[EM] RCIPE version 2

[Kristofer Munsterhjelm]

On 7/22/21 5:31 PM, VoteFair wrote:
> On 7/22/2021 6:04 AM, Kristofer Munsterhjelm wrote:
>  > How about this?
>  >
>  > - Eliminate the candidate with the least number of winning subgroups.
>  > - If there is a tie, break that tie by IRV.
>  > ...
> 
> Isn't the first step basically Copeland's method?

No, because there's no elimination in Copeland (and it doesn't pass 
LIIA). It would just elect the candidate/s with the most winning subgroups.

> That's an ugly "method" that fails to look beneath the surface.
> 
> IRV also fails to look beneath the surface, which is why it too is an 
> "ugly" method.

That leads me to wonder which is the case.

You said you couldn't replace the IRV tiebreaker with minmax elimination 
because IRV is cloneproof and minmax is not -- that clone independence 
was important because it "protects against money-based vote splitting 
tactics". So I found something that invokes IRV's clone independence 
more often.

But then clone independence is not important after all because the 
methods are ugly. I can't quite determine whether clone independence is 
important or not.

>  > But again, the ungrouped mechanic is not cloneproof.
> 
> Being cloneproof is not a goal. The goal is to have a very small failure 
> rate for clone independence.

Then you could check the alternatives by that metric. A method seeming 
ugly may not necessarily have any bearing on the rates of failure.

> Also, electing the Condorcet winner is not a goal. The goal is to have a 
> very small Condorcet criteria failure rate.
> 
> To repeat my concern, attempting to get a zero failure rate will cause 
> other kinds of failure rates to increase.

That's true. You implicitly need some kind of valuation of the different 
failure rates. For instance, if you want LNHarm and LNHelp, you have to 
give up either monotonicity or mutual majority. Which it's going to be 
depends on what values you place on the different criteria.

The same would hold for rates. Say you want to find the method that 
minimizes w * x, where x is the rates of each failure type 
(monotonicity, vote splitting, teaming, crowding, favorite betrayal...). 
Then the weights of the w vector provide a measure of indifference: how 
much of failure type 1 is an acceptable trade for one unit of failure 
type 2?

Or to put it differently: if the method insists on a zero failure rate 
for Condorcet loser, why shouldn't it insist on a zero failure rate for 
Condorcet winner, say? And, equivalently, if "merely a low rate of 
failure" is good enough for the Condorcet criterion (or say, clone 
independence), why is it not good enough for Condorcet loser?

> I'm still willing to consider improvements, but it needs to find a 
> balance between what voters can understand -- both through an animated 
> video and through words -- and what yields low failure rates.
> 
> Again, thank you Kristofer for applying your clear understanding to this 
> revision from RCIPE 1 to RCIPE 2.

You're welcome :-)

-km