Thursday, 24 February 2011

A better rating method?

The contest to find a better method of rating chess players (see my earlier post) has come to an end. It was won by Yannis Sismanis, who utilised a method called stochastic gradient descent. The winner has published a short paper giving the technical details of his method.
Does this mean he has developed a better ratings 'system'? Based on my reading of the paper, I would say not. Simply because the conditions of the contest involved utilising a large set of training data which was then tested against a smaller set of test data. However the training data covered a number of rating periods (ie not just results from the previous rating period), while most ratings system simply use the results from tournaments played in the previous 2,3 or 4 months. (In fact his method had a discounting factor for older games in the data set).
However it does indicate possible improvements in how ratings are calculated. In fact there is now a follow up contest to look at this is issue, once again run by Kaggle. The conditions are slightly more restricted than the previous one, but the winner picks up $10,000 and a trip to Athens to present their system to FIDE.


Anonymous said...

Read it again.
There are two separate prizes.
There are no restrictions on the method used for the $10,00 prize.
There are numerous restrictions in place to win the Athens trip.

The $10,000 goes to whoever gets the lowest binomial deviance score.
The Athens trip goes to the lowest binomial deviance score that meets all the restrictions on the method used.
Of course someone could win both prizes if they meet all the restrictions and still get the overall lowest binomial deviance score.

Mike Barnes said...

All methods have got inconvenients and advantages. Rating used now is quite good, players approve it without to much problem...
Mike from France

Anonymous said...

how are the ratings calculated?