Dynamic stopping criterion - An effective way to interrupt the optimal solution search for MILP

As stated in my previous comment, the dynamic criterion you describe is way too vague to implement as a general feature in CPLEX. CPLEX would have to provide a way to allow the user to submit an arbitrary function to compute the acceptable gap depending on time and other features.

Thank you for your feedback.

We have already implemented the dynamic stopping criterion using callbacks.

Neverthanless, I think that general stopping criterion could be extended supporting also the dynamic stopping criterion.

Luigi Poderico

Can be done using Generic callback.

This dynamic stopping criterion looks a very flexible thing: over time the acceptable gap rises. But how exactly does it rise over time? Is this a linear function? A quadratic one? An exponential one? How exactly the acceptable gap is computed from the elapsed time may be very user specific. So it may be quite hard to cast this into a single parameter. One can also imagine extensions like "reset the gap for every new incumbent found", i.e., give the engine some more time to find good solutions.

All this being said, it seems rather hard to come up with a single dynamic stopping criterion.

On the other hand, the criterion described in the author's note can be easily implemented using callbacks: In a callback you can easily figure out the gap and the elapsed time. From a callback you can also easily abort the solution process. As of version 12.9 we have generic callbacks that guarantee to not have any effect on the optimization process as long as you don't explicitly stop the process. Using the generic callback with the LOCAL_PROGRESS and GLOBAL_PROGRESS contexts it seems rather simple to implement the requested dynamic stopping criterion (and others).