Gilles Sadowski wrote: > Hi. > >> [...] OLS regression (350 - good catch!) also >> obviously needs to be fixed. Thanks to all for testing. > > There is indeed a difference from what CM 2.0 computed but maybe that it is > not a problem after all, and is just to be expected from the implementation > change: some failing tests compare values produced by CM 2.0 to values > produced by the current code allowing a tolerance of 1e-13. When setting the > tolerance to 1e-12, the tests pass. Those tests cannot prove that the values > from CM 2.0 were "better" than the current ones. > > Looking into the issue with Dimitri, the problem initially referred to in > MATH-350 may in fact be related to the user's code. I'm waiting for her > feedback in order to write some (possibly final) comments on JIRA. >
Great. One (non-authoritative, but at least indicative) resource to use is R verification. There is a framework for verification tests comparing results to R in src/tests/R. In some cases for (bivariate) regression, the reference values are also compared against NIST reference data. Phil > > Best, > Gilles > > --------------------------------------------------------------------- > To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org > For additional commands, e-mail: dev-h...@commons.apache.org > --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org For additional commands, e-mail: dev-h...@commons.apache.org
