well, guys, what is so important about the rating anyway? Why don't we just get on with life and focus on improving our chess, not throwing tantrums about this rating system? Since you all say that the rating may not reflect the actual strength of your chess ability, why not just push it out of your mind?
Math People Only!: Changes to how much ratings change...
Sort:
So far I'm thinking lower the RD, raise t, and very slightly change c, although
I am certainly no expert and will continue reading.
zank, RD is not a parameter. If we tweak t,c or q then this would be to lower RDs generally. t or c to make it go up more slowly over time, or q to make it drop by more when a game finishes.
I would like an idea where rating points can't be created or destroyed.
Let's say players X and Y are versing each other. Let's say X's rating is 1678 and Y's is 1825. My idea was that win/loss was depending on the sum of ratings divided by the difference (rounded of course).
Example: if X won, his new rating would be (1678+1825)/abs(1678-1825)
3503/147
24 (rounded) Y would lose 24 points and X would gain Y's points.
So, Player X would have a new rating of 1727 and Y would have a rating of 1801
If the two players where of the same rating the winner would gain 2 points.
If a draw was present, then the higher player would lose a quater of the rating difference and the lower player would gain the higher rated player's points.
For Example: 2000 vs. 2 drawn 2000-(.25(2000-2))
2000-(.25(1998))
2000-500 (about)
1500
if such a draw were possible the 2 rated player would gain a fair amount of his share.
*Use these principles in your games and see how fair they are. I don't expect them to be perfect--it's just an example.
dmerkal wrote:
I would like an idea where rating points can't be created or destroyed.
Let's say players X and Y are versing each other. Let's say X's rating is 1678 and Y's is 1825. My idea was that win/loss was depending on the sum of ratings divided by the difference (rounded of course).
Example: if X won, his new rating would be (1678+1825)/abs(1678-1825)
3503/147
24 (rounded) Y would lose 24 points and X would gain Y's points.
So, Player X would have a new rating of 1727 and Y would have a rating of 1801
If a 1801 played an 1800 and won, his rating would go up by (1801+1800)/abs(1801-1800) = 3601/1 = 3601 points, and his new rating would be 5401.
If an 800 played a 2200 and won, he would get 2 points for his incredible upset.
Well I think the point of rating changes is to get the ratings closer together, so if Player X has a rating of 800 and Player Y has a rating of 2200, and they alternate wins, the two ratings should get closer and closer together, right?
There is a simple solution if you worry about losing too many points to new players with low ratings: don't play them. New players should expect to play players with similar ratings as and until they rise up to their natural level. That's what I did as a newbie: winning my first 14 games, and without anyone losing too many points, because at each step I was playing someone with a similar rating. If I am challenged by a new player with a low rating, I simply refuse and explain why.
[COMMENT DELETED]
i'm very much happy with this type of GLICKO RATING procedure. the best thing is that it is very much scientific & logical..
ExtraBold, we can manipulate the RD by changing the other variables, can't we,
or am I misinterpreting ?
Oct 16, 2009
0
#52
ive just been thinking about this in the last few days, even before i just saw this thread now. Personally i think the RD is too high and it should be much lower for people who have played many games! I play slow and i dont finish a game every week so i have a high RD (110 or something like that). For players who have played over lets say a 100 games or so their rating should atleast be pretty accurate, and i think that if i can lose 250 rating points from losing 3-4 games in a row is very annyoing and aggravating (dont worry i will get over it)
The RD value is the thing i think needs to be changed! Im not sure if the formula for RD value can be made different depending on how many games you ahve played on stuff, or maybe it would be very very hard to do, but i would like it.
Oct 16, 2009
0
#53
Jolly, not sure I agree with this statement, "Glicko assumes that time makes you a worse player." The only assumption I think he makes is that the rating becomes less accurate the longer you don't play (a rated game). It just means the next time you play your rating will fluctuate more drastically than if you were playing every day.
Oct 16, 2009
0
#54
Just to answer some questions about the K value, here is where I got some of my ideas from:
http://www.freechess.org/Help/HelpFiles/glicko.html
search that for K value stuff. I do basically the same thing, K is not a constant, its a calculated value, but then if its less than 16, it is set to 16. However, after talking to Mark Glickman on the phone, he didn't know why we would ever do this, so I have removed the code for now (but have not pushed it live).
Mark suggested we simply lower the c value, but by how much, he did not know. Since its currently .2, I thought maybe lowering it to be .1 would be worth a try. I also asked Mark if using minutes was good (since this is what FICS uses), and he said the more granular the better, so minutes is fine. I asked about the assumption of 5 years passing before a rating is assumed to be completely inaccurate, and he said thats a pretty arbitrary # and that changing it to be 1 year or 10 years is something worth trying. Changing this assumption is whats going to change the c value, since c is calculated by:
where in this case 30 = # of "rating periods" that pass before your RD should reach 350 again. Our rating period is 1 minute, so in this formula, if you are using 5 years, then you have to replace 30 with
2 629 743
and solve for c, which gives you .2 (which is where I got the value).
so yes, all these #s are very arbitrary and somewhat random. :)
Oct 16, 2009
0
#55
Oct 16, 2009
0
#56
jay wrote:
Here is another link for reference, which Mark gave me. Same as his pdf I think:
Yes jay, I read that. What jumped out at me was this statment:
The Glicko system works best when the number of games in a rating period is moderate, say an average of 5-10 games per player in a rating period. The length of time for a rating period is at the discretion of the administrator.
From what erik posted, chess.com uses t = 1 minute, I presume so that RD pretty much updates instantly. Maybe this is less than ideal and is the reason for RD's seemingly creeping up so much.
kurtgodden said
I like meniscus' suggestion of "half rated" games when playing thematic openings and such, when you are not beginning from the standard starting position. (I dislike his other suggestion of a toggle between Elo and Glicko.)
I didn't suggest a toggle. Just the halfrated games.
jay wrote:
Here is another link for reference, which Mark gave me. Same as his pdf I think:
About that document, I read that before, and there is one thing about it that I don't understand ( I may have asked about that before, but I don't recall getting an answer). Glickman gives the following formula for deriving the new RD from the old one:
The very first RD of any player is 350, so the next RD would be the minimum of 350 and the squareroot of 350 squared plus a positive number, which is 350 again. So, by natural induction, the RD never seems to change.
Did I do something wrong?
I see what I did wrong, sorry, there is also the calculation of RD', which can drop below 350.
I suggest a new system that doesn't have the disadvantages of ELO and everything: If you lose the game then your new rating is (old rating) - (opponent's rating divided by 100). If you win then your new rating is (old rating) + (oppenent's rating divided by 100). Ratings are rounded up to the nearest integer. A draw results in a won situation for the weaker player (but still the result is 1/2-1/2) and a lost situation for the stronger player. Unrated players gain their opponent's rating upon a win. Unrated player losing to a rated player is (opponent's rating divided by 2). If one unrated player wins another, then the winner gets no new rating. Draws do not change ratings unless ratings are different.
chess.com needs something unique.
Elo has tons of advantages over Glicko. Glicko assumes that time makes you a worse player. Want your score to go up quickly? Player a lower rated player every few days or weeks and in between play the computer or study. Your score will zoom upwards.
This is just a misunderstanding. Elo is essentially Glicko but with everyone having the same RD. Or perhaps provisional players having a higher RD. Rather than this binary division between provisional and regular players, Glicko allows provisional players to gradually become regular as they play and their RD drops.
Over time, your RD creeps up again - this is not becoming a worse player - your rating stays the same. But we gradually lose confidence in the rating - the likelihood increases over time that when you come back you will have improved or worsened.
The problem of being able to get too many points beating weak players is the problem of RDs being too high. You would get the same problem with ELO using the wrong K value. And that's why we're looking at the parameters.