Different GAMS versions giving very different answers+ weird behavior

[pecenak21]

Hello,

I am running a rather large model in GAMS 24.9 and 25.1, and seeing some weird behavior in the former, but expected behavior in the latter. As you can see in the attached picture, in 24.9 the solution seems to jump from an 86% gap (170398.9 Best Bound / 1257045.9 Best Integer), to reporting 1256452 (Best possbile) as the optimal solution, with a .04% gap. When running this same model, with the same inputs, in 25.1 it actually solves 'correctly' and provides an optimal solution within 0.1% of the expected best possible (170398.9). Also, the runtime is vastly different between the two (as 25.1 has a much lower gap). I have observed this behavior on a few different versions of this model (i.e. different parameter inputs), and can't figure out why this is happening. Can you provide any insight into why the solution jumps like that and changes the optimal solution? Is there a setting I should adjust? Is there a bug in that version, or functionality that changed between the two? Please let me know any other information which is needed to assist in understanding this problem,

Thank you for your time,
Zack

[bussieck]

24.9 and 25.1 use different versions of Cplex (12.7.1. versus 12.8.0.0). Perhaps a bug was fixed that relates to this behavior. You might want to explore the Cplex' fix list at https://www.ibm.com/support/pages/ilog- ... o-fix-list

-Michael

[pecenak21]

Hello Michael,

I am sorry to report this is resurfacing with a much newer version of GAMS (35.0). As you can see below, the solution is stopping at ~50% gap(18283.8 Best Bound / 33390.9 Best Integer), and suddenly jumps to reporting 35858.85 (Best possbile) as the optimal solution, with a .7% gap.

Have you seen this before, or have any general insight into why this is happening. This is the third time I have seen it, and all with different versions of GAMS.

Here are my settings, if it is helpful.

image.png (8.62 KiB) Viewed 2260 times

Zack

[bussieck]

I agree that this is odd, but not necessarily impossible. Cplex has so few open nodes (~140) that finishing up on a particular node might allow Cplex to raise the dual bound significantly. You might try a different solver and see if they confirm the optimal solution (or the primal and dual bounds). If you still suspect something really wrong, send the model plus data to support@gams.com for further inspection.

-Michael

[pecenak21]

Do you have a suggestion for a different solver? I tried using the cbc solver, but it won't solve it even after an hour. I am going to try to let it run overnight. But maybe you have a suggestion or temporary/trial solver license I can use for this purpose?