Problem with Coluna.Algorithm.ColumnGeneration #449
Comments
|
Can you run your code with using Base.CoreLogging, Logging
global_logger(ConsoleLogger(stderr, LogLevel(-3)))and send me the output please ? |
|
I got hundreds of the following message:
and then
|
|
I made some changes in the log levels because it's a little bit messy. Can you update your Coluna version by doing The goal is to understand why Cplex doesn't provide a dual solution after optimizing the master LP. |
|
Here's my code: And here's the output
|
|
The "LP" is solved to optimality but there is no dual solution available. We already had this problem with CPLEX (#315) but I don't have CPLEX anymore to reproduce. Can you reproduce the bug with another solver e.g. GLPK, Gurobi ? |
|
Yes, same error with GLPK. |
|
Can you give me a way to reproduce the bug ? |
|
Here's my code https://github.com/wdscoelho/slicing_Coluna/ You have just to change the related parameters on function create_COLUNA(instance::Instance) in lines 1063-1071 in the my_functions.jl file. |
|
Hi, It's because you're calling the Regression because of #438. (Needs more tests). Thanks for the report. |
|
Hello Guillaume, Thank you for your answer. If I try to solve the instance found in my code https://github.com/wdscoelho/slicing_Coluna, Coluna is not able to increase the lower bound (objective function = minimize). Could you please let me know what you think it's the problem? Thank you in advance. |
|
Hi,
You should take a look at the branch and bound tree to see what happens.
You have to set the parameter `branchingtreefile` of TreeSearchAlgorithm.
‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
Le mercredi, mars 3, 2021 1:38 PM, Wesley da Silva Coelho <[email protected]> a écrit :
… Hello Guillaume,
Thank you for your answer.
So, maybe that's why I'm experiencing a problem to prove the optimality of a solution when using Coluna.Algorithm.TreeSearchAlgorithm.
If I try to solve the instance found in my code [https://github.com/wdscoelho/slicing_Coluna](url), Coluna is not able to increase the lower bound (objective function = minimize).
As you can see in the attached file [Coluna_output.txt](https://github.com/atoptima/Coluna.jl/files/6075969/Coluna_output.txt), Coluna finds the optimal solution within few seconds (I know its value = 140.01) but doesn't improve the lower bound in any interaction after the root node.
Could you please let me know what you think it's the problem?
Thank you in advance.
Wesley
—
You are receiving this because you commented.
Reply to this email directly, [view it on GitHub](#449 (comment)), or [unsubscribe](https://github.com/notifications/unsubscribe-auth/ABNM5BYS2E6PBCYENUYWJ5TTBYUTHANCNFSM4YCEV3AQ).
|
|
Ah ok! Here's a log output: Coluna_log.txt It's very strange cause I let the solver running for hours (more than 15) and it doesn't prove the optimality. This instance takes less than 5 seconds the be solved by Cplex. |
|
Ok, I think you need valid inequalities and a better branching strategy. For the branching strategy, you have to define your own algorithm because the branching callback is not available ( #403) . |
|
Nothing has been improved by applying the modification to the dual-bound strategy. Could it be a matter of precision when calculating the reduced cost? |
|
I do not believe precision is an issue here. It is probably a matter of using valid inequalities and branching strategies. Solver such as Cplex, Gurobi, … automatically applies a lot of valid inequalities (such as MIR or cover cuts) and has a better strategy than branching only on the most fractional solution. Coluna is a framework not a solver. It provides algorithms to try very quickly column generation on your problem and then devices your own branch-cut-and-price algorithm to scale up and hopefully beats the commercial solver. I saw that your root gap is quite large, you just need to improve your formulation with valid inequalities now. |
|
I see, I have several Valid Inequalities that work very well compared to the basic model and without all cuts and pre-solving routines of Cplex, some of them have improved a lot my LP lower bound and the size of the BB tree. I know that Coluna it's not a solver, but it's very strange that, even with a tiny instance of my problem, the framework is not able to converge in less than 12h (I don't know if it solves the instance at all cause I stopped the test after this time limit). Well, thank you anyway for your answers. I'll see what I can do. Wesley |
|
Oh ok ! What is the dual bound at the root node of Coluna with and without your valid inequalities ? For the second run, you are very unlucky because you don't find the optimal solution like you did in the first run. This is because of the primal heuristic. I suggest you to define your model as : Coluna.jl/test/full_instances_tests.jl Line 56 in f5f5553 |
|
Thank you for your help, but the strong branching didn't help. and without it: Coluna_log.txt Either ways, the solver doesn't improve the lower bound after leaving the root node. |
|
The cut callback is not called, you shoud have a message "Robust cut separation callback adds..." |
|
I'm using only a specific valid inequality in these tests and adding it during the creation of my model. For now, I'm not using the callback. |
Hello,
I'm experiencing a problem with Coluna.Algorithm.ColumnGeneration on my model.
If I use the following constructor :
coluna = optimizer_with_attributes( Coluna.Optimizer, "params" => Coluna.Params( solver = Coluna.Algorithm.ColumnGeneration() ), "default_optimizer" => CPLEX.Optimizer )on the example from https://atoptima.github.io/Coluna.jl/latest/user/start/, it works fine. But with my model, I get the following error:
I'm getting the same error with GLPK.
I'm using:
Julia 1.2.0
JuMP 0.21.1
Coluna 0.3.5
Cplex 0.7.3 and ILO CPLEX 12.10
GLPK 0.13
Could you please help with this issue?
Tell me with need my code. I can upload it somewhere because it is complicated enough to write it here.
Thank you in advance.
Wesley
The text was updated successfully, but these errors were encountered: