Originally posted by: utb

Hi,

I am trying to solve a stochastic programing model by using the built-in benders in CPLEX v12.7.1. I have coded the extensive model and I am using BendersUser algorithm for solving where I define the annotations, however I am getting CPLEX error 3019: Failure to solve MIP subproblem(If I use BendersFull, I don't get the error since it uses different annotations). In the explanation of the error, it says "This failure can be due to a limit (for example, an iteration limit) or due to numeric trouble.". I have tried many different values for almost all mip parameters in limits, tolerances etc. but I failed to overcome the error. I also tried to make all my big M parameters as small as possible. Here is the disp>prob>stats of my problem from interactive optimizer:

Objective sense      : Minimize

Variables            :  432474  [Nneg: 432045,  Binary: 429]

Objective nonzeros   :      12

Linear constraints   :  447195  [Less: 391568,  Greater: 105,  Equal: 55522]

  Nonzeros           : 2160011

  RHS nonzeros       :    3083

Variables            : Min LB: 0.000000         Max UB: 1.000000       

Objective nonzeros   : Min   : 1.000000         Max   : 1.000000       

Linear constraints   :

  Nonzeros           : Min   : 2.000000e-09     Max   : 222.0616       

  RHS nonzeros       : Min   : 1.000000         Max   : 78.00000

I have also attached the log files with and without initial solution given by addMIPstart method to the model. What can be the problem?

Thanks,

utb

You have a spread of 11 orders of magnitude in your constraint coefficients. I would not be shocked if that caused numerical stability problems. Any chance that you could rescale some variables to boost those 1e-9 size coefficients without making something else worse?

Originally posted by: utb

Thanks a lot Paul, I have changed the 1e-9 coefficients into 1e-4 to and the instance I have sent above now does not give any error.

Originally posted by: utb

Hi Paul,

Changing the coefficients helped for some of the instances, but for different annotations, I started getting the same error again. I have attached the log file of the new instance which gives the same CPLEX error 3019 and the stats of the problem is below. (Still BendersFull does not lead to error)

Objective sense      : Minimize

Variables            :  174138  [Nneg: 173709,  Binary: 429]

Objective nonzeros   :      12

Linear constraints   :  179298  [Less: 156848,  Greater: 117,  Equal: 22333]

  Nonzeros           :  847120

  RHS nonzeros       :     686

Variables            : Min LB: 0.000000         Max UB: 1.000000       

Objective nonzeros   : Min   : 1.000000         Max   : 1.000000       

Linear constraints   :

  Nonzeros           : Min   : 5.000000e-05     Max   : 219.9408       

  RHS nonzeros       : Min   : 1.000000         Max   : 78.00000

Is there a way to explicitly see the reason of the error? Also, can there be a relation between this error and model size?

Thanks,

utb

Does anything show up in the clone log files?

Does setting CPXPARAM_MIP_Strategy_CallbackReducedLP to 1 help? Alternatively, if you leave it at 0, does setting CPX_PARAM_PRELINEAR to 0 help?

Originally posted by: utb

Hi Paul,

I have coded my model in c++ and cplex doesn't allow me to change CPXPARAM_MIP_Strategy_CallbackReducedLP into 1, it takes 0 as default. For CPX_PARAM_PRELINEAR, I have changed it into 0 but still getting the error. Also nothing written in the clone logs.

Do you have any other suggestions?

Thank you,

utb

Originally posted by: EdKlotz

Try setting the scaling parameter to 1 and turning on the numerical emphasis parameter.   While changing those small coefficients from 1e-9 to 1e-5, you still have a moderately large spread of coefficients.   These settings tend to help in such cases.