Math People Only!: Changes to how much ratings change...

Ok. There has been talk that ratings move up/down too much. It doesn't bother me, but I know it bugs some people. As some of you know, we use Glicko system (http://en.wikipedia.org/wiki/Glicko_rating_system and http://math.bu.edu/people/mg/research/gdescrip.pdf ). Here is what we use to start:

    StartRating = 1200;
    StartRd = 350;
    KValue = 16;
    CValue = .2;
    PValue = .000011;
    QValue = .0057565;

We use 1 minute as the time increment.

What should we change? :D

(Non-related and stupid posts will be deleted.)

FWIW, i think that a target RD for someone who plays a LOT of chess on Chess.com should be 30-50. i play quite a bit and my RD=81.

Perhaps we should lower the initial RD and slightly raise the P value.

The Law Of Large Numbers states:  "the average of the results obtained from a large number of trials should be close to the expected value, and will tend to
come closer as more trials are performed".

In simpler terms, if a coin is flipped it has a 50-50 chance of being either heads or
tails.  At first, one might flip, for example, five heads in-a-row.  But after 1000
flips, the number will be perhaps 507-493.

Jacob Bernoulli first proved this in 1713.

Honestly I have thought of this a lot--an option for "half rated" games would be easy with glicko math. It would come in handy for thematic games (forced openings) and offer less risk, more incentive for higher rated players to compete with players over 400 points below them, which is good for the site. Basically, a half rated game option would be revolutionary as I don't believe anyone has done it. The only difference is Rating change and RD change are both divided by 2.

Cheers

Along the same lines, you might consider having the two scores determine the ratio, but the players can agree to an amount wagered.  As in any gambling situation (usually for money, but here for points), the odds are given by the house, but the player can still decide how much to bet.  If the odds are 2:1, then I can bet $1 to lose $1 or win $2, or I can bet $2 to lose $2 or win $4, and so on.  Likewise here, a formula based on scored might say that Player A can win 3X by winning or X by drawing, and Player B can win X by winning.  Now the two players can determine X, or whoever issues the challenge can do so.

I think we should remove starting rating completely and have players 'ungraded' until they have played 10 games!?.  From those 10 games we should calculate their performance rating and that should become their actual rating from which future rating calculations should take place.

To solve the problem of grades fluctuating rapidly I think it would be a good idea to slow the rate at which the RD fluctuates.

Possible there could be some AI to determine? Take samples from X amount of players who are just starting, have played 100+, and 500+ ?  Also possible here that a lot of losses on time effect this?

I find the emails saying I'm about to lose, come way to late, or not at all :( Therefore I amy lose to a bunch of lower rated's and my rating fluctuates widly.

I agree on NOT computing a rating on new members until several games are played, so if I play a new member which is bound to be 2100 rating in the future and he beats me, I'm not gonna loose 300 points at him if he's still 1200 when I play.

As for other math, I'm not specialist.

Those parameters look quite reasonable. By my calculations it would take 130 days without playing for an RD to rise from 50 to 100, using those parameters - which is reasonable. So I don't quite understand why my RD is 100 today. Unfortunately I haven't recorded a history of it, but my impression is that RDs are either rising much faster than that, or are not falling enough as games are played. Are you sure the implementation is correct?

Is there a min value set at 30?

that's what we're trying to figure out!

Yes that is annoying, especially now in Live Chess as everyone's ratings have been reset to 1200.

erik, are you sure you are using a time increment of 1 minute and not 1 second?

Is there any data you can show us of an example of somebody's change in RD and rating over time?

All I know is that when I played a 1800 player and won, I only got 35 points for the win, yet when I got beat by some 1400 player who had yet to get to his "normal" rating I lost 45. How is that fair?

why can't chess.com use elo?

"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."

Isn't using t = 1 minute very extreme then?

perhaps there should be an option to toggle elo or glicko?

maybe even show both?

Elo has no advantages over Glicko, and would still need the paramaters tweaking to trade off variability against responsiveness. And ELO requires kludges for "provisional" ratings that Glicko incorporates seamlessly.