SDDP.jl
SDDP.jl copied to clipboard
odow
Getting out of local optimal SDDiP
Hi Oscar,
now with SDDiP, we actually have a heuristic to solve the model. We don't proved convergence anymore. I can see that on several occasions, the model is stuck on a local optimal. I have been using BanditDuality to help the model to get out of those local optimal. Using different cuts helps get out of the local optimal. However, StrengthenedConicDuality can make much slower progress.
Any ideas on how to get out of the local optimal? Is there any way to use different duality_handlers depending on the bound stalling?
PS: I also get an issue with BanditDuality when trying to start back the train, it gets stuck with just one type of duality. Example:
1883S -3.526655e+04 2.575658e+03 5.779006e+04 750405 1
1885S -6.197976e+04 2.575544e+03 5.828998e+04 751967 1
1887S -3.414856e+04 2.568755e+03 5.844276e+04 752881 1
1889S -3.993974e+04 2.568373e+03 5.876231e+04 753719 1
1890 -5.145811e+04 2.568373e+03 5.880455e+04 754762 1
1891S -1.481099e+04 2.568067e+03 5.883862e+04 754869 1
1893S -2.585710e+04 2.568097e+03 5.899596e+04 755583 1
1895S -2.168459e+04 2.567502e+03 5.911727e+04 755921 1
1897S -2.168718e+04 2.562634e+03 5.917999e+04 756171 1
1899S -2.104896e+04 2.559852e+03 5.930590e+04 756765 1
1900 -6.025239e+04 2.559852e+03 5.935892e+04 758124 1
1901S -2.302431e+04 2.559217e+03 5.946921e+04 758411 1
1903S -1.432304e+04 2.558576e+03 5.951825e+04 758661 1
1905S -3.676079e+04 2.557187e+03 5.980848e+04 759567 1
1909S -4.680980e+04 2.555810e+03 6.017820e+04 760987 1
-------------------------------------------------------------------
status : time_limit
total time (s) : 6.017820e+04
total solves : 760987
best bound : 2.555810e+03
simulation ci : -1.533718e+05 ± 1.864513e+04
numeric issues : 0
-------------------------------------------------------------------
Starting back training as I don't think we achieved convergence:
-------------------------------------------------------------------
iteration simulation bound time (s) solves pid
-------------------------------------------------------------------
1 -2.247751e+04 2.555810e+03 1.210600e+01 761414 1
2 -3.003325e+04 2.555810e+03 3.303300e+01 762025 1
3 -1.146033e+05 2.555810e+03 1.077620e+02 764276 1
4 -2.030632e+04 2.555810e+03 1.140230e+02 764455 1
7 -6.857285e+04 2.555810e+03 1.704500e+02 766032 1
9 -4.955000e+04 2.555810e+03 2.123530e+02 767198 1
10 -3.825554e+04 2.555810e+03 2.452390e+02 768069 1
13 -4.923447e+04 2.555810e+03 2.830820e+02 769098 1
16 -2.347561e+04 2.555810e+03 3.140410e+02 769911 1
19 -1.949708e+04 2.555810e+03 3.547390e+02 770996 1
20 -4.641142e+04 2.555810e+03 3.917980e+02 771967 1
22 -2.252014e+04 2.555810e+03 4.338440e+02 773057 1
25 -1.398725e+04 2.555810e+03 4.714310e+02 773994 1
26 -4.205176e+04 2.555810e+03 5.093740e+02 774985 1
29 -1.976783e+04 2.555810e+03 5.408240e+02 775806 1
33 -3.568327e+04 2.555810e+03 5.897890e+02 777094 1
35 -1.578067e+04 2.555810e+03 6.226700e+02 777944 1
36 -7.880938e+04 2.555810e+03 6.910770e+02 779639 1
39 -3.428028e+04 2.555810e+03 7.372160e+02 780640 1
41 -4.941172e+04 2.555810e+03 7.725960e+02 781538 1
46 -3.661219e+04 2.555810e+03 8.222400e+02 782745 1
49 -2.477594e+04 2.555810e+03 8.750570e+02 783990 1
53 -4.000302e+04 2.555810e+03 9.189290e+02 785098 1
56 -1.498087e+04 2.555810e+03 9.498160e+02 785839 1
59 -3.442730e+04 2.555810e+03 9.898650e+02 786836 1
61 -4.222634e+04 2.555810e+03 1.045653e+03 788134 1
62 -5.144041e+04 2.555810e+03 1.095805e+03 789329 1
65 -6.342666e+04 2.555810e+03 1.180221e+03 791394 1
69 -1.599701e+04 2.555810e+03 1.212788e+03 792194 1
70 -4.505639e+04 2.555810e+03 1.249560e+03 793073 1
73 -2.727797e+04 2.555810e+03 1.289537e+03 793978 1
76 -4.387685e+04 2.555810e+03 1.331148e+03 794955 1
80 -1.189576e+05 2.555810e+03 1.475348e+03 798267 1
Then I'm stuck in local optimal and the StrengthenedConicDuality is never used again.
