Assume you have optimized a given model instance and want to apply some changes. In that case, the following holds:
- If the original model is feasible:
- Adding a constraint to the primal problem would still allow Gurobi to use warm-starts in the dual space, as that amounts to adding a new decision variable to the dual problem.
- Removing a constraint would effectively relax the primal problem, so Gurobi should be able to use warm-starts in the primal space.
- Adding a decision variable (whose domain includes zero) to the primal problem would still allow Gurobi to use warm-starts in the primal space.
- Removing a decision variable from the primal problem would effectively relax the dual problem, so Gurobi should be able to use warm-starts in the dual space.
- Changing coefficients is not guaranteed to keep the primal or dual basis feasible, and hence Gurobi would not use warm starts at all.
- If the original model is infeasible and then some constraints are removed from it, Gurobi will start from scratch. That said, if the model is an LP and the user wants to compute an IIS, Gurobi will use the information contained in the last basis to guide the search.
Further Information: