, Hero) Tools for Analyzing Shapes of Curves and Surfaces (Fisher, Srivastava)

23. 23 MURI: Integrated Fusion, Performance Prediction, and Sensor Management for Automatic Target Exploitation Agenda 8:30 - 9:00Intro and Overview (Moses) 9:00 – 9:40Optimal, Robust Information Fusion in Uncertain Environments (Willsky) 9:40 - 10:20Sparse Reconstruction and Feature Extraction (Potter) 10:20 - 10:40Break 10:40 - 11:10Bayesian Fusion of Features in Images for Shape Classification (Srivastava) 11:10 - 11:50Graphical Models for Resource-Constrained Hypothesis Testing and Multi- Modal Data Fusion (Fisher) 11:50 - 12:50Lunch - Blackwell Ballroom 12:50 - 1:30Algorithms and Bounds for Networked Sensor Resource Management (Castañón ) 1:30 - 2:00Summary (Moses) 2:00 - 2:30Government caucus ~2:30Feedback and Discussion