wrong result without "-preprocess off"

[raller09]

my .mps-files

20250616_preprocess_wrong_result.zip

lead to a wrong result with preprocess on master.

if i change it to "-preprocess off" the correct solution is found.

Log.txt

Found this as a solution with a missing max constaint was more expensive...

the result is "wronger" since 4386591
(with y2 max constraint the version from 2025-03-18 leads to a correct result. now it is a little off...)

cbc reports: Cgl0014I Postprocessing changed objective from 32768 to 3449.7682 - possible tolerance issue - try without preprocessing

thank you

[jjhforrest]

Initially I thought it was because of semi-continuous variables - but preprocessing seems to work.

However I would claim that the message is correct - it is a tolerance issue. The smallest cost is 0.5 and the largest is 1.0e14 - so a tiny change (i.e. much less than the primal tolerance) in the value of such a variable will cause the difference.

[raller09]

Thank you for your feedback.

I tested the model with neos-server and CPLEX, Gurobi, and HiGHS. These solvers return the correct result but also report issues with the model:

Gurobi:
Warning: Model contains large objective coefficients
Consider reformulating model or setting NumericFocus parameter
to avoid numerical issues.

HiGHS:
WARNING: Problem has excessively large costs: consider scaling the costs by 1e+4 or less, or setting option user_cost_scale to -14 or less

lpSolve also returns the correct result.

The variables with values like 1.0eX are Big-M terms used to find variables which can make the problem infeasible.
I’m struggling to determine the right value here.
Unfortunately, there are 2–3 entries with ratios like 50000000/100% (resulting in 25000 required values), which compute to 0.05%...
This is where the large coefficients come from.

Possible options for preprocessing in CBC: aggregate, equal, equalall, equalallstop, forcesos, off, on, save, sos, stop(aftersaving), strategy, trysos.

I’m currently running it with "-preprocess off", which resolves the value mismatches.

Questions:

Does CBC provide feedback (status, error codes, etc.) indicating that the result might not be optimal?

Is there a better preprocess mode/setting for this scenario?

Thank you for your help and insights!