[EM] Scoring (was Re: OpenSTV 2.1.0 released)

[Juho Laatu]

On 22.9.2012, at 8.03, Michael Ossipoff wrote:

> I'd said:
> 
>> Do you claim that unimproved Condorcet can be
>> defended in a comparison with Symmetrical ICT, or ordinary ICT?
> 
> You replied:
> 
> Definitions of the methods needed in addition to their technical
> properties. I don't exactly know what you mean with unimproved
> Condorcet.
> 
> Unimproved Condorcet refers to what "Condorcet" meant before Improved
> Condorcet was proposed by Kevin Venzke. In other words, unimproved
> Condorcet is Condorcet that isn't Improved Condorcet.

In that case I don't support unimproved Condorcet. That category of methods contains some really stupid methods (e.g. one that elects the most truncated candidate in case there is a top loop).

> 
> Unimproved Condorcet is a broad category that includes every method
> known as "Condorcet" before Improved Condorcet was proposed.
> 
> That includes Beatpath and all of the other Condorcet methods other
> than Improved Condorcet.
> 
> When Kevin first proposed Improved Condorcet, he completed it with an
> Approval count. So he called his proposed method
> Improved-Condorcet-Approval (ICA).
> 
> Later, Chris Benham proposed Improved Condorcet completed instead with
> a top-count. Using Kevin's naming system, I called Chris's method
> Improved-Condorcet-Top (ICT)
> 
> Later, I proposed a modification of ICT that did the same improvement
> at bottom end too. I call that Symmetrical ICT. It could be
> abbreviated SITC.

I might consider some of these methods if there were some real problems with strategic voting and some fixes would be needed to make the elections work propoerly. But as long as I can trust that most voters will vote sincerely, I'd focus on picking a method that picks the best winner with sincere votes.

Some properties of methods that I don't like very much are: 1) truncation based approval, since that encourages voters not to take position on which one of the non-approved candidates should be elected (works against the basic idea of ranked methods of collecting the sincere preferences), 2) use of ties (more than what is sincere), for similar reasons.

Doctors have similar problems. Many medicines are far from harmless. Doctors have to compare the risks of the disease and risks of the medicine. If the medicine is likely to make more harm than help, it should not be used. This means that one should deviate from the method that picks the best winner with sincere votes only if one is certain that otherwise the method would give even worse results because of strategic voting. Obviously you believe that basic Condorcet methods would attract certain strategies to the extent that those methods must be fixed. And I believe that in most societies it is more likely that strategic voting will be marginal.

Btw, here are some links to old proposals that address problems that are close to this case. Just FYI.
http://lists.electorama.com/htdig.cgi/election-methods-electorama.com/2005-April/015788.html => ability to cancel the risk of startegis if there is a risk, but no changes needed if there are no risks
http://lists.electorama.com/htdig.cgi/election-methods-electorama.com/2006-October/018710.html => a complex approach that allows also weak preferences, so you don't need to use flat ties
http://lists.electorama.com/htdig.cgi/election-methods-electorama.com/2006-October/018711.html
Also Condorcet elections that have a second round if the first round has a top loop were discussed around the same time.
But my basic attitude thus is that all these may be "too much medicine" unless there really are serious problems with strategic voting.

> 
> Some criterion compliances:
> 
> A. Condorcet Criterion:
> 
> Here is a way to define the Condorcet Criterion:
> 
> A method passes if:
> 
> 1. It collects ranked ballots. Unlimited rankings are allowed. Voters
> may vote as many pairwise preferences as they wish.
> 
> 2. If a candidate beats each one of the other candidates, then s/he wins.
> 
> [end of definition]
> 
> 
> Unimproved Condorcet meets the Condorcet criterion if "beats" is
> defined in the traditional way, whereby X beats Y iff more voters rank
> X over Y than rank Y over X.
> 
> If the Condorcet Criterion is defined using a meaning of "beats" that
> interprets equal-top ranking in the way that is consistent with the
> preferences, intentions and wishes of the equal-top ranking voter,
> then unimproved Condorcet fails the Condorcet Criterion, and ICT meets
> that criterion
> 
> If "beats" is defined so as to interpret both equal-top and
> equal-bottom rankings consistent with the preferences, intentions and
> wishes of voters voting those rankings, then Symmetrical ICT is the
> method that passes the Condorcet Criterion.
> 
> 
> Preliminary definitions for the definitions of "beats"
> 
> (X>Y) means the number of ballots ranking X over Y
> (Y>X) means the number of ballots ranking Y over X
> (X=Y)T means the number of ballots top-ranking X and Y
> (X=Y)B means the number of ballots bottom-ranking X and Y
> .......You bottom rank a candidate if you rank hir over no one.
> 
> Definition of "beats" that isn't consistent with preferences and
> intent of equal-top or equal-bottom ranking voter:
> 
> X beats Y iff (X>Y) > (Y>X)
> 
> Definition of "beats" that is consistent with preferences and intent
> of equal-top ranking voter:
> 
> X beats Y iff (X>Y) > (Y>X) + (X = Y)T
> 
> Definition of "beats" that is consistent with preferences and intent
> of equal-top ranking voter and equal-bottom ranking voter:
> 
> X beats Y iff (X>Y) + (X=Y)B > (Y>X) + (X=Y)T
> 
> Of course the reason why Symmetrical ICT meets the Condorcet Criterion
> wherein "beats" is defined consistent with the preferences and intent
> of the equal-top ranking voter and the equal-bottom ranking voter is
> because Symmetrical ICT's definition uses that meaning for "beats".
> 
> But note that it is not a matter of re-defining CC so that SITC will
> pass. It's a matter of defining CC consistent with interpreting a
> voter's ballot consistent with hir preferences, intent and wishes.
> 
> B. Later-No-Harm (LNHa):
> 
> Condorcet methods, Approval, and Score fail LNHa.
> 
> But ICT and Symmetrical ICT don't fail it nearly as badly as does
> unimproved Condorcet. In fact, I'll venture to say that ICT and SITC
> don't importantly fail LNHa. Note that I'm not speculating about how
> often they'll pass or fail. I'm saying that their failures aren't
> important.That's because there isn't a chicken dilemma. Chicken
> dilemma is the worst kind of LNHa failure.
> 
> C. Later-No-Help (LNHe):
> 
> Approval, Score, and Symmetrical ICT pass LNHe. Unimproved Condorcet
> and ICT fail LNHe.
> 
> LNHe greatly simplifies u/a strategy. SITC's u/a strategy is as simple
> as that of Approval and Score. In unimproved Condorcet, you won't know
> what to do, even in a u/a election. I've recently told you why.
> 
> In ICT and unimproved Condorcet, u/a strategy calls for ranking the
> unacceptable candidates in reverse order of winnability. That
> incentive or need doesn't exist in Symmetrical ICT. In SITC, the u/a
> strategy for unacceptables is to simply not rank them.
> 
> In (so far as I'm aware of) all rank methods that allow equal top
> ranking, in a u/a election, there is a need to equal top rank the
> acceptables. So, in ICT and SITC that is the u/a strategy. That need
> exists in unimproved Condorcet too, but the problem is that moving
> some particular acceptable to top can change the winner from an
> acceptable to an unacceptable. That's why I say that, in unimproved
> Condorcet, you won't know what to do, even in a u/a election.
> 
> D. FBC:
> 
> Approval, Score, ICT and SITC pass. Unimproved Condorcet fails.
> 
> E. Defection resistance:
> 
> ...is had by ICT and Symmetrical ICT, but not by Approval, Score, or
> unimproved Condorcet.
> 
> As I said, not a problem for Approval and Score (and probably not for
> unimproved Condorcet either, for the same reason--though dealing with
> it could be more complicated). But, as I also said, chicken dilemma is
> the nearest thing to a problem that Approval has, and therefore you
> don't significantly improve on Approval without getting rid of chicken
> dilemma.
> 
> I'd said:
> 
>> The Chicken Dilemma is the nearest thing to a problem that
>> Approval has (though it's so well dealt with in Approval that it isn't
>> really a problem).
> 
> You replied:
> 
> I'm afraid it might be.
> 
> [endquote]
> 
> ...except for the long list of reasons why it wouldn't be a problem,
> the list that I've frequently posted during the past several weeks.
> 
> One of the defenses on that list was something that Forest suggested,
> and which I call Strategic Fractional Ratings. You of course must have
> missed my posting of that.
> 
> I've posted so much and so recently about the reasons why chicken
> dilemma won't be a problem in Approval, and, posted specifically,
> about SFR, that I don't think that I should repeat it again this soon.
> 
> On Fri, Sep 21, 2012 at 7:43 PM, Juho Laatu <juho4880 at yahoo.co.uk> wrote:
>> We are about to dive into the details of some methods. I'm not sure if there are still some unanswered questions that I should cover, or my own claims that I did not clarify yet. I'll comment some random points below.
>> 
>> On 22.9.2012, at 1.48, Michael Ossipoff wrote:
>> 
>>> Maybe you meant to compare unimproved Condorcet to Approval (because
>>> you didn't want to compare it to ICT and Symmetrical ICT).
>>> 
>>> Ok. You mentioned the Chicken Dilemma. It exists in Approval and
>>> Condorcet. Unimproved Condorcet doesn't get rid of the Chicken
>>> Dilemma. It's basically the same in both methods.
>>> 
>>> Approval meets FBC. Unimproved Condorcet fails FBC.
>>> 
>>> Exactly how is unimproved Condorcet better than Approval?
>>> 
>>> Condorcet's Criterion?
>>> 
>>> Condorcet's Criterion compliance is meaningless when people are
>>> favorite-burying.
>>> 
>>> Then there's the matter of the highly computation-intensive count that
>>> every rank method has, including the Condorcet methods.
>>> 
>>> Computation-intensive, labor-intensive count = big count-fraud opportunity.
>> 
>> It should be enough if you can record (digitally) the content of the ballots in a reliable way. >Computations should not lead to fraud since they can be easily double checked.
> 
> By whom :-)  

I think I already said that the computations (from digitized ballot content to results) could be checked either by anyone or by some nominated entities (if ballot content is not published to protect privacy).

> With Approval, representatives of various parties are
> observing a handcount in which approvals are tallied. The tallies are
> added up at the conclusion of the count, in front of everyone.
> 
> In a rank method, what do you have? A digitally-recorded record that
> will be checked...how and by whom? That's a problem even if storage
> security is adequate (guards, locks, cameras, etc.)
> 
> Not quite the same thing.
> 
> You said:
> 
> If the content of the ballots is made public, checking is really easy.
> 
> [endquote]
> 
> Nonsense. How do you make public thousands or millions of rankings?

You can already now download (huge) movie files from the net. There wouldn't be as many people interested in downloading election data. Downloading data from Burlington (one city) was not a problem.

> You could give copies to representatives of the parties (who trust the
> copying process). But, for that matter, they also have to trust the
> machine balloting and the process that made the rankings-record.
> 
> There are paper rankings store somewhere?

Certainly, original paper ballots shall be available for checks and recounts (for an agreed period of time).

> See above.
> 
>> 
>>> 
>>> ...and it also means machine balloting and computerized count. That
>>> means an even more greatly-enhanced count-fraud opportunity.
>> 
>> Machine balloting is a risk in all methods
> 
> Not all methods have as much need for machine balloting.
> 
> You said:
> 
> (if votes are only bits, and there is no paper copy). Also complex
> ballots like ranked and rated ballots can be implemented on paper
> quite well (= without too bad limitations). Computerized count is not
> a problem if reliable source data is available for checks.
> 
> [endquote]
> 
> See above, about the checks of the stored ballots. And remember that,
> with Approval, the tallying and the actual determination of vote
> totals, can be done in front of everyone, during the public count.
> Though ballots can still be stored, everyone has seen them tallied and
> the tallies counted. Everyone has seen the answer determined.
> 
> You said:
> 
> I stick to my comment on Condorcet methods (see above).
> 
> [endquote]
> 
> Stick away. Then, there's nothing more to be said. The arguments have
> already been said, and can be compared, and their merits evaluated, by
> anyone.
> 
> You said:
> 
> Maybe someone should arrange a real-life political Condorcet election
> so we could see how extensive favourite burial there will be and how
> much that will influence the results.
> 
> [endquote]
> 
> There isn't much information available about that, as yet. There have
> been Internet polls, but how are we to know who is favorite-burying,
> if we don't know who everyone's favorite is.

We can ask people. The resulta are not exact, but very informative. Or if there are groups that recommend some strategies, someone will hear about them. Some strategies can be visible in the ballots.

> I happened to observe the
> Condorcet Internet voting of a friend whose preferences I knew. That's
> an unusual exception. In that instance, she favorite-buried, ranking,
> below all the Democrats, the candidate whose policies she preferred to
> theirs.
> 
> I had another conversation with someone who insisted that approving
> Favorite would count against helping Compromise outpoll Worst. Because
> the method was simple Approval, it was possible to eventually show her
> that how she rates Favorite has no effect on the result of approving
> Compromise and not Worst. The lesson from that: People are inclined to
> favorite bury, until you show them that there's no possible need to.

I know at least one person that is inclined to bury and probably would bury :-).

> Because the method was Approval, I could show her that. I wouldn't
> have been able to reassure her about unimproved Condorcet, because the
> reassurance would be a lie.

Maybe you told her that in Condorcet elections it is in her interest to always bury :-).

> 
> With so few people I've spoken to about that, or observed voting,
> those two are a very high percentage. Those two experiences, together,
> would be quite unlikely if favorite-burial inclination weren't common
> with new voting systems such as unimproved Condorcet.
> 
> You said:
> 
> Otherwise it seems to be just your guess against mine.
> 
> [endquote]
> 
> A fair and probable conclusion based on the available evidence.
> 
> But what if it were just guesswork? Does that mean that we should take
> the guess that might cause widespread favorite-burial, and act on that
> guess?

Tust the experts, keep risks manageable, test with trials.

> 
> Look: Favorite-burial incentive, and the chicken dilemma, are so
> easily avoided--why would you want to keep them (and then hope that
> they wouldn't be a problem). It can be guaranteed that they won't be a
> problem if they don't exist.

True, except that there are also irrational voters. But don't take medicine that harms you more than the disease.

> 
> You said:
> 
> I don't believe that the existence of a (theoretical) FBC
> vulnerability would automatically lead to widespread favourite
> betrayal.
> 
> [endquote]
> 
> We've already been over that. It's easily demonstrated (I showed you
> several times) that the typical Democrat voter votes as if s/he
> regards elections as u/a, with Dem acceptable and Repub unacceptable.
> It's been repeatedly shown to you that, with that person's beliefs and
> assumptions, hir optimal strategy will include favorite-burial, if her
> favorite isn't the Democrat.
> 
> What little evidence I've encountered revealed a favorite-betrayal
> inclination in unimproved Condorcet, and even in Approval, till I
> explained why it isn't needed. She eventually understood the
> explanation, but didn't initially believe it. Favorite-burial
> inclination, with new voting systems, including unimproved Condorcet
> has been observed, in a high percentage of cases. For both of those
> observations to be rare exceptions would be highly improbable.
> 
> 
> You said:
> 
> Again I refer to Burlington IRV elections as an argument why this
> probably would not happen even in the U.S. (The last sentence is about
> the U.S. The others are generic.)
> 
> [endquote]
> 
> Municipal elections are nothing like presidential elections. A lot
> more is at stake in a presidential election, or other election for
> federal office.
> 
> Besides, IRV's FBC failure upset a lot of people, so much so that IRV
> was thrown out.

My understanding is that Condorcet criterion or favourite burial or any other (method internal) strategic concerns didn't play any major role in the minds of the voters when IRV was thrown out. Probably most of the voters know very little about the strategic propoerties of different methods.

> 
> Suppose they hadn't been allowed to throw IRV out. Do you think that
> people wouldn't be shown the advantage of favorite-burial, from their
> Burlington experience?

How would you expect them to vote in the next IRV election? (you can assume reasonably similar candidates)

Juho


> 
> Mike Ossipoff

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.electorama.com/pipermail/election-methods-electorama.com/attachments/20120922/0f923fb7/attachment-0004.htm>