[EM] FairVote comment on Burlington dumping IRV

[Abd ul-Rahman Lomax]

At 04:17 PM 7/3/2013, Jameson Quinn wrote:


>2013/7/3 Abd ul-Rahman Lomax <<mailto:abd at lomaxdesign.com>abd at lomaxdesign.com>
>
>EMAV is not a known method, it's brand new, I just announced it, and 
>this is a general post, not about details about the specific method.
>
>
>I find it encouraging that you make a distinction between which 
>methods you'd mention on a "general" and a "details" post. My 
>current push with MAV is intended for "general" posts; I'm not 
>trying to suppress "details" discussion.

No problem with you being encouraged, but ... "general post" really 
mean that this wasn't about Bucklin or EMAV or MAV or any other 
specific method, and I simply mentioned "Bucklin" as a possibility 
for Burlington in a discussion about FairVote, and that name was used 
there for the reasons I stated: FairVote has created much content 
about "Bucklin." Mostly misleading.




>Q2. Is there anything that would convince you to switch to saying "MAV"?
>
>
>Not in that context, not yet.
>
>
>You gave a qualified response, when I wanted an unqualified one. 
>What would it take in some other context, or later? A million dollars?

That might do it. How many times do I need to say it? There are 
limits, you know.

>BR data showing it's better than EMAV for some voter model?

Sure, if the voter model was reasonably realistic. By the way, EMAV 
produces the same results as MAV and ordinary Bucklin under many 
circumstances, and when it doesn't, it's Range, which has superior 
BR. So I would not hold my breath for those BR statistics showing 
what you propose.

The caveat: I'm suspecting from what is below that Jemeson has some 
models of voter behavior that I find quite strange. Some aspects of 
this do reflect common opinion among voting systems activists, 
particularly the idea that Score strategy requires extreme ratings. 
That has never been shown, and neglects the complexity of voter 
motivations, i.e., the *values* that drive strategy.

Naive game theory is well-known to come up with paradoxes, when real 
people have no trouble with supposedly difficult choices.

>Survey data showing voters love it?

Sure. Real voters in real elections or imaginary voters in imaginary 
elections? Real voters participating in an imaginary election? I have 
suggested in the past that we do focus groups on certain issues 
before committing organizationally to some position or strategy.

This is the CES, so "focus groups" are not about deciding what is 
"best," from the point of view of voting system behavior, but rather 
what *communicates* to voters and what appeals to them. What do they 
need to know? What do they want? What do they stand for?

>
>
>
>Comment: To me, "Bucklin" is not a system, but a class of systems; 
>at a minimum, it would include all different Progressive-era systems 
>which were called "Bucklin" at the time, but to me, it includes all 
>descending-approval-threshold-until-majority systems (including for 
>instance MJ, GMJ, MCA, and MAV.)
>
>
>My comments were referring to the class of systems, but also 
>specifically to Bucklin -- which primarily means those early systems 
>-- and FairVote propaganda was about "Bucklin."
>
>Of course I'd love to promote EMAV, but promotion is not my primary goal.
>
>The subject post was written to review Richie's response to the 
>Burlington debacle, and traditional Bucklin -- say, three-rank, 
>mandatory single votes in first and second rank, almost exactly the 
>same as some of the old implementations -- would have fixed the 
>Burlington problem easily. That does *not* mean that this would be ideal.
>
>As to MAV, I'd support it if the "regression" were *necessary.* I 
>don't see it as that, and the fallback to higher preferences clearly 
>moves away from maximizing expressed utilities.
>
>
>1. MAV's "regression" is no more artificial than original-Bucklin's 
>"inclusion".

That's bizarre. Bucklin simply counts the votes in each rank, seeking 
a majority, and adds them in sequence. When it finds a majority, it 
completes, candidate with the most votes wins. MAV does not use the 
last-amalgamated vote, but backs up ("regresses") to the previous 
count. It has the data, then ignores it, using it *only* to 
distiguish a candidate as approved.

Bucklin is "instant runoff approval." It simulates a series of 
approval elections with declining approval cutoff. It's extremely 
easy to understand this, and multiple approvals always existed as a 
possiblity in Approval, and they actually happened with Bucklin. 
Nobody challenged that has being improper in any way. That fits with 
tradition with respect to multiple ballot questions, the common usage 
where multiple approvals are possible. The candidate with the most 
votes wins. That's consident with the Arizona constitution. (The only 
question will be about those uncounted votes at lower preferences, 
where they exist.)

It is simpler to explain ordinary Bucklin, because the winner rule 
does not change. Most votes.

MAV requires "backing up."

Now, from the point of view of utility expression, there is something 
very different going on. A Bucklin ballot is *not* an approval 
ballot, where all approvals are equal in representation of utility. 
It's obvious that the higher ranks are more preferred, that the voter 
will have higher satisfaction with the election of the higher-ranked candidate.

Basic Bucklin amalgamation, used in all three methods until a 
majority is found or multiple majorities are found, suppresses that 
preference information. The ballot is obviously a range ballot, if it 
is ER allowed in all ranks, and ranks can be skipped. So range data 
is available.

MAV does not use the full range data, qua range.

In many elections, regular Bucklin, MAV, and EMAV will produce the 
same result. The difference appears in two places: if there is a 
multiple majority, and if there is no majority. The first can be 
expected to be relatively rare in a mature system. However, there is 
a fear of it, that shows up in LnH fear.

Ordinary Bucklin maximizes this, because the votes collapse to equal value.
MAV minimizes it, because a multiple majority is then backed up to 
consider only the higher preferences, so the voter's lower preference 
vote that created a multiple majority is taken out and the voter's 
higher preference stands.
EMAV treats the situation with balance. It's very much like Range 
voting, and I'd expect EMAV strategy to be very similar.

Bucklin, in general, was not designed to maximize utility. It was 
designed to be majority-seeking. EMAV keeps that, and continues to 
respect the right of decision by a majority, but then uses range 
amalgamation, known to be superior as to BR, where there are multiple 
majorities or majority failure.

Voters have a choice whether or not to add those lower preferences. 
How they do that, quite sure, will be correlated with their actual 
preference strengths. Sane Bucklin strategy would involve voting a 
sane Range ballot, essentially.

>  For instance, any Bucklin method could be described using 
> antiapproval-style bottom-up vote summing, in which case 
> original-Bucklin would be the one with a "regression" and MAV would 
> be a straightforward count.

Try it and see if you can describe it simply. "Bottom up" does not 
describe a series of elections, each one depending on the previous 
one. Your process is essentially backing up to create the past from 
the future. You are stretching for proof, Jameson.

>2. By the same token, it's also no further from maximizing expressed 
>utilities than original-Bucklin.

Ah, MAV probably errs in the opposite direction. That is it 
under-utilizes the approval data from the multiple majority rank, 
while MAV over-utilizes it.

Jameson, I'm not getting that you actually understand EMAV. You may 
understand this or that about it, but not the historic and 
utility-sum arguments.

Maybe it will help if I mention that I've long considered Bucklin as 
a stepping-stone to Range, that I'd see Bucklin sequential 
amalgamation eventually disappearing, replaced with something simpler.


>3. The purpose of the MAV rule is that it is strategically 
>relatively "safe" to add a second-to-bottom rating.

That's correct. Without "backup," the danger is that the first 
preference is suppressed, made equal with the lower preference. MAV 
strips the second preference. EMAV gives the lower preference a 1/2 
or 1/4 vote (if the higher preference was at top rating).

The equal vote only happens if no other majority is found, if the 
lower preference has a majority and nobody else does at that rank.

But I actually don't like this. I'd rather use the pure range data 
from a range ballot. I rather do this in a runoff system, or what 
might be more likely to happen, with nonpartisan elections, an open 
range primary with top two plus Condorcet (if exists) nomination for 
the general election, followed by a range ballot for the general election.

EMAV is an intermediate method. It straddles the majority and range worlds.

>The only way it could cause preferred candidate to lose is if it 
>caused a multiple majority AND the other candidate had more higher 
>ratings. This is both an unlikely circumstance, as it would mean 
>that the other candidate's supporters were being significantly less 
>strategic than your faction; and, in the rare cases where it does 
>occur, acceptable from a utilitarian point of view, as in this case 
>it's probably the other candidate who's the utility maximizer. It's 
>also perfectly clear from an expressive point of view; that is, in 
>the rare cases where doing so was strategically suboptimal, it will 
>be easy to "read" the likely result of a repeated election with 
>strategic voters on all sides, and to choose whether that is 
>preferable than the weak cooperation.

We need to study some examples.


>I understand that it was frustrating for you that I appeared to 
>support MAV, for a short time, but I think that we were pandering to 
>some shallow arguments, that we don't need to avoid the "chicken 
>dilemma," and that using the range ratings adequately addresses the concern.
>
>
>What evidence would convince you that you are wrong about the chicken dilemma?

For me to answer that, I need a statement that is wrong. What is it? 
I've written a lot about the chicken dilemma. What, specifically, is 
allegedly "wrong"?

Above, I wrote that "we don't need to avoid the "chicken dilemma." I 
consider the chicken dilemma to represent the essential choices that 
voters make when there are more than two candidates. It's not 
avoidable without damaging the amalgamation process. Now, have I 
proven this rigorously?

No. But every attempt I've seen to eliminate the chicken dilemma does 
apparently damage social utility.

>  It's fine if you ask for evidence that you think doesn't exist, 
> but if you think your argument is unfalsifiable by any possible 
> evidence, something is wrong.

What's the argument?

In my ontology, "wrong" is neither right nor wrong. It is generally a 
disempowering judgment that does not take responsibility for choices. 
And that judgment is itself only a stand, not an absolute truth.



>I.e.:
>
>original Bucklin: with a multiple majority, all at or above the 
>found majority rating are collapsed to approval. Same with majority 
>failure. The result is that a lower preference may count *the same* 
>as a higher one. It's the "approval problem"
>
>MAV: under the multiple majority at a lower preference rule, the 
>system ignores the lower preference votes, using only higher-level 
>approval information. It does count the lower preference votes, but 
>not to distinguish between those candidates. I haven't done so, but 
>I could show some problem scenarios. It solves the "approval 
>problem," but at the cost of apparent expressed utility.
>
>
>Here's some exactly symmetrical arguments for MAV:
>
>MAV: with a multiple majority against, all at or below the found 
>majority rating are collapsed to antiapproval.

I'm lost immediately. Disoriented. Someone just turned my planet 
upside down and wants me to consider the laws of nature from that 
perspective. Sure, it should be equivalent. But, Jameson, I'm not going there.

>  Same with majority failure. The result is that a higher preference 
> may count *the same* as a lower one. It's the "approval problem"
>
>Original Bucklin: under the multiple majority at a higher preference 
>rule, the system ignores the higher antipreference votes, using only 
>lower-level antiapproval information. It does count the higher 
>antipreference votes, but not to distinguish between those 
>candidates. I haven't done so, but I could show some problem 
>scenarios. It solves the "approval problem," but at the cost of 
>apparent expressed utility.

Wasted.

I have no doubt that original Bucklin, even properly analyzed -- 
which Warren did not do -- loses some expressed utility. I'd expect 
MAV to do so in the opposite direction. EMAV would fall in the middle.


>EMAV: The system uses all the votes in the two cases (multiple 
>majority and majority failure.) Thus lower preference votes do 
>count, but only at deprecated value. The difference between full 
>preference and minimal approval is 1/2 vote. The difference between 
>full prefeence and the below-approval rank that is above maximum 
>opposition is 3/4 vote. So EMAV is intermediate to original Bucklin and MAV.
>
>
>No. GMJ or MJ are intermediate.

GMJ and MJ are *also* intermediate. Jameson, in pursuit of attempting 
to make me wrong, you are losing perspective.

>  EMAV throws the Bucklin principle out the window,

Silly. It uses the Bucklin principle, flipping to Range only in the 
ambiguous situations.

>  which means that in many or most elections, even if you have 
> perfect knowledge of other voters, no non-extreme ballot is 
> strategically optimal.

You have made that up about Range. I can take that apart easily. 
Shall we look at that.

This is a comment about Range, not about EMAV or Bucklin.

>  Thus, EMAV will encourage many more voters to exaggerate than in 
> other rated Bucklin systems, including MJ, GMJ, MAV, MCA, Original 
> Bucklin, Bucklin//Condorcet, etc. The quality of the ballot 
> information, and its usefulness for analysis (be it score, 
> Condorcet, or other) will be lower.

This is FairVote thinking, ultimately. It assumes that voters are 
controlled by a deathly fear of "harming their favorite." Real voters 
liberally voted lower preferences in the contested public elections. 
There was one exception that I know of, I think there was a single 
San Francisco election where they didn't. Election conditions vary.

The same argument could be used against Approval, for sure, and 
against Range, in general. EMAV is designed as a transitional system, 
to introduce a range ballot and to allow its evaluation under certain 
circumstances as Range.

I have challenged, many times, the whole concept of "exaggeration" as 
applied to Range voting. Voters will vote their preferences. Some 
will not understand at first, but that won't lead them *far* astray. 
There is a whole history of the claim of "exaggeration" as respects 
Range. It is not an honorable history. It's essentially an oxymoron.

This is the foundation of it, right with approval: the voter 
supposedly approves of two candidates, but votes only for one, 
because they prefer that one.

There is no absolute condition called "approval." Approval is a 
*choice,* so what is really being said is that the voter chooses two 
candidates, but only chooses one.

Yes, there may be strategic pressure; the voter makes the choice by 
balancing *preference strength* with *election probabilities.*

This is the ordinary process of choice in life. We do it all the time.

The more I discuss it, Jameson, the more I'm convinced, so far. If I 
had time, I'd do the BR studies myself. But I don't.