2011/6/25 Kristofer Munsterhjelm <km_el...@lavabit.com> > Jameson Quinn wrote: > >> >> >> 2011/6/25 Kristofer Munsterhjelm <km_el...@lavabit.com <mailto: >> km_el...@lavabit.com>> >> >> > So I quickly hacked together something to run it through my old >> > multiwinner code, but I'm getting unusual results. Could someone >> > check that they get what I'm getting? >> > > {C101 C102 C103 C104 C105 C106 C108 C109 C110} for birational >> voting >> > {C101 C102 C103 C104 C105 C106 C108 C109 C110} for Range PAV >> (integral) >> > {C101 C102 C103 C105 C106 C108 C109 C110 C116} for STV >> > {C103 C106 C108 C109 C110 C111 C113 C115 C116} for Meek STV >> > {C103 C106 C109 C110 C111 C113 C114 C115 C116} for Schulze STV >> > {C103 C106 C109 C110 C111 C113 C114 C115 C116} for QPQ >> >> >> How are you handling ties in the STV methods? Just eyeballing it, it seems >> that your results are skewing towards the high-numbered candidates (who, in >> this election, seem to be weaker candidates - perhaps the ordering is in the >> order that they submitted statements, so the least-organized candidates come >> last?). And when I look at your ranked inferences, you do indeed put the >> highest-numbered candidate first, so I'm wondering if you're actually >> (mistakenly?) using "lexically-last" as an arbitrary tiebreaker for ballot >> equalities. >> >> Again, that's just from eyeballing, so I could be wrong. >> > > I thought I was using the method of first difference as a tiebreak, but > apparently not. That method breaks the tie in favor of the candidate who > first has a higher score. > > So I made a randomization preround (permuting the candidate-number > assignment instead of assigning 0 to C101, 1 to C102 etc), and the results > changed:
Wait.... is that a global randomization, used across all votes? If it is... or in fact, even if it isn't... I suggest you do what Warren suggested, and run it several times, with different random seeds, to see if the results are reasonably stable. > > > {C101 C102 C103 C104 C105 C106 C108 C109 C110 } for birational voting > {C101 C102 C103 C104 C105 C106 C108 C109 C110 } for Range PAV > {C101 C102 C103 C104 C105 C106 C108 C109 C110 } for STV > Heh. As I suspected, STV came up with the same list as AT-TV. > {C102 C103 C104 C105 C106 C107 C109 C110 C116 } for Meek STV > {C102 C103 C104 C105 C106 C107 C109 C114 C116 } for Schulze STV > {C102 C103 C104 C105 C106 C107 C109 C110 C116 } for QPQ, > Those results also look much more reasonable to me. > > so there seems to be a lot of ties for the ranked methods to deal with. In > retrospect, it makes sense, because there weren't enough ratings for voters > to be able to rate every candidate uniquely. > > Well, that's what I get for using old code whose limits I've forgotten, I > suppose! > > I know the feeling.
---- Election-Methods mailing list - see http://electorama.com/em for list info