Solutions of upper problems and lower bounding problems are obtained by solving two separate local optimisation problems, using two different third-party solvers, each with their own tolerances and numerical issues.
This is the tolerance within which we consider the solution of these two problems to be “the same”. It’s a bit more complicated than that because this number is really the seed to some other complicated error handling formulas, but let’s go with this explanation for now.
Try increasing this tolerance if:
- Your problem is proven infeasible during branch-and-bound and you really think it shouldn’t be
- You globally converge to a sub-optimal value
Try decreasing this tolerance if:
- Convergence is too slow and the node count is insane