On Sun, 2020-09-27 at 11:32 +0200, Manuel Muñoz Márquez wrote: > I agree with you, Andrew, but the problem is when the output is not a > real number. > > Suppose that you have to decide which of the project that are planning > a big company will be done in the next year. Little difference in > computation may lead to a solutions that are far enough one from the > other providing very different sets of projects to be done. Is this > admissible? Is this desirable?
If the instance has a unique optimum, only it will be reported on any platform; however, the solution may slightly differs on different platforms, but only within a tolerance. The case you're talking about may happen only if the instance has multiple optima. In this case on one platform the solver may report one optimal solution while on other platform another, completely different optimal solution. But since all these solutions are optimal, they all are *equivalent* to each other, so you can choose any of them. If, nevertheless, you think that some optimal solution is not that you expect, this may only mean that you missed some essential constraints. > > So on Monday, Wednesday, and Friday that you work on your laptop you > said that "Project A" should be done, but on Tuesday, Thursday, that > your work on your supercomputer cluster you said that "Project A", of > course, should no be done. > > I'm not speaking if that is possible or not, I known that in the > general case it is not, but for me it is a desirable behavior, and I > think that the software should be as close to that behavior as > possible. > > El sáb, 26-09-2020 a las 15:40 +0300, Andrew Makhorin escribió: > > > > In case of TeX it is important to provide identical output on any > > platform, and to attain this Don implemented all calculations using > > rational arithmetic. Though this approach can be used to solve, say, > > linear algebra problems, it is impractical in general case. A > > desirable > > behavior of a computer program that solves a problem with a > > numerical > > method is to provide the result with sufficient accuracy for a > > reasonable time, not more. Engineers and scientists (unlike > > mathematicians and maybe accountants) as a rule are not interested > > in > > more than 5-6 correct decimal places in numerical results. > > > > > > > > > >