Originally posted by: Tulkas

Hello,

I generate some local cuts thanks to C++ UserCutCallback mechanism, exploiting the decisions CPLEX has already taken throughout the branch and cut algorithm. In other words, at any given node, I use only the variables so that getLB(var) == getUB(var) +/- EPSILON to perform some deductions which are specific to my MILP model.

Imagine the constraints:
(1) x1 boolean
(2) x2 boolean
(3) x1 + x2 <= 1
and the current state:
LB(x1) = LB(x2) = 0
UB(x1) = UB(x2) = 1

Thanks to other variables, I deduce x1 = 1 and so execute addLocal(x1 == 1).

MY PROBLEM:

At the next call of my callback, I have the following state:
getValue(x1) = 1
getValue(x2) = 0 (neither my cut, nor my constraints (1)..(3) are violated)

but I still have something like:
LB(x1) = LB(x2) = 0
UB(x1) = UB(x2) = 1
or:
LB(x1) = 1 / LB(x2) = 0
UB(x1) = 1 / UB(x2) = 1

I consenquently miss some new possible deductions or worse, I deduce wrong things...

QUESTION:

Is it possible to set bounds of variables (x1 and x2) inside a UserCutCallback or using other tricks ? I would like to have:
LB(x1) = 1 / LB(x2) = 0
UB(x1) = 1 / UB(x2) = 0

In other words, is it possible to "propagate" constraints after any call of my deuction routine ?

Thanks a lot,
Tulkas
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: SystemAdmin

Adding a (local) cut will not necessarily affect the bounds that CPLEX stores.
If you can fix variables by your deduction algorithm then you are probably better off using a branch callback rather than a usercut callback. In the branch callback you could just create a single new node with all the fixings/bound strengthenings you could deduce. This will affect all bounds in the subtree rooted at the newly created node.

Or did I get you wrong and the fixings you deduce are globally valid and you want to change the global bounds for variables?
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: Tulkas

My deductions are not global. I'll try to use branch callback and keep you in touch !

Thanks a lot.
Tulkas
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: Tulkas

I might be wrong but I have the feeling BranchCallback do not eliminate nodes I would like to prune thanks to my deductions. Considering the same example presented above, if I deduce x=1, I jump to node:

LB(x1) = 1 / LB(x2) = 0
UB(x1) = 1 / UB(x2) = 0

thanks to makebranch(...) method, but I think CPLEX might still generate unwanted nodes later...

Example of undesirable node:
LB(x1) = 1 / LB(x2) = 0
UB(x1) = 1 / UB(x2) = 1 (see above for more details)

Am I Wrong ?

Thanks a lot,
Tulkas
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: SystemAdmin

> I might be wrong but I have the feeling BranchCallback do not eliminate nodes I would like to prune thanks to my deductions. Considering the same example presented above, if I deduce x=1, I jump to node:
>
> LB(x1) = 1 / LB(x2) = 0
> UB(x1) = 1 / UB(x2) = 0
>
> thanks to makebranch(...) method,
>
When you do that, do you only specify LB(x1)=1 to makebranch() or do you also specify UB(x2)=0 for the new node?

> but I think CPLEX might still generate unwanted nodes later...
>
> Example of undesirable node:
> LB(x1) = 1 / LB(x2) = 0
> UB(x1) = 1 / UB(x2) = 1 (see above for more details)
>
Are you sure that this node is a descendant of the node you created above? Maybe this "unwanted" node is in a completely different part of the tree?

> Am I Wrong ?
>
Most probably yes. If you have UB(x2)=0 (for binary x2) for a node in the tree then CPLEX will never create a descendant node with UB(x2)=1 for that node.
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: Tulkas

I think it works (I specify both LB(x1)=1 and UB(x2)=0) :-)

Thanks a lot !
Tulkas
#CPLEXOptimizers
#DecisionOptimization