Skip to main content

Solver behaviour when giving fixed variable a start value

Answered

Comments

1 comment

  • Jaromił Najman
    Gurobi Staff Gurobi Staff

    Dear Sebastian,

    The start solution you provide triggers a "initial solution phase". As can be seen from the two logs, this phase is triggered before the presolve step. What happens in your case, is that completing and computing an initial solution takes a significant amount of time and the found solution does not improve the presolve step (both presolved models are of same size and #nonzeros). Now going into the B&B tree, one run has an initial solution and the other does not. If we cross out the time required for the "initial solution phase" then the run with a starting solution takes ~1200 seconds while the second run without an initial solution takes ~1700 seconds. The downside in this case is that completing the partial solution takes a significant amount of time and the performance tradeoff is bad. To not alter your code too much, you could try experimenting with the StartNodeLimit and SubMIPNodes parameters to limit the time spent to complete a partial MIP start.

    Best regards, 
    Jaromił

    1

Please sign in to leave a comment.