|
In testminqpunit_generallcqptest(), under "SMALL-SCALE TESTS: many tests for small N's", with solvertype = 1, under "Convex optimization problem with known answer", with n = 8 and akind = 1:
gtrial = b + a × (x1 - xorigin) + rep.lagbc + rep.laglc × rawc.
For i ∈ [0,n): the test is for |gtrial[i]| ≤ 0.001, but is only flagging if |gtrial[i]| > 0.001. However, gtrial[i] may be NAN, because rep.laglc occasionally has components equal to NAN. So it is possible for the test to fail, and yet fail to be flagged, resulting in an "ok" for the test - a false positive.
It may be the case that your algorithm allows rep.laglc to have NAN values and treats them as infinitesimals, but this should be examined more carefully. The false positive occurs about 5% of the time that the minqp test in test_c is run.
More generally, most of the tests in test_c flag on reverse logic for real-number comparisons, and may therefore yield false positives if NAN's occur. In our local version, we eliminated the possibility by making all the tests positive logic on real number comparisons. So far, this is the only false positive that's been uncovered, as a result.
|
Login
Register
FAQ
Search
