Hi there,

i had a interesting last week about the ranking system on K2Z.

After 7 days without playing, you lose 1 point/day….

The goal of this is not to force people to play.

In K2Z i had several exemples of 2 players, that played against each other more than 50 times reached a 1600+ ranking (because of bonus)

then disapeared. and no news for months and months.

Without this system, those players would be in the top 20

six months after their last games.

So i wanted a system that show the best and active players.

with no zombies.

Any remark is welcome :)

\pvs

A remark from the Twixt statistician ;-)

777 unique player names have played Twixt on Little Golem. At the current time, 303 of them are listed in Little Golem's Twixt player list. Among the missing is the venerable dog “Snoopy”, who would still be in the Top 10 (as #10) if he were still around.

I don't know how long it takes to become de-ranked on Little Golem. I couldn't find it in the FAQ. Does anyone know? Snoopy's last game ended on May 5, 2004, so it must take less than a year. If it's 6 months, then Pit will vanish in about 6 weeks.

Perhaps a kinder, gentler way to handle inactivity is to calculate some kind of uncertainty on the rating. With a high uncertainty, you get a “?” next to your rating. You start out in this state of uncertainty, eventually lose the “?“, but if you don't play for a while, it comes back. And then “?” players would not normally be listed in the rankings.

Regardless of how you handle it, I think that the ELO rating system is a masterpiece, and therefore one should be very careful and respectful when tweaking it.

It's one month of inactivity - from all games. Pit will not have a problem as he has chess games going on.

I am still on the rating lists here, even though it's well over one month since I played my last game. Maybe it has something to do with logging in also.

i use ELO system on K2Z.

it may be a good idea to extend the inactivity period

from 7 days to 30 days for example.

Philippe, I believe you use the mathematics of ELO as part of your rating system, and your system has the outward appearance of an ELO rating system, but it's not really an ELO rating system. An ELO rating system attempts to measure the skill of the players. Modifications to ELO systems almost always attempt to improve the accuracy of that skill measurement in some way. But your modifications undermine the skill measurement. Hence, not really an ELO system.

That's not to say you have to have an ELO system.

But that's off-topic.

You say, “i wanted a system that show the best and active players.” That's 2 variables - best, and active. For 2 variables, I recommend 2 numbers. Whether the second number manifests itself as a disappearance, or the addition of a question mark, or some other marker, or column, it's undeniably an orthogonal variable to skill.

By putting the two together, you have a number that reflects a combination of skill and activity, which makes it both, and in a way, neither. Two 1650 players can't be confident that they're evenly matched, because one could be downsliding from 2000 after 357 days of playing on Little Golem only, while the other is a newbie pumped up on bonus points.

Not to single you out, Philippe, I have an issue with the way Little Golem, Yahoo! Games, Playsite, and other board game sites handle their meta-game. It's a classic user interface design error to copy the physical limitations of a physical system when you don't have to. The ELO system was designed for Chess tournaments, where the ratings would be written on a board. If you'd like to know your win expectancy against your current opponent, during a physical competition, you generally have to do it yourself. But here, you're sitting in front of a computer! And your win expectancy against another player would be much, much more meaningful than these 4-digit rating numbers.

I think that's about all I have to say on this topic.

Ludoteka has a new client and a new rating system. Your “rating” is a value from 1 to 99, indicating what percentage of total players you should be stronger than. The server tells you at the start of each game, what your chances are of defeating your opponent. (For rated games, anyway.)

i see your point alan…. difficult to translate a strenght or win expectancy to a number.

i have a problem with new strong players. for example when

klauss did register, he was 1500. its not fair to let klauss beat 1600+ players because they will lose many points so i set his ranking to 1950 just behind david. (i did not want to put him at no 1). We all know klauss strenght so its not a problem.

New players start at 1500., if they dont play enough, they will stay at 1500+ or 1600+ whereas their level is more like 1800+ or more. if they play someone real 1500 they will crush him.

Im not sure its clear. anyway the problem is some people play 10 or 50 times more than others.

its not easy to give a ranking that reflect the real level and win expectancy.

the ELO system is a convergent algorithm but it needs a minimum number of iterations.

may be a weighted win/loss ratio could be a good idea…..

On www.mindports.net (that's where I play Havannah), they have an ELO system with a separate class for newcomers. I don't know the formulas but I noticed the points for new players seem to be a lot more *elastic*. The points settle a bit after 20 or 25 games, i'm not sure… It seems a good Idea. It certainly helps me to estimate if my opponent's ELO is reliable!

It might be a good solution for you also.

(The system is explained in detail on the site)

I would like to have a rating system with following features: If you play against a player who has only played few games, your rating shouldn´t change much. And the more games you have played, the less your rating should be changed

Why shouldn't my rating change much, when I've played a lot of games? If I suddenly become better because I learned some new patterns in Twixt or read about the long-chain-rule in Dots&Boxes, the rating should increase to reflect my current strength at this game, and not show my skill during my first n matches where I had no clue about a game.

I want to have an exact rating (as far as possible), I don´t want to know, if i have a good or a bad day. If you get stronger, your rating also will increase if it is not only a lucky streak. The problem with the current rating is that if there are 2 persons A and B who play a lot of games, the rating of A gets unreasonably high if he is better than B

Here on little golem the newbies win or loose normal points from a game while the opponent ranking is not affected too much. After some finished games when the rating of the newbie seamed to be stabilized the points won/lost by the newbie opponents are the same with the newbiew's.

Regarding the 1 point per day loose I am the most affected since I played from the beginning and after a while when I come back I get a loosy rank.

no major changes yet about the ranking. i have to think about it. its not easy to have an accurate method that works for active and non active players.

ELO system converge if the player is active. For low activity players no method can insure a ranking you can trust.

i made a minor change, the inactivity period is 30 days now.

anyway, who really cares about his ranking ?

Ranking may be too visible, and discourage newbies.

The important thing, for me, is an estimate of my relative strength against a potential opponent.

It might also be nice to have some indication of the accuracy of the estimate, and my personal win-loss record against the opponent.

on the top 100 page, there is a the ranking and also the number of wins and loss.

The number of game played gives you the accuracy of the ranking.

i could store the head to head stats, but it will be meaningfull only for the top5 players that play the most…

of course i wont take in account the games vs the bot.

Yes, I know you have overall win-loss stats, but there's no way to tell if a player's been playing against strong players, or weak ones.

While it wouldn't be “accurate” to know I have a “1-0” record against someone, it would still be interesting. They may want revenge :-) It may be more meaningful if this stat only queried the last 1 year of games, though.

the small difference you want in rating when you play newbies, and the start of newbies is handled differently on differnet sites.

On goldtoken.com and jijbent.nl, they take for the first 15 games: the average of the rating points you get from that game: for a win you get the opponents rating +400 (jijbent:+200), for a draw the opponents rating for a loss -400 of the opponents rating. (which can be recalculated as the average opponent rating + 400w , where w is (the number of wins minus number of losses)/(number of games)). This is called provisional rating. Here the number of highest influence of a game is normally 32 (a win against in a even match changes the rating 16). When playing AGAINS a newbie it is much less.