[EM] Outcome Design Goals

[Abd ul-Rahman Lomax]

At 09:22 AM 7/1/2013, David L Wetzell wrote:
>Some thoughts.
>
>1. You need to consider the difference between Cardinal and Ordinal Utility.
>You presume the existence of Cardinal utility.

First of all, who is "you." David is writing to an entire list. *Who* 
"presumes"?

In the study of Bayesian Regret, we do *not* presume the existence of 
cardinal utility, we *assume it.* It's a device for studying voting 
systems, to see how they perform when we know *summable cardinal utilities.*

>   Ordinal utility can be monotonically positively transformed so 
> long as it preserves the order.  For example, if the original scale 
> is between 0 and 100 then one could randomly generate a real number 
> a from the normal distribution, transform it by taking b=E^a, and 
> then transform the utility to become X^b *100^(1-b).  This would 
> not change the rank-orderings of candidates, but it would change 
> the approval or score-ratings given to candidates and it'd muddy 
> the water about your example.

Well, we aren't looking at the example. However, there is a radical 
misunderstanding of how cardinal utilities are used in studies. The 
choice of representational range is a study option, and it's 
arbitrary. Utilities may be represented in the range of 0 to 1, or 
using any real number, positive or negative. What matters is the 
assumption: these are *absolute* utilities and the scale is *assumed* 
to be commensurable. Often "heaveh-hell" utilities are used. These 
represent *actual motivational* preference. I.e, assumed. The 
absolute utilities that underly a study assume commensurability, and 
incommensurable utilities would be adjusted so that they are commensurable.

This system makes a basic assumption that all voters are equal, 
equally deserve to be pleased. So the *full range* of pain/pleasure 
for these voters is made equal.

The utilities in the example that David is referring to seem to have 
been utilities on a scale, perhaps, of 0 to 100. These are utilities 
that have been transformed to representation on a scale of 0 to 100. 
They are no longer absolute utilities in the sense described, but we 
can still posit such utilities, for simplicity, and see how a voting 
system performs.

Take-home point: we are not "presuming" commensurable utilities, we 
are "assuming" them for purposes of study. David is falling into a 
classic error, *presuming* that to be useful, utility must be real. 
What we are doing when we assume a set of utilities is saying that 
*if there are utilities like this,* then this result is better than that.

Something else is done. If the value stated is *actual range votes,* 
then those will have been translated through voter strategy. They can 
no longer be assumed to represent commensurable utilies. Without 
knowing the example used, and how it was used, I can't know what was 
being discussed.

>2. In real life, parties/candidates choose their 
>candidates/positions to enhance their likeability, so what is taken 
>as exogenous in most of our thought experiments are in fact 
>endogenous.  This is also another reason why it's hard to grow the 
>number of competitive candidates, because good ideas from a 
>not-so-competitive candidate will tend to get coopted by 
>already-competitive candidates.

David is taking a *very simple situation* and making it quite complex 
and difficult-to-understand. He's just making up a story about what 
happens, which might have some effect, and it might not. He is 
aiming, it appears, to make some claim about the numbers of 
candidates, claiming that it is "hard to grow" them, when we don't 
give a fig about "growing the number of candidates." It will happen, 
that's what we know. It's like clockwork, under some election 
conditions. People like to run for office, and it's commonly easy to 
get on the ballot in nonpartisan elections. David is mixing up 
partisan and nonpartisan elections, smearing it all together as if 
they were the same.

Presently in many places we see minor candidates, and minor party 
candidates, with few votes. They do this in spite of the spoiler 
effect and vote-splitting. Top two runoff, in some places, causes 
massive spawning of candidates/parties. It's obvious. IRV will do the 
same, and so will any method that addresses the minor spoiler effect. 
The major effect is vote-splitting, which becomes a serious problem 
for IRV as a minor party approachs or just exceeds parity. Since IRV 
will encourage the growth of minor parties, that can be predicted -- 
and it would happen with other systems as well -- we can see, then, 
that IRV is likely to eventually create the conditions for its own failure.

FairVote, had they been capable of intelligent strategic thinking, 
would *never* have chosen Burlington as a place for IRV. Bad Idea. 
But Terry Bouricius was there, and they had LWV Vermont in their 
pockets. They were screaming when the move was being made to drop 
IRV. Quite simply, they were not listening. They had stuffed their 
fingers in their ears, they were not going to recognize a problem 
with their baby.

>3. It seems that your self-described magical certainty as to voter 
>prefs is at worse chimerical and at best a useful fiction.

Again, this is that basic analytical error. David is not 
understanding what he sees. He is seeing an "assumption" of voter 
preferences, i,e., of underlying utilities -- which then define and 
create preferences and preference strengths. It is not a "magical 
certainty," it is an assumed condition. It is not intended to be an 
actual representation of real-world utilites, merely as a possible 
pattern. Unless it was. If this was properly done, for that, the 
assumption would, again, be explicitly stated.

>   It also is similar to the work by Warren Smith using 
> <https://www.google.com/webhp?sourceid=chrome-instant&ion=1&ie=UTF-8#sclient=psy-ab&q=warren+smith+bayesian+regret&oq=warren+smith+bayesian+regret&gs_l=hp.3..0i22i30.102.153444.0.153747.35.30.0.0.0.1.365.5918.0j20j8j2.30.0.epsugrpqhmsignedin%2Chtma%3D120%2Chtmb%3D120..0.0.0..1.1.17.psy-ab.FsjX8OuCjSc&pbx=1&bav=on.2,or.r_cp.r_qf.&bvm=bv.48572450,d.aWc&fp=e16824eab702e431&ion=1&biw=1366&bih=659>Bayesian 
> Regret, which also presumes cardinality of utilities.

No, it *assumes* this. Arrow made this error years ago, but he 
clearly moved beyond it. Simulated commensurable utilities are used 
to test the performance of voting systems with known utilities. The 
real assumption, involved in our thinking that these studies have 
real application, is that a voting system will then perform similarly 
with unknown utilities.

If a voting system can't perform well with clearly commensurable 
utilities (such as payoff values in some lottery or cash benefits, 
etc.) then why would we expect it to perform well when we don't know these?

It's important, of course, to be very careful about how simulated 
utilities are translated to actual votes. Again, that's all invented. 
Is it realistic? We debate about that. But Bayesian Regret 
simulations are the only relatively objective tool we have to study 
voting system performance. There are also, in a similar fashion, Yee 
diagrams. It is obvious why IRV supporters attack the simulations: 
they show how bad IRV really is.

We knew from simulations that what happened in Burlington was a real 
possibility. FairVote denied it, claimed that the scenarios that 
would lead to monotonicity failure and the like were terminally rare 
and therefore irrelevant. And so it actually happened, as could have 
been predicted. There was nothing magical about it.

The assumption of commensurability is fundamental to democracy, it's 
behind the one-person, one-vote principle, that values the 
contribution of each voter equally.

>   I think that there might be scope for probabilistic valuing of 
> candidates so that in examples like what you give, a certain 
> candidate has a certain probability of getting elected.

That can all be incorporated into simulations. Better study of 
strategic voting in Range and other methods has yet to be done, the 
work so far has been primitive.

>[...]

>4. If there is a multi-dimensional issue-space (plus a 
>je-ne-sais-quois character dimension) then these inevitably would 
>tend to get collapsed into a single-dimension and the issue is in 
>large part how the different dimensions get collapsed, plus the 
>inevitable problems with noise being propagated by the campaigns and 
>voters putting less time/energy into researching all candidates so 
>and so forth.

It can all be simulated. However, the point is not to accurately 
simulate real elections, but to simply come up with some reasonable 
distributions of preferences and voting strategy. It doesn't have to 
be complex to work.

>   I favor IRV in part because it lets non-competitive candidates 
> bring up otherwise neglected issues and subvert some of the noise 
> spewed by the main candidates.

Any advanced voting system will accomplish what is real about this. 
Much of the prediction about how IRV will, for example, stop negative 
campaigning, is just made up. It does not appear to correspond with 
reality. So two relatively compatible candidates, minor candidates, 
campaign together. That's happened, but it isn't common. And the 
major negative campaigning comes from frontrunners. They don't attack 
the minor candidates whose supporters they might want to woo, they 
attack the other frontrunner.

>These are things that are hard to model, they are quite indeterministic.

David, you are presuming there is a need to model real behavior. You 
are using this to deny a presented model, an example, I'm assuming, 
of method behavior, given certain assumptions. Sometimes I address 
these by looking at what utilities might underlie give voting 
patterns in examples. Sometimes I assert that a scenario is extremely 
unlikely or even that it is self-contradictory. However, here, David, 
you are attempting to undermine the entire approach of using assumed 
utilities to study voting system performance.

It's a head-in-the-sand approach.

>   The same is likely true for other rules, but their probabilities 
> of getting adopted in the short-run are low and my args are geared 
> not at proving IRV is objectively better than the other rules, just 
> that the diffs between the alternatives are less strong than often 
> purported and so it makes more sense to focus on short-term 
> probabilities of adoption in the short-run at least.

I prefer the continuation of plurality to the adoption of IRV. I 
would *not* oppose STV multiwinner implementations, I expect. Why do 
I prefer plurality? Because Plurality is honest. It's obvious what is 
happening. Voters know the ropes, they know that by voting for Nader, 
they might allow Bush to win. Nobody is surprised by that. But IRV 
Republicans in Burlington, did they know that by voting sincerely for 
the Republican, they were effectively causing the Progressive to win? 
Their worst choice?

When this became clear, I assume, there was then a natural alliance 
between the Democrats and the Republicans, to defeat the Progressive. 
That's why IRV was dumped. The Republicans were more hurt than the 
Democrats, but the Democrats, had the method been fair, would have 
won that election, so they were also hurt. More than half the 
electorate, up to about two-thirds, got a poor result. Roughly 
one-third got the worst result, and another third, roughly, got a 
poor result. Anyone who thought the Progressive was the best result 
would have voted for the Progressive, which I'm stating at roughly 
one-third. The Democrat was the compromise candidate, lesser in 
direct support than the Progressive and Republican, but pretty much 
the first or second choice of everyone. IRV -- like top two runoff -- 
is famous for this problem.

>5. If some people feel strongly on an issue, there are other ways to 
>manifest this preference other than what is given them in the ballot 
>that serve to move swing/low-info voters.  This reduces the mandate 
>to incorporate the capability to express strength of preference in the ballot.

This means something? Mandate? Looks like David is here dissing the 
expression of preference strength, which is what differentiates 
cardinal methods from ordinal methods. But any sane interpersonal 
negotiation system will investigate, in one way or another, 
preference strength. You and I disagree. Okay, *how strongly* and on 
what specific issues? Maybe I can give some weak preference on 
*this*, where you have a strong preference, whereas you give up 
similarly on this *other* issue.

People socialized to cooperate and collaborate readily disclose 
preference strength. People socialized to get as much as possible for 
themselves, who fear that if others know their preference strengths, 
may want to conceal them unless *forced* to reveal them. 
Later-no-harm criterion.

>This is not unlike why I stated that I have epistemic diffs with 
>Kristofer M.  He doesn't have access to much real world experimental 
>data and so to compensate

I'm seeing only noise brought in from that "real world experimental 
data." He means that he's more experienced, without actually having 
"data." Just his collection of judgments and opinions.