7 Jan
2013
7 Jan
'13
5:06 p.m.
On Mon, Jan 7, 2013 at 5:35 AM, Morten Groftehauge
Is it because the quick tests are checking the result against the results that you got earlier and the algorithm has been changed between 3.3 and 3.4?
It can also be due to system and compiler differences - the compiler optimization can have very different numerical effects, especially in stochastic methods. Many of our regression tests are tuned to run on a specific machine, which is unfortunate but we have no way otherwise to check if the actual answer matches what's expected (as opposed to merely checking if the program runs without crashing). -Nat