What happened to the ratings?

Hey, what did they do with the stats. I just went to prove to someone that James had a crappy record, and I find that he is at zero....and I'm at zero. New look, new design. Whats that all about? INFO? http://206.65.59.246/VTES/member.tpl -- Aaron The Nosferatu Stuff

On Tue, 23 Jul 2002 21:49:58 -0400, "Nosferatu Stuff" <roans...@yahoo.com> wrote: >Hey, what did they do with the stats. I just went to prove to someone that >James had a crappy record, and I find that he is at zero....and I'm at zero. >New look, new design. Whats that all about? INFO? >http://206.65.59.246/VTES/member.tpl New set, new stats. As requested. When the person who volunteered to deal with the ratings is allowed to begin entering tournament results again, you will have a much better picture of the players and of their histories. You will also have a much better picture of the people running your events as you will be able to tell at a glance if they sent in the tournament results on time or not. Three cheers for Conrad & the ratings updater!! Carpe noctem. Lasombra http://www.TheLasombra.com

"The Lasombra" <thela...@hotmail.com> wrote in message news:3d3e2c0...@text.giganews.com... > On Tue, 23 Jul 2002 21:49:58 -0400, "Nosferatu Stuff" > <roans...@yahoo.com> wrote: > > >Hey, what did they do with the stats. I just went to prove to someone that > >James had a crappy record, and I find that he is at zero....and I'm at zero. > >New look, new design. Whats that all about? INFO? > > >http://206.65.59.246/VTES/member.tpl > > > New set, new stats. > As requested. The timing is just a coincidence but one that I'm happy about. The old rating system required that each tournament be calculated in chronological order. If tournament "C" didn't turn in their Archon reports in a timely manner, tournaments "A" and "B" couldn't yet be updated. Also, the old rating system required a lot of manual work. The number of tournaments has grown as so much that keeping-up with the ratings had the potential to overwhelm those processing them. Lastly, as was mentioned in this newsgroup, there were flaws with the old rating calculation itself. The new system is stats-based. It is not dependent on chronological processing of results. It is something that we can maintain even as we experience more growth. -Robert Robert Goudie Chairman, V:EKN rob...@vtesinla.org

Robert Goudie wrote: > > The old rating system required that each tournament be calculated in > chronological order. If tournament "C" didn't turn in their Archon reports in a > timely manner, tournaments "A" and "B" couldn't yet be updated. Also, the old > rating system required a lot of manual work. The number of tournaments has > grown as so much that keeping-up with the ratings had the potential to overwhelm > those processing them. Lastly, as was mentioned in this newsgroup, there were > flaws with the old rating calculation itself. > > The new system is stats-based. It is not dependent on chronological processing > of results. It is something that we can maintain even as we experience more > growth. I don't understand. Are they not going to do Elo any more? Fred

"Frederick Scott" <freds64_at_...@removethis.com> wrote in message news:3D3E4D5C...@removethis.com... [ quoted text not captured ] Nope. No more Elo. Players will be ranked based on their Game Wins per Game. Its not much more complex than that. Tie breakers, minimum games played, and other details can be found by following the link to the Ratings FAQ from the main ratings page. http://206.65.59.246/VTES/member.tpl -Robert

Robert Goudie wrote: > > > > The new system is stats-based. It is not dependent on chronological > > > processing of results. It is something that we can maintain even as > > > we experience more growth. > > > > I don't understand. Are they not going to do Elo any more? > > Nope. No more Elo. Players will be ranked based on their Game Wins per Game. > Its not much more complex than that. Tie breakers, minimum games played, and > other details can be found by following the link to the Ratings FAQ from the > main ratings page. Arg! Makes it a lot less meaningful. Well, it _was_ meaningless before due to fact that the selection of constants meant that the most very recent games were all that mattered. This has the opposite problem, that a game you played as a rank beginner is just as important as yesterday's game. But what's really going to prevent me from putting much stock in this one is that it makes no adjustment whatsoever for quality of competition. A game between you, your two children and your neighbor's dog (playing a CL deck, no doubt) will count the same as playing in Continental Finals. Not worthless, I suppose. At least we'll know more or less know who's top dog in each of the various localities... Fred

"Frederick Scott" <freds64_at_...@removethis.com> wrote in message news:3D3E5426...@removethis.com... > Robert Goudie wrote: > > > > > > The new system is stats-based. It is not dependent on chronological > > > > processing of results. It is something that we can maintain even as > > > > we experience more growth. > > > > > > I don't understand. Are they not going to do Elo any more? > > > > Nope. No more Elo. Players will be ranked based on their Game Wins per Game. > > Its not much more complex than that. Tie breakers, minimum games played, and > > other details can be found by following the link to the Ratings FAQ from the > > main ratings page. > > This has the opposite problem, that a game you played as a rank beginner is just > as important as yesterday's game. Think baseball stats. Career averages. ERA. Batting Avg. etc. None are adjusted for quality of opponent, weighted based on improvements in construction of equipment, altitude of the parks, etc. > Not worthless, I suppose. At least we'll know more or less know who's top dog in > each of the various localities... And limited rankings too! :) -Robert

Robert Goudie wrote: > > "Frederick Scott" <freds64_at_...@removethis.com> wrote in message > news:3D3E5426...@removethis.com... > > Robert Goudie wrote: > > > No more Elo. Players will be ranked based on their Game Wins per > > > Game. Its not much more complex than that. Tie breakers, minimum games > > > played, and other details can be found by following the link to the > > > Ratings FAQ from the main ratings page. > > > > This has the opposite problem, that a game you played as a rank beginner is > > just as important as yesterday's game. > > Think baseball stats. Career averages. ERA. Batting Avg. etc. None are adjusted > for quality of opponent, weighted based on improvements in construction of > equipment, altitude of the parks, etc. The most common ones discussed are annual statistics, not lifetime. (At least, that's my perception.) Will these be reset from year to year? The other things matter relatively little compared to the VEKN situation. Even with unbalanced schedules, all the teams in the league will play a certain amount of games against one another, and in each other's home stadiums. American League stats are seldom compared directly to National League, as the two are essentially different games. Contrast these vs. VEKN ratings, where many players will seldom or never play anyone except those in their home cities. Some will travel occasionally. And Jeff Thompson will eventually play on other planets when he runs out of new cities to visit. Like the previous ratings, these numbers will have some arguable meaning but on whole, not really tell us very much. Fred

Frederick Scott wrote: :: :: But what's really going to :: prevent me from putting much stock in this one is that it makes no :: adjustment whatsoever for quality of competition. A game between :: you, your two children and your neighbor's dog (playing a CL deck, :: no doubt) will count the same as playing in Continental Finals. I'm very happy with this new system. With the previous one it was very difficult to be on top of the rankings if you were a seasoned player. I was the french player(*) who entered the biggest number of tournaments (47) and i was also on the bottom of the rankings, not really due to my poor playing skills, but because a lot of other guys had had less involvement in tournaments (say, 15 to 25 participations). A lot of players also won a handful of tournament then sat on the top three and refused to play anymore to keep the (illusionnary) head of the pack. Now things will change. The more you play, the more your odds to gain places in the ranking. And it's quite good. :: Not worthless, I suppose. At least we'll know more or less know :: who's top dog in each of the various localities... mmm... expect big surprises here ;) :: Fred (*) no, Mr Rebstock, you're not a french player anymore ;)

"Robert Goudie" <rrgo...@earthlink.net> wrote: > And limited rankings too! :) When does this go into effect? Will gencon be able to be rated? If so I want to rate the draft at WON! :) I love draft. Ratings are still pretty pointless, but I guess its good that we got a reset. Is that going to happen every year? Maybe then it should be lifetime stats, and this years stats. I think baseball does something like that? [ quoted text not captured ]

"Nosferatu Stuff" <roans...@yahoo.com> wrote in message news:ahm56l$t3i1j$1...@ID-125246.news.dfncis.de... > > "Robert Goudie" <rrgo...@earthlink.net> wrote: > > > And limited rankings too! :) > > When does this go into effect? Will gencon be able to be rated? If so I > want to rate the draft at WON! :) I love draft. It's in effect now. GenCon will be rated. WoN Draft sounds fine to me. -Robert

"The Lasombra" <thela...@hotmail.com> wrote in message news:3d3e2c0...@text.giganews.com... > On Tue, 23 Jul 2002 21:49:58 -0400, "Nosferatu Stuff" > <roans...@yahoo.com> wrote: > > >Hey, what did they do with the stats. I just went to prove to someone that > >James had a crappy record, and I find that he is at zero....and I'm at zero. > >New look, new design. Whats that all about? INFO? > > >http://206.65.59.246/VTES/member.tpl > > > New set, new stats. > As requested. > Great...and right after I win my first tournament ever...

In message <ahl1c3$tk0cu$1...@ID-125246.news.dfncis.de>, Nosferatu Stuff <roans...@yahoo.com> writes: >Hey, what did they do with the stats. I just went to prove to someone that >James had a crappy record, Me? 2948 last time I checked it. Message-ID: <Vk9xS1Qb...@gratiano.zephyr.org.uk> Life keeps getting in the way, so it shouldn't have changed since then, which is a pity as I'd been doing a bit better at that point. My cards live in Birmingham (where I am currently), but I spend over half the year in Cambridge, without a car because of University regulations. As a result, getting home is about 4 hours by train due to the wonders of Dr Beeching (Cambridge->Birmingham New Street->Sutton Coldfield - allow about 20 minutes for delays on the Cambridge->Birmingham leg). Then most of the tournaments are down south, which is even more travelling, so going to tournaments during term-time - unless it's sealed deck or pre-built and I brought it with me to Cambridge - which is a pretty huge chunk of time to take out when I have commitments in Cambridge (some of which I'm trying to whittle down). And then outside of term-time, life intervenes again and the two obvious tournaments outside of term-time which I could play in are the European Qualifiers and GenConUK - both of which I judge. Why do I judge them? Well, I'm a purely average player. My rating would be a bit higher than it is, if I could play a bit more regularly, but I'm never going to be a top-ranked player - the main reason for which is that I don't adapt fast enough during play. It's a problem I have in a number of such games - like Civilisation and such. I want to play how I want to play, and I don't see the problems as fast as I need to to be a significantly above average player. I can do most other things reasonably enough. Since I only get to play in tournaments infrequently, fiddling with decks to oblivion can get a bit frustrating. So I end up playing with decks I haven't had chance to test enough, which is frustrating. Hopefully, at some point, I'll get my life back. Or work out how to move thousands of cards back and forth to Cambridge easily. -- James Coupe PGP 0x5D623D5D You don't need to hear it but I'm dried up EBD690ECD7A1FB457CA2 and sick to death of love. 13D7E668C3695D623D5D

reyda wrote: > > Frederick Scott wrote: > > :: > :: But what's really going to > :: prevent me from putting much stock in this one is that it makes no > :: adjustment whatsoever for quality of competition. A game between > :: you, your two children and your neighbor's dog (playing a CL deck, > :: no doubt) will count the same as playing in Continental Finals. > > I'm very happy with this new system. > > With the previous one it was very difficult to be on top of the rankings if > you were a seasoned player. I was the french player(*) who entered the > biggest number of tournaments (47) and i was also on the bottom of the > rankings, not really due to my poor playing skills, but because a lot of > other guys had had less involvement in tournaments (say, 15 to 25 > participations). I doubt that. I'm pretty sure it was because you weren't doing very well in your recent performances. There was no penalty for playing lots of games in the previous system. Indeed, with the constants set as they were, it hardly made much difference how many games you'd played. All that mattered was how you'd done and who you'd played in your most recent few games. I certainly wasn't defending the previous ratings systems. I'm just complaining that a previously flawed system has been exchanged - and with a great deal of labor and fanfare it appears - with a new flawed system. And when a simple change to the old system would have fixed the problem. It's the sort of thing that makes me bang my head on my computer table. > A lot of players also won a handful of tournament then sat > on the top three and refused to play anymore to keep the (illusionnary) head > of the pack. Now things will change. The more you play, the more your odds > to gain places in the ranking. I'm sorry, you're seeing something in these rankings I'm not seeing. Ranking is dependent on game wins PER GAME. Actually, the problem you cite still exists. If someone gets a few excellent results and cares about their ranking, they would do well to retire immediately. Fred

Frederick Scott wrote: :: reyda wrote: ::: ::: Frederick Scott wrote: (snip) :: I doubt that. I'm pretty sure it was because you weren't doing very :: well in your recent performances. There was no penalty for playing :: lots of games in the previous system. Indeed, with the constants :: set as they were, it hardly made much difference how many games :: you'd played. All that mattered was how you'd done and who you'd :: played in your most recent few games. I admit my performances were not very good lately... But sometime in tournament you get crushed by a very skilled predator. Why should other players who have nothing to do with your defeat (except by not helping you) gain ranking points on your dead body ? I had some tournaments where i did good performances (last year, in LA, went to the finals) but it still didn't raise my ranking... (snip) ::: A lot of players also won a handful of tournament then sat ::: on the top three and refused to play anymore to keep the ::: (illusionnary) head of the pack. Now things will change. The more ::: you play, the more your odds to gain places in the ranking. :: :: I'm sorry, you're seeing something in these rankings I'm not seeing. :: Ranking is dependent on game wins PER GAME. Actually, the problem :: you cite still exists. If someone gets a few excellent results and :: cares about their ranking, they would do well to retire immediately. If he retires, someone with the same amount of VP but more participations will go above him on the ranking =) that's that simple.

Frederick Scott wrote: > I'm pretty sure it was because you weren't doing very well in > your recent performances. There was no penalty for playing lots of games in > the previous system. Indeed, with the constants set as they were, it hardly > made much difference how many games you'd played. All that mattered was > how you'd done and who you'd played in your most recent few games. You're still on the idea that the constants effected that problem? As previously countered, they do not. Changing the K value (that's the 32 in the "Change in Rating" formula) changes the magnitude of the swings, but doesn't change anything w.r.t. relative ratings (except inasfar as it allows roundoff error to creep in). Changing the S value (that's the 400 in the exponent of the denominator of the probability formula) has a similar effect. Starting at 3000 and then swinging between 3100 and 2900 based on recent performance is no different than starting at 3000 and then swinging between 4000 and 2000 or between 3006 and 2994 based on recent performance (under the assumption that the constant is the same for everyone so everyone is swining in the same range). Elo systems inherently are not affected by number of games played. They're zero-sum systems. If you win your rating goes up. If you lose your rating goes down. How much it goes up or down depends on the ratings of your opponents and the constants. > I certainly wasn't defending the previous ratings systems. I'm just > complaining that a previously flawed system has been exchanged - and with a > great deal of labor and fanfare it appears - with a new flawed system. And > when a simple change to the old system would have fixed the problem. > > It's the sort of thing that makes me bang my head on my computer table. Please supply values for the constants (or whatever your "simple change" is) that would fix that particular flaw. -- LSJ (vte...@white-wolf.com) V:TES Net.Rep for White Wolf, Inc. Links to V:TES news, rules, cards, utilities, and tournament calendar: http://www.white-wolf.com/vtes/

"Frederick Scott" <freds64_at_...@removethis.com> wrote in message news:3D3EFCF4...@removethis.com... > > I certainly wasn't defending the previous ratings systems. I'm just > complaining that a previously flawed system has been exchanged - and with a > great deal of labor and fanfare it appears - with a new flawed system. And > when a simple change to the old system would have fixed the problem. > > It's the sort of thing that makes me bang my head on my computer table. Well, any rating system will probably be flawed in one respect or another. If there's a perfect rating system I don't know what it would be. But the reason the ELO system was done away with wasn't *just* because it was flawed in weighting most-recent games too heavily. It was also (so I'm told) a huge amount of work to run the ELO system: there was a lot of manual inputting of results, plus it required results to be entered in order of event date (so if results went unreceived, ratings couldn't be updated until that tournament organizer sent them in, unless the event was to be entirely ignored). I think it'd be cool to have an ELO-rating (with improved constants) available in addition to the "wins per game" that we're apparently going to be seeing soon. But I can certainly live without it, if it would be too much effort to be practical. Even an improved-constant ELO system would have flaws of its own. And, the new system can (I would think) easily be expanded to include more statistics, like "current season GW percentage" and so on. Josh buried in this bare bed

"LSJ" <vte...@white-wolf.com> wrote in message news:3D3EFE95...@white-wolf.com... > Frederick Scott wrote: > > All that mattered was > > how you'd done and who you'd played in your most recent few games. > > You're still on the idea that the constants effected that problem? > > As previously countered, they do not. Changing the K value (that's > the 32 in the "Change in Rating" formula) changes the magnitude of > the swings, but doesn't change anything w.r.t. relative ratings > (except inasfar as it allows roundoff error to creep in). I thought you had agreed that the constants did affect the problem. If either factor is flattened, that makes it easier for your rating to go up again after winning, because the system doesn't assume you "ought to" win as often - probably appropriate in a game with a substantial luck factor as well as a "hot or cold" effect in that if you win a five-player game, you beat four people, but if you lose one you may have lost to four people. Curt Adams even did some empirical testing and found that reducing the "K" helped quite a bit in ranking players more consistently with their "underlying skill". http://groups.google.com/groups?selm=20020131013456.28722.00000519%40mb-cv.a ol.com http://groups.google.com/groups?selm=20020204235941.11248.00002036%40mb-cg.a ol.com Josh not that i'm hugely attached to elo ratings

Joshua Duffin wrote: > Curt Adams even did some empirical testing and found that > reducing the "K" helped quite a bit in ranking players more > consistently with their "underlying skill". > > http://groups.google.com/groups?selm=20020131013456.28722.00000519%40mb-cv.a > ol.com > > http://groups.google.com/groups?selm=20020204235941.11248.00002036%40mb-cg.a > ol.com The K is a scale. Changing from 32 to 4 is, mathematically, exactly the same as changing the 3000 starting point to 24000 and the 400 to 3200 and rounding the results to the nearest multiple of 8. I suppose the key could lie in the ratio of K to S. But I have no idea what sort of argument would mathematically bear that out. [ quoted text not captured ]

"LSJ" <vte...@white-wolf.com> wrote in message news:3D3F09F1...@white-wolf.com... > Joshua Duffin wrote: > > Curt Adams even did some empirical testing and found that > > reducing the "K" helped quite a bit in ranking players more > > consistently with their "underlying skill". > > > > http://groups.google.com/groups?selm=20020131013456.28722.00000519%40mb-cv.a ol.com > > > > http://groups.google.com/groups?selm=20020204235941.11248.00002036%40mb-cg.a ol.com > > The K is a scale. > > Changing from 32 to 4 is, mathematically, exactly the same as > changing the 3000 starting point to 24000 and the 400 to 3200 and > rounding the results to the nearest multiple of 8. I don't think it matters whether you do or don't change the 3000. You could equally well make it 0; the only time "existing rating" is used is to figure the difference between two players' ratings, so any starting point is equivalent. As you say (and Curt Adams wrote in those posts), changing K or S should have equal effects when changed in the same ratios. > I suppose the key could lie in the ratio of K to S. But I have > no idea what sort of argument would mathematically bear that out. S tells you what your "win probability" is supposed to be when you play against some opponent. K tells you how much your rating can change - which means how much your expected win probability can change - from any single result. In combination, they determine how much a single result affects "how good we think you are", eg how much winning once should imply that you will win again. The smaller the S/K ratio, the more any individual result affects your expected win probability. The lower the S/K ratio, the less a single result affects your expected win probability. In my attempts to understand the formula, it seems that the old formula produced so much volatility because it implied that a person with a rating 200 points above starting position should beat a person at starting position 75.9% of the time, gaining 8 points if he won or losing 24 if he lost. In a game with this much luck/seating-position/etc influence that's not (IMO) realistic. Increasing the S/K ratio (to eg 3200/32) would reduce that expected win percentage to 53.6%, and gaining 15 points or losing 17. It's possible that that's too much reduction, but it's hard for me to tell without seeing actual outcomes. You might say that if S/K was 3200/32, all that would do is make people's ratings fluctuate in a narrower band, but Curt Adams' results seem to indicate that that's not the case - that in fact increasing the S/K ratio would allow "good" players' ratings to more accurately reflect that they are "better" than the "bad" players. Josh unsung and unsaid

LSJ wrote: > > I suppose the key could lie in the ratio of K to S. But I have > no idea what sort of argument would mathematically bear that out. (I'm assuming we're using "K" to represent the multiplier and "S" to represent the divisor of the difference in player ratings.) That is precisely what I've been saying all this time. No one has convinced me that I'm wrong about it and I'm pretty sure most of the people who've followed my reasoning have agreed with it. The ratings swing, per result, is based on the differences between the two players' ratings and the constants. If you reduce the value of K with respect to S, you reduce the influence of the difference between the two players' ratings in each ratings swing. This is good because it allows ratings to go higher (for the better players) and lower (for guys like me) before the difference in ratings swings starts to allow me get enough extra points back for times I get lucky to make up for all the drubbings I take from being bad. The reason this is important is that it means a single recent game means less to the overall ratings. The downside of reducing this ration is that it takes more results for the players (especially the extremely good and bad ones) to get to their true places on the spectrum. Reduce the ratio too far and it will take hundreds of games, further and it will take thousands of games. As the ratio approaches zero, the effect of differences between two players' ratings will not matter and ratings will never reach equilibrium (assuming "K" itself is not zero). So you don't want to overdo it. However, the nature of the game being what it is, and the way the results are gained (six results in a four-player game, ten in a five- player game), I would argue that it's appropriate that it take a while to reach equilibrium. As it was, we were reaching equilibrium in two or three games and bouncing so much higher and lower than our true approximate ratings that we were constantly bouncing above and below players of much differing abilities. By the way, the 3000 number is really meaningless. You can make it anything you want. It's really just window-dressing to compare to the "real numbers" - which are the positive and negative rating changes the formula spits out. I would say you want to set it higher than the worst negative rating would ever be. In our old system, since ratings almost never went much below -300, 500 would have been fine. If you were to reduce "K", 500 would still be fine. If you increased "S", 500 times the increase of the "S" factor would be fine. Fred

reyda wrote: > > Frederick Scott wrote: > :: reyda wrote: > ::: > ::: Frederick Scott wrote: > (snip) > :: I doubt that. I'm pretty sure it was because you weren't doing very > :: well in your recent performances. There was no penalty for playing > :: lots of games in the previous system. Indeed, with the constants > :: set as they were, it hardly made much difference how many games > :: you'd played. All that mattered was how you'd done and who you'd > :: played in your most recent few games. > > I admit my performances were not very good lately... > But sometime in tournament you get crushed by a very skilled predator. Why > should other players who have nothing to do with your defeat (except by not > helping you) gain ranking points on your dead body ? I had some tournaments > where i did good performances (last year, in LA, went to the finals) but it > still didn't raise my ranking... The one has little to do with the other. In a multiplayer game, people will gain advantages they don't deserve at times due to an interaction they had nothing to do with. Can't be helped. But your good performances likely helped your rankings just fine. It's just your ratings then raised for a moment making one bad slip-up cost you far more than you deserved - a flaw that fixing the constants will solve. > ::: A lot of players also won a handful of tournament then sat > ::: on the top three and refused to play anymore to keep the > ::: (illusionnary) head of the pack. Now things will change. The more > ::: you play, the more your odds to gain places in the ranking. > :: > :: I'm sorry, you're seeing something in these rankings I'm not seeing. > :: Ranking is dependent on game wins PER GAME. Actually, the problem > :: you cite still exists. If someone gets a few excellent results and > :: cares about their ranking, they would do well to retire immediately. > > If he retires, someone with the same amount of VP but more participations > will go above him on the ranking =) > that's that simple. Again, I think you are mistaken. AGAIN, ranking appears to me to be dependent on game wins PER GAME. Not total game wins or total victory points. Fred

>And, the new system can (I would think) easily be expanded to >include more statistics, like "current season GW percentage" >and so on. But, GW percentage doesn't tell you anything meaningful about how good someone is. So, expanding the stats isn't going to be of any more use. GW percentage has too much to do with how often games are won at all to be any sort of meaningful indicator of ability.

"James Coupe" <ja...@zephyr.org.uk> wrote > > Why do I judge them? Well, I'm a purely average player. My rating > would be a bit higher than it is, if I could play a bit more regularly, > but I'm never going to be a top-ranked player - the main reason for > which is that I don't adapt fast enough during play. It's a problem I > have in a number of such games - like Civilisation and such. I want to > play how I want to play, and I don't see the problems as fast as I need > to to be a significantly above average player. I can do most other > things reasonably enough. This is one reason I find JOL to be satisfying actually... I can play at my own pace and think about my actions and reactions for an hour if I want before replying. But I think this has only helped my game in real-time. I think being used to a slower pace has made me more aware of the situation at any given table or something. It's like I can scan the table and see potential threats more clearly than I used to or something... that sounds weird but I can't explain it better than that. I suppose it's almost like those Tai Chi people knowing how to defend themselves in *other* than slow-motion :) Cheers, WES

"Joshua Duffin" <jtdu...@yahoo.com> wrote in message news:ahmusl$u053o$1...@ID-121616.news.dfncis.de... > > But the reason the ELO system was done away with wasn't *just* > because it was flawed in weighting most-recent games too heavily. > It was also (so I'm told) a huge amount of work to run the ELO > system: there was a lot of manual inputting of results, plus it > required results to be entered in order of event date (so if > results went unreceived, ratings couldn't be updated until that > tournament organizer sent them in, unless the event was to be > entirely ignored). Right. The old ELO system, with adjusted constants, would still have all the other problems. Not much good having a "perfect" rating system if its too cumbersome to maintain. -Robert Robert Goudie rob...@vtesinla.org Chairman, V:EKN

Robert Goudie wrote: > > "Joshua Duffin" <jtdu...@yahoo.com> wrote in message > news:ahmusl$u053o$1...@ID-121616.news.dfncis.de... > > > > It was also (so I'm told) a huge amount of work to run the ELO > > system: there was a lot of manual inputting of results, plus it > > required results to be entered in order of event date (so if > > results went unreceived, ratings couldn't be updated until that > > tournament organizer sent them in, unless the event was to be > > entirely ignored). > > Right. The old ELO system, with adjusted constants, would still have all the > other problems. Not much good having a "perfect" rating system if its too > cumbersome to maintain. I'm not sure why it should be too cumbersome to maintain. To me it seems like the system ought to be smart enough to rewind and refigure results sent out of order. Of course, it has to be programmed to maintain histories, but that appears to be part of the new system anyway. The programming's a little more complex but nothing major. I flummoxed. What's supposed to be more cumbersome about ELO than the new system?!? Ultimately, if there is some real reason the old system was tougher than the new one then there's nothing to be done if WW just simply doesn't want to do it anymore. But it seems silly, doing it for two or three years when the old system was spitting out meaningless numbers and giving up when something worthwhile could actually be done with the data. I guess I can't help but let my disappointment show through, here. This set of statistics they're proposing to show and call "ratings" are just completely uninteresting to me. To give you an idea of how much quality of opponents (the part that's totally ignored) matters, one of the guys in the final in Tortured Confession: Phoenix played the ASSAMITE PRECON DECK, for crying out loud! Contrast this with the quality of the regular players in Los Angeles and you'll understand why I think statistics taken that ignore context are tell us little of value. Fred

"Frederick Scott" <freds64_at_...@removethis.com> wrote in message news:3D3F98AB...@removethis.com... > Robert Goudie wrote: > > > > "Joshua Duffin" <jtdu...@yahoo.com> wrote in message > > news:ahmusl$u053o$1...@ID-121616.news.dfncis.de... > > > > > > It was also (so I'm told) a huge amount of work to run the ELO > > > system: there was a lot of manual inputting of results, plus it > > > required results to be entered in order of event date (so if > > > results went unreceived, ratings couldn't be updated until that > > > tournament organizer sent them in, unless the event was to be > > > entirely ignored). > > > > Right. The old ELO system, with adjusted constants, would still have all the > > other problems. Not much good having a "perfect" rating system if its too > > cumbersome to maintain. > > I'm not sure why it should be too cumbersome to maintain. To me it seems > like the system ought to be smart enough to rewind and refigure results > sent out of order. For the system to be that smart...well, first of all that would require a system to start with. :) I'm surprised we maintained the ruse so long. There was no database, no system, and no histories were kept. Individual brave souls took care of updating the ratings. One at a time, and in order. All the while, harrassing coordinators to turn-in their stuff, correcting mistakes, and manually entering results when players didn't have Archon spreadsheets to send. BTW, a huge thanks to Todd Bannister for doing this for so long. His devotion and steadfastness was incredible. > Ultimately, if there is some real reason the old system was tougher than > the new one then there's nothing to be done if WW just simply doesn't want > to do it anymore. Another misconception. The ratings were started by the V:EKN in the absence of real support by WotC. We've always done it on our own. WW actually supported us quite a bit in creating this system. However, it is still maintained by volunteers. > But it seems silly, doing it for two or three years when > the old system was spitting out meaningless numbers and giving up when > something worthwhile could actually be done with the data. I started doing the ratings in, I believe, 1998. There were maybe 5-10 a month. I used to get behind. :) [ quoted text not captured ]

Joshua Duffin wrote: > > "LSJ" <vte...@white-wolf.com> wrote in message > news:3D3F09F1...@white-wolf.com... > > Joshua Duffin wrote: > > > Curt Adams even did some empirical testing and found that > > > reducing the "K" helped quite a bit in ranking players more > > > consistently with their "underlying skill". > > > > > > > http://groups.google.com/groups?selm=20020131013456.28722.00000519%40mb-cv.a > ol.com > > > > > > > http://groups.google.com/groups?selm=20020204235941.11248.00002036%40mb-cg.a > ol.com > > > > The K is a scale. > > > > Changing from 32 to 4 is, mathematically, exactly the same as > > changing the 3000 starting point to 24000 and the 400 to 3200 and > > rounding the results to the nearest multiple of 8. > > I don't think it matters whether you do or don't change the 3000. Correct. [ quoted text not captured ] That makes sense. [and snip more stuff that makes good sense] Actually, according to many, the ratio should be pretty close to zero, due to the table interaction/seating/other activity beyond the players' control. Probably even 1:100 is too large. But, of course, then players would complain that, after sweeping 4 games in a tournament and watching their rating go up by 5 points, that the system doesn't "swing" enough. [ quoted text not captured ]

Joshua Duffin wrote: > In my attempts to understand the formula, it seems that the old > formula produced so much volatility because it implied that a > person with a rating 200 points above starting position should > beat a person at starting position 75.9% of the time, gaining > 8 points if he won or losing 24 if he lost. In a game with this > much luck/seating-position/etc influence that's not (IMO) > realistic. > > Increasing the S/K ratio (to eg 3200/32) would reduce that > expected win percentage to 53.6%, and gaining 15 points or losing > 17. It's possible that that's too much reduction, but it's hard > for me to tell without seeing actual outcomes. > > You might say that if S/K was 3200/32, all that would do is > make people's ratings fluctuate in a narrower band, but Curt > Adams' results seem to indicate that that's not the case - that > in fact increasing the S/K ratio would allow "good" players' > ratings to more accurately reflect that they are "better" than > the "bad" players. Solid argument (as I said in the previous follow-up). But that doesn't explain Curt's results. If Curt had used a formula for his win expectation that matched the 32:400 formula's win expection, then, regardless of the real-world validity of the 32:400 formula, the statistical analysis should've born out the parallel. But it didn't. Instead it indicated that his formula for expectation more closely (but not completely) matched a 4:400 formula. http://groups.google.com/groups?selm=20020131013456.28722.00000519%40mb-cv.aol.com Quote: > To produce the desired set of two-player paired outcome probabilities from > multiplayer games, I calculate a performance for each player of 10^(Skill/D). > The chance of each player winning the game is their performance divided by the > sum of performances of all players in the game. So the statistical results "showing" that the 4:400 formula more closely follows the skills of the players is actually just showing that it more closely follows the above-cited method of analysing the skills of the players. It's still useful information, but it doesn't disprove the validity of the 32:400 formula. Some other preformance calculation could've been found that more closely matches the 32:400 formula. That is, if Curt had calculated the chance of each player "winning the game" (not sure if that means sweeping or what) so that it worked out to 1/(10^((X-Y)/400)+1) for each pairing of player X and player Y, then it would've supported the current formulation (and would've been just as arbitrary in doing so as his current method was in supporting a 4:400 ratio). Calulating the percentage of each player "winning the game" as merely that player's rating divided by the sum of all ratings at the table would support yet another rating calculation method. And so on. Curt - do you still have the program you used to get those results? Even though it won't matter now that the stats-based method is in place, I'd still like to take a look at it. [ quoted text not captured ]

> With the previous one it was very difficult to be on top of the rankings if > you were a seasoned player. I was the french player(*) who entered the > biggest number of tournaments (47) I thought that *I* was... But you're right, Reyda, you beat me by ONE game =) > > > > (*) no, Mr Rebstock, you're not a french player anymore ;) Don't agree with that, Pierre is definitely a french player, even on english land =) Miller Delmardigan http://sabbatinfrance.free.fr/en

Would it not be more useful to use VP/Games Played instead of Wins/Games Played to determine ranking? This assumes that the better players will consistently gain VPs over any given series of tournaments whereas the less skilled players will only gain VPs in fits and spurts. The number of wins could then be used as a tie breaker. I may be wrong (due to limited tournament experience) but this seems to be more representative of how tournaments are ranked. My $0.02 -- Dave Sysop: Ottawa City Grid The Black Sun: <http://www.ncf.carleton.ca/~bl740/blacksun.html> "...And never fear. Fear is for the enemy. Fear and bullets..." - J. O'Barr

In message <trN%8.2238$Ky3.2...@newsread2.prod.itd.earthlink.net>, Robert Goudie <rrgo...@earthlink.net> writes: >> > Right. The old ELO system, with adjusted constants, would still have all the >> > other problems. Not much good having a "perfect" rating system if its too >> > cumbersome to maintain. >> >> I'm not sure why it should be too cumbersome to maintain. To me it seems >> like the system ought to be smart enough to rewind and refigure results >> sent out of order. > >For the system to be that smart...well, first of all that would require a system >to start with. :) It's potentially doable. Throw in each tournament, in date order. Keep the details saved. Press the big button, calculate the rankings, which parses through each tournament in date order. When an older tournament comes in, plug it in, press the big button, new rankings. When all the sanctioned tournaments before a certain date are in, throw away the details and start with a new baseline for all later calculations. The difficulties - for me - would come with integrating it with the Archon spreadsheet, since I know nothing about interoperating with Excel and the like under 'doze. Still, it's potentially doable. [ quoted text not captured ]

"Curevei" <cur...@aol.commetal> wrote in message news:20020724194600...@mb-cu.aol.com... > >And, the new system can (I would think) easily be expanded to > >include more statistics, like "current season GW percentage" > >and so on. > > But, GW percentage doesn't tell you anything meaningful about how good > someone is. So, expanding the stats isn't going to be of any more use. Sure it does. It tells you how often they win games. Why isn't that meaningful? Because of the "2 VP minimum, ties don't win" restriction? But that applies to everyone everywhere, so should (in theory) balance out. > GW percentage has too much to do with how often games are won at all to > be any sort of meaningful indicator of ability. Well, clearly the assumption is that games should be won. If games are not won, then yes, there's a problem with this scoring method in that, where games are not won, everyone scores a loss. From the rating-system POV, though, everyone *deserves* a loss if no one wins. If this happens a lot more in some areas than others, scores won't be comparable across areas. But that's more or less true anyway, since these GW scores will mainly reflect people playing against others in their area. And it was more or less true of the old system too - yes, the ELO system can "shift points" from one region to another when people travel to play in other regions, but probably not enough points or fast enough to make comparisons truly meaningful. Josh meaningless

"James Coupe" <ja...@zephyr.org.uk> wrote in message news:wK+YLVSP...@gratiano.zephyr.org.uk... [snip] > When all the sanctioned tournaments before a certain date are in, throw > away the details and start with a new baseline for all later > calculations. > > The difficulties - for me - would come with integrating it with the > Archon spreadsheet, since I know nothing about interoperating with Excel > and the like under 'doze. Still, it's potentially doable. I'm sure it's doable. I think the problem is not having anyone to do it. If there were a volunteer to code something that could extract data from Archon spreadsheets and automatically calculate ratings like this, I imagine the VEKN would be more than happy to have it. So, volunteers? :-) Josh probably don't even need to throw away the details ever, unless you have a very small hard drive or something

"Frederick Scott" <freds64_at_...@removethis.com> wrote in message news:3D3F98AB...@removethis.com... > > I guess I can't help but let my disappointment show through, here. This > set of statistics they're proposing to show and call "ratings" are just > completely uninteresting to me. To give you an idea of how much quality > of opponents (the part that's totally ignored) matters, one of the guys > in the final in Tortured Confession: Phoenix played the ASSAMITE PRECON > DECK, for crying out loud! Contrast this with the quality of the regular > players in Los Angeles and you'll understand why I think statistics taken > that ignore context are tell us little of value. On this specific example, probably no rating system can ever take into account the quality of opposing *decks*. Opposing *players*, yes, in theory. But if Rob Treasure shows up with a deck made of 90 Save Face, the ELO system will still assume that you should get lots of points for beating him. On the general topic of "quality of opposition", yes, the ELO system is supposed to take that into account. But even there it's far from perfect; everyone starts at 3000 for example, so even if they're terrible players and play terrible decks they'll be worth a fair amount of points for quite a while (and for longer than they used to be, with "improved" constants). Ultimately, this game may be luck- and seating-dependent enough that no ratings system can possibly tell us all that much about who's the best. The "percent of games won" system has the advantage that it's easy to calculate and won't (after a few games) fluctuate very wildly. It does have the disadvantage of being hard to move at all after a while, but that can (I think) be easily solved by tracking a "current season's GW percentage" as well. Likewise, issues with the GW score itself could be addressed by adding a "VPs per game" stat, a "tournaments won" stat, and so on. Josh multidimensional

"LSJ" <vte...@white-wolf.com> wrote in message news:3D3FF332...@white-wolf.com... > > Solid argument (as I said in the previous follow-up). > But that doesn't explain Curt's results. Thanks. :-) > If Curt had used a formula for his win expectation that matched the > 32:400 formula's win expection, then, regardless of the real-world > validity of the 32:400 formula, the statistical analysis should've > born out the parallel. I'm not sure that's exactly what he did. As I read his article, it appears that he assumed a "true skill value" for each of his 30 "test players". That "skill value" was expressed in terms of the 32:400 formula and set along a normal distribution with a particular standard deviation (90) derived from looking at actual ratings results (the ratings being an attempt to estimate "true skill"). He then set the "chance of winning" for each game to be based on the "true skill" of each player in the game. However, the amount their *ratings* shift after the game is based on their current *rating*, ie the estimate of their skill. > But it didn't. Instead it indicated that his formula for expectation > more closely (but not completely) matched a 4:400 formula. > > http://groups.google.com/groups?selm=20020131013456.28722.00000519%40mb-cv.a ol.com > Quote: > > To produce the desired set of two-player paired outcome probabilities > > from multiplayer games, I calculate a performance for each player of > > 10^(Skill/D). The chance of each player winning the game is their > > performance divided by the sum of performances of all players in the > > game. > > So the statistical results "showing" that the 4:400 formula more closely > follows the skills of the players is actually just showing that it more > closely follows the above-cited method of analysing the skills of the > players. Right. But his method of describing the skills of the players - I think - is to assume a normal distribution of skills. Whether the standard deviation he gave that distribution is appropriate is hard to tell, but it was at least derived from historical results. > It's still useful information, but it doesn't disprove the validity of > the 32:400 formula. Some other preformance calculation could've been > found that more closely matches the 32:400 formula. Probably true, but if I'm understanding it correctly, the version that would support the 32:400 formula would involve a higher standard deviation of skills rather than a different method of calculating each player's chance of winning. > That is, if Curt had calculated the chance of each player "winning the > game" (not sure if that means sweeping or what) so that it worked out > to 1/(10^((X-Y)/400)+1) for each pairing of player X and player Y, then > it would've supported the current formulation (and would've been just > as arbitrary in doing so as his current method was in supporting a > 4:400 ratio). I'm not sure that's correct. But I'm not sure I follow the math enough to say so for sure, either. :-) BTW, "The chance of each player winning the game is their performance divided by the sum of performances of all players in the game. The #2 player is then determined by a similar procedure among the remaining players, and so on. Each pairwise outcome is determined by the first time either of the players beats the remaining players; each player beats each other with the appropriate expected chance." I'm not sure exactly what that means about who beats who, but I think it doesn't just assume one sweeping player per game. > Calulating the percentage of each player "winning the game" as merely > that player's rating divided by the sum of all ratings at the table > would support yet another rating calculation method. Right. We may not know enough about the underlying "true chances" of winning to come up with any really good model of who should win how often. Josh asleep in the deep

In message <ahp6m9$ubd7a$1...@ID-121616.news.dfncis.de>, Joshua Duffin <jtdu...@yahoo.com> writes: >I'm sure it's doable. I think the problem is not having anyone >to do it. If there were a volunteer to code something that could >extract data from Archon spreadsheets and automatically calculate >ratings like this, I imagine the VEKN would be more than happy to >have it. > >So, volunteers? :-) It'd take me (more than) a bit of research, but I could do it if no-one else can. A friend's been doing similar parsing of Word docs, hopefully I could cadge some of his thoughts. [ quoted text not captured ]

Joshua Duffin wrote: > > "LSJ" <vte...@white-wolf.com> wrote in message > news:3D3FF332...@white-wolf.com... > > > > Solid argument (as I said in the previous follow-up). > > But that doesn't explain Curt's results. > > Thanks. :-) > > > If Curt had used a formula for his win expectation that matched the > > 32:400 formula's win expection, then, regardless of the real-world > > validity of the 32:400 formula, the statistical analysis should've > > born out the parallel. > > I'm not sure that's exactly what he did. As I read his article, > it appears that he assumed a "true skill value" for each of his > 30 "test players". That "skill value" was expressed in terms > of the 32:400 formula and set along a normal distribution with > a particular standard deviation (90) derived from looking at > actual ratings results (the ratings being an attempt to estimate > "true skill"). It was expressed in "ratings units" (or vice-versa), correct. > > He then set the "chance of winning" for each game to be based > on the "true skill" of each player in the game. However, the > amount their *ratings* shift after the game is based on their > current *rating*, ie the estimate of their skill. Right. And there's where the assumptions creep in to skew the results. The Elo system computes a chance of X beating Y at a table (via the formula stated in the VEKN rules). Curt used a different formulation: Quoth Curt: > To produce the desired set of two-player paired outcome probabilities > from multiplayer games, I calculate a performance for each player of > 10^(Skill/D). The chance of each player winning the game is their > performance divided by the sum of performances of all players in the > game. End Quoth Curt. > > So the statistical results "showing" that the 4:400 formula more closely > > follows the skills of the players is actually just showing that it more > > closely follows the above-cited method of analysing the skills of the > > players. > > Right. But his method of describing the skills of the players - > I think - is to assume a normal distribution of skills. Whether > the standard deviation he gave that distribution is appropriate > is hard to tell, but it was at least derived from historical > results. That's a separate issue. Whether you assign the players' skill levels by that method, or just assign them 100, 200, 300, and so on, might matter at some point on some level, but it's not part of the problem in computing win expectancy. > > It's still useful information, but it doesn't disprove the validity of > > the 32:400 formula. Some other preformance calculation could've been > > found that more closely matches the 32:400 formula. > > Probably true, but if I'm understanding it correctly, the version > that would support the 32:400 formula would involve a higher > standard deviation of skills rather than a different method of > calculating each player's chance of winning. I disagree. A version whose win expectancy matched the win expectancy of the ratings system would yield a better parallel of observed results. > > Calulating the percentage of each player "winning the game" as merely > > that player's rating divided by the sum of all ratings at the table > > would support yet another rating calculation method. > > Right. We may not know enough about the underlying "true chances" > of winning to come up with any really good model of who should > win how often. Another very good, but separate, point. Even if the "skill level" being modeled incorporated deck choices (including the meta-game-ascertaining insights of the player), playing skill, table talking skills (intimidation, cajoling, deal-making, etc.) and all, it would still fail to account for seating position problems, "random" (beyond meta-game) actions/decks made/played by one's opponents, etc. So that one's "win-expectancy" could never be accurately modeled by comparing skill values alone, given the large uncertainties introduced by the latter aspects of multi-player gaming (and another reason not to put too much weight on any rating/ranking system). [ quoted text not captured ]

James Coupe <ja...@zephyr.org.uk> wrote in message > Robert Goudie wrote: >>For the system to be that smart...well, first of all that would require >>a system to start with. :) > > It's potentially doable. [clip] Well sure its doable. I was just making it clear to Fred that its not a minor change to the existing database and rating system. There is no existing ELO database to modify. -Robert

>> But, GW percentage doesn't tell you anything meaningful about how good >> someone is. So, expanding the stats isn't going to be of any more use. > >Sure it does. It tells you how often they win games. Why isn't >that meaningful? Because of the "2 VP minimum, ties don't win" >restriction? But that applies to everyone everywhere, so should >(in theory) balance out. Because how often someone wins games doesn't correlate that well to how well someone does, first. Would you equate two players who both won the same number of games but where one had 1.5 VPs in every other game played and the other had 0 VPs in every other game played? >> GW percentage has too much to do with how often games are won at all to >> be any sort of meaningful indicator of ability. > >Well, clearly the assumption is that games should be won. If >games are not won, then yes, there's a problem with this scoring >method in that, where games are not won, everyone scores a loss. >From the rating-system POV, though, everyone *deserves* a loss if >no one wins. Why? Not winning a game when no one else does either doesn't display lack of skill. (Mmm, many negatives are fun.) How often games are won at all varies depending upon prevailing style - offensive v. defensive - and relative skill of the players. My experience is that the better the players, the less chance a game will be won. >If this happens a lot more in some areas than others, scores >won't be comparable across areas. But that's more or less true >anyway, since these GW scores will mainly reflect people playing >against others in their area. And it was more or less true of >the old system too - yes, the ELO system can "shift points" from >one region to another when people travel to play in other regions, >but probably not enough points or fast enough to make comparisons >truly meaningful. Okay, the old system was crap. Why put in a new one that's no better? I don't see why it's so hard to come up with a system that has some relevance to skill. Here's one: You get 1 point for winning a tournament. You get another for winning a major. System keeps track of how many events you've played in (otherwise, you wouldn't know someone's average) and how many points you have. Too simple? Fine. But, I'd be a lot more impressed with someone who has a bunch of these points than someone who has a high GW%. And, isn't that what a rating system is supposed to accomplish - determining who is better than someone else?

"Curevei" <cur...@aol.commetal> wrote in message news:20020725141844...@mb-mg.aol.com... > > >Sure it does. It tells you how often they win games. Why isn't > >that meaningful? Because of the "2 VP minimum, ties don't win" > >restriction? But that applies to everyone everywhere, so should > >(in theory) balance out. > > Because how often someone wins games doesn't correlate that well to > how well someone does, first. Would you equate two players who both > won the same number of games but where one had 1.5 VPs in every other > game played and the other had 0 VPs in every other game played? Well, I agree that more metrics would be better. I suggested "VPs per game" and "tournaments won" (or "tournaments won per tournament") in my reply to Fred Scott too. That said, the 1.5 VP and the 0 VP player could both be coming in second in all their other games with those scores. And from a certain point of view, not winning makes you a loser regardless of how much you don't win by. (You'll find that many of the truths we cling to depend greatly on our own point of view. ;-) > >From the rating-system POV, though, everyone *deserves* a loss if > >no one wins. > > Why? Not winning a game when no one else does either doesn't display > lack of skill. (Mmm, many negatives are fun.) From the rating-system POV, if no one wins, everyone has lost. That's sort of nonsensical from a human POV in that we expect every game to have a winner, but not entirely; there are some games where everyone can lose. Now VTES is one of them. :-) > How often games are won at all varies depending upon prevailing style - > offensive v. defensive - and relative skill of the players. My > experience is that the better the players, the less chance a game will > be won. Mine isn't. Hence my comments about regionality... :-) Style can change to fit scoring methods, if people care about scoring. Isn't that what people feared/hoped would happen with the change to the GW system for tournaments? > Okay, the old system was crap. Why put in a new one that's no better? Cause it's a lot easier to maintain. Also, it's supposed to provide a more stable metric, even if not everyone likes the new metric. > I don't see why it's so hard to come up with a system that has some > relevance to skill. The game win percentage has *some* relevance to skill... > Here's one: > You get 1 point for winning a tournament. You get another for winning > a major. > System keeps track of how many events you've played in (otherwise, you > wouldn't know someone's average) and how many points you have. > > Too simple? Fine. But, I'd be a lot more impressed with someone who > has a bunch of these points than someone who has a high GW%. And, isn't > that what a rating system is supposed to accomplish - determining who is > better than someone else? This system is so simple that it could be created just from Jeff Thompson's TWD webpage. (Modulo the small tournaments.) Which isn't a bad thing at all, to my mind. But only a few people can be ranked at all by it, which is a pretty major drawback. Not, IMO, enough of one to dismiss it altogether, but I'd want to have several other stats to go along with it. Which seems to be the point I keep coming back to: more stats is better. Rankings by each of them independently would be even more better. Josh can it be done?

"LSJ" <vte...@white-wolf.com> wrote in message news:3D402CA8...@white-wolf.com... > Joshua Duffin wrote: > > > > Probably true, but if I'm understanding it correctly, the version > > that would support the 32:400 formula would involve a higher > > standard deviation of skills rather than a different method of > > calculating each player's chance of winning. > > I disagree. A version whose win expectancy matched the win expectancy > of the ratings system would yield a better parallel of observed > results. Actually, I suspect it *wouldn't* better match the results we observed in practice over the past few years (since apparently the underlying chances of winning do not match the ELO-calculated chanes). It *should* better match what the ELO scoring system expects to happen, I think you're right about that. Though again, I'd have to see results of simulations; I can't feel sure about it just from turning over the formulas in my head. [skill vs other factors] > So that one's "win-expectancy" could never be accurately modeled by > comparing skill values alone, given the large uncertainties introduced > by the latter aspects of multi-player gaming (and another reason > not to put too much weight on any rating/ranking system). Well, it can always be modeled by including a large enough "random factor". But yes, if there's that much "randomness" involved it certainly implies that we shouldn't take ratings too seriously. Also, we should have more axes to rank people on, because the more things we look at, the more likely that everybody will find at least one of them interesting. :-) Josh axiomatic

Joshua Duffin wrote: > "LSJ" <vte...@white-wolf.com> wrote in message > > I disagree. A version whose win expectancy matched the win expectancy > > of the ratings system would yield a better parallel of observed > > results. > > Actually, I suspect it *wouldn't* better match the results we > observed in practice over the past few years (since apparently > the underlying chances of winning do not match the ELO-calculated > chanes). It *should* better match what the ELO scoring system > expects to happen, I think you're right about that. Right. By observed, I mean output of the system (matching the skill-level input to the system). [ quoted text not captured ]

Oh yeah, one more thing about ratings: Someone had suggested (I think it was Gomi) a "tournament-size-adjusted placement" stat. I think that one would also be great, though if the current rating technology amounts to "Robyn with a calculator" it might be somewhat annoying to calculate. Works like this: placement stat = 1 - (finishing rank/# of participants) It doesn't reward winning bigger events more than smaller ones - kind of the opposite, sort of - but winning big events, as Gomi said at the time, gets you a reputation all by itself. Josh right ben? ;-)

LSJ wrote: (re: rating numbers swinging to slowly or too quickly) > > But, of course, then players would complain that, after sweeping 4 > games in a tournament and watching their rating go up by 5 points, > that the system doesn't "swing" enough. Hey, that's people for you. No one's totally happy. But the important part is that the numbers behave in as reasonable manner as possible. Don't forget, what was really causing the problem of the old system wasn't how small or large the swing number was exactly but how large it was compared to the effective range of possible ratings. When ratings seldom go higher or lower than 2750 to 3250, a swing of greater than 100 points in a single sitting is surely pretty reckless. The problem is caused by how quickly a guy gets up to a point where he's risking three times as many points as he might potentially win. Fred

David Faguy wrote: > > Would it not be more useful to use VP/Games Played instead of Wins/Games > Played to determine ranking? This assumes that the better players will > consistently gain VPs over any given series of tournaments whereas the > less skilled players will only gain VPs in fits and spurts. The number of > wins could then be used as a tie breaker. What's the difference? The problem with either is that neither adjusts for quality of competition. I'm not sure what problem you're trying to solve but it's not the main one. Fred

Joshua Duffin wrote: > > Well, it can always be modeled by including a large enough > "random factor". But yes, if there's that much "randomness" > involved it certainly implies that we shouldn't take ratings > too seriously. More precisely, it means we have to take a larger sample of results to allow the randomness to even out over time. FAILING THAT (as obviously there's a limit to how many games you can get people to play), we shouldn't take ratings too serious. But I intuitively suspect a "constant-corrected" ELO system should be reasonably informative for players that have a few dozen games under their belt. Fred

In article <ahpn46$ut1n6$1...@id-121616.news.dfncis.de>, [ quoted text not captured ] Sure it does reward bigger events, in that 1-(1/50) is a higher number than 1-(1/10), no? Not sure how you'd composite this ranking over several events -- I'm suspicious of just averaging them, but that certainly is simple, at least. gomi -- Blood, guts, guns, cuts Knives, lives, wives, nuns, sluts

Robert Goudie wrote: > > "Frederick Scott" <freds64_at_...@removethis.com> wrote in message > news:3D3F98AB...@removethis.com... > > I'm not sure why it should be too cumbersome to maintain. To me it seems > > like the system ought to be smart enough to rewind and refigure results > > sent out of order. > > For the system to be that smart...well, first of all that would require a system > to start with. :) I'm surprised we maintained the ruse so long. There was no > database, no system, and no histories were kept. Individual brave souls took > care of updating the ratings. One at a time, and in order. All the while, > harrassing coordinators to turn-in their stuff, correcting mistakes, and > manually entering results when players didn't have Archon spreadsheets to send. OK, sure. Let's not overstate your case here. I don't see the harassment or mistake-correcting nor manually entering results going away just because a different system is in place. In short, I agree about the "brave souls" thing but it appears we still need them. The only thing that's changed is the need for consecutive entry of tournament results. Fred

James Coupe wrote: > > >For the system to be that smart...well, first of all that would require a system > >to start with. :) > > It's potentially doable. > > Throw in each tournament, in date order. Keep the details saved. Press > the big button, calculate the rankings, which parses through each > tournament in date order. > > When an older tournament comes in, plug it in, press the big button, new > rankings. > > When all the sanctioned tournaments before a certain date are in, throw > away the details and start with a new baseline for all later > calculations. That was precisely what I was envisioning. The only question in my mind is about the calculation time. If you've got the number crunching power to push button and spit out provisional results that require computing one or two dozen tournaments (back to the last hole) without too much trouble, then it should work just fine. If there's a question about it then you need to do away with provisional results and just calculate up to first hole. Even so, you can still certainly enter results as they come in. Fred

Joshua Duffin wrote: > > "Frederick Scott" <freds64_at_...@removethis.com> wrote in message > news:3D3F98AB...@removethis.com... > > > > I guess I can't help but let my disappointment show through, here. This > > set of statistics they're proposing to show and call "ratings" are just > > completely uninteresting to me. To give you an idea of how much quality > > of opponents (the part that's totally ignored) matters, one of the guys > > in the final in Tortured Confession: Phoenix played the ASSAMITE PRECON > > DECK, for crying out loud! Contrast this with the quality of the regular > > players in Los Angeles and you'll understand why I think statistics taken > > that ignore context are tell us little of value. > > On this specific example, probably no rating system can ever > take into account the quality of opposing *decks*. Opposing > *players*, yes, in theory. Yes, quite true. It makes a difference whether mighty Casey shows up at a tournament gunning to kill or just to have fun with his new trick deck. In theory, a player's rating should in part reflect his proclivity to do one or other, but that's small calculation to victims who were in Casey's path the day he decides to turn mercenary. > On the general topic of "quality of opposition", yes, the ELO > system is supposed to take that into account. But even there > it's far from perfect; everyone starts at 3000 for example, so > even if they're terrible players and play terrible decks they'll > be worth a fair amount of points for quite a while (and for > longer than they used to be, with "improved" constants). Imperfect, true. But "far from perfect" depends on how often this happens. You're only unrated in your first game (or the first game of each instance of adopting a new disguise, I guess), after that the ratings should start to do the right thing. I don't consider this to be a major problem. > Ultimately, this game may be luck- and seating-dependent > enough that no ratings system can possibly tell us all that > much about who's the best. The "percent of games won" system > has the advantage that it's easy to calculate and won't (after > a few games) fluctuate very wildly. It does have the > disadvantage of being hard to move at all after a while, but > that can (I think) be easily solved by tracking a "current > season's GW percentage" as well. I guess it depends on what you want. As ratings, these clearly suck. As statistics, they might be interesting to some - though the lack of context makes them mostly random digits to my eyes. Fred

>Which seems to be the point I keep >coming back to: more stats is better. I don't agree. Only good stats are worthwhile and only in small quantities. I like baseball stats, but most are bogus. They only serve to distract from worthwhile ones. Besides, it's more work for someone to keep track of a multitude of official stats.

"Curevei" <cur...@aol.commetal> wrote in message news:20020725180323...@mb-fz.aol.com... > >Which seems to be the point I keep > >coming back to: more stats is better. > > I don't agree. Only good stats are worthwhile and only in small quantities. I dunno. More stats are definitely more interesting. :) Anyway, waaaaaaaaaay back at the beginning, didn't LSJ (or Robert) mention something about tiebreakers and the like? I firmly believe that we won't just be going on GW/games played. Robert wrote: "Players will be ranked based on their Game Wins per Game. Its not much more complex than that. Tie breakers, minimum games played, and other details can be found by following the link to the Ratings FAQ from the main ratings page. http://206.65.59.246/VTES/member.tpl" I dunno. Works for me. Xian

"Joshua Duffin" <jtdu...@yahoo.com> wrote in message news:ahp772$unsp5$1...@ID-121616.news.dfncis.de... > *players*, yes, in theory. But if Rob Treasure shows up with > a deck made of 90 Save Face, the ELO system will still assume > that you should get lots of points for beating him. Goddamit, Josh! You totally blew my seekrit-maximize-GenCon-ratings-boost tech! Well, that and the no more ELO thing. I was totally going to play Tremere with 40 Save Face, 40 Apportation, and 10 Walk of Flame in every tournament prior to the NA Championships! Arrrgh! ;) Xian maximizin'

In message <3D406E64...@removethis.com>, Frederick Scott <freds64_at_...@removethis.com> writes: >That was precisely what I was envisioning. The only question in my mind is >about the calculation time. If you've got the number crunching power to >push button and spit out provisional results that require computing one or two >dozen tournaments (back to the last hole) without too much trouble, then it >should work just fine. It shouldn't be that difficult. None of the ratings calculations are generally *that* difficult. [ quoted text not captured ]

"Xian" <xi...@waste.org> wrote in message news:ahq1bg$uciai$1...@ID-123937.news.dfncis.de... > Anyway, waaaaaaaaaay back at the beginning, didn't LSJ (or Robert) mention > something about tiebreakers and the like? I firmly believe that we won't just > be going on GW/games played. First tie-break is VPs. Second tie-break is career games played. -Robert

"Frederick Scott" <freds64_at_...@removethis.com> wrote in message news:3D406D2E...@removethis.com... [ quoted text not captured ] Sure. All of the above issues don't go away with the new system. On the other hand, that's not an exhaustive list of the problems with the old system either. :) -Robert

al...@club-internet.fr (Francois Alix) wrote in message news:<b9f814b6.02072...@posting.google.com>... > > (*) no, Mr Rebstock, you're not a french player anymore ;) > > Don't agree with that, Pierre is definitely a french player, even on english land =) Nope, sorry we have 'claimed' him now. He is lost to you for ever and no, you don't have any say in this Pierre :o) Rob

James Coupe wrote: > > In message <3D406E64...@removethis.com>, Frederick Scott > <freds64_at_...@removethis.com> writes: > >That was precisely what I was envisioning. The only question in my mind is > >about the calculation time. If you've got the number crunching power to > >push button and spit out provisional results that require computing one or two > >dozen tournaments (back to the last hole) without too much trouble, then it > >should work just fine. > > It shouldn't be that difficult. None of the ratings calculations are > generally *that* difficult. If anyone wishes to pursue this (as James has mentioned he might), then once a working system is unveiled, I'm sure it wouldn't be too much trouble to have the Archons that WW receives forwarded to the maintainer of the "improved" rating/ranking system. [ quoted text not captured ]

In message <3D4136ED...@white-wolf.com>, LSJ <vtesrep@white- wolf.com> writes: >> It shouldn't be that difficult. None of the ratings calculations are >> generally *that* difficult. > >If anyone wishes to pursue this (as James has mentioned he might), then >once a working system is unveiled, I'm sure it wouldn't be too much trouble >to have the Archons that WW receives forwarded to the maintainer of the >"improved" rating/ranking system. I could. However, if people want to implement a different ratings system anyway, there's probably little point to it all. [ quoted text not captured ]

"Curevei" <cur...@aol.commetal> wrote in message news:20020725180323...@mb-fz.aol.com... > >Which seems to be the point I keep > >coming back to: more stats is better. > > I don't agree. Only good stats are worthwhile and only in small > quantities. I guess this depends on taste; some people like stats in mass quantities and others don't. :-) What's a "good" stat seems to be hugely subjective in the realm of VTES. > I like baseball stats, but most are bogus. They only serve to distract > from worthwhile ones. > > Besides, it's more work for someone to keep track of a multitude of > official stats. True, but given that (apparently) not everyone can be satisfied by one single stat, and that (as I understand it) it wouldn't be too much more work to calculate rankings for several (easily-derived- from-Archon) stats, it seems worthwhile to me. As far as "the most universally meaningful single stat" goes, I'd guess that the "tournament-size-adjusted-placement" stat is probably it. But even that has its issues. "placement stat" = 1 - (tournament ranking/# of tournament participants) so: 1st out of 50 = 0.98 1st out of 10 = 0.90 50th out of 50 = 0.00 10th out of 10 = 0.00 49th out of 50 = 0.02 9th out of 10 = 0.10 For multiple events you could average your "placement stats" together, or do something else more clever, if there is such a thing. The 'issues' I see with this stat are: It is entirely linear, so there's only 0.02 difference between 1st and 2nd in a 50-player tournament, or 0.10 between 1st and 2nd in a 10-player. It only fairly slightly rewards "winning bigger tournaments" vs "winning small tournaments" (0.98 vs 0.90). It gives the same result whether the winner of the event "won" by being first-ranked (potentially tied and resolved by coin-flip) going into the finals and then timing out in the finals with tied VPs, or by sweeping three tables to get to the finals and then sweeping the final round too. Of course, all of these things are benefits from another point of view. ;-) Which is why it seems to me that no one stat can hope to be universally hailed as "the right one". Josh off the scale on the bogometer

"James Coupe" <ja...@zephyr.org.uk> wrote in message news:5tvqSVfI...@gratiano.zephyr.org.uk... > In message <3D4136ED...@white-wolf.com>, LSJ <vtesrep@white- > wolf.com> writes: > > >If anyone wishes to pursue this (as James has mentioned he might), then > >once a working system is unveiled, I'm sure it wouldn't be too much > >trouble > >to have the Archons that WW receives forwarded to the maintainer of the > >"improved" rating/ranking system. > > I could. However, if people want to implement a different ratings > system anyway, there's probably little point to it all. Well, I can't speak for all people (ha!), but: at least me and Fred Scott would love to see an "improved" ELO system. (It might not sound like it from how much I've talked up the "other" methods of rating people, but I like the ELO system too.) Josh eloriffic

In message <ahrj2o$v5fl3$1...@ID-121616.news.dfncis.de>, Joshua Duffin <jtdu...@yahoo.com> writes: >> >If anyone wishes to pursue this (as James has mentioned he might), then >> >once a working system is unveiled, I'm sure it wouldn't be too much >> >trouble >> >to have the Archons that WW receives forwarded to the maintainer of the >> >"improved" rating/ranking system. >> >> I could. However, if people want to implement a different ratings >> system anyway, there's probably little point to it all. > >Well, I can't speak for all people (ha!), but: at least me and Fred >Scott would love to see an "improved" ELO system. (It might not >sound like it from how much I've talked up the "other" methods of >rating people, but I like the ELO system too.) Are there any actual bona-fide games like Chess (i.e. games with a long fixed (or close to) rules-set and play set, because they don't get expansions released for them) which are genuinely multi-player? Whilst Diplomacy's multi-player system sucks in many ways, is there any long term rating system, based on who went out first, who had fewest units/locations at the end, or whatever? Or any similar games. Say, Star Fleet Battles (Mr PDB6, are you there?)? Ratings systems aren't something I know that much about. Are there any decent primers anywhere? [ quoted text not captured ]

James Coupe wrote: > Are there any actual bona-fide games like Chess (i.e. games with a long > fixed (or close to) rules-set and play set, because they don't get > expansions released for them) which are genuinely multi-player? > > Whilst Diplomacy's multi-player system sucks in many ways, is there any > long term rating system, based on who went out first, who had fewest > units/locations at the end, or whatever? Or any similar games. Say, > Star Fleet Battles (Mr PDB6, are you there?)? > > Ratings systems aren't something I know that much about. Are there any > decent primers anywhere? Try Pete's Ludicrously Overcomplicated Points System for 2- to 6-player games (which doesn't consider opponents' skill/ratings levels): http://www.personal.dundee.ac.uk/~pjclinch/PLOPS.htm It might have some ideas of use, at any rate. [ quoted text not captured ]

LSJ wrote: > James Coupe wrote: > > Ratings systems aren't something I know that much about. Are there any > > decent primers anywhere? > > Try Pete's Ludicrously Overcomplicated Points System for 2- to 6-player > games (which doesn't consider opponents' skill/ratings levels): > > http://www.personal.dundee.ac.uk/~pjclinch/PLOPS.htm > > It might have some ideas of use, at any rate. Or try the "RACE" section of http://wolff.to/area/calcexp.html It looks a bit like the the old VEKN system, except with the Elo part replaced by a simpler pairwise system, but it might work. For a primer (decent? I don't know), try http://www.gamasutra.com/features/20000209/kreimeier_pfv.htm It seems tailored for online deathmatch style games, but certainly has enough theory explanation built in to be useful. [ quoted text not captured ]

"James Coupe" <ja...@zephyr.org.uk> wrote in message news:USwRtap3...@gratiano.zephyr.org.uk... > > Are there any actual bona-fide games like Chess (i.e. games with a long > fixed (or close to) rules-set and play set, because they don't get > expansions released for them) which are genuinely multi-player? Bridge, but it probably doesn't count because it's a team game. Poker, but I don't know if it has ratings. Its money-betting system makes it "self scoring" in an admirably simple way that VTES probably can't emulate. :-) Hearts; again I don't know if it has ratings. Likewise spades, rummy. Monopoly - might have ratings? Tennis - doubles matches might have some applicability? But it is still a "two sides" thing. > Whilst Diplomacy's multi-player system sucks in many ways, is there any > long term rating system, based on who went out first, who had fewest > units/locations at the end, or whatever? Or any similar games. Say, > Star Fleet Battles (Mr PDB6, are you there?)? Star Fleet Battles is normally only played two-player in tournament settings, as far as I know. Or two-side (team) if not, I'd think. > Ratings systems aren't something I know that much about. Are there any > decent primers anywhere? Nor do I... The ones LSJ pointed to look interesting; I haven't finished reading them yet though. Josh mala fide

"Frederick Scott" <freds64_at_...@removethis.com> wrote in message news:3D4070A3...@removethis.com... > Joshua Duffin wrote: > > > On the general topic of "quality of opposition", yes, the ELO > > system is supposed to take that into account. But even there > > it's far from perfect; everyone starts at 3000 for example, so > > even if they're terrible players and play terrible decks they'll > > be worth a fair amount of points for quite a while (and for > > longer than they used to be, with "improved" constants). > > Imperfect, true. But "far from perfect" depends on how often this > happens. You're only unrated in your first game (or the first game > of each instance of adopting a new disguise, I guess), after that > the ratings should start to do the right thing. I don't consider > this to be a major problem. But with "improved constants" they'll start doing the right thing fairly slowly. Plus, if you look at the VEKN player database lots of the players in it have (or had) very few rated games under their belts. In the first twenty results for "USA" (not a random sample, of course), thirteen of the entries have 0-3 games played. Unknown quantities galore. :-) Of course you (and I) might expect that if somebody has never played a tournament game, they're probably not going to be all that great at the tournament game. Though I'm not sure what my point is, here. > I guess it depends on what you want. As ratings, these clearly > suck. As statistics, they might be interesting to some - though > the lack of context makes them mostly random digits to my eyes. Well, I'm not yet convinced that "GW percentage" does suck as a rating method. How often you win tournament games seems to me quite intuitive as a measure of how good a tournament player you are. Sure, it doesn't give a bonus for winning the final, supposedly-most-important game, but the mere fact of having won an additional game does improve the winner's rating. Josh player of games

"Xian" <xi...@waste.org> wrote in message news:ahq1gt$uph1q$1...@ID-123937.news.dfncis.de... [ quoted text not captured ] Come on. I cleverly said "Rob Treasure". No one was going to expect it from *you*. You would have been fine if you hadn't just gone and spilled it all over the newsgroup. ;-) Josh besides, you can still use your "save face" tech to impress everyone at the week of nightmares with how much your decks are going to suck...

rrgo...@earthlink.net writes: >Nope. No more Elo. Players will be ranked based on their Game Wins per Game. >Its not much more complex than that. Tie breakers, minimum games played, and >other details can be found by following the link to the Ratings FAQ from the >main ratings page. If you want to use game wins for rankings, straight percentages are a bad system since variability is so high with even moderate numbers of games. The highest rated players will usually not be the best but rather the luckiest of those who've just crossed the minimun games threshold. Better would bet a Bayesian approach - assume players start with an expected skill from some baseline distribution. Modify the distribution as the player gets games and report the mean of each player's distribution as his rating. Curt Adams (curt...@aol.com) "It is better to be wrong than to be vague" - Freeman Dyson

Joshua Duffin wrote: > > "Frederick Scott" <freds64_at_...@removethis.com> wrote in message > news:3D4070A3...@removethis.com... > > Joshua Duffin wrote: > > > > > On the general topic of "quality of opposition", yes, the ELO > > > system is supposed to take that into account. But even there > > > it's far from perfect; everyone starts at 3000 for example, so > > > even if they're terrible players and play terrible decks they'll > > > be worth a fair amount of points for quite a while (and for > > > longer than they used to be, with "improved" constants). > > > > Imperfect, true. But "far from perfect" depends on how often this > > happens. You're only unrated in your first game (or the first game > > of each instance of adopting a new disguise, I guess), after that > > the ratings should start to do the right thing. I don't consider > > this to be a major problem. > > But with "improved constants" they'll start doing the right > thing fairly slowly. Sure. > Plus, if you look at the VEKN player database lots of the players in > it have (or had) very few rated games under their belts. In the > first twenty results for "USA" (not a random sample, of course), > thirteen of the entries have 0-3 games played. Unknown quantities > galore. :-) Yea, but the people who haven't played many games haven't played many games. No, that's not just my fingers keystroke-repeating themselves because my mind has gone out to lunch (for the third time today, before 11 AM no less). I mean the influence of such things depends on how often it happens and I expect if you multiply people by games, you discover that those how have played a dozen games or more (arbitrary number plucked from mid-air) influence the rankings a lot more often than those with less games. Obviously, you can wipe the players with 0 games right off the list, as they haven't done a thing to the rating system. > > I guess it depends on what you want. As ratings, these clearly > > suck. As statistics, they might be interesting to some - though > > the lack of context makes them mostly random digits to my eyes. > > Well, I'm not yet convinced that "GW percentage" does suck as a > rating method. How often you win tournament games seems to me > quite intuitive as a measure of how good a tournament player > you are. Sure, it doesn't give a bonus for winning the final, > supposedly-most-important game, but the mere fact of having > won an additional game does improve the winner's rating. Winning games is always good. But without the adjustment for skill, there's no saying how good. My experience is that some games are just a heckuva lot easier to win than other, merely due to the competition. Certain players will undoubtedly eternally play relatively mediocre competition. Perhaps through no fault of their own, but it still affects their winning percentage. Fred

vte...@white-wolf.com writes: >Quote: >> To produce the desired set of two-player paired outcome probabilities from >> multiplayer games, I calculate a performance for each player of 10^(Skill/D). >> The chance of each player winning the game is their performance divided by the >> sum of performances of all players in the game. >So the statistical results "showing" that the 4:400 formula more closely >follows the skills of the players is actually just showing that it more >closely follows the above-cited method of analysing the skills of the >players. Actually in 2-player games my system produces identical outcome probabilities to the ELO system. It just generalizes to multiplayer systems. I did make a mistake in the stuff I posted. I assumed 20 *games* per player on average. Actually it should have been 20 *seatings* per player. The optimal S/K ratio should be about 9:400, not 4:400. 4:400 is optimal for player who have played in 100 games on average. 32 is still way too high, though. [ quoted text not captured ]

CurtAdams wrote: > > vte...@white-wolf.com writes: > > >Quote: > >> To produce the desired set of two-player paired outcome probabilities from > >> multiplayer games, I calculate a performance for each player of > 10^(Skill/D). > >> The chance of each player winning the game is their performance divided by > the > >> sum of performances of all players in the game. > > >So the statistical results "showing" that the 4:400 formula more closely > >follows the skills of the players is actually just showing that it more > >closely follows the above-cited method of analysing the skills of the > >players. > > Actually in 2-player games my system produces identical outcome probabilities > to the ELO system. It just generalizes to multiplayer systems. Right. And it generalizes to multiplayer in a manner different than the manner in which the VEKN rules generalized to multiplayer, as near as I can tell by the oft-quoted quote above. [ quoted text not captured ]

jtdu...@yahoo.com writes: >I'm sure it's doable. I think the problem is not having anyone >to do it. If there were a volunteer to code something that could >extract data from Archon spreadsheets and automatically calculate >ratings like this, I imagine the VEKN would be more than happy to >have it. Long-term, the best thing to do is to keep the raw data as much as possible. People can go back and reanalyzed as they see fit later that way. In 10 years we'll probably have AI that can read handwriting so you could just dump file drawers of tournament reports into a scanner, so even hardcopy is useful. [ quoted text not captured ]

"CurtAdams" <curt...@aol.com> wrote in message news:20020726130729...@mb-dd.aol.com... > > I did make a mistake in the stuff I posted. I assumed 20 *games* per > player on average. Actually it should have been 20 *seatings* per > player. The optimal S/K ratio should be about 9:400, not 4:400. > 4:400 is optimal for player who have played in 100 games on average. > 32 is still way too high, though. I'm not sure if I'm "getting" the difference between "seatings" and "games". But I think you might be saying: you ran 600 games for 30 players thinking "20 games per player". But since you were running 5-player games you actually selected players for the games 3000 times = 100 games per player. Is that right? If so I guess I see the problem - you don't want to assume that people will play in 100 games each to reach a somewhat-appropriate rating since historically only a few people have played that many rated games. :-) Josh too many games not enough players in fact while we're at it, could your system be rerun with, say, 300 players and 1200 games? might be interesting...

"LSJ" <vte...@white-wolf.com> wrote in message news:3D416CC2...@white-wolf.com... > > For a primer (decent? I don't know), try > http://www.gamasutra.com/features/20000209/kreimeier_pfv.htm > > It seems tailored for online deathmatch style games, but certainly > has enough theory explanation built in to be useful. Wow, that last one is *fascinating.* Plus, it discusses the meanings of "T" and "max_gain" (our S and K)! I highly recommend it to anyone interested in ratings methods, or specifically in Elo-like methods (what it's focused on, as its premise seems to be that Elo-like methods are the "best" ratings for its purposes). Josh impressed

On Fri, 26 Jul 2002 12:28:13 -0400, "Joshua Duffin" <jtdu...@yahoo.com> wrote: >"Frederick Scott" <freds64_at_...@removethis.com> wrote in message >news:3D4070A3...@removethis.com... >> I guess it depends on what you want. As ratings, these clearly >> suck. As statistics, they might be interesting to some - though >> the lack of context makes them mostly random digits to my eyes. > >Well, I'm not yet convinced that "GW percentage" does suck as a >rating method. How often you win tournament games seems to me >quite intuitive as a measure of how good a tournament player >you are. Sure, it doesn't give a bonus for winning the final, >supposedly-most-important game, but the mere fact of having >won an additional game does improve the winner's rating. > I don't know if the Powers That Be of V:EKN can or should agree to this, but I for one would be VERY interested in more robust tournament play statistics of various sorts. The easiest way to implement this, it seems to me, would be to have a volunteer(s) archive Archon reports and make the results available for database queries through a web site. Or just post 4 or five different sets of stats on a regular basis... For example, I would be interested in knowing (1) How often a given player makes the wins/makes final/scores GW/scores VP in a tournament (2) How often a given CLAN (defined as 50% of crypt) wins/makes the finals/scores GW/scores VP (3) The average size of the field in tournaments in a given region/time period (4) Which players appear in tournaments in the most different venues The perfect situation for me would be "Josh is a .500 player when facing opponents who seeded higher than him in a final, but wins .68 when he is top seed. He has never come less than 3rd when scoring more than 0 VP in a final." is an achievable result. I would be willing to archive the results, and even to do some work on designing a database structure, but implementation of any kind of interface is beyond my pitiful skills. Mark Woodhouse Prince Of Minneapolis (aka Data Geek)

vte...@white-wolf.com writes: >> Actually in 2-player games my system produces identical outcome probabilities >> to the ELO system. It just generalizes to multiplayer systems. >Right. >And it generalizes to multiplayer in a manner different than the manner >in which the VEKN rules generalized to multiplayer, as near as I can tell >by the oft-quoted quote above. No; each 2-player matchup within an n-player game produces the expected outcome probabilities. [ quoted text not captured ]

jtdu...@yahoo.com writes: >I'm not sure if I'm "getting" the difference between "seatings" >and "games". But I think you might be saying: you ran 600 games >for 30 players thinking "20 games per player". But since you >were running 5-player games you actually selected players for >the games 3000 times = 100 games per player. Is that right? If >so I guess I see the problem - you don't want to assume that >people will play in 100 games each to reach a somewhat-appropriate >rating since historically only a few people have played that many >rated games. :-) Ding! Right on the money, Josh. One odd thing came up in the sims, though. If the rating movements are "too small" then the ratings aren't accurate. Oddly, though, the rankings actually *improve* a little bit! Any movement was adquate for ranking purposes, down to 0.2. I expect this is only due to my assumption that each players is expected to play the same number of games. If an ELO-like rating system is reimplimented I'll redo the sims with differing play frequencies and look for an optimum. Still, it's interesting - the cost of undershoot is less than one would expect. [ quoted text not captured ]

CurtAdams wrote: > > vte...@white-wolf.com writes: > >> Actually in 2-player games my system produces identical outcome > probabilities > >> to the ELO system. It just generalizes to multiplayer systems. > > >Right. > >And it generalizes to multiplayer in a manner different than the manner > >in which the VEKN rules generalized to multiplayer, as near as I can tell > >by the oft-quoted quote above. > > No; each 2-player matchup within an n-player game produces the > expected outcome probabilities. Oh. The oft-quoted part indicates otherwise: [BEGIN OFT_QUOTED TEXT] > The chance of each player winning the game is their performance divided by > the sum of performances of all players in the game. [END] [ quoted text not captured ]

[ quoted text not captured ] So your system didn't round ratings to the nearest integer? That (obviously) will help improve performance. [ quoted text not captured ]

"Joshua Duffin" <jtdu...@yahoo.com> wrote in message news:ahrtei$vnoea$1...@ID-121616.news.dfncis.de... > > "Xian" <xi...@waste.org> wrote in message > > Goddamit, Josh! [snip] > Come on. I cleverly said "Rob Treasure". No one was going to expect > it from *you*. You would have been fine if you hadn't just gone and > spilled it all over the newsgroup. ;-) D'oh! :) > besides, you can still use your "save face" tech to impress > everyone at the week of nightmares with how much your decks > are going to suck... Boy, are you going to laugh when I actually do. :) Xian

vte...@white-wolf.com writes: >> No; each 2-player matchup within an n-player game produces the >> expected outcome probabilities. >Oh. The oft-quoted part indicates otherwise: >[BEGIN OFT_QUOTED TEXT] >> The chance of each player winning the game is their performance divided by >> the sum of performances of all players in the game. >[END] The ELO system only specifies the outcome of each two-player match. It actually has no explicit system for determining the "winner" of a game and theoretically would permit impossibilities (e.g A beats B beats C beats D in a 4-player game). The way I did it (or an equivalent) is the only way you can get the numbers right if you assume games generate ordinal outcomes. Technically, I suppose you're right; the ELO system does permit non-transitive outcomes and mine doesn't. But since V:tES doesn't permit intransitive outcomes, is that a problem? [ quoted text not captured ]

vte...@white-wolf.com >So your system didn't round ratings to the nearest integer? >That (obviously) will help improve performance. Yes, but the 32-shift system I was comparing it to was also done without rounding. The improvement is relatively small, though, as the 32-shift puts in so much noise the additional rounding error is minimal. Obviously rounding on a 4-shift is a big problem but that's not relevant - the computations are simple enough, and if somebody has a conniption about invisible rounding you can always go the (equivalent) other way and increase the range to 1400 or so. [ quoted text not captured ]

[ quoted text not captured ] I don't know. It's just that, by the VEKN system, the chance of each player winning the game is *not* their performance divided by the sum of the performances of all other players in the game, in general, TTBOMK. [ quoted text not captured ]

"LSJ" <vte...@white-wolf.com> wrote in message news:3D427901...@white-wolf.com... > CurtAdams wrote: > > > The ELO system only specifies the outcome of each two-player match. > > It actually has no explicit system for determining the "winner" of a > > game and theoretically would permit impossibilities (e.g A beats B > > beats C beats D in a 4-player game). The way I did it (or an equivalent) > > is the only way you can get the numbers right if you assume games > > generate ordinal outcomes. > > > > Technically, I suppose you're right; the ELO system does permit > > non-transitive outcomes and mine doesn't. But since V:tES doesn't > > permit intransitive outcomes, is that a problem? > > I don't know. It's just that, by the VEKN system, the chance of each player > winning the game is *not* their performance divided by the sum of the > performances of all other players in the game, in general, TTBOMK. I'm not 100% sure that my approach here is correct, and I'm pretty sure it's incomplete. But, suppose you have a 5-player game under "the old system". A: rating 3400 B: rating 3200 C: rating 3000 D: rating 2800 E: rating 2600 Expected chance of each player beating all other players, according to the "chance = 1/(10^(diff/400))+1" formula: A: 0.75975 * 0.90909 * 0.96934 * 0.99009 = 0.66283 B: 0.24025 * 0.75975 * 0.90909 * 0.96934 = 0.16085 C: 0.09091 * 0.24025 * 0.75975 * 0.90909 = 0.01509 D: 0.03065 * 0.09091 * 0.24025 * 0.75975 = 0.00051 E: 0.00990 * 0.03065 * 0.09091 * 0.24025 = 0.00001 (rounded) Adding up these probabilities, I only get a 0.83929 or 83.93% chance of any player beating all other players. I don't know if this means that the formula expects there to be a tie for first the other 16.07% of the time, or if it just isn't feasible to use that formula this way. (The former seems somewhat unlikely given how much more likely it is for player A to win than for anyone else to win.) Looking at it from the perspective of a single player seems to fit the formula better: chance for various outcomes for A: A > BCDE : 0.66283 B > A > CDE : 0.20961 C > A > BDE : 0.06629 D > A > BCE : 0.02096 E > A > BCD : 0.00663 BC > A > DE : 0.02096 BD > A > CE : 0.00663 CD > A > BE : 0.00209 total for these: 0.99600 or 99.6%, IOW this probably covers all the reasonably-likely possible outcomes for A (according to the formula). These probabilities might be usable for Curt's program, but I don't know if they would be adequate or not? They don't specify, for example, who beats whom among the non-A players. Josh unsure

"Frederick Scott" <freds64_at_...@removethis.com> wrote in message news:3D4183F4...@removethis.com... > > Yea, but the people who haven't played many games haven't played > many games. No, that's not just my fingers keystroke-repeating > themselves because my mind has gone out to lunch (for the third > time today, before 11 AM no less). I mean the influence of such > things depends on how often it happens and I expect if you multiply > people by games, you discover that those how have played a dozen > games or more (arbitrary number plucked from mid-air) influence > the rankings a lot more often than those with less games. Obviously, > you can wipe the players with 0 games right off the list, as they > haven't done a thing to the rating system. Aha, good point there, I hadn't thought of that. Though, if you have a lot more 3-game-players than 50-game-players (and I expect you do), and they generally lose points in those three games, they could contribute "ratings creep" to the people who do play a lot - ie there will be "more points to go around" among the "play in lots of tournaments" crowd since those people will have picked up extra points from the ones who only ever played in one tournament and lost points in those. I'm not sure if that issue is really related, but it is another "flaw" (probably unavoidable, though correctable-for with the "ratings decay" idea) in the old Elo system. (re ratings decay - as described in the article LSJ mentioned at http://www.gamasutra.com/features/20000209/kreimeier_pfv.htm, if you're concerned about maintaining the "zero-sum nature" of an Elo-type system, you can/should "decay" ratings toward the starting point over time (not sure if they were saying you should do this for everybody or just for people who haven't played a rated game within the last X days/weeks/months). So, you assume that a high-rated player who hasn't played for a while isn't actually as good as his rating, and a low-rated player who hasn't played for a while isn't as bad as his rating - that is, since our information about them is old and therefore potentially out of date, we assume that they have become closer to "an average player" since then.) [re GW percentage scoring] > Winning games is always good. But without the adjustment for > skill, there's no saying how good. My experience is that some > games are just a heckuva lot easier to win than other, merely > due to the competition. Certain players will undoubtedly eternally > play relatively mediocre competition. Perhaps through no fault of > their own, but it still affects their winning percentage. This is true, and that article was pretty convincing (to me) that an Elo-type system is a really good way to rate people compared to less-"handicap"-based scoring systems. Josh never knew Elo was a person

Joshua Duffin wrote: > > chance for various outcomes for A: > > A > BCDE : 0.66283 > B > A > CDE : 0.20961 > C > A > BDE : 0.06629 > D > A > BCE : 0.02096 > E > A > BCD : 0.00663 > > BC > A > DE : 0.02096 > BD > A > CE : 0.00663 > CD > A > BE : 0.00209 I'm not seeing where you account for the possibilities of passing Kindred Restructures and Dramatic Upheavals mid-game. Did I miss the "political deck coefficient"? :) Fred

"W. Mark Woodhouse" <shan...@tcinternet.net> wrote in message news:3d41a6c3....@news.tcinternet.net... [re having more analysis-of-tournament-data available] > The perfect situation for me would be "Josh is a .500 player when > facing opponents who seeded higher than him in a final, but wins .68 > when he is top seed. He has never come less than 3rd when scoring > more than 0 VP in a final." is an achievable result. I agree. That would rock. Especially for me to be winning that much. :-) > I would be willing to archive the results, and even to do some work on > designing a database structure, but implementation of any kind of > interface is beyond my pitiful skills. It's probably feasible, the "forward Archons to somebody who wants to maintain a comprehensive database" part. I don't know who'd be up for doing interface, though. Josh not a coder

vte...@white-wolf.com writes: >I don't know. It's just that, by the VEKN system, the chance of each player >winning the game is *not* their performance divided by the sum of the >performances of all other players in the game, in general, TTBOMK. It's just a method for generating the VEKN probabilities. Just like using a pseudorandom number generator to make a die roll. [ quoted text not captured ]

jtdu...@yahoo.com writes: >Adding up these probabilities, I only get a 0.83929 >or 83.93% chance of any player beating all other players. >I don't know if this means that the formula expects there to >be a tie for first the other 16.07% of the time, or if it >just isn't feasible to use that formula this way. (The former >seems somewhat unlikely given how much more likely it is for >player A to win than for anyone else to win.) You're assuming that every possible player-player interaction is independent. The missing 16.07% is the intransitive outcomes (e.g. [ quoted text not captured ]

CurtAdams wrote: > > vte...@white-wolf.com writes: > > >I don't know. It's just that, by the VEKN system, the chance of each player > >winning the game is *not* their performance divided by the sum of the > >performances of all other players in the game, in general, TTBOMK. > > It's just a method for generating the VEKN probabilities. Just like > using a pseudorandom number generator to make a die roll. Yes, but it doesn't, TTBOMK, generate (match) the VEKN probabilities. It generates some other, possibly similar, probabilities. [ quoted text not captured ]

LSJ wrote: > > CurtAdams wrote: > > > > vte...@white-wolf.com writes: > > > > >I don't know. It's just that, by the VEKN system, the chance of each player > > >winning the game is *not* their performance divided by the sum of the > > >performances of all other players in the game, in general, TTBOMK. > > > > It's just a method for generating the VEKN probabilities. Just like > > using a pseudorandom number generator to make a die roll. > > Yes, but it doesn't, TTBOMK, generate (match) the VEKN probabilities. It > generates some other, possibly similar, probabilities. I don't really understand what your concern is, here. I realize that there's a lot of different capacities and aptitudes people have: deckbuilding, card playing, diplomacy, and so forth. Things will cause players to do better in certain situations relative to how well they do in others, such as being better at (or having decks build better for) winning the two-man game at the end vs. surviving the initial oustings to get into that position. Ultimately, I'm not sure why the rating system has to account for all these. There's a very simple thing that the VEKN Elo system tests: on average, will player A do better than, worse than, or as well as player B in terms of victory points. That's all it has to do and it's not a test that really cares which aptitudes cause which result. As long as the prediction is basically accurate over time, that's fine. Is there something about the nature of Curt's model that you fear is skewing his results enough to make his prediction of the optimal constants substantially off? If so, how? Fred

Frederick Scott wrote: > Is there something about the nature of Curt's model that you fear is skewing > his results enough to make his prediction of the optimal constants > substantially off? If so, how? Yes. As stated, his estimation for the chance of a given player winning is not the same as what the VEKN Elo system computes as that player's chances. To use the psuedo-random number for a die roll as an exagerrated example, it would be like using a pseudo-random integer from 1 to 10 and returning 1 for 1 2 for 2 3 for 4 and 5 4 for 6 and 7 5 for 8 and 9 6 for 10 [ quoted text not captured ]

jtdu...@yahoo.com writes: >Adding up these probabilities, I only get a 0.83929 >or 83.93% chance of any player beating all other players. >I don't know if this means that the formula expects there to >be a tie for first the other 16.07% of the time, or if it >just isn't feasible to use that formula this way. (The former >seems somewhat unlikely given how much more likely it is for >player A to win than for anyone else to win.) You're assuming that every possible player-player interaction is independent. The missing 16.07% is the intransitive outcomes (e.g. A beats B beats C beats A) that can't happen in a V:tES game. The ELO definition is designed for 2-player and is indeterminate when applied to multiplayer, since hypothetically there can be correlations amongst player outcomes (if A beats B and B beats C, A must beat C). If you assume players must come out in a transitive ranking but there are no other correlations (e.g. A beating B doesn't change C's chance against D) then you'll come out with something isomorphic to my system. I haven't proven it, but it seems pretty obvious. (Sorry about the abortive post. #$%&*$# AOL) [ quoted text not captured ]

vte...@white-wolf.com writes: >Yes, but it doesn't, TTBOMK, generate (match) the VEKN probabilities. It >generates some other, possibly similar, probabilities. It *does* generate identical probabilities. Calculate the chance of A beating B using my system and it's identical to the VEKN system. [ quoted text not captured ]

CurtAdams wrote: > > vte...@white-wolf.com writes: > > >Yes, but it doesn't, TTBOMK, generate (match) the VEKN probabilities. It > >generates some other, possibly similar, probabilities. > > It *does* generate identical probabilities. Calculate the chance > of A beating B using my system and it's identical to the VEKN system. That what's the business about summing them up and dividing by the number of players? (see the oft-quoted text). [ quoted text not captured ]

"LSJ" <vte...@white-wolf.com> wrote in message news:3D472B61...@white-wolf.com... > CurtAdams wrote: > > > > vte...@white-wolf.com writes: > > > > >Yes, but it doesn't, TTBOMK, generate (match) the VEKN probabilities. It > > >generates some other, possibly similar, probabilities. > > > > It *does* generate identical probabilities. Calculate the chance > > of A beating B using my system and it's identical to the VEKN system. > > That what's the business about summing them up and dividing by the > number of players? (see the oft-quoted text). I think what Curt's saying is: the VEKN Elo-based scoring system doesn't have any inherent method for calculating the chance of winning the game for each player in a multiplayer game. It only has a formula for determining the expected chance of one player beating one other player. If you use Curt's "performance" stat to calculate the chance of one player beating one other player - ie look at the oft- quoted text and assume a two-player game - it's equivalent to the VEKN expected-chance formula. But as he explained in response to the message where I tried to calculate "chances of beating all other players" using the VEKN formula, there are about 16% of outcomes "unaccounted for" after calculating "chance for A to beat D-E, chance for B to beat A&C-E, etc". This is, he proposes, because the VEKN formula assumes that it's possible for A to beat C and C to beat B and B to beat A. It seems plausible to me, given that the probabilities I was getting for "A beats all, B beats all, etc" weren't going to reach 1 even when adding "A ties with B, A ties with C, etc". And since the VEKN system has no way to tell us anything about the results within non-A players if we just look at A's chances against each other player, we have to (I think) use some other system to interpret what their "skills" mean in terms of chances for each outcome. So: I'm not 100% sure that Curt's method of generalizing to multiplayer is the best way to match the VEKN expected-win formulas. But it seems very plausible, given that it does match for the 2-player case and that as far as I can see there is no way to use the exact VEKN formulas to produce a result. Josh might hack more numbers later

Joshua Duffin wrote: > > "LSJ" <vte...@white-wolf.com> wrote in message > news:3D472B61...@white-wolf.com... > > CurtAdams wrote: > > > > > > vte...@white-wolf.com writes: > > > > > > >Yes, but it doesn't, TTBOMK, generate (match) the VEKN probabilities. > It > > > >generates some other, possibly similar, probabilities. > > > > > > It *does* generate identical probabilities. Calculate the chance > > > of A beating B using my system and it's identical to the VEKN system. > > > > That what's the business about summing them up and dividing by the > > number of players? (see the oft-quoted text). > > I think what Curt's saying is: the VEKN Elo-based scoring > system doesn't have any inherent method for calculating the > chance of winning the game for each player in a multiplayer > game. It does once you define what "winning the game" you're talking about. That's why I'm asking for clarification. What "chance of winning the game" (what "winning" is being computed)? Simply summing and dividing will not produce the same chances as are actually calculated by the VEKN system. If one begins with the assumption that it cannot be done meaningfully, then why do it? [ quoted text not captured ]

"LSJ" <vte...@white-wolf.com> wrote in message news:3D4992C9...@white-wolf.com... > Joshua Duffin wrote: > > > I think what Curt's saying is: the VEKN Elo-based scoring > > system doesn't have any inherent method for calculating the > > chance of winning the game for each player in a multiplayer > > game. > > It does once you define what "winning the game" you're talking about. > > That's why I'm asking for clarification. > > What "chance of winning the game" (what "winning" is being computed)? I'm confused. How are you saying the VEKN system can compute a chance for "winning the game"? Can you supply a formula? As far as I can tell, the VEKN system *doesn't* calculate a chance for "winning the game" (= "scoring the highest number of VPs in the game, beating all other players"). > Simply summing and dividing will not produce the same > chances as are actually calculated by the VEKN system. I can't see where the VEKN system calculates chances for "winning the game". It only calculates pairwise chances for beating players individually. > If one begins with the assumption that it cannot be done > meaningfully, then why do it? It seems to me still meaningful to model chances of winning, based on what the ratings imply about skill, even if the existing system doesn't provide us enough guidance to use only it as the basis for the model. No? Josh modeling clay

LSJ wrote: > > Joshua Duffin wrote: > > I think what Curt's saying is: the VEKN Elo-based scoring > > system doesn't have any inherent method for calculating the > > chance of winning the game for each player in a multiplayer > > game. > > It does once you define what "winning the game" you're talking about. > > That's why I'm asking for clarification. > > What "chance of winning the game" (what "winning" is being computed)? If you mean, "What chance does a given player have of receiving more victory points than any other player in a game of Jyhad, I guess *I'm* confused why this is relevant. (At least why it's directly relevant.) The ratings formula says nothing about predicting the outcome in that manner. It only rates players' chance of finishing with more victory points than one another, not what their chances are of finishing with more points than any other player. Fred

Joshua Duffin wrote: > > "LSJ" <vte...@white-wolf.com> wrote in message > news:3D4992C9...@white-wolf.com... > > Joshua Duffin wrote: > > > > > I think what Curt's saying is: the VEKN Elo-based scoring > > > system doesn't have any inherent method for calculating the > > > chance of winning the game for each player in a multiplayer > > > game. > > > > It does once you define what "winning the game" you're talking about. > > > > That's why I'm asking for clarification. > > > > What "chance of winning the game" (what "winning" is being computed)? > > I'm confused. How are you saying the VEKN system can compute > a chance for "winning the game"? Can you supply a formula? Not without knowing what Curt means by "winning the game" , no. The formula should be derivable from the Elo pairwise information, yes. > As far as I can tell, the VEKN system *doesn't* calculate a > chance for "winning the game" (= "scoring the highest number > of VPs in the game, beating all other players"). The VEKN system doesn't calculate since it isn't needed for the ratings, true. Curt's system *does* calculate it (using a formula that does not match the VEKN system), and uses it, according to the quote, to determine who "wins" the game. Thus the (apparent) discrepancy that leads to the concern about the validity of the results of statistical analysis on the data produced from that calculation. > > Simply summing and dividing will not produce the same > > chances as are actually calculated by the VEKN system. > > I can't see where the VEKN system calculates chances for > "winning the game". It only calculates pairwise chances for > beating players individually. Right. > > If one begins with the assumption that it cannot be done > > meaningfully, then why do it? > > It seems to me still meaningful to model chances of winning, > based on what the ratings imply about skill, even if the > existing system doesn't provide us enough guidance to use > only it as the basis for the model. No? No. Unless by some miracle of cancellations (or possibly some definitions I haven't thought of) deriving the formula for "winning the game" from the pairwise Elo forulae comes out to "sum and divide". -- LSJ (vte...@white-wolf.com) V:TES Net.Rep for White Wolf, Inc. Links to V:TES news, rules, cards, utilities, and tournament calendar: http://www.white-wolf.com/vtes/

Frederick Scott wrote: > > LSJ wrote: > > > > Joshua Duffin wrote: > > > I think what Curt's saying is: the VEKN Elo-based scoring > > > system doesn't have any inherent method for calculating the > > > chance of winning the game for each player in a multiplayer > > > game. > > > > It does once you define what "winning the game" you're talking about. > > > > That's why I'm asking for clarification. > > > > What "chance of winning the game" (what "winning" is being computed)? > > If you mean, "What chance does a given player have of receiving more > victory points than any other player in a game of Jyhad, I guess *I'm* > confused why this is relevant. (At least why it's directly relevant.) > The ratings formula says nothing about predicting the outcome in that > manner. It only rates players' chance of finishing with more victory > points than one another, not what their chances are of finishing with > more points than any other player. Right. Hense the concern of the accuracy of a statistical analysis based on a system which uses what Curt (not what I) mean by "chance of winning the game". [ quoted text not captured ]

"LSJ" <vte...@white-wolf.com> wrote in message news:3D49B753...@white-wolf.com... > Joshua Duffin wrote: > > > I'm confused. How are you saying the VEKN system can compute > > a chance for "winning the game"? Can you supply a formula? > > Not without knowing what Curt means by "winning the game" , no. > The formula should be derivable from the Elo pairwise information, > yes. But it's not. The pairwise expectations, as Curt pointed out, don't take into account the correlations. For example, when I tried to calculate Elo expectations for each player's chance of beating all other players in a 5- player game, I did: A beats B * A beats C * A beats D * A beats E. But that's not right, because if A beats B and C, then D and E *cannot* both beat A. One of them could beat A (if A got 1 and B and C each got 0, D could get 2 and E get 1), but one of them would have to at best tie A. That is, the chances aren't entirely independent. The Elo formulas also don't give a chance of tying (just a result for what happens if people *do* tie), so would leave possible results unaccounted for. (You're right that Curt's analysis also has this problem, in that it apparently also ignores tying - ties are actually broken for tournament finals by rolling a die, but aren't for individual games, so this is an issue.) And in fact, every possible actual 5-player result has at least two players tying (this is also true for 4-player games, I believe): 5/0/0/0/0 4/1/0/0/0 3/2/0/0/0 3/1/1/0/0 2/2/1/0/0 2/1/1/1/0 1/1/1/1/1 3.5/0.5/0/0/0 2.5/1.5/0/0/0 1.5/1.5/1/0/0 2.5/0.5/0.5/0/0 1.5/1.5/0.5/0/0 1.5/0.5/0.5/0.5/0 0.5/0.5/0.5/0.5/0.5 Elo pairwise expectations can't ever tell us what the chance of that happening is. What we need is a way to figure out VP distribution and then adjust ratings based on that. The "way of figuring out VP distribution" probably ought to be based on "skill" but it can't just use the Elo expectations because they don't exist for actual allowed VP distributions. As far as I can tell, there is literally no way to model a VTES game using only the pairwise Elo expectations. You have to make outside assumptions somewhere to come up with a model of possible outcomes. Curt's model is actually pretty close to the "ignoring correlations of results" probabilities from the Elo expectations: Curt's "skill" model (again for the 3400/3200/3000/2800/2600 set of players A-E): A wins: 0.68594 B wins: 0.21691 C wins: 0.06859 D wins: 0.02169 E wins: 0.00686 Elo expectations (note these do not sum to 1 like Curt's do, because of the "correlated results" problem): A wins: 0.66283 B wins: 0.16085 C wins: 0.01509 D wins: 0.00051 E wins: 0.00001 E's chance of winning, for example, is probably vastly lower than it should be because E has such a small chance of beating each player, especially the highest-"skilled". But if E *does* beat the highest-skilled player, E automatically at least ties with the other three players. Anyway... I don't think I have my head wrapped around the probability stuff enough to come up with a really cogent justification of a method of applying "skills" to estimate game results. But I hope this shows that the Elo expectations alone can't be used to produce such a method either. (This message got heavily re-edited since I ended up working through all this reasoning while writing it; hopefully I don't end up repeating myself too much or leaving out crucial information. :-) Josh at length

Joshua Duffin wrote: > > "LSJ" <vte...@white-wolf.com> wrote in message > news:3D49B753...@white-wolf.com... > > Joshua Duffin wrote: > > > > > I'm confused. How are you saying the VEKN system can compute > > > a chance for "winning the game"? Can you supply a formula? > > > > Not without knowing what Curt means by "winning the game" , no. > > The formula should be derivable from the Elo pairwise information, > > yes. > > But it's not. > > The pairwise expectations, as Curt pointed out, don't take > into account the correlations. > > For example, when I tried to calculate Elo expectations for > each player's chance of beating all other players in a 5- > player game, I did: > > A beats B * A beats C * A beats D * A beats E. > > But that's not right, because if A beats B and C, then D > and E *cannot* both beat A. One of them could beat A (if > A got 1 and B and C each got 0, D could get 2 and E get 1), > but one of them would have to at best tie A. > > That is, the chances aren't entirely independent. True. But that isn't to say that you cannot compute the chances of player A winning* given the chances of player A beating players B,C,D, and E. That just means that the formula (if it exists) is non-trivial. * whatever Curt means by this. You first have to find out exactly what Curt meant by this to even attempt to find the formula. Hence my asking. > The Elo formulas also don't give a chance of tying (just a > result for what happens if people *do* tie), so would leave > possible results unaccounted for. (You're right that > Curt's analysis also has this problem, in that it apparently > also ignores tying - ties are actually broken for tournament > finals by rolling a die, but aren't for individual games, so > this is an issue.) Again, that's why I'm asking what the "sum and divide" is used for. > And in fact, every possible actual 5-player result has at > least two players tying (this is also true for 4-player > games, I believe): Yes, but immaterial to the "chance of A winning the game". > Elo pairwise expectations can't ever tell us what the chance of > that happening is. What we need is a way to figure out VP > distribution and then adjust ratings based on that. The "way > of figuring out VP distribution" probably ought to be based on > "skill" but it can't just use the Elo expectations because they > don't exist for actual allowed VP distributions. Just because the formulation isn't trivial doesn't mean it cannot be derived. It may in fact not be derivable, but it's far too early in the analysis of the subject (at least as far as I can tell) to pronounce it so. [ quoted text not captured ] I'm not sure how you computed the "Elo expectations" (especially after you claim it cannot be done), but, assuming you've done do correctly, then this is exactly the problem. Why not use the Elo expectations instead of the sum-and-divide expectations? > Anyway... I don't think I have my head wrapped around the > probability stuff enough to come up with a really cogent > justification of a method of applying "skills" to estimate > game results. But I hope this shows that the Elo expectations > alone can't be used to produce such a method either. Then how did you come up with your Elo expectations, if that isn't possible? [ quoted text not captured ]

thela...@hotmail.com (The Lasombra) wrote in message news:<3d3e2c0...@text.giganews.com>... > On Tue, 23 Jul 2002 21:49:58 -0400, "Nosferatu Stuff" > <roans...@yahoo.com> wrote: > > >Hey, what did they do with the stats. I just went to prove to someone that > >James had a crappy record, and I find that he is at zero....and I'm at zero. > >New look, new design. Whats that all about? INFO? > > >http://206.65.59.246/VTES/member.tpl > > > New set, new stats. > As requested. > > When the person who volunteered to deal with the ratings is allowed to > begin entering tournament results again, you will have a much better > picture of the players and of their histories. You will also have a > much better picture of the people running your events as you will be > able to tell at a glance if they sent in the tournament results on > time or not. > > > Three cheers for Conrad & the ratings updater!! Three cheers indeed, but... (forgive me if this has been handled, I've been on vacation and cannot spend the time to read all the posts in the meantime) My statistics are already wrong (as far as i can determine from the information given to me), as, I think, are some of the other Dutch players (more difficult to check). Evidence: I have played in 2 constructed santioned tourneys, taking 0 vp in the first and 2.5 and a TW in the second (no other VP that tournament). The info at WW says I played 6 career games (correct), of which 3 constructed (incorrect), with 0 TW (incorrect) and 1 VP (incorrect). My name, for those who wish to check, is Tobias op den Brouw. How do I go about correcting this? Tobias Deventer

"LSJ" <vte...@white-wolf.com> wrote in message news:3D4A9B12...@white-wolf.com... > Joshua Duffin wrote: > > > That is, the chances aren't entirely independent. > > True. But that isn't to say that you cannot compute the > chances of player A winning* given the chances of player > A beating players B,C,D, and E. That just means that the > formula (if it exists) is non-trivial. Non-trivial enough that I can't even picture it, yes. :-) > * whatever Curt means by this. You first have to find > out exactly what Curt meant by this to even attempt to > find the formula. I thought it was obvious that by "winning the game" he meant "winning the game by the game's terms", ie getting a higher VP total than any other player. Which is to say, beating players B through E. Coming in second then means beating the third-through- fifth ranked players; coming in third means beating the fourth and fifth ranked players; coming in fourth means beating the fifth ranked player. > > The Elo formulas also don't give a chance of tying (just a > > result for what happens if people *do* tie), so would leave > > possible results unaccounted for. (You're right that > > Curt's analysis also has this problem, in that it apparently > > also ignores tying - ties are actually broken for tournament > > finals by rolling a die, but aren't for individual games, so > > this is an issue.) > > Again, that's why I'm asking what the "sum and divide" is > used for. As far as I can tell, it's to rank the players in order to produce pairwise results to apply the ratings-change formula to. Curt's "sum and divide" looks to me like it ranks the players ordinally from 1st through 5th finish and each player loses to every player he doesn't beat. Hence it ignores ties, just like the Elo expectations... > > And in fact, every possible actual 5-player result has at > > least two players tying (this is also true for 4-player > > games, I believe): > > Yes, but immaterial to the "chance of A winning the game". No, it's not, because the chance of A winning the game depends on the chance of each possible game outcome. > > The "way > > of figuring out VP distribution" probably ought to be based on > > "skill" but it can't just use the Elo expectations because they > > don't exist for actual allowed VP distributions. > > Just because the formulation isn't trivial doesn't mean it > cannot be derived. It may in fact not be derivable, but it's > far too early in the analysis of the subject (at least as > far as I can tell) to pronounce it so. Heh. Well, I suppose you're right, but I don't see how to make the leap from "expected chance for player 1 to beat player 2 based on their ratings" to "expected chance for player 1 to tie with player 2". > > As far as I can tell, there is literally no way to model a > > VTES game using only the pairwise Elo expectations. You have > > to make outside assumptions somewhere to come up with a model > > of possible outcomes. Curt's model is actually pretty close > > to the "ignoring correlations of results" probabilities from > > the Elo expectations: > > > > Curt's "skill" model (again for the 3400/3200/3000/2800/2600 > > set of players A-E): > > > > A wins: 0.68594 > > B wins: 0.21691 > > C wins: 0.06859 > > D wins: 0.02169 > > E wins: 0.00686 > > > > Elo expectations (note these do not sum to 1 like Curt's do, > > because of the "correlated results" problem): > > > > A wins: 0.66283 > > B wins: 0.16085 > > C wins: 0.01509 > > D wins: 0.00051 > > E wins: 0.00001 > > I'm not sure how you computed the "Elo expectations" (especially > after you claim it cannot be done), but, assuming you've done > do correctly, then this is exactly the problem. Why not use the > Elo expectations instead of the sum-and-divide expectations? As I wrote above the listed probabilities, the Elo-based ones are "ignoring correlations of results". I may have deleted the lines where I described the method I had attempted there, sorry. It was: chance of A winning = chance of A beating all other players = Prob(A beats B) * Prob(A beats C) * Prob(A beats D) * Prob(A beats E) The "wrongness" here is that if A beats B and C, then A *cannot* lose to D and E; A could lose to D and tie with E (if A got 1, B and C got 0, D got 2, E got 1) but can't lose to both. This correlation of results isn't accounted for in simple multiplicative probability calculations (naturally). Also, there's no Elo expectation for "A ties with E" to plug in, even if I figured out correlations and the limitations they impose on possible results. (This "no expectation of tie" seems to me the "making it impossible" sticking point for using only Elo expectations to create a model. If you can see a solution for it, please explain...) > > But I hope this shows that the Elo expectations > > alone can't be used to produce such a method either. > > Then how did you come up with your Elo expectations, if that > isn't possible? By getting them wrong. :-) Josh dazed and confusticated

Joshua Duffin wrote: > "LSJ" <vte...@white-wolf.com> wrote in message > > I'm not sure how you computed the "Elo expectations" (especially > > after you claim it cannot be done), but, assuming you've done > > do correctly, then this is exactly the problem. Why not use the > > Elo expectations instead of the sum-and-divide expectations? > > As I wrote above the listed probabilities, the Elo-based ones > are "ignoring correlations of results". > > I may have deleted the lines where I described the method I > had attempted there, sorry. It was: > > chance of A winning = > chance of A beating all other players = > Prob(A beats B) * Prob(A beats C) * Prob(A beats D) * > Prob(A beats E) > > The "wrongness" here is that if A beats B and C, then A > *cannot* lose to D and E; A could lose to D and tie with > E (if A got 1, B and C got 0, D got 2, E got 1) but can't > lose to both. Ah. Right. Then the "true" derivation would have to take the correlations into account. So while your probabilities are possibly better than the sum-and-divide ones, there would still be room for improvement. [ quoted text not captured ]

"LSJ" <vte...@white-wolf.com> wrote in message news:3D4AA870...@white-wolf.com... > Joshua Duffin wrote: [I thought that] > > chance of A winning = > > chance of A beating all other players = > > Prob(A beats B) * Prob(A beats C) * Prob(A beats D) * > > Prob(A beats E) > > > > The "wrongness" here is that if A beats B and C, then A > > *cannot* lose to D and E; A could lose to D and tie with > > E (if A got 1, B and C got 0, D got 2, E got 1) but can't > > lose to both. > > Ah. Right. Then the "true" derivation would have to take > the correlations into account. So while your probabilities > are possibly better than the sum-and-divide ones, there > would still be room for improvement. I don't think mine *are* better, though, because they're clearly wrong. They don't sum to 1. I'm not sure the Elo-expectations-formula approach would be worth the effort (since it looks like *so* much effort) to make a model out of even if a solution to "no expectation for ties" could be found, either. At least, I'm probably not going to put the effort in, as I don't even see any chance of it working in the first place. :-) To my eyes, then, a related approach (since Curt's version used the same "skill expression" as the Elo expectations) that produces fairly similar probabilities to a naive Elo-expectation-approximation is about as good a model as we're likely to see. If further analysis were desired, probably the thing to do would be to take actual game results from as many Archon spreadsheets as possible, plug them into an Elo ratings calculator with various choices for the constants, and put together some graphs of the results. If anyone (James Coupe?) manages to whip up an "Elo machine" and get the game results, maybe we can see that at some point. :-) Josh met my match

Joshua Duffin wrote: > > "LSJ" <vte...@white-wolf.com> wrote in message > news:3D4AA870...@white-wolf.com... > > Joshua Duffin wrote: > > [I thought that] > > > > chance of A winning = > > > chance of A beating all other players = > > > Prob(A beats B) * Prob(A beats C) * Prob(A beats D) * > > > Prob(A beats E) > > > > > > The "wrongness" here is that if A beats B and C, then A > > > *cannot* lose to D and E; A could lose to D and tie with > > > E (if A got 1, B and C got 0, D got 2, E got 1) but can't > > > lose to both. > > > > Ah. Right. Then the "true" derivation would have to take > > the correlations into account. So while your probabilities > > are possibly better than the sum-and-divide ones, there > > would still be room for improvement. > > I don't think mine *are* better, though, because they're > clearly wrong. They don't sum to 1. But their ratios are better. If you want to sum them to one, then OK - just normalize them (in the same ratios so that they sum to one) (or, once we find out how this "chance of winning" is being used, do something more approriate to handle the tie cases). > I'm not sure the Elo-expectations-formula approach would be > worth the effort (since it looks like *so* much effort) to > make a model out of even if a solution to "no expectation > for ties" could be found, either. At least, I'm probably > not going to put the effort in, as I don't even see any > chance of it working in the first place. :-) Well, since even tracking the Elo ratings was too much effort, you've probably got a point there. [ quoted text not captured ]

vte...@white-wolf.com writes: >> I think what Curt's saying is: the VEKN Elo-based scoring >> system doesn't have any inherent method for calculating the >> chance of winning the game for each player in a multiplayer >> game. >It does once you define what "winning the game" you're talking about. Not really. ELO only specifies what the pairwise outcomes are. Winning could mean lots of things. Even if you define it as "beating all other players" ELO is still ambiguous as it doesn't mention anything about transitivity or correlations between players. You can cook up a lot of screwball systems which still meet the ELO pairwise outcomes. Curt Adams (curt...@aol.com) "It is better to be wrong than to be vague" - Freeman Dyson

CurtAdams wrote: > > vte...@white-wolf.com writes: > > >> I think what Curt's saying is: the VEKN Elo-based scoring > >> system doesn't have any inherent method for calculating the > >> chance of winning the game for each player in a multiplayer > >> game. > > >It does once you define what "winning the game" you're talking about. > > Not really. ELO only specifies what the pairwise outcomes are. > Winning could mean lots of things. Even if you define it as > "beating all other players" ELO is still ambiguous as it doesn't > mention anything about transitivity or correlations between > players. You can cook up a lot of screwball systems which > still meet the ELO pairwise outcomes. <blink> So, back to the original question, what, exactly, does your system think it is doing when it computes the chance that "Player A will win the game" per the oft-quoted quote, and how does that figure get used? [ quoted text not captured ]

vte...@white-wolf.com writes: >Just because the formulation isn't trivial doesn't mean it >cannot be derived. It may in fact not be derivable, but it's >far too early in the analysis of the subject (at least as >far as I can tell) to pronounce it so. But in this case the formula *cannot* be derived. Josh nicely posted two different systems (mine and completely uncorrelated results) which both meet the ELO definitions (the chance of player X beating player Y is exactly as specified) and yet results are quite different. ELO does not specify correlations and you *must* know them to determine the mp outcome. "Uncorrelated" might be a reasonable additional assumption except that it's patently wrong since it permits intransitive outcomes. It would still be an additional assumption, though. [ quoted text not captured ]

jtdu...@yahoo.com writes: >The "wrongness" here is that if A beats B and C, then A >*cannot* lose to D and E; A could lose to D and tie with >E (if A got 1, B and C got 0, D got 2, E got 1) but can't >lose to both. A profound problem with any ELO extension (including mine) is that it ignores ties. With 2-player it's not a problem, you just pretend the game didn't happen. But with multiplayer it's a problem; you can't throw out the entire game since other pairs may have decisive outcomes. So, you need a tie probability, which ELO doesn't have. [ quoted text not captured ]

CurtAdams wrote: > > jtdu...@yahoo.com writes: > > >The "wrongness" here is that if A beats B and C, then A > >*cannot* lose to D and E; A could lose to D and tie with > >E (if A got 1, B and C got 0, D got 2, E got 1) but can't > >lose to both. > > A profound problem with any ELO extension (including > mine) is that it ignores ties. With 2-player it's not > a problem, you just pretend the game didn't happen. "Standard" (e.g., Chess Federation) Elo system doesn't. It takes the tie result and proceeds accordingly. > But with multiplayer it's a problem; you can't throw > out the entire game since other pairs may have > decisive outcomes. So, you need a tie probability, > which ELO doesn't have. But it handles ties just fine. [ quoted text not captured ]

CurtAdams wrote: > > vte...@white-wolf.com writes: > > >Just because the formulation isn't trivial doesn't mean it > >cannot be derived. It may in fact not be derivable, but it's > >far too early in the analysis of the subject (at least as > >far as I can tell) to pronounce it so. > > But in this case the formula *cannot* be derived. > Josh nicely posted two different systems (mine > and completely uncorrelated results) which both > meet the ELO definitions (the chance of player > X beating player Y is exactly as specified) If I knew how your system used the "winner" computation, I might be able to see that. So, what does your system do after calculating the "winner" of the game? [ quoted text not captured ]

>And in fact, every possible actual 5-player result has at >least two players tying (this is also true for 4-player >games, I believe): > >5/0/0/0/0 >4/1/0/0/0 >3/2/0/0/0 >3/1/1/0/0 >2/2/1/0/0 >2/1/1/1/0 >1/1/1/1/1 > >3.5/0.5/0/0/0 >2.5/1.5/0/0/0 >1.5/1.5/1/0/0 >2.5/0.5/0.5/0/0 >1.5/1.5/0.5/0/0 >1.5/0.5/0.5/0.5/0 >0.5/0.5/0.5/0.5/0.5 Life Boon. I always remember how it can create bizarre results due to having to overwrite The Archon one time because it didn't like people's VP totals.

"Curevei" <cur...@aol.commetal> wrote in message news:20020802141608...@mb-cb.aol.com... > > Life Boon. I always remember how it can create bizarre results due > to having to overwrite The Archon one time because it didn't like > people's VP totals. Ah yes, you're right. Life Boon only creates a new "possibility", though, when it grabs someone's half-VP for timing out. So, additional timeout possibilities would be: 4/0/0/0/0 3/1/0/0/0 2/2/0/0/0 2.5/1/0.5/0/0 etc. I think they would still all have the property "at least two people have tied", though, because it looks like grabbing the extra half-point will always cause a new tie- pair if it eliminates the old one. Josh can't remember ever actually seeing a life boon nick a half-VP

"LSJ" <vte...@white-wolf.com> wrote in message news:3D4ABED5...@white-wolf.com... > CurtAdams wrote: > > > A profound problem with any ELO extension (including > > mine) is that it ignores ties. With 2-player it's not > > a problem, you just pretend the game didn't happen. > > "Standard" (e.g., Chess Federation) Elo system doesn't. > It takes the tie result and proceeds accordingly. No, I think he means it doesn't *predict* ties, so you ignore that they can happen when making predictions. The chess Elo system doesn't predict ties either. It can supply an expected chance for Player 1 to beat Player 2, and 100% minus that chance is the chance for Player 2 to beat Player 1. But it has no expectation for the chance of a draw. > > But with multiplayer it's a problem; you can't throw > > out the entire game since other pairs may have > > decisive outcomes. So, you need a tie probability, > > which ELO doesn't have. > > But it handles ties just fine. In results, sure. In modeling possible games, no. Josh win place or show

"CurtAdams" <curt...@aol.com> wrote in message news:20020802130808...@mb-fx.aol.com... > vte...@white-wolf.com writes: > > >Just because the formulation isn't trivial doesn't mean it > >cannot be derived. It may in fact not be derivable, but it's > >far too early in the analysis of the subject (at least as > >far as I can tell) to pronounce it so. > > But in this case the formula *cannot* be derived. > Josh nicely posted two different systems (mine > and completely uncorrelated results) which both > meet the ELO definitions (the chance of player > X beating player Y is exactly as specified) and yet > results are quite different. ELO does not specify > correlations and you *must* know them to determine > the mp outcome. "Uncorrelated" might be a reasonable > additional assumption except that it's patently wrong > since it permits intransitive outcomes. It would still > be an additional assumption, though. I think that in theory, the best assumption to make would be to write down all possible VTES-game outcomes, and then figure out the probability of each one for each game you simulate. This is cumbersome in itself since if I remember right there are 120 possibilities (for a five-player game) just for games that *don't* time out. :-) Then, once you have the list of outcomes, you still have no way to figure probabilities "correctly" for any of them because there is no prediction for the "these players tie" piece of each outcome. Sigh. Scott, did you read that article you posted the link to earlier? This one? http://www.gamasutra.com/features/20000209/kreimeier_pfv.htm It seemed to come close to saying something useful about choosing parameters for a multiplayer Elo system (though by multiplayer there they seem to mean "two players or more" and are concerned pretty much only with pairwise rankings/ ratings). But then it stops short, for me, of implying anything definitive. It almost seems to say that you need data on player skill to choose the constants. How annoying. :-) Josh can't hack it

Joshua Duffin wrote: > > "LSJ" <vte...@white-wolf.com> wrote in message > news:3D4ABED5...@white-wolf.com... > > CurtAdams wrote: > > > > > A profound problem with any ELO extension (including > > > mine) is that it ignores ties. With 2-player it's not > > > a problem, you just pretend the game didn't happen. > > > > "Standard" (e.g., Chess Federation) Elo system doesn't. > > It takes the tie result and proceeds accordingly. > > No, I think he means it doesn't *predict* ties, so you > ignore that they can happen when making predictions. Hard to extract such a meaning from what he wrote, but whatever. > The chess Elo system doesn't predict ties either. It can > supply an expected chance for Player 1 to beat Player 2, > and 100% minus that chance is the chance for Player 2 to > beat Player 1. But it has no expectation for the chance > of a draw. Um. Right. > > > But with multiplayer it's a problem; you can't throw > > > out the entire game since other pairs may have > > > decisive outcomes. So, you need a tie probability, > > > which ELO doesn't have. > > > > But it handles ties just fine. > > In results, sure. In modeling possible games, no. Right again. [ quoted text not captured ]

Joshua Duffin wrote: > Scott, did you read that article you posted the link to > earlier? This one? > > http://www.gamasutra.com/features/20000209/kreimeier_pfv.htm Yes. > It seemed to come close to saying something useful about > choosing parameters for a multiplayer Elo system (though by > multiplayer there they seem to mean "two players or more" > and are concerned pretty much only with pairwise rankings/ > ratings). But then it stops short, for me, of implying > anything definitive. It almost seems to say that you need > data on player skill to choose the constants. How annoying. > :-) Right. [ quoted text not captured ]

vte...@white-wolf.com writes: >So, back to the original question, what, exactly, does your system >think it is doing when it computes the chance that "Player A will >win the game" per the oft-quoted quote, and how does that figure >get used? With a little algebraic manipulation the chance of player A beating player B is: 10^(SkillA/D)/(10^(SkillA/D)+10^(SkillB/D)) ) (Here skill refers to the player's true rating) This is equivalent to throwing a dart at a board with area 10^(SkillA/D) marked for A and 10^(SkillB/D) marked for B. The player whose section gets hit is the winner. What I do is generalize to a n-player dartboard. To pick the top player, throw a dart at the board and pick that player as the winner. Remove that section of the board and repeat. Each 2-player outcome is settled the first time the dart hits one of the player's section. Obviously, this is the same as the ELO system. Making each dart determine all remaining pairwise encounters for the hit player averts intransitive outcomes. Having no overlap or section size alterations during the process avoids any other correlations, which would be counterintuitive. [ quoted text not captured ]

On 2 Aug 2002 08:14:53 -0700, tobiasop...@hotmail.com (Tobias) wrote: >My statistics are already wrong (as far as i can determine from the >information given to me), as, I think, are some of the other Dutch >players (more difficult to check). All of the statistics for this year are being re-entered. When all of the games up through April 25th have been entered, then your constructed games played should match your career games played if all of your tournaments took place this year. If you had games/tournaments from previous years, they will not be a part of the new statistics. >How do I go about correcting this? You don't. Give the people working on it some time, and realize that most of the people involved in the programming and updating of the statistics have either already left for GenCon, or they will be leaving at the beginning of next week. Don't expect any major updates to the statistics before the 19th or so (minor updates are happening every day though). Carpe noctem. Lasombra http://www.TheLasombra.com

jtdu...@yahoo.com writes: >Then, once you have the list of outcomes, you still have >no way to figure probabilities "correctly" for any of them >because there is no prediction for the "these players tie" >piece of each outcome. Although if you only concern yourself with the one winner, you can dodge the tie problem (timeout and no-winner can be excluded) [ quoted text not captured ]

> > How do I go about correcting this? > > Tobias > Deventer To explain the weirdness of the ratings: FIRST: The "rank" calculator is just.....broken....something. There are plenty of us with 10 games to qualify for rank and yet only 3 players in the world are currently "ranked", with Kevin Scribner being 1st. (Which I find hysterical.) The team is looking in to it and that should be working shortly. Uh, after GenCon probably. SECOND: Currently, career games for anyone who's played in a tournament this year are wrong. This should be corrected in few days. I am currently catching up converting to Archon 1.1e and inputting all 103 tournaments from Jan 2 to April 30 that I had already entered. The new data base carried over the "career games" total and every time I enter a tournament it adds those games to the career total. After the converting is complete, I will then go back and re-adjust everyone's career games to match exactly what they were as of April 30. Tobias, in your specific example I entered a tournament for you that occurred for you on January 12. According to the tournament report that your prince turned in you had .5 vp in round 1 & .5 vp in round 3. The site reflects that. I don't have the results of your Tilburg tournament on July 13. Maybe you can nudge your prince for me. Hope this helps, Robyn Tatu Ratings Soddess vtesr...@white-wolf.com

rob...@mindspring.com (robyn tatu) wrote in message news:<56d4bf85.02080...@posting.google.com>... > > > > How do I go about correcting this? > > > > Tobias > > Deventer > > To explain the weirdness of the ratings: > > FIRST: The "rank" calculator is just.....broken....something. There > are plenty of us with 10 games to qualify for rank and yet only 3 > players in the world are currently "ranked", with Kevin Scribner being > 1st. (Which I find hysterical.) > The team is looking in to it and that should be working shortly. Uh, > after GenCon probably. Good on you! Note that I am thankful for the time you are putting into this. > SECOND: Currently, career games for anyone who's played in a > tournament this year are wrong. This should be corrected in few days. Right... ok. No hurries, was just wondering. > > I am currently catching up converting to Archon 1.1e and inputting all > 103 tournaments from Jan 2 to April 30 that I had already entered. The > new data base carried over the "career games" total and every time I > enter a tournament it adds those games to the career total. After the > converting is complete, I will then go back and re-adjust everyone's > career games to match exactly what they were as of April 30. Hope things go smoothly for you. > Tobias, in your specific example I entered a tournament for you that > occurred for you on January 12. According to the tournament report > that your prince turned in you had .5 vp in round 1 & .5 vp in round > 3. The site reflects that. That's interesting. AFAIK, I should only have .5 vp in round 3, and not in round one, but memory might be rusty. I can now understand where the number comes from, thanks. > I don't have the results of your Tilburg tournament on July 13. > Maybe you can nudge your prince for me. I would, but as has been commented upon, even though players showed, that tournament did not happen. You can remove it from your list. > Hope this helps, It certainly did, thanks. Tobias Deventer

CurtAdams wrote: > > vte...@white-wolf.com writes: > > >So, back to the original question, what, exactly, does your system > >think it is doing when it computes the chance that "Player A will > >win the game" per the oft-quoted quote, and how does that figure > >get used? > > With a little algebraic manipulation the chance of player A beating player B > is: > > 10^(SkillA/D)/(10^(SkillA/D)+10^(SkillB/D)) ) > > (Here skill refers to the player's true rating) > > This is equivalent to throwing a dart at a board with area > 10^(SkillA/D) marked for A and 10^(SkillB/D) marked > for B. The player whose section gets hit is the winner. > > What I do is generalize to a n-player dartboard. To pick the > top player, throw a dart at the board and pick that player > as the winner. Remove that section of the board and repeat. Thanks for the elaboration/repetition. I understand the process now, at any rate. [ quoted text not captured ]