Comparison of minimizers for individual test problems

As in the summarized results (Comparing Minimizers), the best possible results with Mantid minimizers are given a top score “1” (lowest error, or lowest run time).

The following color codes are used to distinguish different levels of performance, note the irregular (increasing) difference between level boundaries):

  • best or within 10% (ranking < 1.1)
  • within 33% over the best (1.1 < ranking < 1.33)
  • within 75% over the best (1.33 < ranking < 1.75)
  • within 200% over the best (1.75 < ranking < 3)
  • over 200% of best (ranking > 3)

The integer numbers 1 and 2 next to the problem names denote the starting point (all NIST problems are usually tested for two different starting points). By clicking on the names of the test problems (leftmost column) you can see the original definition of the problem.

Comparison in terms of accuracy

For details on the ranking approach and the color codes see above and the summarized results. When interpreting these results note that the median shown in the summarized results is calculated excluding undefined values. Undefined accuracy values are obtained when the minimizers fail to produce any result and are shown in the detailed results as “nan”. For the more difficult groups of problems the proportion of undefined values is higher for some minimizers. This can bias the statistics shown in the summary table in favor of the minimizers that fail more often, as some of the the hard problems are excluded from the median instead of contributing to a higher median (lower ranking).

Alternatively, see the summary and detailed results when using weighted least squares as cost function.

Accuracy for individual NIST problems, “lower” difficulty

  BFGS Conjugate gradient (Fletcher-Reeves imp.) Conjugate gradient (Polak-Ribiere imp.) Damping Levenberg-Marquardt Levenberg-MarquardtMD Simplex SteepestDescent
Misra1a 1 156.5 156.7 156.7 1 1 1 1 156.7
Misra1a 2 1.829 2.253 2.253 1 1 1 1 2.253
Chwirut2 1 1.067 1.013 1 1 1 1 1.068 1.066
Chwirut2 2 1.049 1.002 1 1 1 1 1.197 1.004
Chwirut1 1 1.097 1.001 1 1 1 1 1.112 1.097
Chwirut1 2 1.017 1.007 1 1 1 1 1.109 1.017
Lanczos3 1 7.683e+06 5.433e+04 4.963e+04 1 1 1 973.7 7.797e+04
Lanczos3 2 2.652e+05 6.206e+04 6.264e+04 1 1 1 1.807e+04 1.285e+05
Gauss1 1 5.6 1.745 1.044 1 1 1 1 5.6
Gauss1 2 1.655 9.182 9.182 1 1 1 1 9.168
Gauss2 1 7.272 7.341 7.341 1 1 1 1.022 7.272
Gauss2 2 2.081 3.754 3.754 1 1 1 1 3.393
DanWood 1 72.41 1 1 1 1 1 1.011 55.52
DanWood 2 1.244 1 1 1 1 1 1.031 1.97
Misra1b 1 96.76 96.97 96.97 1 1 1 1 96.97
Misra1b 2 8.521 16.43 16.43 1 1 1 1 16.43

Accuracy for individual NIST problems, “average” difficulty

  BFGS Conjugate gradient (Fletcher-Reeves imp.) Conjugate gradient (Polak-Ribiere imp.) Damping Levenberg-Marquardt Levenberg-MarquardtMD Simplex SteepestDescent
Kirby2 1 423 25.07 25.37 1 1 1 956.3 510.9
Kirby2 2 2.554 1.717 1.68 1 1 1 1.396 121
Hahn1 1 4.508e+04 45.51 217 69.31 1 1 200.6 4962
Hahn1 2 4000 35.48 211.6 1 1 1 157.2 4896
MGH17 1 1.419e+04 1.42e+04 1.42e+04 nan 1 754.5 1.421e+04 1.413e+04
MGH17 2 2473 1.414 1.422 1 1 1 4179 1762
Lanczos1 1 1.561e+18 1.107e+16 1.01e+16 151.7 1 21.29 2.182e+15 1.583e+16
Lanczos1 2 1.559e+17 3.652e+16 3.687e+16 1 1 28.59 1.064e+16 7.518e+16
Lanczos2 1 5.553e+09 3.939e+07 3.591e+07 1 1 1 7.765e+06 5.631e+07
Lanczos2 2 1.917e+08 4.491e+07 4.533e+07 1 1 1 1.308e+07 9.22e+07
Gauss3 1 15.19 1.815 1.568 1 1 1 1.036 15.19
Gauss3 2 3.147 2.44 2.763 7.912 1 1 1.25 8.944
Misra1c 1 115 115.8 115.8 1 1 1 1 115.8
Misra1c 2 6.595 6.702 6.702 1 1 1 1 6.702
Misra1d 1 12.67 19.62 19.62 1 1 1 1 19.62
Misra1d 2 1.03 1.939 1.939 1 1 1 1 1.939
ENSO 1 1.259 1.025 1.024 1.524 1 1.524 1.014 1.202
ENSO 2 1.033 1.018 1.029 1.087 1 1.087 1.028 1.052

Accuracy for individual NIST problems, “higher” difficulty

  BFGS Conjugate gradient (Fletcher-Reeves imp.) Conjugate gradient (Polak-Ribiere imp.) Damping Levenberg-Marquardt Levenberg-MarquardtMD Simplex SteepestDescent
MGH09 1 7.705 25.02 25.02 1.817e+04 3.342 1 20.62 24.19
MGH09 2 1.62 1.621 1.62 256.6 1 1 1.025 1.604
Thurber 1 228.3 175.3 173.2 2.693 1 1 176.5 177.2
Thurber 2 422.3 20.02 40.38 1 1 1 4.43 368.9
BoxBOD 1 8.366 8.366 8.366 nan 8.366 1 8.366 1.515
BoxBOD 2 33.46 8.366 8.366 1 1 1 8.366 6.177
Rat42 1 13.33 13.27 13.27 4.759e+06 1 1 16.48 193.6
Rat42 2 1.331 1.213 1.302 1 1 1 1.429 1.449
MGH10 1 6.367e+05 6.362e+05 6.344e+05 2.984e+47 1 5.306e+05 6.422e+05 6.37e+05
MGH10 2 4.08e+04 664.2 664 1.394e+51 1 41.54 4.368e+04 4.132e+04
Eckerle4 1 478.1 478.1 478.1 366 1 1 478.1 478.1
Eckerle4 2 31.01 1.02 1.026 1 1 1 1 1.003
Rat43 1 39.21 1.02 1.02 nan 1.136 1 51.2 21.8
Rat43 2 1.07 1.001 1 1 1 1 1 1.001
Bennett5 1 1294 648.7 1.028 1 1 1 1646 1307
Bennett5 2 690.3 3.347 1.165 1 1 1.065 1572 698.8

Comparison in terms of run time

For details on the ranking approach and the color codes see above and the summarized results. Note that this comparison is approximate and was performed on a particular platform (Ubuntu) and machine.

Alternatively, see the summary and detailed results when using weighted least squares as cost function.

Run time for individual NIST problems, “lower” difficulty

  BFGS Conjugate gradient (Fletcher-Reeves imp.) Conjugate gradient (Polak-Ribiere imp.) Damping Levenberg-Marquardt Levenberg-MarquardtMD Simplex SteepestDescent
Misra1a 1 1.042 1.03 1.046 1 1.116 1.011 1.768 1.95
Misra1a 2 1.008 1.036 1.107 1.005 1.011 1 1.421 2.223
Chwirut2 1 1.165 19.75 19.91 1 1.066 1.075 1.287 10.02
Chwirut2 2 1.213 22.14 22.13 1.014 1.049 1 1.359 10.83
Chwirut1 1 1.691 50.26 13.62 1 1.073 1.187 1.354 21.98
Chwirut1 2 1.763 8.55 26.32 1 1.145 1.038 1.624 24.16
Lanczos3 1 1.304 3.878 4.103 1 2.617 5.288 5.068 12.83
Lanczos3 2 1.379 4.23 2.776 1 1.053 1.028 1.728 12.84
Gauss1 1 1.427 7.891 45.86 1 1.181 1.047 7.806 56.56
Gauss1 2 2.962 1 1.004 1.308 1.493 1.372 9.91 73.93
Gauss2 1 1.991 1 1.005 1.306 1.487 1.376 12.36 73.53
Gauss2 2 2.526 1 1.003 1.312 1.496 1.39 11.63 74.15
DanWood 1 1.079 1.601 1.527 1 1.006 1.034 1.158 5.121
DanWood 2 1.029 1.223 1.208 1 1.036 1.03 1.102 5.405
Misra1b 1 1.047 1.096 1.091 1.022 1.163 1 1.62 2.519
Misra1b 2 1.142 1.108 1.255 1.073 1 1.024 1.495 2.948

Run time for individual NIST problems, “average” difficulty

  BFGS Conjugate gradient (Fletcher-Reeves imp.) Conjugate gradient (Polak-Ribiere imp.) Damping Levenberg-Marquardt Levenberg-MarquardtMD Simplex SteepestDescent
Kirby2 1 1.463 37.51 37.56 1 1.058 1.027 2.042 21.9
Kirby2 2 1.355 38.81 39.63 1.025 1.104 1 2.745 23.81
Hahn1 1 1.884 41.9 21.86 2.089 1.006 1 1.53 23
Hahn1 2 1.298 52.93 22.61 1 1.128 1.08 5.577 28.7
MGH17 1 1 1.184 1.169 1.448 11.24 5.252 1.553 9.74
MGH17 2 1.185 14.82 15.27 1 1.283 1.454 1.421 12.09
Lanczos1 1 1.268 3.7 3.995 1 2.5 5.376 4.295 12.5
Lanczos1 2 1.392 4.178 2.727 1 1.064 1.018 1.742 12.84
Lanczos2 1 1.317 3.918 4.009 1 2.495 5.654 4.272 12.6
Lanczos2 2 1.401 4.34 2.706 1 1.056 1.043 1.759 13.1
Gauss3 1 1.336 55.55 49.51 1 1.136 1.056 8.739 51.18
Gauss3 2 1.661 45.89 11.86 45.69 1.116 1 7.838 46.25
Misra1c 1 1.121 1.09 1.024 1.059 1 1.008 1.535 3.129
Misra1c 2 1.095 1.098 1.085 1 1.019 1.019 1.476 2.974
Misra1d 1 1.062 1.066 1.082 1 1.008 1.018 1.399 2.411
Misra1d 2 1.075 1.096 1.084 1 1.008 1.094 1.374 2.042
ENSO 1 2.804 185.3 173.5 1 137.4 1.046 20.45 129.8
ENSO 2 5.554 183.9 112.1 1.022 138 1 16.54 118.1

Run time for individual NIST problems, “higher” difficulty

  BFGS Conjugate gradient (Fletcher-Reeves imp.) Conjugate gradient (Polak-Ribiere imp.) Damping Levenberg-Marquardt Levenberg-MarquardtMD Simplex SteepestDescent
MGH09 1 1.149 1 1.038 1.658 1.486 2.59 1.214 1.537
MGH09 2 1 1.058 1.041 1.81 1.057 1.212 1.728 3.476
Thurber 1 1 6.162 24.11 1.518 1.145 1.183 1.255 11.6
Thurber 2 1.098 15.26 25.03 1.517 1 1.065 4.53 13.11
BoxBOD 1 1.041 1.097 1.114 1.5 1 1.333 1.297 4.698
BoxBOD 2 1.026 1.076 1.095 1.002 1 1.062 3.725 4.483
Rat42 1 1.18 1.509 1.407 1 1.039 1.169 1.167 5.347
Rat42 2 1 7.957 1.876 1.005 1.018 1.012 1.161 5.295
MGH10 1 1 2.99 2.939 1.231 6.398 5.026 1.048 3.473
MGH10 2 1 2.035 2.433 1.925 5.985 5.84 1.122 6.092
Eckerle4 1 1 1.088 1.088 1.258 1.171 5.581 1.127 1.39
Eckerle4 2 1.014 1.644 1.699 1 1.049 1.033 1.197 3.224
Rat43 1 1.115 15.44 3.682 1.431 1 1.14 1.314 7.362
Rat43 2 1.002 1.548 4.422 1.019 1 1.051 1.183 8.667
Bennett5 1 1.173 40.41 4.457 1.017 1 15.08 1.181 13.89
Bennett5 2 1.233 43.28 4.892 1 17.72 14.75 1.246 14.83