[EM] Strong methods (was Re: 3 or more choices - Condorcet)

[Michael Ossipoff]

On Sun, Sep 30, 2012 at 4:56 AM, Kristofer Munsterhjelm
<km_elmet at lavabit.com> wrote:
> On 09/29/2012 10:49 PM, Juho Laatu wrote:

[...]

[in answer about what a strong method is]:

> ...and not giving weird results
> or having weird result dynamics that could be used to discredit the method.

You mean like failing Participation, Consistency, IIAC, etc.?  :-)

Rank methods have lots of problems. I've mentioned a few already:

1. There are so many rank-counts that there will never be agreement on
which one to propose, adopt or enact.

2. Rank methods are count-computation-intensive, and therefore offer
more count-fraud opportunity, especially since their
computation-intensiveness necessitates machine balloting and
computerized count

3. Failure of such "embarrassment criteria" as Participation and
Consistency is a criticism vulnerability in enactment campaigns.
Opponents can say, "This bizarre and ridiculous result couldn't happen
with Plurality. Do we want that?"

Approval, and probably Score, doesn't do anything wrong or
criticizable that Plurality doesn't do. Approval is nothing other than
an improvement on Plurality. A clear and unqualified improvement.

That can't be said for the rank methods.

By the way, such "strong" methods as Beatpath, Ranked-Pairs, River,
Kemeny, etc., have several undesirable and avoidable problems:

FBC failure

easy successful burial strategy

Chicken dilemma

Later-No-Help failure

------------------------------------------

Symmetrical ICT avoids those problems. I recommend it for
informational polling. But, for any purpose, it would be better than
the "strong" methods listed above.

>
> In practice, that means: is cloneproof, passes independence of as much as
> possible (independence of Smith-dominated alternatives, say), and is
> monotone.

That's the problem that I was speaking of earlier: People at EM are
inclined to say what criteria are important, without justifying them.

Monotonicity is, of course, an important embarrassment-criterion. The
others need justification.

> River would be even better than Ranked Pairs, since River passes
> independence of Pareto-dominated alternatives and RP doesn't, but River is
> even less known than Ranked Pairs.

In the overall scheme of things, what would be the dire societal
consequences of failing Independence of Smith-Dominated Alternatives,
or Independence of Pareto-Dominated Alternatives?



> I've put strong in quote marks because I know others may disagree with my
> priorities.

Exactly. That's why criteria need to be justified.



>FairVote obviously doesn't consider the "having weird result
> dynamics" part important as long as the strangeness can't be exploited by
> deliberate strategy.

Advocates of the "strong" methods you listed seem, too, to be in
denial about that, among other problems of those methods,
and of rank methods in general.

> The first is consistency with itself. Nonmonotone methods do badly here. The
> intuitive idea is that if a method is not monotone (say), then that means
> that its concept of what is better is lacking - it's like someone who says
> "I'm closer to the city" after traveling in the wrong direction. It's
> important to make clear that whether or not these inconsistencies can be
> exploited through strategy is not really important.

Inconsistencies like that are serious criticism-vulnerabilities to
enactment proposals.

...like when you and a few other people show up late to vote, and
thereby cause the defeat of someone whom you've all ranked over
everyone else.

You try to gloss over that by saying, "Though it fails Participation,
it doesn't fail Mono-Add-Plump, and _that's_ the important thing."

No, you want that to be the important thing because you want to
propose a Condorcet method, and so you retreat to Mono-Add-Plump.


>The danger is that a
> perfectly innocent election will find itself on the wrong side of an
> inconsistency and so the result will be either inferior (as a result) or
> less legitimate (because people will say "WTF is going on here?").
> Of course, there are some such inconsistencies we have to accept if we want
> Condorcet.

_You_ will have to accept them, and you do accept them. But that
doesn't mean that the public will accept them when an enactment
proposal's opponents aggressively present them to the public.

I don't think that the embarrassment-criteria are important, except as
criticism-vulnerabilities. But it's a mistake to be in denial about
the importance of criticism-vulnerabilities.

Non-vulnerability to an army of clones seems important, and I don't
think anyone would deny that.

Likewise avoiding the split-vote problem that you refer to below.

> The second is resistance to noise and strategy. Independence of clones fit
> here, as well as independence of X (Smith-dominated alternatives,
> Pareto-dominated alternatives, weak IIA). The resistance may protect against
> strategy - cloneproof methods keep parties from running an army of identical
> candidates - or improve the outcome when there is no strategy - e.g. by not
> being affected by the liberal parties' vote-splitting in a replay of the
> 1988 South Korean presidential election.
>
> The third is quality of the outcome under honesty, according to some metric
> or desired logic.

That's where it gets more arbitrary, of course.


> It's hard to say which metric one should pick,

Exactly.


> unfortunately, and for Ranked Pairs (and Schulze), there's probably no
> simple metric that the method optimizes. Furthermore, the logic one uses for
> rated methods probably wouldn't directly fit onto rank methods (because
> utilities are either unknown or not applicable).

Sincere rating, in the form of trying to make your ratings match the
candidates' utilities to you, as the ideal of Score and MJ, is an
illusory notion. There's nothing "insincere" about max-rating all of
the acceptable candidates, or all of the candidates whom you like and
trust.

>
> I'm not sure where Condorcet compliance fits into the categories above,
> either. Perhaps it's the third, in a sort of deontological logic that says
> "do whatever you want, but if there's a candidate that would win every
> runoff, elect him". Perhaps it's a consistency criterion, where the people
> expect X to win outright if he can win every runoff. Or maybe it's "doing
> without strategy what the voters could do with enough coordination in other
> methods", easing the burden on those voters - or a way to have the method
> resist single-group repeal efforts, where electing the CW ensures that if
> the supporters of a loser tries to repeal or complain, there will always be
> a greater group of supporters to defend the method, no matter who that loser
> is.

The Condorcet Criterion can be justified in all of those ways. I don't
think it's necessary as a consistency criterion. I don't think
enactment opponents are going to criticize Approval or score for not
meeting a criterion failed by Plurality.

The strategic justification of CC is the important one, and the one
that I've often advocated.

Your last justification, above is good, but I don't think it's
necessarily more helpful than Approval's electing the most approved
candidate, or Symmetrical ICT electing the most top-rated unbeaten
candidate. (or the most top-rated candidates if no one is unbeaten)

The Condorcet Criterion is a valuable luxury. A luxury, but certainly
not a necessity. And its value is conditional upon voters not having
serious disincentives to sincere ranking. That's where ICT and SICT
come in. CC is worthless if people don't feel free to rank sincerely.

But you discussed CC in the right way, discussing its justification.
That's what EM discussion needs to be about: Justification of the
criteria that we use.

Mike Ossipoff