Hello. Le lun. 30 août 2021 à 15:02, Alex Herbert <alex.d.herb...@gmail.com> a écrit : > > Hi, > > This test for the SimplexOptimizer is not robust.
Indeed; sorry for that. Perhaps those "standard" test functions (being purposefully "difficult" problems) should be moved to some "integration tests" module (or to the "examples" module; thus leaving as "unit" tests only "simple" functions... > It uses random seeding to > run optimisation problems. If these do not work from the chosen start point > then the test fails. Up to 10 repeats of the test are allowed. Sometimes > this is not enough. > > The tests originate from a long time ago in CM. Keeping the tests ensures > that the library continues to support these problems and avoids regressions. Yes; but it is not sure that the old test suite ever exercised the "difficult" nature of the test functions (possibly because only "lucky" start points were selected). > > Gilles has been working on the optimizers and may be able to provide more > details. Recent work has been successful in enabling many previously > ignored test cases to be reintroduced to the test suite. However all this > good progress has not totally eliminated spurious failures. Failures can be caused by a wrong implementation or by an inherent weakness of the algorithm(s) that may or may not show depending on the input data (start point, initial simplex, ...). Gilles > For the time being you can ignore failures in these tests which are likely > unrelated to changes you are performing. Any PRs raised against the master > branch are evaluated in the context of the change. Any failures of these > tests during the CI build will be noted and often just restarting the CI > build will pass the next time round. > > Alex > > >> [...] --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org For additional commands, e-mail: dev-h...@commons.apache.org
