probabilistic solution discovery for network reliability
play

Probabilistic Solution Discovery for Network Reliability - PowerPoint PPT Presentation

Probabilistic Solution Discovery for Network Reliability Optimization Jose E. Ramirez-Marquez Assistant Professor SSE Stevens Institute of Tech. Hoboken, NJ Short Academic Bio Graduated from Universidad Nacional Autonoma de Mexico with


  1. Probabilistic Solution Discovery for Network Reliability Optimization Jose E. Ramirez-Marquez Assistant Professor SSE Stevens Institute of Tech. Hoboken, NJ

  2. Short Academic Bio • Graduated from Universidad Nacional Autonoma de Mexico with Bachelor in Actuarial Science • Obtained Ph. D., M.Sc. in Industrial & Systems Engineering and M.Sc. in Statistics from Rutgers University • Started working at Stevens Institute of technology Fall 2004 • Project 1: New Methods for network Reliability Analysis and Optimization

  3. New Methods for Network Reliability Analysis & Optimization • ARMY Research & Development Center, Picatinny Arsenal, NJ • PM: “we need tools that can be incorporated into our process to measure network operational effectiveness” • ACADEMIC: research on machine learning techniques to analyze network reliability

  4. Quick background on Network reliability analysis • Reliability: Probability a network (collection of components) works for a specified period of time under specified operating conditions • Analysis: • understand how network components interact (build reliability block diagram or network graph) • obtain component reliability data • use technique to obtain network reliability

  5. Tool development- actual research • TOOL- employee with minimal background on reliability must be able to analyze network. • existing methods- computationally expensive & not easily applicable without background. • proposed approach- provide interval for network reliability via a data classification method.

  6. Data Classification Technique • CART- classification and regression technique developed by salford networks • customer profiling, fraud detection, credit card scoring, etc... • METHOD- sifts through data and isolates significant patterns and relationships: • determines a complete tree with minimal misclassification, Assigns terminal nodes to a class outcome

  7. Reliability Via Data Classification Methods (1) • Step 1: Monte Carlo Simulation of Network • generate component states • determine network state • Step 2: Data Analysis and Rule Transformation • set arc states as predictor variables & network state as target variable • rules indicate component interaction related to network behavior

  8. Reliability Via Data Classification Methods (2) • Step 3: Analysis of Generated Rules • Extracted rules are analyzed for validity using a Monte-Carlo simulation • Step 4: Reliability Approximation • Valid rules are used to approximate the reliability of the network via: B l ≤ R ≤ B u

  9. Results • Results obtained showed a narrow bound for the network reliability • With increased arc reliability it is not necessary to obtain all of the minimal cut and path vectors • New method is very practical for medium sized and large networks • Extendable to many network configurations & behavior

  10. Network Reliability Optimization • Problem-network design via strategic allocation of components & redundancy to: • determine types & copies of components assigned to subnetworks to maximize a network function • Selection of components among different choices • Known reliability, cost, weight, etc… • Constraints should be satisfied

  11. Techniques for Reliability Optimization • many approaches to generate very good solutions for most component allocation problems • integer & dynamic programming, simulated annealing, tabu search • genetic algorithms, ant colony,… • very good results tend to depend on algorithms with much parameter tuning • few methods address diverse network structures or behavior

  12. Probabilistic Solution Discovery Optimization (1) • Machine learning algorithm with three main iteration steps: • Step 1: random generation of network configuration via vector of component appearance probabilities (two parameters: # of configurations, probabilities) • appearance probability defines the frequency of a component appearing in the final design • equivalent to step 1 in slide 7

  13. Probabilistic Solution Discovery Optimization (2) • Step 2: (two parameters: Simulation runs, penalty factor) • reliability analysis of network configuration (via method presented in slide 7) • computation of penalty for deviation from constraining target. • Step 3 : (one parameter: sample size) • Update component appearance probabilities using a sample of the configurations generated in steps 1& 2.

  14. Example- cost minimization with reliability constraint 1 Arc x 12 x 13 x 14 x 15 x 23 x 24 x 25 x 34 x 35 x 45 2 5 Cost 3 2 5 4 6 2 2 5 3 4 5 8 4 5 3 6 5 2 2 9 Rel . 0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8 Iteration 1 4 3 appearance u 12u 13u 14u 15u 23u 24u 25u 34u 35u 45u probabilities 1 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 ˆ ( ) ( ) ( ) h h h u = 1 x 12u x 13u x 14u x 15u x 23u x 24u x 25u x 34u x 35u x 45u R x u C x u C x u generated 1 1 1 0 1 1 0 0 1 1 1 0.8984 262 264.1722 2 1 0 0 1 1 1 1 1 1 0 0.8992 282 283.2015 solutions 3 1 1 0 1 0 1 1 1 0 1 0.9173 279 283.9117 4 1 1 1 1 1 0 1 0 0 1 0.8957 281 287.1202 5 1 1 1 1 1 0 1 0 0 1 0.8932 281 290.6788

  15. Example- cost minimization with reliability constraint 1 Arc x 12 x 13 x 14 x 15 x 23 x 24 x 25 x 34 x 35 x 45 2 5 Cost 3 2 5 4 6 2 2 5 3 4 5 8 4 5 3 6 5 2 2 9 Rel . 0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8 0.8 Iteration 10 4 3 appearance probabilities u 12u 13u 14u 15u 23u 24u 25u 34u 35u 45u 1 0 . 5 0 . 5 0 . 5 0 . 5 0 . 5 0 . 5 0 . 5 0 . 5 0 . 5 0 . 5 2 0.65 0.75 0 . 6 0.75 0.65 0.55 0.75 0 . 5 0 . 6 0.65 3 0.75 0 . 7 0.55 0.85 0.75 0.65 0 . 7 0 . 6 0 . 6 0 . 9 4 0.75 0.55 0.35 1 0.75 0 . 7 0.55 0.65 0.65 0.95 5 0.85 0 . 4 0 . 3 1 0.85 0.65 0.25 0 . 6 0.75 1 6 1 0.35 0.15 1 0.95 0.55 0 . 2 1 0 . 8 1 7 1 0 . 5 0 1 1 0.35 0 . 2 1 0.95 1 8 1 0 . 6 0 1 1 0 0 . 4 1 1 1 9 1 0.15 0 1 1 0 0.85 1 1 1 1 0 1 0 0 1 1 0 1 1 1 1

  16. Results for networks with known optima Best Mean Median Problem n l r ij R 0 C v C v [9] Cost b Cost u 1 5 1 0 0.80 0.90 2 5 5 2 255.00 0.0000 0.0000 2 5 1 0 0.90 0.95 2 0 1 3 204.60 0.0309 0.0000 7 2 1 0.90 0.90 7 2 0 6 3 746.00 0.0276 0.0000 7 2 1 0.90 0.90 8 4 5 6 4 852.00 0.0126 0.0185 7 2 1 0.95 0.95 6 3 0 7 5 661.00 0.0561 0.0344 8 2 8 0.90 0.90 2 0 8 4 6 215.70 0.0315 0.0211 8 2 8 0.90 0.90 2 4 7 6 7 259.10 0.0314 0.0183 8 2 8 0.95 0.95 1 7 9 5 8 181.70 0.0284 0.0228 9 3 6 0.90 0.90 2 3 9 6 9 246.30 0.0356 0.0497 9 3 6 0.90 0.90 2 8 1 7 10 294.60 0.0474 0.0340 9 3 6 0.95 0.95 2 0 9 7 11 216.20 0.0569 0.0839 12 a 1 0 4 5 0.90 0.90 1 5 4 7 167.10 0.0791 0.0618 13 a 1 0 4 5 0.90 0.90 1 9 7 7 210.00 0.0448 0.0095 1 0 4 5 0.95 0.95 1 3 6 8 14 145.00 0.0618 0.0802

  17. Results for networks with unknown optima Worst Best Cost Mean Cost Problem n l r ij R 0 Cost GA ANN ANN SDA GA [9] SDA SDA [9] [10] [10] 344.6 1 15 105 0.90 0.95 317 304 268 307.6 281.4 301 956.0 2 20 190 0.95 0.95 926 270 200 281 233.1 262 1651.3 3 25 300 0.95 0.90 1606 402 331 421.8 348.1 367 ˆ ( h ) R x u ˆ ( h ) Problem Worst Cost Configuration R x u Best Cost Configuration 4,5,8,10,13,15,22,26,37,39,46, 4,5,8,15,22,25,26,38,46,47,53, 1 47,53,58,65,69,72,74,77,79,84, 0.9570 58,65,69,72,73,77,79,84,88, 0.9505 88,92 92,101 4,8,29,33,36,42,43,58,66,79, 4,8,11,33,35,42,44,56,66,79,80, 80,92,96,102,106,119,131, 2 92,102,106,123,131,140,146, 0.9525 0.9550 137,140,146,154,160,165,177, 151,154,168,172,188,189 188,189 6,14,32,37,40,51,57,63,70,71, 3,17,32,40,51,54,63,73,97, 87,97,129,137,149,156,174,176, 102,114,119,137,141,160,176, 3 0.9027 0.9003 187,193,206,207,209,215,225, 177,195,207,209,210,225,229, 233,235,242,246,268,274,300 233,242,243,257,269,274,300 Solutions Searche d Fraction Problem Search Spac e SAMPL E out of [9] GA [9] SDA 1.02 × 10 31 1.40 × 10 5 4.50 × 10 3 1 0.032 3 0 0 7 1.02 × 10 57 2.00 × 10 5 6.00 × 10 3 2 0.030 3 0 0 7 2.10 × 10 90 4.00 × 10 5 1.20 × 10 4 3 0.025 400 7

  18. Future work & questions • extend technique to larger reliability problems • all terminal reliability with redundancy • implement algorithm outside reliability area: • orienteering problem • single & group • completely characterize algorithm parameters for any problem size

  19. Reliability Block Diagram & network Graph S 2 S 1 S K 1 1 1 2 2 2 . . . . . . . . . . . . n n n

  20. Step 1 Arc States Probabilities 4 S i1 S i2 P i1 P i2 i 1 8 1 0 9 0.3 0.7 2 0 6 0.3 0.7 3 5 3 0 4 0.3 0.7 7 4 0 5 0.3 0.7 2 9 5 0 8 0.3 0.7 6 0 4 0.3 0.7 6 7 0 2 0.3 0.7 8 0 8 0.3 0.7 Figure 1: Demand = 6 units 9 0 7 0.3 0.7 ϕ u x u x 1 x 2 x 3 x 4 x 5 x 6 x 7 x 8 x 9 Max Flow x 1 9 6 4 5 8 4 0 8 7 12 1 x 2 0 0 0 0 0 6 0 5 2 8 7 x 3 0 0 0 0 4 0 0 4 2 8 7 x 4 9 6 4 0 8 4 2 0 7 6 1 x 5 9 6 4 5 0 4 2 0 7 6 1 x 6 9 6 0 5 0 4 2 8 0 5 0 x 7 9 6 4 5 8 0 2 8 7 10 1 x 8 0 0 0 0 9 0 4 5 8 0 2 x 9 4 2 0 6 1 9 6 0 5 8 7 x 10 0 2 0 2 0 9 6 4 5 8 7

Recommend


More recommend