A lot of good ideas have died on the vine because clones were ignored. The three best ones are Borda, Copeland, and Kemeny.
The key to fixing them is to make appropriate use of lotteries that re-distribute the probability of an alternative to the members of the clone set that takes its place. We have mentioned a few of them and have shown how to use them to fix Borda and Copeland in recent posts to this EM list. Some people will quit reading at the first mention of probabilities because they think the only use for probabilities is to introduce randomness into the determination of the outcome ... but that confuses the probabilities with the random variables; for example let X be one or zero depending on the outcome of a coin toss ... heads or tails respectively. The X is a Bernoulli random variable with parameter p, the probability of "success" and q equal to 1-p, the probability of "failure." If the coin is "fair" then both p abd q are equal to 50 percent. These parameters p and q are not not random variables ... any more than the fraction 1/2 is a random variable. Once all ballots have been counted, the probability that a random ballot ranks alternative A in first place is just a well defined number ... the fraction or percentage of the ballots that rank X first. Just because we use probabalistic language to define a number doesn't mean that number is going to make our method outcomes random. In all of my recent posts referring to lotteries the only purpose of the lotteries has been to define deterministic probabilities for use in deterministic methods. So in this post to be definite let P(X) be the probability that X will be chosen by randomly drawn ballots .. if the first ballot is ambiguous, additional ballots are drawn at random to narrow down to one alternative ... this is a well defined Markov process, so for any given set of marked ballots, and candidate X the probability P(X) is a well defined number between zero and one, a number that can be calculated by elementary arithmetic. Without further ado, let's de-clone Kemeny. Kemeny finds the ranking of the candidates that minimizes the of the distances from that ranking to the ballot rankings. The problem with classical Kemeny is that it uses a clone dependent distance metric, simply counting the number of adjacent pairs in one order that need to be swapped to transform it into the other order. When you replace one alternative with a clone set it increases the number of needed transpositions, thus distorting the metric. Here's how to do it right: 1. Construct the pairwise matrices M1and M2 for the respect orderings (rankings) of the alternatives: The entry in row i column j of M1 is a 1 or a zero respectively, depending on whether the first ranking ranks alternative i ahead of j. M2 is defined similarly from the second ranking. 2. Let M be (M1 -M2) minus its transpose. 3. Remove all of the minus signs from the entries of M and divide it by two to get the matrix A. 4. Let D be the matrix product bAb', where b is the benchmark lottery in the form of a row vector and b' is its transpose. This number D is the decloned distance between the two rankings. Let's say that the distance between two ranked preference ballots is the distance D between their rankings as defined above. The pairwise matrices used above in no way required complete rankings, so the distance is well defined between any two ballots, complete or not! Kemeny is just one potential use of a decent metric on the ballots. How would you use it for designing a PR multiwinner method?
---- Election-Methods mailing list - see https://electorama.com/em for list info |
On 04/03/2021 08.50, Forest Simmons wrote:
> A lot of good ideas have died on the vine because clones were ignored. > The three best ones are Borda, Copeland, and Kemeny. > > The key to fixing them is to make appropriate use of lotteries that > re-distribute the probability of an alternative to the members of the > clone set that takes its place. We have mentioned a few of them and have > shown how to use them to fix Borda and Copeland in recent posts to this > EM list. I think this is related to a pattern that often shows up: that when you generalize certain methods that reweight candidates, you end up reweighting the voters instead. Consider, for instance, party list. Traditional Sainte-Lague type party list works by electing a candidate from the party with the greatest (adjusted) support, then recalculating that party's adjusted support to be penalized by 1/3, 1/5, 1/7 etc. Now suppose you want to generalize this to Approval. Then there's no easy way to adjust the parties' support because a voter may vote for many different parties at once. So what does proportional approval voting do? It reweights the voters instead. However much overlap there may be between candidates, a voter is always a single voter. So this sort of "probability of voters doing X" perspective may be useful more generally. ---- Election-Methods mailing list - see https://electorama.com/em for list info |
Free forum by Nabble | Edit this page |