Optimization of Containers Inspection at Port-of-Entry

Transcription

1 Optimization of Containers Inspection at Port-of-Entry E. A. Elsayed, Christina Young, Yada Zhu Department of Industrial and Systems Engineering Mingyu Li, Minge Xie Department of Statistics Rutgers University

2 Introduction Containers arrive at a port-of-entry for inspection n attributes tested independently Sensors used to classify attributes of a container as safe (d=0) or suspicious (d=1) based on selected threshold values (T i for station i) Overall accept/reject decision based on specified Boolean function of station decisions 2

3 Simple Container Inspection Procedure Containers arriving at port 95~97% Segregation 3~5% to be inspected Station 1: Image Analysis Fail Result Storage area: Containers to Pass be picked up by receivers Station 2: Physical & Visual Check Pass Result Station 3: PRD & RIID Fail CES Central Examination Site Result Fail Pass 3

4 Inspection Problem Description Threshold levels affect the decisions and probabilities of misclassification at each station Sequence of inspection stations affects the expected cost and time of inspection per container Inspection policy specifies sequence (S) and T i values Objective Minimize expected cost of inspection, cost of container misclassifications, and expected inspection time Decision Variables Sensor threshold values and sequence of stations

5 Modeling Approach Assume the unit s true state (x) is 0 or 1 A container attribute value comes from a mixture of two normal distributions, depending on the unit s true state Let r represent the sensor reading returned from a unit, assumed equal to the attribute value 2 r x = 0 Norm( μ0, σ0 ) 2 r x = 1 Norm( μ, σ ) 1 1 For some value T i at station i If measurement r i > threshold level (T i ), then decision for station i, d i =1 5 If r i T i, d i =0

6 Modeling Approach Assume prior distribution of suspicious containers is known: π = P x = 1 = 1 P x = 0 ( ) ( ) Subscript i is used to indicate association with station i Assume parameters of the two normal distributions are known: μ, μ, σ, σ 0i 1i 0i 1i 6

7 Probabilities of Error ( i = 0 x = 1) P( r T x 1) P d = = i Ti μ 1i =Φ σ1 i i ( i = 1 x = 0) P( r T x 0) P d = > = i T i 0i = 1 Φ i μ σ 0i 7

8 Cost of Misclassification The individual results of stations are combined according to defined system Boolean function to reach an overall inspection decision to accept or reject a container This system decision D may or may not agree with the container s true status Probability of false accept, PFA = P ( D = 0 x = 1) Probability of false reject, PFR = P ( D = 1 x = 0) Two sources of container misclassification cost: c FA = cost of false acceptance (undesired cargo) c FR = cost of false rejection (manual unpack) Total cost: = π + ( 1 π) C PFA c PFR c F FA FR 8

9 Cost of Inspection c i = cost of using i th sensor Expected cost of inspection depends on probability of passing (or failing) sensor i : ( ) ( )( ) ( ) pi = P di = 0 = P di = 0 x= 0 1 π + P di = 0 x= 1 T ( ) i μ 0i Ti μ 1i = 1 π Φ + πφ σ0i σ1 i π q = 1 p = P( d = 1 x = 0)( 1 π ) + P( d = 1 x = 1) i i i i T ( ) i μ 0i Ti μ 1i = 1 π 1 Φ + π 1 Φ σ0i σ1 i π 9

10 Optimum Inspection Sequence The sequence in which stations are visited affects the expected cost of inspection A sequence which minimizes this cost is known as an optimum sequence Theorem 1: For a series Boolean decision function, inspecting attributes i, i = 1,2, K, n in sequential order is optimum (minimizes expected inspection cost) if and only if:. c / q c / q Kc / q n n A container is suspicious if decision for any station i is 1. In other words: F( d, d, K, d ) = d d... d ( ) 1 2 n 1 2 n 10

11 Optimum Inspection Sequence Theorem 2: For a parallel Boolean decision function, inspecting attributes i, i = 1,2, K, n in sequential order is optimum (minimizes expected inspection cost) if and only if: c / p c / p Kc / p In other words: Fd (, d, K, d) = d d... d n n ( ) 1 2 n 1 2 n 11

12 Total Expected Cost Example: parallel Boolean decision function Cost of misclassifications: C F = (False acceptance cost) + (False rejection cost) n n Ti μ 1i Ti μ 0i = πcfa 1 1 Φ + ( 1 π) c FR 1 Φ i= 1 σ1 i i= 1 σ0i 12

13 Total Expected Cost Example: parallel Boolean decision function Cost of inspection: n i 1 CI = c1 + qj ci i = 2 j = 1 n i 1 Tj μ 0j Tj μ 1j = c1 + c i ( 1 π) 1 Φ + π 1 Φ i = 2 j = 1 σ 0j σ 1j Total expected cost: c = C + C total I F 13

14 Inspection Time Time for a container to complete inspection at station i denoted t i Optimum sequence with regard to total expected inspection time can be found with similar method to cost Optimum Cost Sequence Similar Methods Optimum Time Sequence 14

15 Inspection Time t i could be a function of threshold T i, approximated from data Higher Threshold More Pass Decisions Fast Inspection 15

16 Objectives Multi-Objective Problem min { c, t } Sequence, Threshold Minimize the total expected cost including inspection cost and misclassification cost, c = C + C Minimize the total expected inspection time Some trade-off between objectives Goal: find solutions located along Pareto front Find min of weighted objective function fw for various 1, w ( ST, ) = w 2 1ctotal + w2ttotal, w2 = 1-w1 weights Each solution is a Pareto optimal point for multiobjective problem Take advantage of optimal sequence theorem to improve efficiency of algorithms total total total I F t total 16

17 Optimum Sequence for Weighted Objective Given fixed weights, optimum sequence theorem can be adapted Change objective from c i to w 1 c i +w 2 t i For parallel Boolean, minimum sequence condition: wc 1 1+ wt 2 1 wc 1 2+ wt 2 2 wc 1 n + wt 2 n... p p p 1 2 Condition for series Boolean: wc 1 1+ wt 2 1 wc 1 2+ wt 2 2 wc 1 n + wt 2 n... q q q 1 2 n n 17

18 Modified Weighted Sum Algorithm Computationally expensive to solve minimization of weighted objective function Easier if sequence is known Apply optimum sequence given thresholds to compute fw1, w( T) = minf 2 w1, w( ST, ) S 2 Solve min fw 1, w( T ) T 2 Avoiding consideration of all potential sequences improves the efficiency of the algorithm 18

19 Solution Methods Three methods developed and implemented to compare optimality of results Grid search- complete enumeration of discrete threshold values Two methods involve repetitions with various weights to solve min fw, w( T ) and generate 1 2 Pareto-optimal solutions Matlab fmincon function Genetic algorithm Output graphed (Pareto frontier) Time vs. cost expectations T 19

20 Numerical Example Three station system using parallel Boolean decision function Cost fixed for each station, c i = 1 Prior π = Distribution parameters µ 0 = [0 0 0] µ 1 = [1 1 1] σ 0 = [ ] σ 1 = [ ] Cost parameters c FA = c FR = 500 Time related a = [ ] b = [ ]

21 Comparison of Three Solution Methods 12 Grid Search Genetic Algorithm Parallel - Grid Search Parallel - Genetic Algorithm fmincon Parallel - fmincon TIV-set1 TIV-set C o s t Time

22 Pareto-Optimal Solutions Output: Graph of time vs. cost trade-off curve Optimal sequence of sensors Threshold level for each sensor Examples of solutions T 1 T 2 T 3 Sequence Cost Time

23 Conclusions Port-of-entry container inspection problem was formulated to determine optimum threshold levels of sensors by minimizing total expected cost and time Estimate threshold-dependent probabilities of false accept and false reject to calculate expected cost of false classification Sequence of inspection affects expected cost and time of inspection but not probabilities of error Compare three approaches to multi-objective problem GA provides dependable set of Pareto solutions 23

24 Acknowledgment and Website This research is conducted with partial support from ONR grant numbers N , N and NSF grant number NSFSES to Rutgers University For further information and related papers, please visit the DIMACS website 24