[EM] The easiest method to 'tolerate'

[Kevin Venzke]

Hi Kristofer,

----- Mail original -----
De : Kristofer Munsterhjelm <km_elmet at t-online.de>
À : Kevin Venzke <stepjak at yahoo.fr>; EM list <election-methods at electorama.com>
Envoyé le : Vendredi 1 juillet 2016 10h48
Objet : Re: [EM] The easiest method to 'tolerate'

>> Unrelated: That link says Chris Benham invented MCA. I'm pretty sure
>> that's wrong, and that it was either Forest, or else it didn't have a
>> clear inventor. But I am struggling to find any archives for this list
>> older than 2015. Hopefully we have not managed to lose nearly 20 years
>> of posts...?
>
>http://permalink.gmane.org/gmane.politics.election-methods/184
>

>says it's Forest.

Ok, good that that is there.
>> Generally I don't think that much is based on culture when it comes 
>> to comparative politics. If something is at stake, one should look
>> and keep looking for the practical reasons why actors do what they
>> do.
>
>Yes, I understand that, and I would be inclined to believe that voters
>generally won't employ strategy in the kind of thresholded utility
>setting I mentioned in the previous post (independent of culture). I'm
>mostly mentioning culture because my inclination seems to be greatly at
>odds with, say, Ossipoff's. He focuses a lot on compromising strategy,
>as well, which seems (sort of) to be based on how prevalent
>lesser-of-two-evils voting is in the US. Also, Norway uses party list PR
>and I have the impression that tactical voting isn't really a thing here
>when talking about parties above the threshold, even though
>Plurality-based party list PR has a compromise incentive between seats.
>
>So I could be wrong about culture playing a part. But then some other
>explanation has to account for my observations, or I have to be wrong

>about some of them as well.

My inclination would be to ask whether the stakes are high enough. A party list's share of the vote is the main thing that will have any consequence, isn't it?

(I recognize that "stakes aren't high enough" would be an argument in favor of tolerating some vulnerability.)
>
>If, on the other hand, voters generally do strategize in threshold
>scenarios where there's nothing to lose or gain until a certain level of
>strategy, then we're in rather more trouble. One could easily argue that
>Condorcet would lead to inferior outcomes because voters would betray
>their favorites in that manner. See e.g.
>http://rangevoting.org/IRVStratPf.html :


Yes but I think there's a qualitative difference between Condorcet's compromise incentive and the way that individual ratings under median/mean rating mean nothing to the would-be strategist. Under Condorcet your relative rankings at least get counted. Under median rating you're just trying to pull numbers up or down, and you probably will never have adequate information to say e.g. that the rating you need a candidate to hit is 7/10.

>>> It's always advantageous if the method also makes sense based on how it
>>> works, but I'd rather have an opaque method with good compliance than
>>> vice versa. Borda is very simple, but its extreme teaming incentive
>>> makes it of little use in competitive elections.
>> 
>> I feel like the value comes not just from understanding how it works
>> but being comfortable with how the method behaves as an agent for the
>> voter. I've been wanting to write a post for several months on a few
>> topics but I haven't found an excuse or even an overarching idea. This
>> is one piece though. It's common to think of the method as an agent
>> representing the entire electorate, which figures out the best possible
>> result for everybody. But we could also try to imagine a given method's
>> rules in terms of how we would define an agent representing an
>> individual voter, such that the interactions of the agents produce the
>> method's results. Think of the agents as legislators talking among
>> themselves before making a final vote.
>
>I see what you mean. IRV fits well as an individual DSV method. It's
>much harder to do that with Condorcet because Condorcet is "global". I
>imagine you could make a sort of probabilistic variant where you vote on
>candidates in turn, legislature style, but it wouldn't be strategically
>resistant.
>
>E.g. something like
>
>1000 times:
>- order the candidates in random order
>- ask "do you prefer X or Y" where X is the first candidate and Y is the
>second
>- whoever wins is then compared to the third in line, and so on.
>
>Pick the winner who won most often.
>
>Then each agent participates by voting for the proper candidate in each
>proposal, and the agenda is set to be random so that nobody has an
>unfair advantage. But this would most likely reduce to Copeland and so

>be full of ties.

I see the eliminations in IRV (in this DSV conception) as representing that the agent has permission to give up on a candidate and propose the next compromise. I think we can interpret that to exist in some sense in Condorcet methods (like when MAM locks defeats) but it fails somewhat for me that a 90:10 A>B win is taken as lack of viability for B without any comment intended on the viability of A. That seems different from IRV, and prone to mischief.

I think a method with constantly adjusting approval cutoffs would probably satisfy Condorcet and also make intuitive sense. Not a very novel idea I know.
>Another option is BTR-IRV where the two last by Plurality count get an
>elimination round and the pairwise loser is eliminated.
>
>Bucklin could be considered a method where each voter has a reluctance
>to voice his support for candidate X, and then as time passes, he adds

>inn support he's more reluctant to voice until someone gets a majority.

Yes Bucklin works really well in this context. I probably prefer Bucklin to median rating purely because it (and its shortcomings) feel more clear.
>I suppose that Borda-elimination doesn't feel as well-suited as an
>individual DSV method because it's not clear why one is using Borda. If
>IRV proponents are coming from Plurality, it makes sense to use
>Plurality logic, even though from a logical perspective, it is not
>altogether clear why Plurality should be good at finding losers if it's

>bad at finding winners.

Actually I think Borda-elimination is fine in this context, and Borda alone is a fine estimate of strength. (That doesn't mean some of the side-effects are tolerable.) I feel similarly about Plurality: I don't think that an IRV advocate should say that Plurality is bad at finding winners but good at finding losers. I think he would want to say that Plurality is a meaningful measure and it's improved by taking the measure more than once.

>It's too bad we can't just search over the whole space of
>"natural-sounding" rules and find the best one. What does or doesn't
>constitute a natural-sounding or pleasing rule is not very clearly
>defined, though I do think you're onto something with the DSV explanation.
>
>> There are obvious reasons why we wouldn't want to just use first
>> preferences to tell us about apparent viability. But the first
>> preferences are the ones we can definitely trust. If we were to accept
>> that as important (even if just for sake of argument) I wonder if it
>> would lead somewhere.
>
>My brute force attempt to find strategically resistant Condorcet rules
>for three-candidate elections did find something that used first
>preferences.
>
>If there's no cycle, elect the CW. Otherwise, let A>B>C>A and call the
>candidate we're considering A without loss of generality. Then A's score
>is fpA - fpC. The candidate with highest score wins.
>
>First preferences can be important since you can't bury with them.

>They're not all that important under honesty, however.

Hmm I didn't remember the exact rule you found. I think the rule I just suggested is more like 0 - fpC. I can't see a "philosophical" reason not to bring in fpA though. And both methods elect B in my example.

Sort of wondering whether I could come up with an iterating method that would pick B, or if this burial idea is a completely separate thing. I feel like it is: i.e. if you just use a constantly adjusting approval cutoff (based on tentative winners) then there isn't a natural means of discovering which preferences might be bogus.

--

Totally unrelated but one thing I think about is two-round methods and whether we can do anything interesting with them, by "sharpening the questions" for voters. Here's a method that doesn't quite work for me, but it's really close:

1: Have a first round vote. Ideally this doesn't collect full rankings but is still a decent estimate at finding the best candidate. Approval is probably best. Call the winner X.
2: Don't eliminate anybody. Have a second round with an approval ballot and all voters do is vote for candidates they think are better than X. Alternatively they can check a box saying nobody is better than X. But X does not themselves receive any votes.
3: If anybody gets majority approval, the candidate with the most approval wins. Otherwise X wins. Effectively, X has an assumed approval score a bit less than the barest (mathematically possible) majority.

The (obvious) point over some other methods is that it's trying a little harder to find a majority. And more importantly, trying to make sure the winner wouldn't be voted down by any majority. It's also burial-proof (if you don't count first-round push-over as burial). We (try to) avoid majority cycle controversy by not collecting a full pairwise matrix.

Downsides are: by electing a candidate who beats the tentative winner, one can perceive a disadvantage to being the tentative winner, even if we don't know exactly what the sincere rankings were: A monotonicity issue. And secondly, I'm concerned that in sufficiently simple scenarios it could frequently happen that the second round just reverses the outcome of the first round, due to variations in turnout.

(I'd also say I'm disappointed that the second round's purpose could be defeated by chicken dilemma issues, and that even the second round considered independently can't be said to satisfy FBC.)

Kevin