Remi, you mentioned how the other algorithms predicted well and guessed that it's because the great majority of games are between experienced players whose strength is not changing much. I also feel that the existing KGS ratings work well for those players already. So how about focusing on how the various algorithms perform in the case of improving players. I think it would be interesting to simulate game results of various improving players and show how the different rating algorithms work.
For example: Suppose a player's true strength is 1500 for some time, and then he suddenly improves to 2000. Both before and after he plays a fixed number of games per day (say 10). Show a graph of what each rating algorithm would think his rating is over time. Many people complain that the KGS algorithm does not move fast enough for a case like this. Also the last paragraph of section 4 talks about how the model does not account for the different ability of new players to change (improve) their ratings compared to older players. Could you vary the parameter 'w' based on the player's current rating? (Assume players with low ratings are capable of improving more quickly than strong players). I don't know enough about the math to know if this would blow up the computation time or if that's simply impossible. On Tue, Apr 8, 2008 at 5:37 PM, Rémi Coulom <[EMAIL PROTECTED]> wrote: > Hi, > > This is my CG2008 paper, for statisticians: > > Whole-History Rating: A Bayesian Rating System for Players of > Time-Varying Strength > > Abstract: Whole-History Rating (WHR) is a new method to estimate the > time-varying strengths of players involved in paired comparisons. Like > many variations of the Elo rating system, the whole-history approach is > based on the dynamic Bradley-Terry model. But, instead of using > incremental approximations, WHR directly computes the exact maximum a > posteriori over the whole rating history of all players. This additional > accuracy comes at a higher computational cost than traditional methods, > but computation is still fast enough to be easily applied in real time > to large-scale game servers (a new game is added in less than 0.001 > second). Experiments demonstrate that, in comparison to Elo, Glicko, > TrueSkill, and decayed-history algorithms, WHR produces better > predictions. > > http://remi.coulom.free.fr/WHR/ > > Feedback is welcome. > > Rémi > > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ >
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
