detection, tracking and sensor management in …• dynamic strategies for avoiding capture •...
TRANSCRIPT
Valueof
Information
1st year review. UCLA 2012
VOI
Year 2 review. ARL, Sept 9 2013
Kickoff
ARO MURI on Value-centered Information Theory for Adaptive Learning, Inference, Tracking, and Exploitation
Valueof
Information
1st year review. UCLA 2012
ARO MURI on Value-centered Information Theory for Adaptive Learning, Inference, Tracking, and Exploitation
VOI
Year 2 review. ARL, Sept 9 2013
Detection, Tracking and Sensor Management in Adversarial Setting
Co-PI Emre ErtinThe Ohio State University
Valueof
Information
1st year review. UCLA 2012
VOI
Year 2 review. ARL, Sept 9 2013
Kickoff
1st year review. UCLA 2012
Valueof
InformationVOI
Year 2 review. ARL, Sept 9 2013
Valueof
Information
1st year review. UCLA 2012
VOI
Year 2 review. ARL, Sept 9 2013
Kickoff
1st year review. UCLA 2012
Valueof
InformationVOI
Year 2 review. ARL, Sept 9 2013
Three major axes of progress
• VoI Measures for Adversarial Scenarios– Sensor Placement for Detection in Adversarial Setting– Robust Tracking and Measurement Design for Gaussian Mixture Models
• Resource Exploitation– Sequential Testing with Quantized Belief States – Learning Quantizers for Sequential Testing
• Information Driven Learning– Learning statistical manifolds for geometry of target signatures– Estimation of Fisher Information for statistical manifolds
Valueof
Information
1st year review. UCLA 2012
Kickoff
VOI
Kickoff
1st year review. UCLA 2012
Valueof
Information
Kickoff
VOI
Design for Adversarial Opponent
• Sensor systems do not compete against nature but against adversaries with opposing interests. Adversarial action extends to:• Countermeasures for avoiding detection • Dynamic strategies for avoiding capture• Coordinated team actions
• We focus on sensor control and sensor system design to maximize VoI assessed in an adversarial setting
• Our approach combines Design for adversary (target) and Design for uncertainty (Sensor, Noise) • Optimal sensor control strategies to maximize information rate• Adversary models that attempts to minimize information rate to the observer• Omniscient adversary models are employed to provide performance guarantees
Valueof
Information
1st year review. UCLA 2012
Kickoff
VOI
Kickoff
1st year review. UCLA 2012
Valueof
Information
Kickoff
VOISensor Selection in Adversarial Setting
Sensor 1
Sensor 2
Sensor K
Observer
(q1m, p1)
(q2m, p2)
(qKm, pK)
H0H1(m)
• Surveillance game between target and observer with opposing objectives.
• The observer is choosing an open loop randomized strategy to choose sensor observations to maximize probability of detection.
• The adversary (target) is using an open loop randomized control strategy over the available evading actions to minimize the probability of being detected.
• Min-max composite detection problem • Focus on asymptotic detection
performance under fixed false alarm rate
Valueof
Information
1st year review. UCLA 2012
Kickoff
VOI
Kickoff
1st year review. UCLA 2012
Valueof
Information
Kickoff
VOISearching for Optimal Strategies
We show that the asymptotic performance in this game is given by convex combination of KL divergence for each sensor strategies
Min-max strategies for this surveillance game satisfying the saddle point property
We derive an interior method for computing optimal survaillance strategies, using log barrier terms to express inequality constraints
J (r, s⇤) J (r⇤, s⇤) J (r⇤, s)
J (r, s) =X
k
rkD�pk|qk(s)
�
Valueof
Information
1st year review. UCLA 2012
Kickoff
VOI
Kickoff
1st year review. UCLA 2012
Valueof
Information
Kickoff
VOIOptimizing Sensor Placement
• We propose a sensor placement strategy for maximizing the observer advantage in the subsequent surveillance game
• Envelope theorem provides an efficient way to compute the gradient
dJ (r⇤(x), s⇤(x), x)
dxk=
@L(r, s,�, x)@xk
����r⇤,s⇤,�⇤
= �r
⇤k
X
j
p
kj
q
kj (s
⇤)
X
m
@q
kj (m)
@xks
⇤m
x
⇤= argmax
x
J (r
⇤(x), s
⇤(x), x)
Valueof
Information
1st year review. UCLA 2012
Kickoff
VOI
Kickoff
1st year review. UCLA 2012
Valueof
Information
Kickoff
VOIOptimizing Sensor Placement
Valueof
Information
1st year review. UCLA 2012
Kickoff
VOI
Kickoff
1st year review. UCLA 2012
Valueof
Information
Kickoff
VOITracking with Initial State Uncertainty
• Adversarial strategies often concentrate on finite support. We consider robust filtering for Gauss-Markov model with linear observations with initial state uncertainty modeled using GMM
• We assume prior distribution of mixture proportions is unknown and consider estimators for target position and show that for the class of linear estimators the least favorable mixing prior and minimax robust filter form a saddle point and therefore can be computed using a convex program in LFM prior.
Valueof
Information
1st year review. UCLA 2012
Kickoff
VOI
Kickoff
1st year review. UCLA 2012
Valueof
Information
Kickoff
VOIMeasurement Matrix Design with GMM Priors
• Next, we consider robust observation matrix design with GMM prior under Frobenius Norm Constraints to optimize minmax LMMSE
• Using the saddle point property of the least favorable prior and min-max filter
• Inner optimization problem is a convex program, which provides a closed form gradient for the observation matrix iterates
Z0 = H0X0 +N0
Valueof
Information
1st year review. UCLA 2012
VOI
Year 2 review. ARL, Sept 9 2013
Kickoff
1st year review. UCLA 2012
Valueof
InformationVOI
Year 2 review. ARL, Sept 9 2013
Sequential Testing with Quantized Belief States (Poster)
Teng & Ertin GlobalSIP2013
W w
†
p (w)
†
p(y t |w)
†
{y t}
Finite state machine design for learning control strategies for sequential decision problems. The expected cost for any decision strategy in the form of a finite state machine can be calculated using markov chain properties
For a decision rule with L bits of memory, likelihood space is divided in 2L+2 regions. The optimal design is equivalent to Bayes updates with quantized belief states satisfying equilibrium conditions.
The optimal boundaries in the likelihood space is shown to be the critical points of the lower envelope for expeted cost of immediate decision and continue sampling
We alternatively calculate the intersections and the centroids to learn the optimal quantizer.
Valueof
Information
1st year review. UCLA 2012
VOI
Year 2 review. ARL, Sept 9 2013
Kickoff
1st year review. UCLA 2012
Valueof
InformationVOI
Year 2 review. ARL, Sept 9 2013
Learning Target Manifolds and Performance Prediction (Poster)
Ertin GlobalSIP2013
We study classification and pose estimation from training data. Target signatures in high dimensional space form articulation manifolds. We focus on statistical characterization of these manifolds.
Linear dimensionality reduction reveals target geometry and results in tractable noise statistics
Recover GeometryAngles, Distances and Curvature
For labeled data, parametric fits to the manifold surface can be used to construct classifiers and to perform calculus on the manifold and performance prediction.
Valueof
Information
1st year review. UCLA 2012
VOI
Year 2 review. ARL, Sept 9 2013
Kickoff
1st year review. UCLA 2012
Valueof
InformationVOI
Year 2 review. ARL, Sept 9 2013
Ongoing and future focus areas and collaborations
VoI Measures in Adversarial Setting:
• Nonlinear min-max filtering for GMM Models
• Information Theoretic robust measurement matrix design for multimodal sensors
Collaborations:
• Kronecker Factorization of Interference Covariance for decentralized tracking (OSU/UM)
• Non-parametric methods for estimation of Fisher Information for Target Manifolds (OSU/UM)
• Confirmation Bias and Divergent Beliefs Human Subjects: Role of Quantized Beliefs (OSU/UCSD).
• Validation experiments on the SDR testbed (UM/OSU/ASU/MIT).
Valueof
Information
1st year review. UCLA 2012
VOI
Year 2 review. ARL, Sept 9 2013
Kickoff
1st year review. UCLA 2012
Valueof
InformationVOI
Year 2 review. ARL, Sept 9 2013
Publications in year 2• N. Sugavanam and E. Ertin, "Sensor selection and placement in adversarial
environments," Proceedings of the IEEE Global Conference on Signal and Information Processing, December 2013
• D. Teng and E. Ertin, "Optimal quantization of likelihood for low complexity sequential testing," Proceedings of the IEEE Global Conference on Signal and Information Processing, December 2013
• E. Ertin, "Manifold learning methods for wide-angle SAR ATR," Proceedings of the 2013 International Conference on Radar
• J. Gao and E. Ertin, "Contactless sensing of physiological signals using wideband RF probes," Proceedings of the 46th Asilomar Conference on Signals, Systems, and Computers, November 2013
• M. R. Riedl, L. C. Potter, and E. Ertin, "Augmenting synthetic aperture radar with space time adaptive processing," Proceedings of SPIE, vol. 8746, Baltimore, May 2013
• C. Rossler, E. Ertin, and R. L. Moses, "Waveform diversity and optimal change detection," Proceedings of the 46th Asilomar Conference on Signals, Sysyems, and Computers, November 2012