Voting system R&D (Re: 2017 update to the SPI voting algorithm for Board elections)

[Barak A. Pearlmutter]

> You are advocating range voting.  I remain convinced that range voting
> is a terrible voting system, because all but the most tactically aware
> voters will cast hopelessly ineffective ballots.  This criticism
> applies less to approval voting, but approval voting still involves a
> lot of guesswork for voters.  Many people will cast ineffective
> approval ballots.

Why do think that is the case? I do not see any evidence for it.

Range votes, or in their most simplified form approval votes, are at
least monotonic. You know that giving someone a higher score, or
marking them APPROVE, cannot lead to their losing when they would
otherwise have won.

This is not the case with IRV or STV. Tactical voting is
extraordinarily important with IRV and STV, and at worst a very minor
detail with range voting or approval voting.

> But the key point, as discussed, is that SPI is poorly equipped to
> analyse voting systems.  SPI is full of technologists.  We largely
> lack political scientists, electoral officials, constitutional
> engineers, and historians; we're probably even short of game
> theoreticians.
>
> (I have observed that technologists, particularly some computer
> people, have acquired a kind of hubris that means they think they are
> good at everything, and don't recognise the difficulty, complexity, or
> value, of other fields of learning.  It's fine to be a polymath, but
> that mostly means knowing how much there is you don't know.)
>
> We should defer the question of voting systems to well-regarded civil
> organisations for whom these questions are the primary focus, and who
> are thereofore more competent: that means voting reform groups.
> Almost uniformly, such groups recommend STV for multi-winner
> elections.[0]

I have looked at the actual technical data supporting those
recommendations, and I find the particular groups you advocate to be
very unconvincing in this regard. I'm not saying everyone needs to be
an expert and actually look at the data. But surely *someone* should!

> The Secretary's current practice is to publish the tally sheets which
> enable re-analysis.  Do we really need that to be Board-mandated ?
> asked this question a few days ago, proposing a paragraph codifying
> existing practce, and there seemed little enthusiasm for it.

Unfortunately, tally sheets, at least as usually defined, do not
contain all the information in the ballots, and are insufficient for
purposes of deep analysis. Debian releases the actual ballots (without
any identifying information of course.) I would suggest SPI follow
suit.

> It is not surprising that AV produces somewhat different answers in
> some close-run DPL elections.  ... The analysis you pointed me to in
> August, of the Burlington 2009 election, IMO is devoid of
> understanding of the political context

You are, I think, being inappropriately dismissive of problems that
actually arise in practice.

In the Burlington election, there were three main candidates: A, B,
and C. With a hefty majority of ballots preferring A to C, and a hefty
majority of ballots preferring B to C, the IRV system declared C the
winner. Do you seriously not realize that this is very bad behaviour?

> Going on to those references you provided in August.  They were to
> people who advocate range voting for single-winner elections.  As I
> said in private email, I find it difficult to take seriously anyone
> who proposes range voting for single-winner elections.

Then and now, I do not understand your antipathy to range voting or
approval voting.

Arrow (you know, "Arrow's Theorem", Nobel prize) advocated range voting.

All voting systems we might consider here do require strategic voting
for maximally effective votes. But this is a much much worse problem
with STV than with RRV. Due to non-monitonicity, maximally effective
STV votes will often have reversals, where A is ranked above B for
strategic reasons even though the voter actually prefers B to A.

Moreover, rank ballots are less expressive than numerical scores. With
a rank ballot, someone can say

    Johnson > Clinton > Trump

but with a range ballot, then can express the strength of their preference

    Johnson: 99
    Clinton: 95
    Trump: 0

Expressing actual preferences this way, assuming people actually do
so, has been shown in simulations to give superior election results.
So either people vote honestly and we get superior results, or they
vote strategically and we still get very good results, the same as
with approval voting.

I really don't understand why you're advocating STV when it has a long
list of serious problems not just theoretically but which seem to
appear quite often in actual practice.

--Barak.