[EM] Absolutely new here

[Kristofer Munsterhjelm]

On 06/16/2013 06:18 PM, Abd ul-Rahman Lomax wrote:
> At 02:02 AM 6/16/2013, Kristofer Munsterhjelm wrote:
>
>> It would work, but the rating variant is better. In the context of
>> ranking, Bucklin fails Condorcet, for instance.
>
> Straight Bucklin does fail Condorcet, of course, as do straight Range
> and Approval. However, we can tell from the fact that Range fails
> Condorcet that there is a problem with the Condorcet Criterion, one of
> the simplest and most intuitively correct of the voting systems criteria.

Your first claim is right, of course. It is also right that I'd never 
want to have Range used as a straight ranked system. Depending on how 
you implement it, ranked Range becomes Borda, Plurality or 
Antiplurality, neither of which are very good.

I don't think your second claim follows, though. Condorcet might be a 
bad criterion to apply to a rated or grade-based method (because it 
doesn't take preference strength into account), but that it is a bad 
criterion in general doesn't necessarily apply. In the case of a ranked 
method, like say the original Bucklin, you don't *have* preference strength.

> The problem also applies to the Majority Criterion. Those criteria do
> not consider preference strength. Practical, small-scale, choice systems
> do, routinely. They do it through deliberative process and repeated
> elections, vote-for-one, seeking a majority. And then, a process that
> can even review a majority choice and reverse it, where preference
> strength justifies it.
>
> Thus a deterministic single-poll method that optimizes social utility,
> and that collects information allowing that, *must* violate the criteria.
>
> And that's a problem, because this is a fundamental principle of
> democracy: no binding choice is made without the consent of a majority
> of those voting on the issue. Some are aware of the "tyranny of the
> majority," but solutions to *that* cannot be found in deciding *against*
> the preference of the majority, *without their consent.* The result is
> minority rule, not broader consensus.

And that's kind of the thing I'm talking about. Ranked ballot methods 
don't have preference strength data. In that case, ruling in favor of 
the majority is better than doing so in favor of the minority. Of 
course, it would be better to not have to make that choice in the first 
place, but my point is that if you consider Bucklin, the ranked method, 
then it has to make a choice one way or another. And between the choice 
of making assumptions that makes it fail Condorcet, and making 
assumptions that makes it pass, the latter is preferrable. That is also 
what ranked Bucklin does.

But my point is that while rated Bucklin (MAV) might be a good method, a 
ranked quantization of it might be worse than other ranked systems that 
exist: simply because in the rated version, the weirdness is less 
relevant since it doesn't have to make assumptions the ranked version does.

> So there is a solution: repeated election. Over the years of considering
> this problem, I've concluded that with the use of advanced voting
> systems, such as Range methods, and good ballot analysis in a first
> round, with a runoff where a majority decision is not clear, such that a
> Condorcet winner in a primary will *always* make it into a runoff, in
> addition to one or more social utility maximizers, it is possible to
>
> 1. Find a majority choice, almost always, in two ballots, with the
> exceptions being harmless.
> 2. Satisfy the Majority and Condorcet criteria.
> 3. Optimize social utility.
>
> These have been considered opposing goals. That is because
>
> 1. Voting systems study has neglected repeated ballot.
> 2. Voter turnout has been neglected.
> 3. The electorate has been assumed, where runoffs have even been
> considered, to be the same electorate with the same opinions. Neither is
> real.

And that would be a system that gets around the problem by not having to 
make an assumption. For instance, when MJ/MAV fails Condorcet, or when a 
Condorcet method picks a Condorcet winner, and you're right about the 
differences in voter turnout fixing things, then the runoff can 
distinguish a weak CW from a strong CW -- or a candidate that fails 
Condorcet for good reason from a candidate that is picked by MJ because 
of artifacts in MJ/MAV itself.

>> It also has some bullet-voting incentive. Say that you support
>> candidate A. You're reasonably sure it will get quite a number of
>> second-place votes. Then even though you might prefer B to A, it's
>> strategically an advantage to rank A first, because then the method
>> will detect a majority for A sooner.
>
> This is somehow assumed to be "bad." That incentive exists if there is
> significant preference strength. Thus "bullet voting" is a measure of
> preference strength, i.e., is useful in measuring social utility. There
> is, however, another cause for bullet voting: voter ignorance (which is
> natural and normal). A voter simply may not know enough about another
> candidate to vote for the candidate. And this is probably the major
> cause of bullet voting, historically, with Bucklin, combined with high
> preference strength.

Or it could simply be partisan gaming of the system. With only ranks, 
and only one round, there's no way to know.

(As I have said before with regard to LNHarm and LNHelp, at least for 
ranked methods, I prefer a method that fails both to about an equal 
degree, to one that passes one and fails the other.)

>> One of the points of the graded/rated variants is to encourage the
>> voters to think in absolute terms ("is this candidate good enough to
>> deserve an A") rather than relative terms ("is this candidate better
>> than that candidate"). If they do, then the method becomes more robust.
>
> If somehow we could extract absolute utilities from the voters, sure.
> However, real-world, people make choices based on relative utility, not
> absolute utility. Imagining a voting system as becoming more "robust,"
> if voters behave utterly unrealistically, depends on a rather strange
> idea of "robust." We *are* machines, but we are programmed to optimize
> among *choices*. Our very assessment mechanisms are relative to what is
> "espected as realistic possibilities."

But doesn't B&L's experimental evidence show that the voters *can* think 
in absolute terms? Not exactly in utility-normalized ones, but in terms 
where the comparison points are set? For grades, the distance between A 
and B might be greater than the distance between B and C, but for MJ, 
that doesn't matter. I don't think it matters for other Bucklin variants 
either, though I am less sure of that.