Re: [EM] Why is wikipedia so biased pro-IRV?
Bob Crossley wrote: I defer to Walabio's point about "Fair Votes" as I know nothing about US politics but I strongly suspect that they are currently getting a lot of support from across the Atlantic. You may have heard that in May the UK will have a referendum on introducing IRV (we call it AV, as you probably know), this being one of the trading-chips in the current Conservative/Liberal-Democrat coalition agreement. I'm not a UK politics expert, but it seems this is a minimal concession, of the sort one would see in negotiation. AV/IRV doesn't really lead to multiparty systems, if Australia is to be any judge. Instead, you get two large parties and one middle sized party (as in Australia's Labor and LibNats), which is an improvement from Plurality, and definitely so from the point of view of the Liberal Democrats (who could become the middle sized party). AV+ or STV/MMP would have been better, but alas. Election-Methods mailing list - see http://electorama.com/em for list info
Re: [EM] Why is wikipedia so biased pro-IRV?
On Fri, Feb 25, 2011 at 2:28 PM, Kristofer Munsterhjelm wrote: > I'm not a UK politics expert, but it seems this is a minimal concession, of > the sort one would see in negotiation. Also, since the concession was for a referendum not the actual policy, it isn't even a 100% concession. Ironically, it allows the Conservative to say that even those arguing for a Yes, don't really want IRV, even though the Conservatives are the reason that PR isn't being offered. However, if it passes, it will strengthen the Liberal Democrats in all future elections. This might lead to them dropping PR as a policy objective, since IRV gives them a boost while PR will support all parties. There was also a suggestion of having PR for the House of Lords, not sure what happened to that. Election-Methods mailing list - see http://electorama.com/em for list info
Re: [EM] ASCII maps
Kevin Venzke wrote: Hi Kristofer, --- En date de : Lun 21.2.11, Kristofer Munsterhjelm a écrit : Then it tries to come up with a nice map that minimizes inaccuracy. You could try using synthetic coordinate algorithms for mapping the distances to 2D. I did that for competing entries in a programming game, once, using the centralized Vivaldi algorithm as described in pdos.csail.mit.edu/papers/vivaldi:sigcomm/paper.pdf. Another option would be to use principal components analysis, but I know less about that. Oh dear. That looks very complicated. I just used the number of scenarios where the methods agreed. So, the maximum distance between two methods is 27. When the methods are plotted I find the Euclidean distance, subtract from the desired distance, and square it. Add all that up and that's the measure of a poor fit. To improve the fit it tries rerolling a few candidates, or tweaking their positions. FWIW, I dug up that old code and made a map of my own using Kendall tau distance between the different methods' output orderings. The ballot generator was a combination of IIC (every ordering equally likely) and a spatial model - every other round used the one, every other the other. The picture is here: http://munsterhjelm.no/km/elections/Methods.png I don't see any obvious axial characteristics. Perhaps methods closer to the bottom are more indecisive, but it ranks Copeland in the "misc Condorcet methods" blob, ahead of VMedian-Ratings which does try to break ties quite a bit more. Or perhaps methods closer to the bottom make use of less information at once, which could explain why the eigenvector-type methods are near the top. As for the abbreviated terminology I use, I'll explain some of the methods: Eliminate[X] is a loser elimination method using X. Eliminate[Plurality] is IRV. AVGEliminate[X] is a below-mean loser elimination method using X. AVGEliminate[Borda] is Nanson. Keener is the Keener eigenvector method. Maximin is the pairwise (but not Condorcet) method where the candidate who beats whoever he beats by the smallest margin, by the greatest, wins. Minmin is the method where X's score is not the greatest win of some candidate Y against him, for an Y maximizing it, but the least win of some candidate Y against him, for an Y minimizing it, negated. Ext-Minmax breaks Minmax ties by considering the next-to-greatest win and so on. Same thing with Ext-Minmin wrt ordinary Minmin. HITS is https://secure.wikimedia.org/wikipedia/en/wiki/HITS_algorithm . NREM-Opt is Warren's optimal positional system (optimal in context of the every-ordering-equally-likely model, in that it produces the least Bayesian regret of all ranked methods under that model). L-R are variations upon Least Reversal. LR-Defense is the simple least reversal method: least sum of opposing pairwise stats is better. LR-Offense is the symmetric version for offense: the candidate with the greatest sum of pairwise victory margins (or wv, etc) wins. LR-Both scores by the former subtracted from the latter. Median Ratings are just that, either *norm*alized to 0-10 or with raw utilities (which is just the rank index for IIC). VMedian-Ratings break ties by considering ever expanding truncated means when the median fails to provide a result. Mode-Ratings is the silly method where the candidate with the greatest mode wins. It is not monotone, and I thought its strategy aspects and monotonicity weirdness would be similar to IRV's. Apparently not, unless the map's a bad fit. Gradual-Set is a group of methods based on the Condorcet-Borda relation that Stephen Turner pointed out. It starts by calculating the Set (Smith, CDTT, whatever) using the pairwise matrix where A beats B by c(A,B), i.e. median values of what Stephen called s(A,B). This will produce a division; some are considered better than others, but since Set is a set, it won't be decisive, so the method expands the median to a truncated mean and breaks ties with the new ordering found by applying the set method to the new matrix. It continues alternating between expanding and refining the output until either all ties have been broken or there are no more data points to use. Smith is the Smith set. Same for CDTT and CGTT, they are what they say. SDom ranks candidates strongly dominated by any other (as by the Independence of Strongly Dominated Alternatives criterion River passes) above those nondominated. MDD is the Majority Defeat Disqualification set: candidates who would be eliminated in MDDA are ranked below those who wouldn't be. Election-Methods mailing list - see http://electorama.com/em for list info
Re: [EM] ASCII maps
Hi Kristofer, --- En date de : Ven 25.2.11, Kristofer Munsterhjelm a écrit : > FWIW, I dug up that old code and made a map of my own using > Kendall tau distance between the different methods' output > orderings. The ballot generator was a combination of IIC > (every ordering equally likely) and a spatial model - every > other round used the one, every other the other. > > The picture is here: http://munsterhjelm.no/km/elections/Methods.png > > I don't see any obvious axial characteristics. Perhaps > methods closer to the bottom are more indecisive, but it > ranks Copeland in the "misc Condorcet methods" blob, ahead > of VMedian-Ratings which does try to break ties quite a bit > more. Or perhaps methods closer to the bottom make use of > less information at once, which could explain why the > eigenvector-type methods are near the top. That's an interesting picture. I don't really understand the proximities at all. I can see FPP near Antiplurality. And not too far away is Random Pair. IFPP is near Plurality as expected, but totally opposite IRV. Were there tons of candidates I wonder? I'm inclined to guess that Maximin(wv) and the three Gradual methods at the bottom give very strange results. I wonder if you ever tried it with some subsets of methods to see if you got different orientation. Or if having fewer points perhaps makes it easier to find a good map. For example the map I showed you with the huge Borda cloud showed a few of the main methods relocated in relation to each other. I currently think it was based on nothing, because I can plot the main 6 methods on notebook paper and the arrangement and distances make total sense. Doesn't appear that anything is being fudged. (On the other hand, Borda as I had it definitely doesn't satisfy Plurality, and such methods don't fit the north-south test well...) Kevin Election-Methods mailing list - see http://electorama.com/em for list info
[EM] WSJ article about AMPAS voting
http://online.wsj.com/article/SB123388752673155403.html Cites our own Warren Smith! Clearly we've been going about advocacy all wrong. Politics is boring, we should appeal to American's fascination with celebrities and sports. Election-Methods mailing list - see http://electorama.com/em for list info
[EM] Some numbers (LNHs, compromise/withdrawal, burial games)
Hello, I have my generator looking over the method DNA and counting scenarios from which we can find a LNHarm failure, LNHelp failure, compromise/ withdrawal incentive (i.e. the faction would have preferred to not vote for their favorite), and two types of burial games. Let's call them truncation-deterred "TD" and reversal-deterred "RD." In a TD game the attacker will hesitate to bury if he fears the defender faction will truncate support for the attacker. In an RD game the attacker would have to fear that the defender will vote for the candidate intended to be a pawn. In both cases the deterrent is that the pawn may be elected, which neither attacker nor defender want. To be clear, this is a burial game: X does not vote for Z, Y wins, Y votes X>Y; and if X changes to X>Z, X wins; and if Y faction then changes their vote (to what, depending on TD vs RD), Z wins. XYZ can be anybody. I get numbers overall and by faction. Even then I realize some of these numbers may not tell a complete story. For example, one method may have greater LNHarm failures than another method, but we don't see what happens with the new preference when LNHarm isn't failed. I'm including numbers for margins methods, but it produces 7 DNA sequences depending on the exact faction size ratios. I produce them all. Here's a list. I won't break out by faction for everything for now. The format is method name, LNHarm (failure scenarios), LNHelp, compromise/ withdrawal incentive, TD burial games, RD burial games. FPP 0 0 18 0 0 IRV 0 0 9 0 0 DSC 0 3 12 1 1 DAC 9 0 6 0 0 Bucklin 12 0 5 0 0 WV 1 4 2 6 0 C//App 4 4 3 5 0 C//IRV 3 0 6 0 0 QR 0 3 7 4 0 MMPO 0 6 2 6 1 BklnVariant 5 2 6 2 0 C//KH 6 0 5 0 0 KH 6 0 7 0 0 margins 0 6 2 6 1 (= MMPO) margins 1 4 4 2 2 margins 0 6 3 5 2 margins 1 4 3 5 1 margins 1 4 2 6 0 (= WV) margins 1 4 4 4 2 margins 2 2 6 2 1 I was surprised that DSC and MMPO have RD games. In DSC it is: AB, B, CB. A wins. B can vote B>C and win. If A thinks the B voters would be lying, they will have to make B voters fear that the vote will be AC. This is unusual from what we usually discuss because the "attacking" faction B is actually defending the CW, and so could be said to be using *defensive* burial. If we take the margins averages (which we probably shouldn't, as they won't occur with equal frequency) margins is by a hair the best LNHarm- failing method wrt LNHarm. WV places second. WV has a bit less LNHelp, less compromise/withdrawal incentive, more TD games, but no RD games. If we are supposed to expect different voting behavior from WV and margins in the zero-info case, I don't think these numbers suggest it. Condorcet//Approval's burial resistance advantage over WV doesn't look all that great here. You get one less burial scenario, in exchange for a compromise scenario and 3 LNHarm scenarios. QR over IRV: With QR you pay a price (LNHelp and TD burial) for two fewer compromise scenarios. Not that *much* improvement but I'm still pleased. Bucklin vs. my variant is interesting. Although I only aimed to give LNHarm to the A faction, the LNHarm failures are cut by over half. Six of Bucklin's 12 failure scenarios involved A's preference, so those are gone, plus one more. The one more is actually a good example of how the numbers can't show a complete picture. In Bucklin, AB B CA elects B, but the B voters can throw it to A by voting BA. In the variant, B doesn't win in the first place, which is (at least on its face) a worse situation for B than when he is prone to a LNHarm failure. Next I am working on a more thorough strategy solver. For a given scenario (in terms of sincere utilities) I want to have the factions play against each other, so that we can actually see which voting scenarios occur when voters are smart. With these results we should be able to say that such-and-such method (or lottery of methods) "maximized utility" for instance, for the given scenario. Or, we could say which methods were most likely to reduce to two candidates in practice. Or, which methods occasionally produced train wreck outcomes due to gaming. (I still want to study nomination strategy, but I'm really stuck on the details. I want to approach strategy, without assumptions, on both the nomination and voting sides. But then I need intelligence on both sides. An intelligence that isn't based on my assumptions is hard to imagine, and if it's some kind of brute force thing it will never finish running.) Kevin Venzke Election-Methods mailing list - see http://electorama.com/em for list info