|
Post by Dubber on Mar 10, 2009 1:16:05 GMT -5
Basically, huge discussions and ELO formula tweaking occurred around March 2004 on ravenblackgames... I'm not sure where the calculation discussion tree ended... but I did find the date when I mentioned I had a higher ELO than Taliesin games.groups.yahoo.com/group/ravenblackgames/message/350ravenblackgames #350 March 3, 2004 Re: ELO ratings
I find it fun that I currently have a better elo than Taliesin. Hopefully no one will bet on me against him with real $$$ :-)' The current calculation seems to favor those who play *a lot* (like myself) over those who actually try to live a life outside Warlocks. -Dubber :-)'
--- In ravenblackgames@yahoogroups.com, RavenBlack <raven@r...> wrote: > >Which scores are used in the formula - the scores when the > >challenge was posted, challenge accepted, or end of game? > > End of game. Unlike ladder, if it turns out during your > game that the person is actually crap (or actually great) > it's appropriate for that to be reflected in the Elo, > because of what Elo is measuring (probability of victory > between two players). > > --RavenBlack
|
|
|
Post by Rycchus on Mar 10, 2009 12:06:00 GMT -5
So it's basically the same table except with the negative values in "you surrender" set to zero. That's what I thought. Shouldn't make much difference to anyone trying to work out ELO differences from the RB/dubber links.
|
|
|
Post by Rycchus on Mar 10, 2009 12:12:23 GMT -5
games.groups.yahoo.com/group/ravenblackgames/message/367RE: [ravenblackgames] We calculation I've now changed it so that you can't gain Elo for surrendering, or lose it for being surrendered to - it cuts off with no change in the event that one of these would be the case. And since I was there, here's the bit of code used to get the expected win-score value. $elowe=exp(-$elodiff*$elodiff/53824); $elowe=($elodiff>0)?(1-$elowe/2) $elowe/2); >Have you seen the Sonas variation RB? It's a much more >simple formula to implement and to my layman's brain it seems >more accurate. *shrug* Not more simple than "already implemented" though, and it's not like accuracy is going to be an issue what with the bit of random factor and/or guesswork inherent to the game. >http://www.chesscenter.com/twic/event/sonas/sonasrat.html >It's a linear function. Not sure if it shows the same problem >of a win possibly losing points. It would have, without the change I just made. The difference between the two is really negligible, except for the Sonas one being a *tiny* amount more accurate (according to the person there trying to get it adopted), and, more importantly, Elo being better-known and already implemented. --RavenBlack
|
|
|
Post by Rycchus on Mar 10, 2009 12:18:08 GMT -5
Also mentioned in games.groups.yahoo.com/group/ravenblackgames/message/360is forced surrenders: >2. Ravenblack, if you want effective K=25, consider using K=30 or so >in the formula.
It's not really 20 rather than 25; it's probably about 21. There are, after all, still some kills even amongst the stronger players. I quite enjoy the side-effects, really - if your rating is 400 higher than someone else's you really should be able to work up a kill (unless they don't play at all and have a forced surrender, in which case it doesn't count in the Elo rankings at all).
I might go for making the scores 1.05 for a kill and 0.95 for a surrender to balance it out a bit, at some point - effectively increasing K, but also reducing the negative effect of having someone rubbish surrender to you.
--RavenBlackIs it true that a FS has no score change or is he just referring to the effect we already know about, that games less than a certain length have no score recorded? Edit: found something else on it. >What about for a forced surrender? No points, both sides?
There are additional conditions for some things that are ignored. A forced surrender isn't always ignored (because otherwise people wanting to maintain their Elo could just wait three days instead of surrendering properly when they're about to lose).
|
|
|
Post by Rycchus on Mar 10, 2009 12:21:17 GMT -5
Also, Yaron notes that you need to adjust upwards when reading the table...
When you consult the table (in Ravenblack's new ELO page) to find out your chances of beating someone, the better player is underestimated (i.e., her real chance of beating the opponent is higher than what the table tells you).
Detailed explanation: Table Skew: The table is actually quite correct, as long as you don't read it as giving your chance of winning, but rather as giving the score you expect. As an extreme example, if player A is 366 points above player B, then the table says that A expects to get a score of 0.9 in an average game against B. In standard ELO, this means a 90% chance of winning. In the 0.9/0.1 system, it could mean 90% chance of killing, but it could also mean 100% chance of winning without a kill. With better players, kills are rare, so a difference of, for instance, 149 points, listed as 0.7, would probably mean 75% win, rather than 70% kill. You just have to add a bit to the number listed for the better player, to get the real probability of winning.
This of course is already accounted for if you're using Dubber's table and not the one on the RavenBlack ELO page.
|
|