[EM] RV comments

[Abd ul-Rahman Lomax]

At 07:20 AM 7/20/2007, Michael Ossipoff wrote:

>First the desirability argument, and then the meaningfulness argument.
>
>Desirability:
>
>The Rangers believe in social utility, its maximization, as the 
>literal be-all and end-all of criteria.

Ossipoff is arguing as if we have a political situation, "our" side 
and "their" side....

If there are a handful of Range Voting supporters who believe as he 
claims, well, I'd be hard put to name more than one. The one of us 
who seems to have such a "belief" is someone quite like Ossipoff: a 
tenacious and highly argumentative individual. Most of us, though, 
see "social utility" as a *measure* of election success, which is far 
more than a criterion. Election criteria, generally, are pass-fail 
measures. A method either satisfies the Condorcet Criterion or it 
does not, at least if the Criterion is sufficiently well-defined.

But SU, properly considered, can *measure* the *relative* success of 
election methods. And, frankly, I don't see, yet, any alternative, 
beyond assuming that this or that Criterion is superior. Yes, there 
is what we might call an SU Criterion, which would be defined, 
perhaps, as "does the method always choose the SU maximizer?"

But no method on the table satisfies that. If, somehow, we could 
force voters to vote true, absolute utility, by some commensurable 
measure -- a rather difficult concept in itself -- then Range would 
*by definition* choose the SU maximizer. The closest we can get to 
this would be "elections" where the outcomes have clear economic 
consequences, with voters who are somehow motivated to vote 
accurately, and even in this, there is a problem: a dollar for a poor 
person is worth more than a dollar for a rich one. How to make the 
utilities commensurable?

But this doesn't mean we can't use utility as a measure, in 
simulations. We do, in fact, use them by assuming that we know the 
true utilities, and we use distributions of utilities that, with 
varying degrees of sophistication and accuracy, reflect what happens 
in the real world. These are simulated voters, and, since we control 
the simulation, we know -- because we assume it -- the underlying, 
honest and accurate utilities.

Then, from this, we predict the voter's behavior in an election, 
using various known strategies, from direct voting of absolute 
utilities (and, yes, there is a way to simulate absolute utilities; 
we essentially equate, between voters, the absolute worst outcome 
with the absolute best outcome, among *all* possibilities, not just 
the particular candidates in the immediate election, for each voter.

In the end, though, this is just a measure, and it is possible that 
there are other measures. There is an "Approval" measure, but it is 
quite problematic in itself. The Approval measure maximizes 
"approval" of the candidate. While I haven't see this from Approval 
advocates, I would define this as the voter would rather elect the 
candidate than see the office go vacant.

However, if we want to test how voters behave with regard to the 
Approval measure, aside from using real-world polls, we would need, 
probably, to do the *same* kind of simulation as was described above, 
because the various utilities generated randomly in the simulation 
can be used to estimate whether or not the voter will approve of the candidate.

However, Approval is, in fact, a Range method. We can expect that, 
generally, the SU maximizer will also maximize Approval, though bare 
Approval, being binary in expression, loses data and thus cannot be 
accurate with respect to *relative approval,* i.e., Approval does not 
discriminate between two candidates where the voter disapproves of 
both of them, thus, if the ultimate contest is pairwise between those 
two, the voter's opinion is utterly disregarded, whereas in Range, it 
is possible that the voter's satisfaction still retains some significance.

Let's see what Ossipoff does:

>  Say, for the moment, we disregard the fact that the SU claims 
> depend on sincere voting, and that sincere voting is nearly always 
> suboptimal in RV.

Ossipoff continually makes this claim. It's false. "Suboptimal" is 
the trick. It is suboptimal, true, from the point of view of the 
individual voter maximizing his or her own personal utility, *in some 
scenarios.* In others, it is clearly optimal to vote "sincerely."

There are a series of problems. First of all, "sincere" vote requires 
some serious work to define. It is far from clear what a "sincere" 
Range vote is. We can defined a "clearly insincere" vote as being one 
where the range ratings reverse preference. But this is never 
strategically forced in Range; if a voter does it, it would be for 
reasons other than trying to get the best outcome, since you never 
hurt your expectations by voting your favorite maximum, or, indeed, 
by voting in reverse preference.

What Ossipoff is talking about is voting "Approval style," in Range. 
But this isn't "insincere," it merely does not disclose the voter's 
full and accurate estimations of personal satisfaction with the 
outcome. If I prefer A>B>C, and those are the only options, it's 
clear that I optimize my expectation by voting A max, B min, but 
where do I rate B? In simulations, we sidestep the problem, 
initially. We have *numbers* for the candidate utilities, as seen by 
the voter, and we *assume* that the scale is linear. That is, we 
scale is forced to be linear; if it were not linear, nevertheless we 
can assume that it is monotonic, i.e., that preferences are, for an 
individual, in sequence such that we never get a Condorcet cycle 
*within an individual voter*. And then we can simply transform the 
utilities from the original "scale" into what is linear, i.e., we can 
add utilities and get a meaningful result.

Where do I rate B? Well, if the B utility is midway between A and C, 
we can define a "sincere" rating of B as 50%. If we have rated A max 
and C min. However, that max and min rating is itself a full 
disclosure of the utilities, the ratings have been normalized to the 
election candidate set, causing loss of absolute utilities.

It never hurts the voter personally to normalize in that way. But it 
might hurt society, overall. Nevertheless, in large public elections, 
we can assume that the ways in which it could possibly hurt would be 
rare, due normally to the kind of fluctuations that can cause damage 
will average out.

But what if the *real* pairwise election is between A and B? And C is 
actually irrelevant, C has no chance of being elected. Well, then I'd 
rate B min also. And in the reverse case, i.e., the real election is 
between B and C, then I'd rate B max. And if the real election is 
between A and C, I might as well rate B sincerely, i.e., 50%, it's 
not going to hurt me.

So far, this is what would be meant by Ossipoff claiming that 
"sincere voting is nearly always suboptimal in Range Voting."

But what if the voter does not know who the two frontrunners are? 
What if the probability of election is equal for all three 
candidates? Well, in this case, the optimal vote is clear: rate B 
sincerely, i.e., at 50%.

So the kind of voting that is optimal depends on the relative 
probabilities of election, as estimated by the voter. Sometimes it is 
what we might call "sincere," but it is *never* insincere. It is 
merely, in some circumstances "magnified," or "truncated." Let me 
explain truncation in this context:

It means that the voter has a set of utilities that range outside the 
election set. Candidates having utilities falling outside the range 
are "truncated" or "rounded off" to the limits of the election 
ratings. This is *not* insincere, rather, the voter is not normalizing.

>  Even then, even in principle, RV advocacy is really only based on 
> a subjective personal opinion.

That's a subjective personal opinion, and not only about election 
methods, but about people, and Ossipoff is a notoriously bad judge of people.

>  The opinion that it's more important to maximize the sum of 
> everyone's happiness than it is to minimize the number of people to 
> whom the outcome is unacceptable.

It's true that it is an opinion; however, it is an opinion that is 
almost the definition of sanity; not caring about the satisfaction of 
others is sociopathic.

Ossipoff is here slipping in "the number of people" argument, which 
is based on the old assumption that preference strength doesn't 
matter, and this is *clearly* false if we look at how small groups of 
people cooperate and coordinate their activities. Ossipoff doesn't 
understand people!

"Unacceptable" is an expression of strong preference. Range, dealing 
with strong preference, is Condorcet compliant. So what in the world 
is Ossipoff talking about?

>  There's a good case for saying that opinion is wrong. Do we really 
> want to make the outcome unacceptable to more people, as long as, 
> by so doing, we increase the benefit for someone already 
> well-benefited more than we reduce it for those to whom we make it 
> unacceptable?

Range maximizes, as well as we can tell, "the number of people who 
find the outcome acceptable." Ossipoff is slipping in his argument by 
avoiding defining what "acceptable" means.

And voters in Range are totally free to vote Approval style; if they 
do not, they are clearly granting permission to elect someone whom 
they rate at more than minimum. If I find a candidate totally 
unacceptable, I'm not going to give that candidate one point! Why should I?

Now, suppose that there are two such candidates, B and C, and they 
are both totally awful, but, hold a gun to my head, I prefer B. But 
what my estimate of election probabilities makes it clear to me that 
either B or C will win? In either Approval or Range, I have a motive 
to vote for B. Range gives me *slightly* more flexibility.

Approval and Range are generally the same method, with Range giving 
the *option* to the voter to vote with more resolution. Whether or 
not this resolution is useful depends on the election context.

However, we do know that if voters vote sincerely, the *overall* 
outcome is improved. Not always the individual outcome; however, this 
is the point that Ossipoff will not mention:

When the individual outcome declines by voting sincerely, it always 
does so by a small amount, not a large amount, and that decrease is 
generally more than matched by an increase with other voters.

Now, if it cost you $1 to create a $1000 benefit for your friends, 
what would you choose?

Elections come and go, and we may get our personal maximum on one and 
not in another. If we act to maximize overall benefit, *usually*, in 
the end, we personally benefit.

If voters vote as Ossipoff would have them vote, we have the tragedy 
of the commons. In each individual election, it seems that I'm 
maximizing my outcome, but if everyone acts that way, we all are losing.

I'd recommend googling Traveler's Dilemma for a very good explanation 
of this, and some real studies have been done on how people behave in 
such circumstances, where individual game theory predicts one kind of 
behavior, and real people behave quite differently, the result being 
that, usually, everyone benefits.

>First let's answer the Rangers' claim that there's no such thing as 
>acceptable or unacceptable-- (because there are only varying degrees 
>of utility).

We don't make that claim. What we would claim, instead, is that 
acceptability and unacceptability are relative. Just like Range 
ratings, the acceptability of a candidate depends on the universe of 
choices possible.

We can use "varying levels of utility" to predict acceptability, to a degree.

>  But the voter hirself can answer that for you, when s/he accepts 
> or rejects a candidate's offer to govern in hir behalf, on hir ballot.

However, if the election rules are going to produce a winner, the 
voter, by rejecting the candidate, may simply be making his or her vote moot.

And, yes, the voter can answer. In Range just as well as in Approval. 
If a candidate is not acceptable, period, the voter should vote 
minimum rating, period. That's what the voter will do in Approval, 
and that is what the voter will do in Range. Same system, really. 
Only difference: the voter is forced to *only* vote the extremes. 
Forced. Coerced. Required.

Why?

It's like IRV claiming that it always elects a majority winner. Sure 
it does. By either discarding votes that are exhausted, or what is 
almost as bad, considering a preference between two candidates 
collectively in last place as a vote *for* the candidate. I'd say 
that was a vote *against* the candidate, but, yes, it is better that 
it be counted; but it doesn't mean that a majority actually supported 
the candidate or found him acceptable.

Range methods may include an explicit Approval cutoff, where the 
voter can clearly define the acceptability level. And we have suggest 
that where majority approval of a candidate is not clear, there be a 
runoff, which is already common practice in analogous situations; 
only here the runoff would be between the Range winner and a preference winner.

I go further than Ossipoff: I claim that no candidate should be 
elected without majority approval. Just as he says we should do, if 
approval is not clear from the original ballots, *ask the voters*! 
But ask them explicitly, not under circumstances where their vote may 
be strategically forced.

(I favor Asset Voting, where a series of runoffs becomes quite 
practical and easy; indeed, it becomes possible to use full 
deliberative process to choose the winner, and this is actually 
better, we can expect, than Range alone at maximizing SU.)

>  As for why the voter does that material act of accepting or 
> not-accepting, that's none of our business.

Hogwash. If the voter did it because there was a gun at their head, 
we care. It wasn't approval. You want approval of a candidate, ask 
the voter in a context where approval is the only possible 
interpretation of the vote.

Approval as a method adds nothing to the power of the voter, the 
voter has exactly the same power in Range. That is what is so bizarre 
about Ossipoff's claims, which he has been making over and over on 
the Approval Voting list. He apparently doesn't dare to do it on the 
Range list, where he knows he will face real consideration and real 
arguments, including some from experts.

>I propose that you give me all your savings. What? You say you'd 
>have to move into the slums? Yes, but my investment broker could 
>double your money for me, and it would allow me to trade up to a 
>better Mercedes. Don't be so selfish. The sum of our combined money 
>would increase when my broker doubles your money for me. I only want 
>the greatest overall summed good for us! Maybe I'd let you ride in 
>my new Mercedes sometimes, to show you how much we've gained.

If I believe you, I should give you the money, if I possibly can. But 
I don't believe you. What an idiotic argument!

No, Range is not suggesting we give all our money to some investment 
broker. Rather, it *allows* voters to express weak preference. It is 
never in their interest to do so if their preference is strong, but 
they are not required to express weak preference, they are the ones 
who decide what preferences are weak -- don't really care -- and strong.

There is no claim, at least not by us, that voters are being selfish 
if they vote "approval style." It's a red herring.

However, we do know that the more accurately voters give the method 
about their true preferences, the more accurately the method can 
optimize overall satisfaction. If a method only allows one preference 
gap, i.e., it is Range 1, then the information which can be provided 
to the method is quite limited and quirky. Approval, Range 1, with a 
runoff is better than Approval alone, because the original votes 
don't have sufficient resolution.

Runoffs also, we think, will encourage voters to vote more 
accurately, but this is not the time to make that full argument.


>Meaningfulness:
>
>SU from ballots means nothing if the ballots aren't sincere, and, as 
>I said, sincere RV voting is suboptimal.

False. False. Tired of explaining.

>Some Rangers have claimed that they found out from 
>(inadequately-described) simulations that, if the percentage of 
>strategizers is below some particular number, then even RV's sincere 
>suckers will be better off than they would have been with Approval. Nonsense.

Because Ossipoff says so, not because he has actually investigated 
how the simulations are done. I'll acknowledge, Warren's information 
about the simulations is spread out, I haven't found it easy to find 
it all in one place. But there *is* a paper on it, which is pretty 
thorough, and, besides, the source code is available and Warren is 
known to answer questions. Why doesn't Ossipoff ask, if he doesn't 
understand how the simulations are done? Instead, he just, without 
knowledge and making a host of assumptions, simply calls them nonsense.

*That's* nonsense! And I don't have time for more of this nonsense 
today. I think it was just gettting interesting, so ... later.