On Mon, 2013-04-29 at 18:48 -0500, Michael Hennebry wrote: > On Sat, 27 Apr 2013, Heinrich Schuchardt wrote: > > > the current definition of the MIP gap is compatible to other optimizers > > like CPLEX. > > To the best of of my knowledge, > the terminal output is designed to be read by humans. > It will not be inerpreted by another program. > There is no compatibility issue. > > > The advantage of the current definition has the advantage of beeing > > meaningful to the > > end user. > > > > If the gap is 1 % and the current MIP solution costs me 1 Mio Euro, I > > know that even > > if there is a better solution, it cannot be better by more than 9901 > > Euro. Your proposed > > definition would not have such economic significance. > > You would not need the 1%. > An optimistic bound is one of the numbers provided on the same line. > > On the other hand, were the objective temperature, > the relative gap would change depending on whether > it was expressed in Fahrenheit or Celcius. > > My denominator would provide a better indication > of how hard it would be to close the gap. > > That said, Andrew has already expressed an unwillingness to change it. > > Heinrich Schuchardt quoted Michael Hennebry: > > I'd suggest that a better denominator would be |best_mip - root_lp| That > > would make the formula immune to shifts and scalings. It would also > > ensure that the gap was never greater than 100% .
The mip gap has the same meaning as the relative error (please see http://en.wikipedia.org/wiki/Approximation_error ), so changing the formula would confuse the user. A requirement to make the result "immune ... and never greater than 100%" looks too artificial. Probably, in order to avoid a meaningless result when the lower and upper objective bounds have different signs, the mip gap should not be computed at all. > > -- > Michael [email protected] > "On Monday, I'm gonna have to tell my kindergarten class, > whom I teach not to run with scissors, > that my fiance ran me through with a broadsword." -- Lily _______________________________________________ Help-glpk mailing list [email protected] https://lists.gnu.org/mailman/listinfo/help-glpk
