[EM] IEVS 2.56 update on findings, status, deficiencies

[Warren Smith]

IEVS 2.56 now includes 11 different utility generators,
based either on random normal numbers, or on positions in an
I-dimensional issue space for I=1,2,3,4,5, and utilities based on
either a function of distance, or dot product, between candidate and
voter issue-stance-vectors.

It also includes over 50 different voting methods.

Also it includes adjustable voter "ignorance" and
adjustable voter strategy-vs-honesty mix.

Unfortunately all that means the damn thing now takes about a week
to run and produces about 100 Megabytes of output. I need to do
something about that.

The results so far [PRELIMINARY] are
that range voting is always at the top of
the heap - meaning its Bayesian regret is the least of all voting
methods (of those programmed, and up to noise comparable to one
standard statistical error) EXCEPT for the following:
* Range+top2runoff (and sometimes approval with top-2-runoff,
and perhaps sometimes HeitzigDFC, but those both appear to be
dominated by range+top2runoff and HeitzigDFC appears to be dominated
by Approval+top2runoff) beat plain range's Bayesian Regret in
circumstances with enough strategic voters in the mix.
Note the top2runoff part is honest voting even for strategic voters.
The crossover point is approximately at a 50-50 honest-strategic
voter mix.
* Top median Rating and maybe MCA are occasionally up there with the
big boys, although usually not and never clearly surpass them.
MCA seems to dominate approval, although it is often exactly the same
(and is more complicated).
* There may be some future addenda to this list as IEVS grinds further.

The really bad voting methods at the bottom of the heap
(usually near the bottom and sometimes with over 3 times as
bad Bayesian Regret as the best method) include
HeitzigLFC, RandomWinner, Hay, RandomBallot, ContinCumul,
LoMedianRank, AntiPlurality, Plurality, VenzkeDisqPlur, Bucklin,
BaseballMVP, Coombs, Dabagh, Nauru, and Top3IRV
(listed in a roughly improving-quality direction).

If there are enough strategic voters, then
Range and Range+top2runoff both give a higher probability than
everything else (including "Condorcet methods") of selecting
a (true-utility-based) Condorcet winner. The crossover point is
approximately at 70-30 honest-strategic voter mix.

Mike Ossipoff has pointed out that the theorem in
http://rangevoting.org/AppCW.html
which explains how (a certain plausible kind of) strategic range
voters can generate a Condorcet winner with 100% probability, is
really part of a larger claim that EVERY voting method, with informed
strategic voters, will generate a Condorcet winner with 100%
probability. (The AppCW file now includes a note to that effect.
This may have been pointed out by Niemi and/or Riker in the 1970s.)
However, Ossipoff's claim is not quite true.
It is true for a large class of (but not all)
voting methods, and for certain kinds of strategic voter behavior
within those methods (how plausible those behaviors then are, is
subject to debate), yes. Anyhow, the IEVS simulation
is finding range and range+top2runoff are ESPECIALLY
good at generating honest-condorcet-winners. The other methods,
contrary to the Ossipoffian all-inclusive view, are not able to
equal range and range+top2runoff in that respect.

Future plans (in the rough order I plan to do them):
1* I want to build a "reality based utility generator" based on
Tideman's dataset of 87 real world elections (plus some more
elections). Your contributions?
2* Add more voting methods. Your suggestions?
3* Add more voter-strategies. Your suggestions?
4* Add stuff to permit rank-order WITH EQUALITIES PERMITTED ballots.
That will also interact with 2 and 3, unfortunately.

If you want to help: the more work you do, the more help it'll be
and the more likely I'll incorporate it.
I.e. providing an election in a NICE FORMAT is a hell of a lot better
than just pointing me to some election data somewhere in the format
from hell. (Burlington and Ireland = hell.)
Nice formats include something like the format
Debian uses for their elections (e.g. see
http://www.debian.org/vote/2005/leader2005_tally.txt ),
or Tideman's format (http://rangevoting.org/TidemanData.html).
Debian format permits equal rankings, Tideman does not.

Providing actual or pseudo-code for a voting method is a hell of a lot
better than just babbling, and the more detailed the pseudo-code, the
better.

The present IEVS code is available as always at
http://rangevoting.org/IEVS/IEVS.c

Warren D Smith
http://rangevoting.org
warren.wds AT gmail.com