[EM] correctionish. Utility normalization issue

[Kevin Venzke]

Hi,

I said in past messages that IRV and related methods were electing
the utility maximizer relatively often in the -30 +24 +34 scenario.

This turns out to be dependent on whether voters normalize their sincere
utilities. I have generally been making voters do this, because
1. it seems more democratic
2. it affects nothing other than utility performance metrics, with no
effect on strategies or anything
3. if sincere utilities aren't on comparable scales then the task of
guessing at the utility maximizer seems unnecessarily futile.

However, on these spectrum-based scenarios normalization seems to make
the outcome too dependent on which candidates are nominated. Here is
a simple example. Say A is at -10, B is at +10, C is at +99. Voters are
at -99, 0, and +99. At a glance it looks like on average, the utility
maximizer will be A or B about evenly. However, if the voters normalize
their sincere utilities and you aggregate that, the maximizer is actually
A. The -99 and +99 voters' dislike for B is slightly greater than the +99
voter's dislike for A.

Maybe someone would disagree, but it seems to me it should be a tie.

After a quick test without normalization, in the -30/24/34 scenario,
it looks like most methods plummet to about 60% efficiency. The best
method is somehow DSC-er, at 73%. It elects A (-30) relatively often...
Not very good sincere Condorcet efficiency.

Kevin Venzke