Re: [EAI] Please - Don't state preferences yet!

[John C Klensin <klensin at jck.com>]

--On Wednesday, 25 July, 2007 10:58 +0200 Arnt Gulbrandsen
<arnt at oryx.com> wrote:

> John C Klensin writes:
>> For example, in a situation like the one we have, there is a
>> strong  argument that one should first examine "I completely
>> hate this"  votes, eliminating any options that are intensely
>> disliked by a  non-trivial number of people, and only then
>> applying a model based on  affirmative preferences to what is
>> left.
> 
> +1
> 
> But please name/suggest a method that has these properties and
> wasn't designed by a naÃve layman.

Unfortunately, I am a long distance from my library, haven't
reread the literature in years (more or less since I stopped
teaching it).  This is also not a good week for me to try to
engage in an extended discussion.  For all I know, something new
and better has been invented, although the reasons most of the
procedures don't work universally can be demonstrated with
rather clear mathematics.

The key is not to find an universally-optimal method, but to
understand that each one measures something slightly different
and that, by choosing the method/measure, one can affect the
outcome.

Charles's recent on-list argument is ultimately close to
tautology, something that would be immediately recognizable if
he stated it as "Condorcet measures what I consider valid,
therefore it works and there are no attacks on it".  

So, _ because I believe that negative preferences are very
important_ --clearly a statement about ideology, not voting/
measurement methods-- I would prefer that we 

(i) pick a threshold for eliminating noise/ outliers.  It turns
out to not make a lot of difference what that threshold is.
There are scaling techniques to reduce the variance introduced
by people who express themselves strongly about everything
versus those who don't, but they are time-consuming and rarely
worth the trouble in this sort of situation.  Given the level of
participation in the WG, I'd pick a number greater than 1 and
less than 5.

(ii) Go through the options and count "I hate this" votes that
can be supported with an explanation/ argument that will pass a
laugh test.   Laugh tests are better administered/validated by
judges whom almost everyone believes are fair than  by voting.

(iii) Drop anything that gets more than the threshold value of
"I hate this" votes.

(iv) From the rest, pick.  It may not make much difference how
one picks: Condorcet works well, Alternate Vote works pretty
well, and lotteries may work adequately well.  Remind me
sometime to tell you what we found out about MIT undergraduate
admissions -- decisions that have far more effect on people's
lives than this sort of thing.  It all depends on how much one
really cares, how much trouble one wants to go to, and the
ideology one holds about what is fair (or about what one wants
to optimize).  And, like just about everything else that has
statistical properties, the results are more satisfying when one
has thousands, or even hundreds, of voters (or whatever) in the
decision universe than when there are a handful.

     john



_______________________________________________
IMA mailing list
IMA at ietf.org
https://www1.ietf.org/mailman/listinfo/ima