Assertion `new_infeasibility < 0' failed in HEkkPrimal.cpp:2813: void HEkkPrimal::shiftBound()

[ambros-gleixner]

Running SCIP with HiGHS in debug mode on https://miplib.zib.de/instance_details_traininstance2.html runs into the above assert.

The attached log traininstance2.txt shows that SCIP initially calls the dual simplex, HiGHS reaches the objective limit, and then switches the primal.

In rebuild() -> correctPrimal() the assert fails because the bound is huge and adding the small perturbation of 1e-6 gets lost. Maybe the perturbation added should be scaled up by the size of the bound, if you want to break degeneracy in this case.

[jajhall]

Thanks, I'll take a look in due course

[ambros-gleixner]

One suggestion, not sure what you think: use a combination of relative and absolute perturbation:

• for bounds with abs. value smaller than X keep as now
• for larger bounds multiply the perturbation with (abs. value of the bound / X) X could be something like 1e4 then, the perturbation would be at least 1e-10 times the bound and would not disappear when added.

[ambros-gleixner]

I implemented this suggestion in pull request #1175 ; it resolves the problem in this issue.

[ambros-gleixner]

What do you think?

[jajhall]

I'll look at it, but will have to be sure that it doesn't lead to other problems or impair general performance

[ambros-gleixner]

Of course.