[EM] [CES #8848] Re: MAV on electowiki

[Abd ul-Rahman Lomax]

At 12:33 PM 6/19/2013, Jameson Quinn wrote:
>My general response to Abd: a number of good points, and a number 
>where I'd disagree. Tough to respond to a kitchen-sink list, so I'll 
>try to prioritize. Unfortunately that means that things I don't 
>respond to could be either "you're right" or "I disagree but don't 
>think it's a productive argument".

Aw, we are building something. Some pieces may be incomplete, that's 
to be expected.

To keep this all in context, this is a discussion of an article:
http://wiki.electorama.com/w/index.php?title=Majority_Approval_Voting&oldid=10169

Current link (has been changed): 
http://wiki.electorama.com/wiki/Majority_Approval_Voting

I see and acknowledge that Jameson has already responded to certain 
comments of mine by editing the article accordingly. Now, to what he wrote:


>2013/6/19 Abd ul-Rahman Lomax 
><<mailto:abd at lomaxdesign.com>abd at lomaxdesign.com>
>...
>
>If one or more candidate has a majority, then the highest majority wins.
>
>
>Arrghh. That is ordinary Bucklin.
>
>
>Just for the first rank, because at that point there's no higher 
>rank to fall back to.

Ah. The full text:

>Voters rate each candidate into one of a predefined set of ratings 
>or grades, such as the letter grades "A", "B", "C", "D", and "F". As 
>with any Bucklin system, first the top-grade ("A") votes for each 
>candidate are counted as approvals. If one or more candidate has a 
>majority, then the highest majority wins.

Yes, it's stated that this is related to the first rank, but, even 
knowing the process, I misread it. Maybe I misread it because I know 
the process....

There is extra language here, and missing language that would make 
the matter clearer. For example, "predefined" is unnecessary, voters 
would not even think of being able to define the categories. So they 
will assume "predefined," and it need not be stated. The neutral 
concept is "categories." That also, by the way, fits with what Dr. 
Arrow has been saying of late.

Further, multiple majorities in the first round is exceedingly 
unlikely. Thus we are leading with what would be a rare exception. 
That's to be avoided.

I suggesting first describing the basic Bucklin amalgamation. This 
*is* a Bucklin method, and has a basic method that is followed. The 
case of multiple majorities is an exception, handled as such. It will 
almost always happen at a lower rank than first rank. First rank was 
called, "first choice," which is decent language.

I'm narrowing down the method, here. I have, for a first proposal, 
eliminated the unapproved category ("D") entirely. It simplifies the 
explanation. I say:

----
Majority Approval Voting (MAV) is an evaluative version of Bucklin 
voting. Voters rate or vote for each candidate, using a set of 
preference categories, from "first choice" down to a lowest-choice 
that is expressed by leaving the vote for the candidate blank. All 
votes excplicitly cast are forms of approval and may directly cause 
the election of the approved candidate.

Voters may place as many candidates as they choose into each category.

The categories are canvassed, being summed as approvals, and as 
needed, through the first rank to the third rank. This process ceases 
when any candidate has accumulated approvals from more than half of 
the voters. If there is only one such candidate, this candidate wins. 
If there is no such candidate, or if there are two or more such 
candidates, the process enters "tiebreak."

If there is no majority-approved candidate, deterministic MAV is a 
plurality method, it can elect without an explicit majority. The 
ranks are evaluated as scores of 4, 3, 2, 1 (and 0 for blanks). The 
scores are summed and the candidate with the highest score wins.

If there are two or more candidates with a majority, the 
majority-approved candidate with the highest sum of scores is elected.
----

This is a compromise between the back-up used in the first MAV 
proposal, and the highest vote standard that would be ordinary 
Bucklin. This causes lower-ranked votes to be devalued, down to a 
half-vote, when there is a majority conflict. It leaves the number of 
expressed ranks as three, keeping it simple. This could be improved 
at the cost of a lower rank: the D rank, which would only be used for 
tiebreak. This method introduces, more explicitly, Score Voting.

What made me uncomfortable about the "backup" method of breaking a 
tie in median vote was the effective disregard of votes cast in the 
multiple-majority producing round.

There is a possible legal problem when votes are cast that are not 
counted. In some places, the pure Bucklin method might be *legally 
required,* because, as in Arizona, the constitution may require a 
preponderance of legal votes, and we would be hard-pressed to claim 
that lower-ranked votes were not legal. Indeed, this may *require 
Range.* Or a runoff system.

>At some point it must be specified how "rank overvotes" are handled. 
>There are possible ways:
>
>1. Ballot is voided. We don't like that!
>2. Vote for candidate is voided. Also that is not great.
>3. Highest rating marked counts. (This may have been traditional 
>Bucklin. The instructions said "don't do it.")
>4. Lowest rating marked counts.
>5. Top and bottom ratings are averaged. (So a rating of A and B 
>would give a rating that could be called A- or B+. GPA contribution 3.5.)
>
>
>6. Vote counts for forming a majority at top rating, but doesn't 
>count as "higher than" until you pass bottom rating. This is 
>actually strategically attractive (if you were in a chicken dilemma, 
>you could give the other side an honest B, but also a 
>chicken-conservative D or even F, and that would basically be saying 
>"I think playing chicken is silly but if it turns out I'm in the 
>minority then game on"), in keeping with the spirit of the method, 
>and in effect very similar to 5.

I don't understand this. I'd like those counting the votes to be able 
to determine the vote, period, right then, no questions, no 
"conditions." Any vote-counting that depends on conditions can then 
fail precinct-summability. Not good. Interpreting the vote as 
intermediate gives it a meaning that *averages* out to something 
close to what an erring voter intended, and that, done deliberately, 
increases voter flexibility.

This is a minor detail, but introducing it from the beginning, in 
footnotes (which would be a sentence in a code implementation), 
establishes something that could be quite valuable later. How to 
increase the resolution of a Range ballot without increasing the 
number of categories. And the whole MAV/backup point is a concern 
made significant by "rapid collapse of approvals" caused by lack of 
refinement in expression.


>But I think that pre-specifying this is premature. These are 
>implementation details, not part of the method.

That's correct. But it should be a part of any core document on MAV. 
Not leading, it's a detail.


>I like the fifth option because it actually would allow voters to 
>give an intermediate rating, thus providing some additional range 
>resolution with no additional ballot complexity, but some cost in 
>canvassing. That additional cost would be small, because most voters 
>would not use it, most voters not needing it.
>
>In amalgamation, then, as to rank, 3.5 would be added in after A but before B.
>
>This would convert a Range 4 ballot to Range 8, with the penultimate 
>bottom rating being missing, unless the bottom rating is explicit.
>
>
>If not, votes at next grade down ("B") are added to each candidate's 
>approval scores. If there are one or more candidates with a 
>majority, the winner is whichever of those had more votes at higher 
>grades (the previous stage). If there were no majorities, then the 
>next grade down ("C") is added and the process repeats; and so on.
>
>Note that if this process continues without a majority until the 
>last grade ("F") is added, no new rules are needed.
>
>
>This is technically correct, because the lowest grade amalgamation 
>*must* show 100%. This assumes an F default, so marking of F is 
>*irrelevant* except as a confirmation.
>
>If Fs are amalgamated with blanks not being treated as F, then F 
>becomes a vote for the bottom-rating candidate. We don't want to 
>create that confusion for a moment, even.
>
>
>I agree that I should be more explicit about F default.
>
>For simplicity, dump the F rating entirely. It is assumed, an 
>unexpressed default rank.
>
>
>I strongly disagree that there should be no explicit F. As redundant 
>as it seems to you, there are definitely voters who get satisfaction 
>from explicitly voting against someone, and I would never design a 
>survey where an explicit answer was always considered to connote 
>more support than an implicit "no".

They should know that it has no effect except as a slight difference 
in how the votes might be reported. Voters are *totally accustomed to 
not-voting for a candidate being rejection.* They will be even more 
experienced with this if they first have seen the simplest Approval 
implementations, i.e, Count All the Votes.


>MAV as a method should not specifically depend on the number of ratings.
>
>
>Absolutely agree. I'll make that even more explicit.

I'm now disagreeing somewhat with myself. We should have a Simple MAV 
method that is maximally simple. The challenge is to make this *also* 
maximally effective, to the extent possible with simpler rules and a 
simpler ballot.

>
>MAV outside of a runoff system is a plurality method (as are all 
>voting systems that can complete without an *explicit* approval of a 
>majority of voters. In treating a midrange vote ("C", 2/4) as an 
>approval, we have already stretched *slightly* beyond that. Taking 
>this down more deeply into D/F waters is pretending a majority that 
>doesn't exist.
>
>
>OK, but in a non-runoff method, sometimes you'll elect without 
>majority support, and it's still fair to say that D is better than F.
>
>I can see more complex amalgamation rules that could use the D 
>ratings. The simplest rule, attempted here, doesn't cut it.
>
>
>Any of those more complex rules is also trying to infer a majority 
>from a plurality. They might sometimes do a slightly better job of 
>that, but my feeling is that it's a false hope to think they'll 
>always succeed, and so KISS.

We would aim the method to *usually succeed.* Indeed, plurality 
methods do exactly this, and with nonpartisan elections, plurality 
*almost always* predicts the ultimate winner, were fuller ballots to 
be cast. And that is an argument, Jameson, for using preponderance of 
the votes. I.e, straight Bucklin. But there are also partisan 
elections. It may be that we will design a series of Simple MAV 
versions, for different environments.

So we'd have Nonpartisan Simple MAV and Partisan Simple MAV, as examples.



>  Since by that point all grades will have been counted, all 
> candidate tallies will reach 100%.
>
>
>In other words, amalgamating the F ratings is *useless*. This treats 
>F as if it were a majority, it's a multiple majority, so 
>*unconditionally* the previous ratings will be used and the F 
>amalgamation will be ignored. So why do we even count it?
>
>Just to make the system seem uniform all the way down?
>
>As I've mentioned, the system should exist in two forms: plurality 
>result, or runoff feeder (which, then, is also in two forms: 
>conditional runoff and unconditional runoff. Unconditional runoff 
>may be better handled with a different system, an explicitly 
>multiwinner one, which has not yet been examined in detail.)
>
>
>  The process above then naturally elects the candidate with the 
> most approvals at the higher grades (D or above); that is, 
> whichever has the fewest F's.
>
>
>And I would say, fewest disapprovals, and disapproval includes D and 
>F. It also includes blanks.
>
>
>Great. Make your own, slightly more-complex system, and give it a 
>different name.

Nope. I'll make a less-complex system, and give it *this* name.

>
>
>
>  This is the best way to resolve such an election using only the 
> information on the given ballots.
>
>
>"Best" is a red flag word when it has not been clearly defined. 
>You'd never be allowed to say "best" on Wikipedia, without that, and 
>maybe even with it. That's called a "peacock word."
>
>
>OK. A way.

Yes.

>
>
>Why would MAV be any "better" than Bucklin?
>
>
>Wait a minute. In this particular case, MAV and "Bucklin" (you mean, 
>Grand Junction) do exactly the same thing: use second-to-bottom rank 
>if there are no majorities.

But with Bucklin, quite clearly, the second-to-bottom rank is an 
approved rank. In the system design just set up, by Jameson, the 
second-to-bottom rank is a disapproved rank, but then, is it?


> From the information on the ballots, the plurality winner in the 
> majority-finding stage is arguably the "best winner." No, we 
> support MAV because of the effect on strategic voting. If we have a 
> Bucklin system with greater ballot resolution, MAV backup may be 
> unnecessary. If voters are more informed, it may be unnecessary.
>
>
>You seem to be under the impression that there's a difference 
>between MAV and "Bucklin" in this case.

If I was, I must have misplaced it. I'm discussing the *entire 
method*, not just the one case. "Plurality winner" as I used the 
term, *includes the multiple majority case.* Plurality does also have 
the meaning of "the most votes."

>
>
>I am not satisfied that we have adequately examined the issues, to, 
>yet, proclaim MAV as the "best" system up from pure Approval.
>
>
>Nope. I am not sure it's "best", and don't really care. I know that 
>the differences between Bucklin methods are tiny, yet we'll make no 
>progress until we settle on one, simple method.

Let's make it maximally simple. We are complicating Bucklin, there 
had better be *strong reason* for that. Bucklin, remember, was *very 
popular.* That is what is obvious from the history. It was killed 
*politically*. In disregard of what the voters actually wanted. 
Bucklin also may have been oversold. That's obvious from the history as well.

>
>
>This point should be clearly made, though: MAV is Bucklin, with an 
>exception only appearing in the handling of multiple majorities. 
>Those will tend to be rare, because, more common, we can expect in 
>some kinds of elections, will be majority failure.
>
>How majority failure is handled is a classic problem. Robert's 
>Rules' recommendation to organizations: Don't Do It! *Never* elect 
>with a mere plurality. The version of IRV that they described 
>*requires* a true majority (and even then, they point out the center 
>squeeze problem.)
>
>The backup concept, treating multiple majorities as possibly 
>indicating over-rapid addition of approvals, is interesting, but 
>looking at this is causing me to warm to a graduated median 
>method.... plus pushing toward finer resolution, which will reduce 
>the number of multiple majorities. *None of this is going to be 
>fully satisfactory, because making a decision with a mere plurality 
>is intrinsically problematic.*
>
>
>Right. And the differences are slim. So let's keep debating these 
>technical points among ourselves, but use common language in our 
>activism. That's what this is about, principally.

Sure. So don't rush judgment, or any consensus won't be real.

>
>
>What I *don't* like with MAV is that a large majority can be ignored 
>vs a bare majority, based on the votes at the higher rating. 
>Graduated median amalgamation can balance this.
>
>
>Right. For honest votes, GMJ is clearly superior. See above for why 
>I don't care.

But we can do better than GMJ, and it might be quite simple. That's 
what I've come to with this discussion, so far. GMJ, essentially, 
attempts to simulate a more graduate introduction of approvals. Yet 
we have range data, which is *superior*. That is, more likely to 
reflect social utility.

Looking at this, keeping the votes at full value creates the problem 
of too-rapid introduction of approvals. Deprecating the votes to 
their utility expression, what the voters actually voted, in setting 
up their instant-runoff approval robot, their ballot, *partially* 
compensates for possible overenthusiastic introduction of additional 
approvals. It also can serve as a first introduction to range amalgamation.

>
>But I have not seen a detailed examination of this issue, and it's important.
>
>
>No, it isn't.

Then, Jameson, you are welcome to sit out the discussion.

>
>The MAV backup may be just a bit *too* simple. There is an 
>alternative, as well, that would pick the candidate with the highest 
>GPA. That would consider *all* the ratings. It's theoretically 
>superior, because it would be Range. The D ratings would count.
>
>Another approach would use pairwise comparison. We *do* have 
>explicit approvals of a majority, if the median vote is C or higher. 
>So, then the ballots would be recanvassed to see how the 
>majority-approved candidates fare against each other. The collapsed 
>approvals are separate again. This would make the method 
>Condorcet-compliant, if I'm correct. And that would be a plus. I'm 
>on board the concept that the Condorcet criterion is not absolute, 
>because of social utility, but without explicit majority 
>confirmation, I dislike finding *against* the preference of a majority.
>
>
>  However, in this and other cases of multiple majorities, a runoff, 
> if feasible, would be a better way to ensure a clean majority win.
>
>
>It's a *way*. Not simply a "better" one. The alternative presented 
>here is *not* a way to "ensure a clean majority win."
>
>
>In my book, not putting your hand in a band saw is a "better" way 
>not to cut it off. That is, "a way" is better than "not a way".

Don't do that.




>This system was promoted and named due to the confusing array of 
>Bucklin and Median proposals.
>
>
>Premature. It has not been promoted. It has been *tentatively* 
>named. This, by the way, would be in a History section of an 
>article, and it only begins to tell the history.
>
>  It is intended to be a relatively generic, simple Bucklin option 
> with good resistance to the 
> <<http://wiki.electorama.com/wiki/Chicken_dilemma>http://wiki.electorama.com/wiki/Chicken_dilemma>chicken 
> dilemma. It was named by a 
> <<http://lists.electorama.com/pipermail/election-methods-electorama.com/2013-June/031938.html>http://lists.electorama.com/pipermail/election-methods-electorama.com/2013-June/031938.html>poll 
> on the electorama mailing list in June 2013.
>
>
>Geez, *recentism* to the 9s.
>
>You know what I think about the "chicken dilemma." (Voting is not a 
>game of chicken, and "dilemma" is merely an ordinary problem in 
>strategy, which only afflicts a small segment of the population, the 
>rest will simply *vote.*) What MAV will do is to, I expect, 
>*slightly increase* the number of additional approvals, because they 
>become *safer*, i.e., less likely to cause the loss of the favorite 
>to the lower ranked candidate.
>
>I prefer, Jameson, that the issue of the exact definition of MAV be 
>*left open* for the time being. Let "Majority Approval Voting" 
>actually be Supermajority Approval Voting," i.e, ultimately approved 
>by a supermajority of voting systems activists.
>
>Not as the "ideal voting system," which it is not, but as the "ideal 
>next step beyond ordinary Approval, in a plurality context." It 
>introduces a ranked/rated ballot. It's simple to amalgamate and 
>understand. In most elections, the "tiebreaker" procedure is not 
>activated. Indeed, in most elections under some fairly common 
>conditions, there will be majority *failure*, not the multiple 
>majorities that the backup procedure handles.
>
>The grades or ranks for this system could be numbers instead of 
>letter grades. Terms such as "graded MAV" or "rated MAV" can be used 
>to distinguish these possibilities if necessary. In either case, 
>descriptive labels for the ratings or grades are recommended. For 
>instance, for the letter grades:
>    * A: Unconditional support
>    * B: Support if there are no other majorities above "C"
>    * C: Support if there are no other majorities above "D"
>    * D: Oppose unless there are no other majorities at all.
>    * F: Unconditional opposition.
>
>This treats D as an approval, so it divides the approved categories 
>into four rather than the three from Bucklin. We will need, 
>ultimately, a single, coherent, simple system to propose and try. 
>Notice that the definitions treat not-oppose as support. In my 
>analysis of the traditional Bucklin ballot, it's clear that the 3rd 
>rank was *minimum approval.* The ordinary meaning of a grade of "D" 
>is a kind of failure.
>
>
>In most schools, you can pass with a "D".

If you don't get too many. Want to show that report card to 
prospective employers? Cs are bad enough.

>
>
>This is a *really complicated* explanation of the rating levels.
>
>
>It's also accurate.

Somewhat. Better than putting your hand in a band saw.

>
>
>As the above labels indicate, support at the middle grades or 
>ratings is not partial, as in 
><<http://wiki.electorama.com/wiki/Score_voting>http://wiki.electorama.com/wiki/Score_voting>Score 
>voting, but conditional.
>
>
>That is correct. If the method terminates above the rating, it is 
>not used. That, by the way, is a flaw, from a social utility point 
>of view. I don't like that voters may be casting votes that are not 
>counted. I'm starting to think toward lines of using range 
>amalgamation in the "tiebreaker." And of suggesting that all votes 
>be counted, even if only for reporting. Incubator effect.
>
>
>Count and report all grades: of course.
>Votes counting fully is a flaw: It gives less social utility for 
>honest votes, but by doing less to encourage exaggerated votes, it 
>could result in better information and thus better social utility overall.

There is a myth out there, called "exaggerated votes." No, voters 
vote their preferences, and "exaggeration" is an oxymoron. It's just 
a mark on a piece of paper, perhaps. If the voter starts screaming at 
the voting place, "You *have to* vote for X and against everyone 
else," so that they cart him off, that would be, perhaps, an 
exaggeration. Or would it? Presumably, that was the way he felt!

>  That is the basic premise of median versus average.

And that is exactly why median can fail to maximize utility. It 
suppresses much voter information. The particular median called 
"majority approval," however, has a *different value,* having to do 
with fundamental democratic process.

>  Like it or not, that is what MAV gives you. And MAV can do a lot 
> better at giving that than "maybe we should do this, or maybe that, 
> no wait I just thought of another thing..."

A few weeks discussion, Jameson, is not this. Hey, if you think we 
are done, you are free to go ahead with whatever. Just without my 
support, perhaps, and that may or may not mean anything. I might even 
support what you do. But do consider the value of *full consensus*. 
It can be well worth seeking.

And by "full consensus," I don't mean that most people simply got 
tired of discussing it and stood aside. Listlessly.

>
>
>
>  That is, the typical ballot will still count fully for or against 
> a given candidate.
>
>
>Yeah. But I want to see a much more detailed analysis of this.
>
>
>The different grade levels are a way to help the voting system 
>figure out how far to extend that support so that some candidate 
>gets a majority.
>
>
>Uh, really? It's more like "so that majority support is found if 
>possible from the votes."
>
>
>OK, sure.

By the way, my objection was to the implication that a majority would 
be found. Obviously, by extending amalgamation down to the bottom 
rating, we could find, not merely a majority, but *unanimity.* And it 
would be meaningless.

>
>And a D vote is *not* "support." It might be a weak stand-aside. 
>"Okay, if the rest of you insist, this one is better than that one." 
>A strong stand-aside would be a C vote. In consider C as neutral, as 
>to preference strength. So-so. Not good, not bad. "Passing," but 
>barely. No honors. But ... might be serviceable.
>
>
>For a strategic voter, the most important ratings are the top ("A"), 
>second-to-bottom ("D"), and bottom ("F").
>
>
>"Most important" is, again, Peacock. Here, Jameson, you are giving 
>your own analysis, not the community's analysis.
>
>
>OK. Section name: "Preliminary and tentative strategic analysis".


And attributed, and others may make contributions.

>
>This is not a signed article, even if your name is in the History. 
>If you want to present your own analysis, *attribute it.* Or 
>attribute analysis to others. I suggest not presenting your own 
>opinions as if they were fact. Sometimes in writing Wikipedia 
>articles, I'd use "weasel words." This is weak language, like, "Some 
>say that ...." or "It could be claimed that ...." or "According to 
>some sources, ...." I really only did that to find quick consensus, 
>not to propose weasel language as stable or desirable. It was 
>typically replacing *strong language* that did not actually reflect consensus.
>
>
>A typical zero-knowledge strategy would be to give the best 30% of 
>candidates an "A", the next 25% a "D", and the bottom 45% an "F".
>
>
>"Typical" according to what standard? What population? I would *not* 
>think this way *at all.*
>
>
>"Strategic" means at a minimum "game-theoretic equilibrium". Your 
>suggestion is certainly not that, so it is not "strategic", it is "honest".

I've done the research, Jameson. Your comment is naive. Game 
theoretical analysis is often hampered by overspecification of the 
goals of the player. Honest voting in the zero-knowledge case can be, 
with some voter goals, the minimum you speak of. This has been 
examined in some detail, Jameson. Warren Smith even signed on to the 
article, and extended it a little.

>
>I would cast a Range ballot, period, using the same strategy, i.e., 
>considering what I know of the preferences of others. With Bucklin, 
>I can vote sincerely for my favorite under almost all conditions, 
>and the loss of voting power is miniscule. If it's a zero-knowledge 
>situation, I'd vote simple, normalized Range. I'm really not worried 
>about the multiple majority problem. My votes will represent *true 
>preference strength*. That only shifts with knowledge of the 
>electorate, where I will then vote Von Neumann - Morgenstern 
>utilities. I.e., to put it simply, which will powerfully choose 
>between realistic possibilities, reserving *some voting power* for 
>the expression of true preference, which is an independent value. 
>(And where partisan elections are involved, that has a long-term 
>effect. I am not just concerned about the *present election.*)
>
>
>Great analysis of why an honest vote is a good idea in this system.

Thanks. I came up with a result that suggested that an optimal range 
vote (i.e., "honest") was *not* strategically harmful. There could be 
*some harm* if the vote was not *accurate*, this was not studied. 
That could be fixed by *exaggerating* the vote to cancel out error, 
I'd suspect. It would not need to be exaggerated all the way to 
max/min range. But the argument I give above suggests a *different 
value* that could swamp the naive game theory of the situation. After 
all, I will have to live with myself and my vote and the 
consequences, whereas the *actual probablility* of my vote affecting 
the outcome, in public elections, is so small as to be ridiculous. 
Game theory often has been used to suggest that voting at all is a 
waste of time, effort, and money. And it would be, if not for these 
additional values.