[EM] Why I think IRV isn't a serious alternative 2

[Kevin Venzke]

Hi,

--- En date de : Dim 14.12.08, Abd ul-Rahman Lomax <abd at lomaxdesign.com> a écrit :
> > >> That's not very generous. I can think of
> a couple of defenses. One would
> > >> be to point out that it is necessitated by
> the other criteria that IRV
> > >> satisfies. All things being equal, I consider
> LNHarm more desirable than
> > >> monotonicity, for instance.
> > >
> > >I, and certainly some experts, consider LNH to
> cause serious harm.
> > >Absolutely, it's undesirable in deliberative
> process, someone who insists
> > >on not disclosing lower preferences until their
> first preference has
> > >become impossible would be considered a fanatic or
> selfish. That's a
> > >trait I'd like to allow, but not encourage!
> > 
> > Well, I said "all things being equal." All
> things being equal I think it
> > is a positive thing that by providing more
> information, you don't have
> > to worry that you're worsening the outcome for
> yourself. Maybe something
> > else gets ruined, but then all things are not equal.
> 
> You don't add the information if you reasonably fear
> that the damage to your desired outcome would be serious.
> You provide it if you think it will increase your expected
> outcome.

I don't understand what this is a response to. LNHarm is a guarantee
that says you do not have to fear any damage.

You must be talking about methods that don't satisfy LNHarm. Yes, that's
right, you don't add the information if you fear it will hurt you, and
you do provide it when when you think it will help. You don't really
think someone would argue with that?

> Where I would agree is that it would be ideal if a voter
> could control LNH compliance. It is possible. This is
> equivalent to the voter taking a very strong negotiating
> stance. But I would not, myself, want to encourage this
> unless the method tested majority failure and held a runoff
> in its presence. And it's a general truth that if there
> is a real runoff, with write-ins allowed, total LNH
> compliance is impossible. Unless you truly eliminate the
> candidate. Never again can an eliminated candidate run!
> 
> Basically, so that I can't "harm" my favorite
> by abstaining in one of the pairwise elections involving
> him, the *method* eliminates him! I'd rather be
> responsible for that, thank you very much.
> 
> > Again, you seem to describe LNH as though it is
> synonymous with the IRV
> > counting mechanism. MMPO and DSC do not render
> preferences "impossible"
> > thereupon "disclosing" more preferences.
> 
> I think this is correct. LNH, however, is strongly
> associated with sequential elimination methods.

Ok, but that doesn't make it effective to criticize LNHarm using 
characterizations that only apply to IRV.

> It's
> possible to reveal lower preferences but to not use them in
> the pairwise election with the additional approval. I've
> not studied all the variations, there is enough to look at
> with forms of Approval and Range.
> 
> When Bucklin is mentioned to knowledgeable IRV proponents
> -- there are several! -- LNH will be immediately mentioned
> as if it were a fatal flaw. But the "harm," as
> I've noted is actually not harm from the ballot but only
> the loss of benefit under one particular condition: the
> voter, by adding a lower preference, *if* there is majority
> failure in previous rounds, has abstained from that
> particular pairwise election while participating in all the
> rest. It should be possible, by the way, to leave the second
> rank in 3-rank Bucklin empty, thus insisting on LNH for one
> more round. That shouldn't be considered an error, but a
> legitimate voting pattern.

It's impossible for me to imagine how you could use this option
effectively.

If I were voting under Bucklin, with equal ranking allowed, I would
vote approval-style, with one exception. If I had some reason to believe
that my favorite candidate is either the majority favorite, or else not
especially viable, then I would use the top two slots.

More complicated scenarios are theoretically possible but I can't imagine
the information would be adequate in real elections to act on them.

> > >> 3. they simulate voter strategy that is
> customized to the method
> > >
> > >That is relatively easy, and has been done.
> > 
> > No, this is the hard one! I don't know if Warren
> has even implemented
> > this for Approval and Range. I don't remember,
> whether the strategic
> > voters simply exaggerate, or actually approve
> above-mean.
> 
> Various strategies have been used.
> 
> "Above-mean" is an *awful* strategy, unless
> it's defined to mean something other than the mean
> utility for all the candidates. 

Above-mean is zero-info strategy.

>"Exaggerate," with
> Approval, is meaningless.
> 
> That strategy was indeed used: from Smith's 2001
> simulation run:
> 
> > 16. Honest approval (using threshhold=average
> candidate utility)
> > 17. Strategic range/approval (average of 2 frontrunner
> utils as thresh)
> > 18. Rational range/approval (threshhold=moving
> average)
> 
> Strategy 16 is awful. That's what Saari assumed as a
> strategy when he gave his example in his paper, "Is
> Approval Voting an Unmitigated Evil?" How's that
> for a nice, academically objective title? The paper does not
> disappoint.
> 
> Strategy 17 is better. Strategy 18 is not described enough
> that I could figure out what it means. 17 is adequate, but
> better strategies can be described, and it's possible to
> devise a zero-knowledge strategy (where the voter
> doesn't actually know the frontrunners) that would work
> better than bullet voting or only approving candidates
> "almost indistinguishable" from the favorite. The
> last strategy, except for the possibility of equal ranking
> effective clones, reduces to Plurality Voting, which
> isn't a terrible system as long as there aren't too
> many candidates and the configurations are those common in
> settled Plurality voting environments. (Hint: two party
> system).

17 would be good if Warren simulated some kind of poll to determine
who the frontrunners are. He selects the frontrunners arbitrarily.

> > For rank ballot methods Warren has implemented the
> same strategy for all,
> > and it is the biggest problem, with the least clear
> solution.
> 
> This doesn't seem to be true. I'm looking at his
> old simulation run, which describes the strategies briefly.
> 
> http://math.temple.edu/~wds/homepage/voFdata

The part that I criticize is the reliance on "same strat as 26." He
judges the performance of his Condorcet method and IRV alike as though
the "strategic" voters all employ a half-baked Borda strategy.

> But, absolutely, Warren's simulation approach needs
> much work.
> 
> > >> 4. they simulate pre-election information
> > >
> > >This is necessary for Approval and Range strategy,
> for sure, so I believe
> > >this has been done.
> > 
> > I don't believe Warren's simulations do this
> for any method. All
> > strategy is either zero-info, or (for rank ballot
> methods) based on
> > random arbitrary info provided uniformly to all
> voters.
> 
> No. Simulations using "poll strategy" involve, as
> described by Smith, simulated polls answered by random
> voters pulled from the complete voter set. That's not
> "random arbitrary info." It's a simulated poll
> of the voters.

Wrote something here but see lower instead.

> The most common Approval Voting strategy is to vote for the
> preferred frontrunner, then for any candidate preferred to
> that candidate. This leaves out intermediate candidates,
> i.e., preferred to the worst frontrunner, but the best
> frontrunner is preferred to that candidate. However, these
> votes are mostly moot, unless the election is close between
> that candidate and the frontrunner, which would require that
> there be something close to a three-way tie.
> 
> In any case, to apply this strategy, the voter needs poll
> data. I've argued that the voter can *estimate* this
> from the voter's own opinion, either alone or together
> with the voter's general estimate of where the voter
> sits in the electorate. This is technically zero-knowledge,
> I'd assert, but it uses the voter's own opinions as
> a sample to estimate election probabilities. This has to be
> right more often that not! -- and this strategy would knock
> Saari's silly Approval voting scenario upside the head!
> 
> It's really crazy to expect that most voters will
> approve above the mean candidate, with no regard at all for
> anything else. That strategy would make Approval highly
> vulnerable to clones, when it probably is not. It would make
> Approval highly vulnerable to irrelevant alternatives, when
> it probably is not.

Approving above the mean candidate is again the zero-info strategy.
If you decide that similar candidates are probably clones, then you're
guessing what other voters think, which is info.

The reason this fact doesn't make Approval vulnerable to clones is that
in a real election you would have info.

> > >It can actually be done, in the simulations, with
> perfect strategy,
> > >though, obviously, if you take this too far, you
> could run into loops, so
> > >I'd guess that the best strategy used would
> assume some uncertainty and
> > >would only iterate so many times, simulating polls
> and then shifts in
> > >votes as a result, then another poll, etc. The
> "polls" would solicit how
> > >the voter intends to vote, and the model can
> assume that the voter can't
> > >hide the information. After all, just how
> complicated do we want to make
> > >the model?
> > 
> > Yes, my simulations are based on polls. Polls are a
> great idea.
> 
> Smith used them....

Wrote something here, but see lower instead.

> > >Heavy use of serious strategization is pretty
> unlikely with ranked
> > >methods, in my opinion, most voters will simply do
> as the method implies,
> > >rank in preference order, and they can do this a
> bit more easily if equal
> > >ranking is allowed.
> 
> Yes. I agree with both of these comments. 

These are your own comments though.

> The problem with
> ranked methods is that, sometimes, they come up with a poor
> result *from sincere rankings,* but this is hugely
> ameliorated, I expect, when equal ranking is allowed. Still,
> there is still the problem of the defective assumption of
> equal preference strength for each ranking, and that limits
> the performance of ranked methods.
> 
> My opinion is that all the benefit of ranked methods can be
> realized within a Range method, with appropriate rules. This
> is best done with an additional round when necessary, and
> this dovetails with runoff voting, then, and the
> desirability of explicit majority approval of any result. In
> other words, voting systems theory, the theory of democracy,
> and the long-standing understanding that top two runoff is a
> major election reform, all come together here. It's
> amazing to me that this wasn't being considered when I
> arrived on the EM list, it's not like it's really
> complicated....
> 
> > Warren's implementation would suggest that he
> strongly disagrees with you
> > on this.
> 
> No, I don't think so. 

Well, I think so, based on the page you just linked to.

> But Warren is quite quirky and
> sometimes cranky. Further, I don't see that he's
> doing serious continuing work on the simulations. He pretty
> much has said to others who criticize his simulations:
> "Fix it, then! Do your own damn simulations, my IEVS
> engine is published, you can use it and tweak it to your
> heart's content."
> 
> He has a point.

I do make my own simulations.

I wouldn't let my lack of desire to improve IEVS stop me from criticizing
it. Hopefully no one will take this personally.

> > >> Some of this isn't difficult, it's
> just again a question of how far you
> > >> take it. Strategic voter behavior needs to be
> made less ridiculous.
> > >> But what kind of strategy should be allowed,
> for (let's say) Condorcet
> > >> methods? If everyone votes sincerely, then
> Range will look bad. So
> > >> clearly the line has to be drawn somewhere
> else.
> > >
> > >No, you'd have to compare sincere Condorcet
> with some kind of sincere
> > >Range.
> > 
> > Look at it this way: To compare methods fairly we need
> to know how
> > strategic voters attempt to be, in the same
> situations, under whatever
> > methods we want to compare. Why compare strategic
> Method A with strategic
> > Method B, if Method B voters would never vote that way
> in reality?
> 
> Well, maybe you are right. To synthesize this, the
> probability of voters using some voting strategy must be
> included in the simulation. In fact, with proper ballot
> design and voter education, I think we would see *more
> sincere* Range voting, than just popping a Range ballot in
> front of them. For starters, voters should be encouraged to
> maintain preference order, where it's distinguishable.
> *How much vote strength* they give to this is another
> matter. They should know that voting Range Borda style may
> not be optimal!

> > What if, in real life, Condorcet voters just don't
> use any strategy?
> > And what if it's also true, that Range voters in
> real life turn the
> > method into Approval?
> 
> Well, the first is reasonably likely. The second isn't.
> Rather, real voters will push Range toward Approval, which
> isn't a bad result! But it won't go all the way
> there, probably not even close.

Well then, if Warren agreed with you, the interesting study would be
just how far Range must approach Approval before Range and (sincere)
Condorcet are tied.

> > In that case, the only useful comparison to be
> > done by the simulations, would happen to be sincere
> (or strategic, no
> > difference) Condorcet vs. strategic Range/Approval.
> And according to
> > Warren's simulations, Range doesn't win, in
> that case.
> 
> Depends. Do you have a page reference? Mostly what I've
> seen doesn't disclose enough details to make that
> conclusion.
> 
> http://math.temple.edu/~wds/homepage/rangevote.pdf

Yes. First of all, I want to note that on page 19, Warren explains that
he did not see the need to do any real polls, because his profiles were
so random that the selection of any pair of frontrunners was as likely 
as any other. The identities of the frontrunners are hard-coded.

On the same page he justifies the use of Borda-style strategy under
IRV and Condorcet as "plausible-sounding." It is amazing how much work
has been done depending on the logic in this single paragraph.

Also, the "moving average" concept is defined starting on page 7 and
doesn't involve a more sophisticated polling mechanism. It basically
means you redetermine the average every time you assign some status to
a candidate (and then leave that candidate out). I could be mistaken.

To answer your question, Warren's relevant remarks on Condorcet "LR" are
on page 25.

Quoting:
>If a large fraction of voters had somehow become
>falsely convinced that honesty is the best voting
>strategy in the Condorcet-LR system – but they
>correctly understood Range voting strategy – then
>since honest-CLR (#4) exhibits smaller (up to 30%
>smaller) regret values than rational-Range (#18),
>CLR would become the best system. This is conceivable,
>since I myself once suffered from that delusion
>about CLR (and also Hare-STV)!
...
>So the only contenders for the Title are Condorcet-LR
>and Range – and CLR can only compete if the voters are
>deluded.

> This is the full original paper, written in 2000.
> 
> First of all, we should realize that advanced voting
> systems encourage more candidates to run. This can cause
> some systems to experience seriously increased regret. So
> I'm going to look at the maximum number of candidates
> used, five. Honest Copeland seems to do the best of the
> Condorcet methods, in the five-candidate elections, average
> regret of 0.14181. (This is the run with issue-based
> utilities, 50 voters).
> 
> Sincere Range, by the way, seems to do better with more
> candidates. Given that San Francisco sees more than twenty
> candidates in the ballot on their elections, this is
> interesting.
> 
> However, here we are comparing with strategic Range,
> though: 0.23232. Strategic Range does worse with many
> candidates (like Copeland), probably because of the
> oversimplified votes that result. Now, that is *fully*
> strategic Range, i.e., all voters vote that way. This is
> highly unlikely. I think I recall seeing that some work has
> been done with mixes. However, I'd assume that real
> Range Voting would roughly in between fully sincere, what
> Smith calls "Honest," and fully strategic. Honest
> Range with 5 candidates has regret of 0.05368.
> 
> While this may not be accurate, if 50% of the voters vote
> honestly, and 50% strategically, we might expect regret for
> the mix of the average: about 0.14. Roughly the same as
> Honest Copeland.
> 
> Now, Smith examined strategic Copeland. The strategy was to
> max rank the preferred frontrunner and to min rank the worst
> frontrunner, and to order the rest honestly. This is a
> simple strategy, I don't know how effective it is for
> the voter, but it's certainly easy to apply, and I think
> it does increase the voter's expected utility. Some
> voters will use it, if it is reasonably rewarding. (I
> don't know if that is true). 

Whether it increases the voter's expected utility depends on how badly
it is needed/rewarded to compromise and bury to that extent. I can't
really comment on Copeland.

> > >> I wonder if you have ever been curious to
> wonder what a "strategic"
> > >>voter is, for a rank ballot method.
> > >
> > >Nah, curiosity killed the cat.
> > >
> > >I've done a fair amount of reading on this,
> but who remembers anything?
> > >Often not me.
> > 
> > Actually this question was specifically about the
> simulations.
> 
> Read the paper. Smith describes it explicitly.

I hope you're saying *you* read the paper rather than asking me to.
I have been trying to tell you what he did.

> > >> Some six months ago I wrote a strategy
> simulation for a number of
> > >> methods. One situation I tested was Approval,
> given a one-dimensional
> > >> spectrum and about five candidates, A B C D
> E.
> > >>
> > >> In my simulation, once it was evident that C
> was likely to win, one of
> > >> either B or D's supporters would stop
> exclusively voting for that
> > >> candidate, and would vote also for C.
> > >
> > >B and D voters are motivated to ensure that C wins
> if their favorite
> > >doesn't. Hence Approval will tend to find a
> compromise. If B or D are not
> > >relevant, can't win, they *may* also vote for
> B or D, so I'm not sure
> > >that the simulation was accurate.
> > 
> > I'm not sure what you mean by this. Voters that
> prefer B or D to C have
> > no reason to not continue voting for B or D.
> > 
> > The issue is that when all the D supporters (for
> example) *also* vote
> > for C, then it isn't possible for D to win. And
> the more that D voters
> > "give up" and vote for both, the less sense
> it makes strategically for
> > the remaining D-only voters to not "give up"
> as well.
> 
> I think something has been missed here. The votes for C are
> added, not when C becomes a frontrunner, but when C becomes
> a frontrunner preferred to another frontrunner. If B or D
> are frontrunners, with C, and if the voters prefer B or D to
> C, they won't vote for C in the most common strategy.
> The example is incompletely explained, I don't know if
> something was missed by me, or it just wasn't there.

I may have supposed that B and D voters sincerely rate C above midrange
(when the ratings for B and D define a range).

> No. If the D voters prefer D, they won't vote for C
> unless there is another candidate more likely to win if they
> don't vote for C. The only situation where this breaks
> down is a three-way tie (three-way close race) between their
> favorite, with C and another candidate less preferred than
> C. In other words, if their candidate can win, most voters
> will not add an additional approval for a likely and
> significantly less-preferred rival. They might do it with
> Bucklin, where it's more like insurance and an easier
> decision.

But a three-way close race is likely to break down, just as it would
under Plurality. Eventually people realize that somebody is in third
place. They will *try* to see that someone is in third place, so that
they know how to participate in the "real" contest.

> > My simulations involve polls. When the polls find that
> the winner will
> > either be B or C, then it's strategically unwise
> to not approve one of
> > them.
> 
> That's correct. However, you were talking about B or D
> voters. If it's a B voter, the strategy means
> "don't vote for C." If it's a D voter, it
> means "Vote for C," assuming that C is preferred
> to B. But in that case, the C vote is probably harmless to
> D, who isn't likely to win anyway, with or without the
> vote.

What I'm saying is that D only has to slip a little in the polls to be
hit by an avalanche of "defecting" supporters. It's just like Plurality.
And just like Plurality it's not very likely that D can make a comeback.

> > At first, the polls report that C will win a lot but
> (due to bullet
> > voters for B and D) there is some chance that B or D
> will win. Eventually
> > the polls (which are subject to some randomness) will
> produce a prediction
> > that D's odds (or B's) are abnormally poor.
> This causes D voters to stop
> > voting only for D, and also vote for C. This almost
> immediately makes D an
> > unviable candidate, and the bullet voters for D
> disappear.
> 
> You mean that they stop indicating in polls that this is
> how they will vote. I don't think that real voters will
> iterate in polls like that..... not with significant
> differences. 

Not sure what you mean. Polls may not work exactly like this, but we
do observe that in most Plurality elections we end up with two candidates.
We use primary elections to remove and institutionalize some of the chaos,
but even if we didn't, we would still end up with two candidates somehow,
because voters are smart and information is easy to come by.

If the voters want, or cannot do better, they can give lousy information
in the polls. Or the polls could just by nature be lousy. The voters will
get lousy information back. But they will still have to base their
strategy on this information.

> Most Approval studies of iterative voting start
> with bullet votes. Then approval thresholds are gradually
> adjusted. If the bullet voters for D disappear, it must mean
> that the voters have concluded that D can't win, hence
> they go for the compromise, C. They will only do this if
> they think that the real pairwise election is not between C
> and D, but between B and C, and they prefer C.

Right.

They don't necessarily have to conclude that D can't win, so much as fear
that they're about to waste their votes if they don't list a second
choice. This just happens to sink D in the process.

> But remember, it starts from bullet votes, pure favorites.
> Plurality has a fairly good ability to predict what a
> preferential voting system -- or Approval system -- will
> come up with, and it only breaks down under certain
> conditions. If the D voters have a significant preference
> for D over C, they will hold out longer, and some of them
> will never compromise. Remember, not all voters will follow
> frontrunner strategy. They don't with Plurality, why
> should they start with Approval?

Well, I'm not using "frontrunner strategy" but "better than expectation"
strategy, since that can be applied more universally.

If D voters are more resilient then it's possible that B will sink
instead. It's not as likely to be C, though, since C has an avenue of
bouncing back that B and D lack.

In this simulation, I don't simulate voters who don't care if their
vote isn't expected to be effective.

> To summarize this, the scenario makes sense only if B, C,
> and D are in a near-tie. If both B voters and D voters
> prefer C over the other of B and D, then C is, indeed, their
> compromise candidate! It's perfectly rational that the B
> and D voters, iterating over polls, increase their support
> for C, but it will never go all the way.

Well it wouldn't be both the B and D factions. You would only add votes
for C if you believe your expectation is dropping. That happens when
your preferred candidate (D) looks to be slipping. The B voters have no
need to compromise that far.

In this situation, D voters who decline to vote in the main contest
are basically "voting for Nader."

> The behavior described seems reasonable, proper, and is
> effective for finding a compromise winner. Is there some
> problem with it?

No, I don't think so. It's pretty good behavior actually. At least on its
face, it would seem that Approval would ruthlessly favor the median
voter's candidate in this kind of scenario.

The big concern is what happens when poll stability can't be achieved.

> Bucklin allows them to maintain their sincere preference,
> but, effectively, vote this way. Some might add C in the
> second rank, some in the third, depending on their
> preference strength. But some will always bullet vote,
> perhaps even most. Real voters don't give up so easily
> as your simulated ones!

I did simulate MCA, and yes, the D voters continued to vote for D as
their favorite (they were not allowed to list multiple favorites, but 
this was simply to make the coding more manageable).

I don't know what you mean by voters not giving up so easily as my
simulated ones. How easy is easy?

I could conceivably program some voters to insist on being sincere.
(In whatever sense that it is not sincere to vote also for C.) But
it seems to me that this type of voter is a bad thing for Approval,
just as it is under Plurality.

Kevin Venzke