How to Fuse Independent Sensors Fast? (Nuclear Detection) Endre Boros MSIS & RUTCOR, Rutgers University Endre.Boros@rutgers.edu Joint work with Noam Goldberg, Paul B. Kantor and Jonathan Word Statement of Problem The Problem: There are several tests that can be applied (document checks, passive and active sensors of several kinds). Find the “optimal” detection policy based on these tests! [Multiple branching + policy mixing!] 2 Assumptions and Complexity • Stochastically independent non-repeated sensors s k • s sensors, k labels at each: >> 2 policies sensors labels policies 4 2 >> 216 4 5 >> 2625 8 2 >> 2256 • Futility of heuristic searches! 3 Move to a decision support model: Minimize total damage over all available policies MinP C(P) + πK(1- Δ(P)) Δ(P), C(P) - detection rate, and operating cost of policy P π (~ 0), K (~very large) - a priori probability of a “bomb”, and expected cost of false negative MaxP {Δ(P) | C(P) ≤ B } 1 0.9 mixing and domination of policies 0.8 Detection 0.7 concave envelope of best policies 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Cost 4 1 Dynamic Programming: Sensor Fusion Fusing sensor k on top of the given policies optimally is a multi-knapsack problem that can be solved by a modified greedy algorithm: Greedy Algorithm Detection Sensor k Cost Dynamic Programming: common concave envelope We then merge the given policies with the best combination of them with sensor k on top – and generate the common concave envelope of all these policies Greedy Algorithm Detection Sensor k Cost Dynamic Programming: Summary • We build the concave envelope of best possible policies constructible from the given set of s sensors. • We solve s2s sensor fusion problems (for up to s ≤ 20) • Each Sensor Fusion can be solved in O(P*B+Plog(P)) time, where B is the number of channels of the top sensor, and P is the number of pure strategies on the effective frontier. What about approximating the output in each step? Stroud and Saeger sensors (4 sensors, ~100 labels each) Time (sec) Number of Policies Max. Relative Error % 1440 52319 0.005% 3.53 204 0.05% 0.61 33 0.5% 1.33 GHz Intel Atom processor 8 Extremal frontier with 33 undominated policies 9 Detection = 81.527% Cost = 0.1977826 units = $11.867 (< $13+) 2 1 R 3 3 R 1 4 R R I 4 4 R R I 1 1 R I I 4 R I I Publications • • • 1. Goldberg, N., Word, J., Boros, E. & Kantor, P. (2008). Optimal Sequential Inspection Strategies. Annals of Operations Research Vol. 187, 2011. 2. Boros, E., Fedzhora, P.B., Kantor, P.B., Saeger, K., & Stroud, P. Large Scale LP Model for Finding Optimal Container Inspection Strategies. Naval Research Logistics Quarterly, Vol. 56 (5), 404-420, 2009. 3. Kantor, P.B. & Boros E. Deceptive Detection Methods for Optimal Security with Inadequate Budgets: the Screening Power Index. Risk Analysis Vol. 30, 2010.