Optimality gap in a mathematical model solver

4.9k Views Asked by At

How is optimality gap obtained from a solver like CPLEX related to the objective function. Is a higher optimality gap desirable?

1

There are 1 best solutions below

0
On

The gap between best possible objective and best found objective is obtained by keeping track of the best relaxation currently in the pool of nodes waiting for further processing. (For details see any book on Integer Programming).

This is a very unique property of a MIP solver that practitioners (like me) use a lot: instead of looking for proven optimal solutions that take a long time to compute we are happy with a solution guaranteed not worse than x% from the optimal solution. This can often speed up solution times by a lot.

Most other solution techniques (like meta-heuristics) do not have this capability and we don't know really how good their solutions are (except maybe using some probabilistic argument).

A big advantage is that often most of the gap is reduced quite early (in other words we find good solutions quickly). It is not unusual to actually find the optimal solution quite quickly but that proving global optimality (that is when the gap becomes zero) takes a very long time.