Lower bound can be better ...

[gregy4]

Hello,
for attached full model, lower bound of objective variable can be better.

In the model are present requirements for individual elements and also summed requirements for groups of the elements to improve performance. When only group requirements are present in a model, found solutions (in a minute) are better than lower bound of full model running for one hour.

So far my experience was that presence of "right" constraint in a model helped with lower bound, but in this case it seems it is not enough and "solution" is to execute solver especially for lower bound.

Is it recommended approach to run a solver for lower bound or some form of relaxation with limited number of constraints that didn't worked for me exists ?

Thanks,
Jan

Ortools 9.2, 32 search workers used

models.zip

[lperron]

You cannot reason that way.

Model m1 is more constrained that model M2. Both are minimization models.

Thus any solution of m1 is a solution of m2.

The the optimal value of m1 is >= optimal value of m2.

You cannot reason about feasible solutions.

2 strategies:
Over constrain a model (like force symmetrical solutions). In that case, you reduce the number of variables. And any solution you find is a solution to the original model.

Under constrain the model (usually remove extensive constraint). Go to optimal. The objective value is a lower bound of the optimal value of the original model.

[gregy4]

I understand it same way, with solutions from simpler model I can deduce only lower bounds of original model.

Thanks

[lperron]

and I cannot load your models.

[gregy4]

Thanks.
It seems that model.export_to_file function is somehow broken in windows. When I transferred source code to debian valid proto files were produced.

models-fixed.zip