[EM] The "Single Contest" method

[Kevin Venzke]

Hi Jameson,

--- En date de : Jeu 21.7.11, Jameson Quinn <jameson.quinn at gmail.com> a écrit :
>>I did an example on paper. In my example the "M" score was almost
>>always zero, which makes me doubt the M*U product will work well in
>>sims. Also, disfavoring a high M score has an obvious favorite
>>betrayal incentive which (at least in SC) is noticeable. 
>
>Can you explain further? I don't see how it has any non-semi-honest 
>incentive at all. Perhaps it does have a bullet incentive, though.

I have two guesses. Both are based on the fact that sometimes, if I
want two candidates I like to be the final pair, I will have to 
disapprove one or the other. I will probably disapprove the weaker one,
because 1) the stronger one is more likely to win as majority favorite,
and 2) the stronger one is more likely to win other possible pairings.
It is true that I will lose the ability to favor the weaker candidate
(if I do) in the runoff, but it may be a good trade in my view.

In any case, I think I see a real psychological annoyance, that one
can't approve two candidates and simultaneously vote for them to be the
finalists.

>>Maybe one could just look at U (like SC does).
>
>But then the threshold is always pushed down.

Hmm, that's a good point. That could be why my U-only automated method
didn't work too well: It turned into approval.

So, I guess I should try using both U and M at the same time. But
without turning it into the equivalent of D.

>How about maximizing D(A) * (D(A) + D(B))? Or D(A) * (D(A) + D(B) + 
>M), which is the same as D(A) * (V-U)?

I'm a bit unclear, didn't you define D as a two-dimensional variable?


Second post:

>Yes, I expected that the sim voters will never be able to trust 
>automation, that for any rated method they'd tend to either compromise
>or truncate excessively.

Well, let's be clear: There's no such thing as "trust." I'm joking. We 
may have a different good reason for the truncation, since I was only 
using U. As you say, this will favor a threshold that gives high 
approval scores. The method will raid lower preferences trying to get
this. The voters obviously will want to defend themselves, whether they
are AI or humans.

Not just "automation" but even "ratings" as opposed to "rankings"
is quite hard to define.

>But it does give me an idea. If you were to use this method in real 
>life, you could have ratings ballots and an empirically-chosen fixed
>threshold. It's equivalent to your ranked version, strategically, but
>gives more info and takes less effort from voters (if B+L are to be
>believed).

You could. Actually, I can't prove that my rank version isn't actually
performed on a ratings ballot. I just assumed it wouldn't be done that
way.

Kevin Venzke