1 / 98

The Traveling Salesman Problem in Theory & Practice

The Traveling Salesman Problem in Theory & Practice. Lecture 12: Optimal Tour Lengths for Random Euclidean Instances 15 April 2014 David S. Johnson dstiflerj@gmail.com http:// davidsjohnson.net Seeley Mudd 523, Tuesdays and Fridays. Outline. The Beardwood-Halton-Hammersley Theorem

len
Download Presentation

The Traveling Salesman Problem in Theory & Practice

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. The Traveling Salesman Problem in Theory & Practice Lecture 12: Optimal Tour Lengths for Random Euclidean Instances 15 April 2014 David S. Johnson dstiflerj@gmail.com http://davidsjohnson.net Seeley Mudd 523, Tuesdays and Fridays

  2. Outline • The Beardwood-Halton-Hammersley Theorem • Empirical Approximations to Asymptopia • Properties of Optimal Tours • Student Presentation by Junde Huang on the missing details of the proof that the Gilmore-Gomory algorithm correctly runs in polynomial time.

  3. Random Euclidean Instances(The Classic TSP Test Case) N = 10

  4. Random Euclidean Instances(The Classic TSP Test Case) N = 10

  5. Random Euclidean Instances(The Classic TSP Test Case) N = 100

  6. Random Euclidean Instances(The Classic TSP Test Case) N = 1000

  7. Random Euclidean Instances(The Classic TSP Test Case) N = 10000

  8. N = 100,000

  9. Random Euclidean Instances • Performance of heuristics and optimization algorithms on these instances are reasonably well-correlated with that for real-world geometric instances. • We can generate many samples, and very large instances. • Instances have convenient statistical properties.

  10. Optimal Tour Lengths (Difference from Mean):One Million 100-City Instances -1e+07 -5e+06 0 +5e+06 Optimal Tour Lengths Appear to Be Normally Distributed

  11. Optimal Tour Lengths (Difference from Mean):Ten Thousand 1000-City Instances -1e+07 -5e+06 0 +5e+06 With a standard deviation that appears to be independent of N

  12. Well-Defined Tour Length Asymptotics • Expected optimal tour length for an N-city instance approaches CN for some constant C as N  . [Beardwood, Halton, and Hammersley, 1959] Key Open Question:What is the Value of C?

  13. The Early History • 1959: BHH estimated C  .75, based on hand solutions for a 202-city and a 400-city instance. • 1977: Stein estimates C  .765, based on extensive simulations on 100-city instances. • Methodological Problems: • Not enough data • Probably not true optima for the data there is • Misjudges asymptopia

  14. Stein: C = .765 BHH: C = .75 Figure from [Johnson, McGeoch, Rothberg, 1996]

  15. What is the dependence on N ? • Expected distance to nearest neighbor proportional to 1/N, times n cities yields (N) • O(N) cities close to the boundary are missing some neighbors, for an added contribution proportional to (N)(1/N), or (1) • A constant number of cities are close to two boundaries (at the corners of the square), which may add an additional (1/N ) • This yields target function OPT/N = C + /N + /N

  16. Asymptotic Upper Bound Estimates (Heuristic-Based Results Fitted to OPT/N = C + /N) • 1989:Ong & Huang estimate C ≤ .74, based on runs of 3-Opt. • 1994:Fiechter estimates C ≤ .73, based on runs of “parallel tabu search” • 1994: Lee & Choi estimate C ≤ .721, based on runs of “multicanonical annealing” • Still inaccurate, but converging? • Needed: A new idea.

  17. New Idea (1995): Suppress the variance added by the “Boundary Effect”byusingToroidal Instances • Join left boundary of the unit square to the right boundary, top to the bottom. 

  18. Toroidal Unit Ball

  19. Toroidal Distance Computations

  20. Toroidal Instance Advantages • No boundary effects. • Same asymptotic constant for E[OPT/N] as for planar instances [Jaillet, 1992] (although it is still only asymptotic). • Lower empirical variance for fixed N.

  21. Toroidal Approaches 1996:Percus & Martin estimate C  .7120 ± .0002. 1996:Johnson, McGeoch, and Rothberg estimate C  .7124 ± .0002. 2004: Jacobsen, Read, and Saleur estimate C  .7119. Each coped with the difficulty of computing optima in a different way.

  22. Percus-Martin(Go Small) • Toroidal Instances with N ≤ 100: • 250,000 samples, N = 12,13,14,15,16,17 (“Optimal” = best of 10 Lin-Kernighan runs) • 10,000 samples with N = 30 (“Optimal” = best of 5 runs of 10-step-Chained-LK) • 6,000 samples with N = 100 (“Optimal” = best of 20 runs of 10-step-Chained-LK) • Fit to OPT/N = (C + a/N + b/N2)/(1+1/(8N)) (Normalization by the expected distance to the kth nearest neighbor)

  23. Jacobsen-Read-Saleur(Go Narrow) • Cities go uniformly on a 1 x 100,000 cylinder – that is, only join the top and bottom of the unit square and stretch the width by a factor of 100,000. • For W = 1,2,3,4,5,6, set N = 100,000W and generate 10 sample instances. • Optimize by using dynamic programming, where only those cities within distance k of the frontier (~kw cities) can have degree 0 or 1, k = 4,5,6,7,8. • Estimate true optimal for fixed W as k  . • Estimate unit square constant as W  . • With N ≥ 100,000, assume no need for asymptotics in N

  24. Johnson-McGeoch-Rothberg(Go Held-Karp) • Observe that • the Held-Karp (subtour) bound also has an asymptotic constant, i.e., HK/n  CHK [Goemans, 1995] , and is easier to compute than the optimal. • (OPT-HK)/N has a substantially lower variance than either OPT or HK. • Estimate • CHK based on instances from N=100 to 316,228, using heuristics and Concorde-based error estimates • (C- CHK)based on instances with N = 100, 316, 1000, using Concorde for N ≤ 316 and Iterated Lin-Kernighan plus Concorde-based error estimates for N = 1000.

  25. Concorde • Not only computes optimal tours – also can compute precise Held-Karp bounds. • Only a pre-release version was available in 1995 when Johnson-McGeoch-Rothberg was written. • Machines are much faster now, cycles are much cheaper, and Concorde is much better.

  26. Running times (in seconds) for 10,000 Concorde runs on random 1000-city planar Euclidean instances (2.66 Ghz Intel Xeon processor in dual-processor PC, purchased late 2002. Range: 7.1 seconds to 38.3 hours

  27. Running times (in seconds) for 1,000,000 Concorde runs on random 1000-city “Toroidal” Euclidean instances Range: 2.6 seconds to 6 hours

  28. The New Data • Points chosen uniformly from a 10Mx10M grid • Solver: • Latest (2003) version of Concordewith a few bug fixes and adaptations for new metrics • Primary Random Number Generator: • RngStreampackage of Pierre L’Ecuyer. See • “AN OBJECT-ORIENTED RANDOM-NUMBER PACKAGE WITH MANY LONG STREAMS AND SUBSTREAMS,” Pierre L'ecuyer, Richard Simard, E. Jack Chen, W. David Kelton, Operations Research 50:6 (2002), 1073-1075

  29. Toroidal Instances

  30. Euclidean Instances

  31. Standard Deviations N = 100 N = 1,000

  32. 99% Confidence Intervals for OPT/Nfor Euclidean and Toroidal Instances

  33. 99% Confidence Intervals for (OPT-HK)/Nfor Euclidean and Toroidal Instances

  34. Gnuplot Least Squares fit for the Percus-Martin values of N --OPT/N = (C + a/N + b/N2)/(1+1/(8N)) C = .712234 ± .00017 versus originally claimed C = .7120 ± .0002

  35. Least Squares fit for all data from [12,100] --OPT/N = (C + a/N + b/N2) C = .712333 ± .00006 versus C = .712234 ± .00017

  36. Least Squares fit for toroidal data from [30,2000] -- OPT/N = (C + a/N + b/N2) C = .712401 ± .000005 versus C = .712333 ± .00006

  37. What is the right function? Power Series in 1/N – (Suggested by Percus-Martin) Justification: Expected distance to the kth nearest neighbor is provably such a power series.

  38. What is the right function? OPT/sqrt(N) = Power Series in 1/sqrt(N)) Justification: This is what we saw in the planar Euclidean case (although it was caused by boundaries).

  39. What is the right function? OPT/sqrt(N) = = (1/sqrt(N) · (Power Series in 1/N)

  40. What is the right function?

  41. Effect of Data Range on Estimate[30,2000], [60,2000], [100,2000], [200,2000], [100,1000] 95% Confidence Intervals C + a/n.5 C + a/n.5 + b/n C + a/n.5 + b/n + c/n1.5 C + a/n + b/n2 + c/n3 C + a/n.5 + b/n.1.5 C + a/n.5 + b/n.1.5 + c/n2.5

  42. The Winners? C + a/n + b/n2 + c/n3 C = .71240 ± .000005 C = .71240 ± .00002

  43. Does the HK-based approach agree? Question

  44. CHK = .707980 ± .000003 95% confidence interval derived using C + a/N + b/N2 functional form

  45. C-CHK = .004419 ± .000002 95% confidence interval derived using C + a/N + b/N2 functional form

  46. HK-Based Estimate C-CHK = .004419 ± .000002 + CHK = .707980 ± .000003 C = .712399 ± .000005 Versus (Conservative) Opt-Based Estimate C = .712400 ± .000020 Combined Estimate? C = .71240 ± .00001

  47. “Explaining” The Expected Optimal Tour Length • The fraction of optimal tour edges that go to kth nearest neighbor seems to be going to a constant ak for each k.

  48. Fraction of Optimal Tour Edges 1st Neighbor (44.6%) 2nd Neighbor (26.0%) 3rd Neighbor (13.6%) 4th Neighbor (7.1%) 5th Neighbor (3.9%) 6th Neighbor (2.1%) 7th Neighbor (1.2%) 8th Neighbor (0.66%) 9th Neighbor (0.37%) 10th Neighbor (0.21%) 11th Neighbor (0.12%) 19th Neighbor (.00014%) 20th Neighbor (.00008%)

  49. “Explaining” The Expected Optimal Tour Length • The fraction of optimal tour edges that go to kth nearest neighbor seems to be going to a constant ak for each k. • If dk is the average distance from a city to its kth nearest neighbor, then dksqrt(N) also seems to be going to a constant for each k.

  50. (√N)·(Average distance to kth Nearest Neighbor) 20th Neighbor 17th Neighbor 14th Neighbor 11th Neighbor 8th Neighbor 6th Neighbor 5th Neighbor 4th Neighbor 3rd Neighbor 2nd Neighbor 1st Neighbor

More Related