[EM] replies to Ossipoff re Range Voting; explanation of latest RV results

[Abd ul-Rahman Lomax]

At 04:50 AM 2/12/2007, Dave Ketchum wrote:
>Warren is investing significant effort in simulations - GREAT, for 
>there is hope for learning something useful with such.
>
>BUT, Warren is CLAIMING be making valid comparisons among methods 
>via the simulations.  When we find that there are differences 
>between the simulated voting and real voting, the claimed results of 
>the simulations become suspect:
>      HOW NEAR do the simulations come to being valid measures?  The 
> bit about voting equal liking is just one difference that needs 
> evaluating as part of evaluating the simulations.

Warren has a habit of, shall we say, jumping to conclusions. He will 
state something indicated by evidence as if it has been proven. One 
thing is clear, though. He listens to criticism, and, quite often, he 
moves his opinions based on it.

Nevertheless, Warren has a point. He states conclusion based on thin 
evidence, yes. But he does this when there, essentially, is *no* 
other evidence. In this particular field, we have a whole lot of 
theory that is not grounded in either real elections *or* 
simulations. Most of the methods we favor have either never been 
tried, or they have been tried only in very limited circumstances.

Warren's simulations are limited, yes. But they are all we have. Or 
almost all. Brian Olson ran some simulations, he reported. I haven't 
seen comparisons. Yes, the assumptions in the simulations are very 
important. Warren has, he claims, written IEVS to be extensible, to 
use various kinds of assumptions about voting patterns. Others are 
specifically invited by him to use the tool, *especially* if they 
don't like the assumptions in the simulations he has run.

>I DO NOT buy the claim that permitting equal ranking makes ranked 
>methods close to Approval, for the heart of these methods is the 
>ability to do UNequal ranking.

As seems too common, Mr. Ketchum has misunderstood the claim. I 
wrote: "... most of us would prefer equal ranking (which actually 
turns ranked methods into something closer to Range, or at least to Approval)."

There is no doubt but that Ranked methods which allow equal ranking 
are "closer to ... Approval" because a voter may vote under them 
*exactly* as the voter votes in Approval, and, further, the election 
outcome would generally be identical if all voters vote this way. 
Having the ability to rank, obviously, is an additional capacity.

Generally, my opinion is that, generally, absent issues of voter 
confusion and capacity, the best method will be one which collects 
the most information about voter preferences. However, there is a 
limit. That is, it is possible to collect more information than is useful.

And this is the big debate about Range. How useful is it to be able 
to rank two candidates that one, in fact, would rate equally? In 
Bayesian Regret terms, rank, under those conditions, is of vanishing 
significance.

This leads us to the next big debate about Range: how will Range 
behave under conditions of strategic voting, where there is 
significant compression of ratings (or what I've called 
magnification). The way that strategic voting works under Range is 
different from how it works with pure ranked methods. With pure 
ranked methods, order reversal is necessary, but with ranked methods 
which allow equal ranking, strategic voters *can* avoid order 
reversal. But then they are voting exactly as they would with Range!

>How does a voter decide whether to rank equally, assuming the method 
>permits this?  That is up to the voter, who will sometimes see two 
>candidates as equally worthy and then be thankful for being able to 
>vote accordingly.

Absolutely. Look, the point was that allowing equal ranking can be 
expected to improve election outcomes as measured by Bayesian Regret. 
Sure, voters may use it to indicate equal approval, but what is more 
interesting here is how strategic voters will use it, if they wish to 
avoid downranking their favorite. With equal ranking allowed, they 
can use Approval strategy. Which, of course, has the claimed problem 
of Approval, that there is no way to indicate ranking while voting equally.

Well, duh!

Actually, there are ways. For example, Range could allow a precedence 
indication where candidates were equally ranked. However, it gets 
complicated, and, this is the point that has been made:

There is no indication that this benefits society, overall.

The goal of utility measurement is to maximize total utility of the 
election to society. There is a limiting assumption that is made, 
generally, which is that all voters are equal. Range does allow 
voters to cast weak votes, which would allow better anticipation of 
true utility, but generally, we assume that voters will cast full 
votes, which means that they rate one or more candidates at minimum 
and one or more at maximum. The effect of this is to assume that 
society is equally benefited by utility maximization for one voter 
vs. another, even if one voter doesn't really care that much and the 
other has a lot at stake.

But this assumption is pretty basic to democracy. Range does *allow* 
voters to indicate that all their preferences are weak. The cost of 
this, of course, is that their vote is weak. But that, of course, 
would be precisely what they want! Weak preferences *should* be weak 
votes. This is *exactly* what sincere Range does.

Somehow I've seen writers here assume that strong preferences are 
somehow morally reprehensible, or that it is morally repugnant to 
grant strong preferences more power. What these people are thinking 
of, of course, is fanaticism or strongly partisan belief and action. 
But we shouldn't forget that strong preference can also come simply 
from increased knowledge.

And when I've mentioned that here, I've been attacked for elitism. 
No, I'm far from an elitist. What I'm claiming, in fact, is that the 
only person who should be allowed to decide that my preferences are 
strong or weak, and thus to have strong or weak effect on outcome, is 
me. Or you. Or any voter.

It is totally obvious that failure to consider preference strength 
can warp election outcomes, and this is the problem with pure ranked 
methods. It is *somewhat* ameliorated by allowing equal ranking, 
which is a way of expressing a weak preference, so weak that it is to 
be disregarded.

Now, it is obvious that any high-resolution Range ballot allows 
ranking. If the resolution is great enough, the practical effect of 
indicating rank while maintaining vote strength is as if there were 
no difference. Range 100 is probably quite adequate for this; note 
that Warren would advocate higher resolution Range if it were 
considered practical. What in the world would a rating of 99.9% vs 
100% mean, other than some intuitive sense of one being better than 
another, but not by much that it would be a big deal to me if my 
"favorite" among these lost. If one really wants absolute, 100%, full 
vote strength, doesn't want to chance that even one thousandth of a 
vote could flip the outcome, then Range ballots could provide a means 
of indicating ranking with equal rating.

But the question is whether or not this would provide a significant 
benefit to society? I'd suggest that considering elections from the 
point of view of utility leads us to conclude that it would not 
provide any measurable benefit.

Personally, I'd like to encourage voters to vote sincerely. Sincere 
Range maximizes utility precisely because it sums expressions of 
utility from each voter. The only question is the behavior of Range 
under conditions of insincere, strategic voting, and there is a huge 
amount of hot air on this subject. And only one person who has 
studied it in a more grounded way. Warren Smith.

Note that, if I am correct, Condorcet himself received criticism of 
his method based on a comment about strategic voting. He dismissed 
it, saying that his method was for honest voters.

What is the ideal method for honest voters? I'd say that if we can't 
agree on that, we aren't even going to get close to agreement about 
what is best when voters opt to vote strategically.

For reasons that I won't detail at this time, though, my expectation 
is that Range will behave well in the presence of strategic voting. 
Any significant harm done is likely to fall upon the strategic 
voters, not on the honest ones. The true measure of *strategic* 
voting would be voting where Range votes actually reverse sincere 
rank. Where you want your second favorite to lose so badly that you 
rate him at zero while rating less-favored candidates higher because 
you fear that they will win. This kind of voter is quite likely to 
get burned. And, I'd suggest, deserves it. The voter lied, and the 
consequence of that lie might fall upon him.

Range never forces a voter to lie, merely, under severe conditions, 
to submerge relatively minor differences in favor of large ones. If 
the election is Adolf Hitler, Gerald Ford, and Hubert Humphrey, and I 
fear that Hitler might win, how important is it that Ford win over 
Humphrey or vice-versa? The fact is that I really should *fully* 
support both of them in that environment. Aiming for minor values in 
the presence of major threats is not a great idea.

If Hitler is known to have no reasonable chance of winning the 
election, *then* I can be choosy! I don't really worry about the 
American Nazi candidate for President.... at least not at present!