forum.alglib.net

About a 5% frequency of failure detected in the minbleic module, under the "feasibility test" unit test -- where it reaches the "Feasibility problem" section of the test. It may fail to find a feasible point, so that the minbleic iteration will not even start. It has been seen in the minbleic test (*converr can be flagged, with the iteration report returning -3.)

Run the subtest in isolation at a high volume.

Hi!

I am aware of this issue. The problem is that unit test sometimes generates nearly degenerate set of constraints, which should not happen in real life scenario.

However, probability of failure is far less than 0.1% (according to last full run of test suite) and heavily depends on combination of compiler flags being used. You were just "lucky" to trigger it. Unfortunately, this particular test generates other interesting combinations of constraints, which we want to check, so we can not simply remove it. And fixing it turned out to be quite tricky, given the fact that failure rate is very low. Speaking short, it is problem of unit test, not of the solver itself (it can solve well defined problems).

It is planned to rewrite some critical parts of the BLEIC solver (in particular, feasibility detection code), so it was decided that this test should wait for the refactoring of the optimizer. And then a decision will be made regarding it.

The failure probability for the unit test, itself, was gauged to be around 5%. The routine tested is executed a large number of times, so I assumed that the failure probability for the iteration routine would be several orders of magnitude smaller (a tentative estimate was under 1/100000).

What other subpackages are affected?

The particular combination I found was this:
unit test, p = 4, n = k = 5.

x0: +1.972865, -0.397756, +1.521437, -0.817839, +0.083507
xs: +0.457828, -0.302526, +0.833163, +1.843420, -0.702891
xc: +0.432256, +0.372770, +0.826043, +0.161946, +0.453414
ct: +1, -1, +1, 0, +1
bl: 0,0,...,0
bu: 1,1,...,1
c
+0.089912, +0.718084, -0.063148, -0.793405, -0.152167
-0.198801, -0.886876, +0.201788, +0.810620, +0.056276
+0.868877, -0.389735, -0.457104, -0.714402, +0.345693
-0.799495, -0.586873, -0.369640, +0.303995, -0.605595
-0.766605, -0.804217, -0.247915, +0.888507, -0.135991
with column 6 filled with the sum of the first 5 columns

Since it is an artifact of the tester, then I will probably make repairs in my local copy of the library to address the problem and later synchronize to whatever you come out with in your later versions.