Originally posted by: SystemAdmin

Hi everybody,

I have just started to compare the local search implementation of my own against ILOSOLVER.
The thing is that as far as I could understand the objective function must be written explicitly in terms of problem variables. Now, I have a hard-to-calculate objective function which is nonlinear and NMIP by itself. I am happy with an approximation of the objective value which I can obtain from a specialized heuristic in polynomial time instead of solving the ugly NMIP.
My question: is there any ways, any callback etc to enable user to assign a value to a candidate/incumbent rather than letting this to be calculated from a closed form formulation in terms of problem variables?

regards,
Shahin
#CPOptimizer
#DecisionOptimization

Originally posted by: ChrisBr

Hello Shahin,

We need clarifications to answer your question.
Could you be more precise about what you are using or want to use?
You seem to be using CP, so maybe a custom constraint can be used?
Please help us to help you!
Thanks.

Chris.
#CPOptimizer
#DecisionOptimization

Originally posted by: SystemAdmin

Thanks Chris for your comment.
I am just trying to develop a local search let say TS in CP while in spite the common practice I do not have a closed form for my objective function. If solver gives me a candidate I can calculate the objective but can not put it in the objective function of solver.
Is there any possibility like a solve callback or whatever else such that I can see what neighbor the solver is going to move towards and I myself calculate the objective function and install it when the move actually happened?

I can have a permutation based representation solution in heuristic but my objective is another BIG problem that I can solve it efficiently if the candidate is known. hoe can I inject my objective value to solver.

I hope that clarifies ;)
#CPOptimizer
#DecisionOptimization

Originally posted by: SystemAdmin

Dear Shahin,

I am not sure to understand the model you want to solve: you have variables X, a set of constraints C(X), and an objective function f(X). You can model C(X) but you cannot easily model f(X). However, you can compute a value for f(X) once all variabled of X are fixed. Is that correct ?

In this case you need to define your own constraint c(y, X) that is equivalent to y = F(X). A basic implementation would be to wait until all variables X are fixed to fix y to the objective value. A smarter way to go would be to infer a range for y even when a part of the variables X are fixed, and - converely - to be able to reduce the set of possible values for variables in X when having a range on y.

There are several examples in the Solver manual that shows how to define specialized constraint. I particular, you can look at the chapter "Defining New Constraints" in the user manual.

Regards,

Philippe
#CPOptimizer
#DecisionOptimization

Originally posted by: SystemAdmin

Many Thanks Philippe for your comment. I try to see how this can be done.
#CPOptimizer
#DecisionOptimization