4

Suppose I have an integer program model in the form of a minimization. I noticed that Gurobi (my solver) often finds a very good upper bound (i.e., feasible solution) whereas it takes a significant time to improve the lower bound to reduce the optimality gap.

Here is my question. Is there a way to obtain a probabilistic statement about the optimality? For instance say

$$prob\{f^u - f^\star > \epsilon\} \le \gamma$$

where $f^u$ is an upper bound and $f^{\star}$ is the global objective value.

Similarly, an inequality in this form is also desirable:

$$prob\{f^u - f^{\ell} > \epsilon\} \le \gamma$$

where $f^{\ell}$ is the lower bound.

Any thoughts or suggestions will greatly be appreciated!

Rodrigo de Azevedo
  • 23,223
user2512443
  • 445
  • 2
  • 12

1 Answers1

4

Unfortunately, no. There is really nothing else you can say other than that the optimal solution is between the two bounds.

Presumably you could estimate these probabilities experimentally, but I assume that is not what you’re asking.

LarrySnyder610
  • 1,131
  • 7
  • 16
  • Thanks for your feedback. I agree it seems very far fetched to be able to come up with a probabilistic statement given the nature of branch-and-bound procedure. That said, I'm curious to know how to estimate these probabilities experimentally, as you pointed out. Any suggestions? – user2512443 May 15 '19 at 00:00
  • I just meant, solve the model a bunch of times for a bunch of randomly generated instances, and empirically estimate the distribution of the gaps. But this seems prohibitively computationally intensive. – LarrySnyder610 May 15 '19 at 00:36
  • 2
    I agree with this answer in general. One exception: if Gurobi generates a feasible solution from a randomized rounding method (or Gurobi's solution is better than a solution which you independently generate from a randomized rounding method), then you can sometimes write down bounds on $\epsilon$ and $\gamma$. For instance, using the probability distribution which arises in the proof of Goemans-Willliamson rounding, you should be able to do this. – Ryan Cory-Wright May 15 '19 at 03:48