Sie sind auf Seite 1von 12

Chess Rating

Ann Miura Eric Carr

Chess Rating
Goal Paired comparison scaling to make a ranking of all active players and to serve as a basis for selection of players for training and tournaments Description Rating and scoring of two-player game Paired-comparison scaling Extrapolation of player relative strength from small number of games Geographically distributed players Generalize diverse set of backgrounds, styles and strategies into one number

Interesting Facts
Kasparov does not hold either of two World Titles in Chess but is universally regarded the best chess player in the world by virtue of his number 1 ranking (2813 rating in Oct 2004)

Rating Scale

ELO/FIDE is most widely accepted ranking system (FIDE - Fdration Internationale des checs)
Source: http://www.ex.ac.uk/~dregis/DR/bcftable.html

FIDE Algorithm (Elo)


Base awards and penalties around difference in rating of 2 players and index of how established those ranks are A difference of 200 rating points in Chess would mean that the stronger player has an expected score of 0.75 Score represents probability of winning + 0.5 * probability of draw Index considers how many and how significant games were to calculate rank (modeled as K in FIDE system) Given Provisional rating with less than 30 games played Create probabilistic likelihood of win/loss of two opponents based around rating difference and index Intuitively, larger upset results in more points being transferred Against equally ranked opponent, win equals ~ 5-12 points Large upset generally rewards up to ~ 25 points
Source: http://www.fide.com/official/handbook.asp?level=B0210

FIDE Algorithm (Elo): Simplified Example


1613

A vs. A vs. A vs. A vs. A vs.

B
1609

LOSS DRAW WIN WIN LOSS

C
1477

1613

D
1388

1613

E
1586

1613

B
1720

A is penalized for having a worse performance than should be expected given his opponents. If he had scored two wins, one loss, and two draws, for a total score of three points, that would have been slightly better than expected, and his new rating would have been (1613 + 32*(3 - 2.867)) = 1617.

1613

Score: 0+0.5+1+1+0 = 2.5 E_Score: 0.506+0.686+0.785+0.539+0.351 = 2.867 New Rating: 1613+32*(2.5-2.867) = 1601
Source: http://www.wordiq.c om/definition/ELO_rating_system

FIDE Algorithm - Likelihoods

Conversion of rating difference into likelihood of winning for each player H - higher ranked player L - lower ranked player

No History Rating - Ru
Rating Player with no Rating history: Tournament Play
Scores > 50%

Ru = Rc+12.5 * each .5% above 50%

Determine Average Rating of Competition Rc

Scores 50%

Ru = Rc

Scores < 50%

Swiss/Team: Ru = Rc + dp Round Robin: Ru = Rc + dp * n/(n+1)

Rc = R a for Unrated Players Ra = Tournament Average Rar = Avg Rating of all Rated Players dpa = determine p for rated players, then determine dp for each, then avg Ra = R ar - d pa * n/(n+1)

Rn = Game Weighted Average Of Ru Results

Avg Rating > 1800 ?

Published Unrated Player Rating Rn

Source: http://www.fide.com/official/handbook.asp?level=B0210

History Rating - Rn
Rating Player with Rating history: Tournament Play
Determine Rc For Tournament D = R - Ra Determine PD W = Achieved Score We = Expected Score
We = PD * n for Swiss or individual play

Change in Rating DR = K * (W - We)

Difference in rating above 350 is treated as 350 for rating purposes Rc = (Ra (n+1) - R) / n R = Rating of Player n = Number of Players Ra = avg rating of all tournament players Calculating new R n : Rn = R 0 + SDR R0 = Old Rating SDR = Sum DR for each new event Source: http://www.fide.com/official/handbook.asp?level=B0210

K = Development Coefficient Players not ranked with R n < 1801 K = 25 until player has completed 30 games K = 15 while Player Rating < 2400 K = 10 once published Player Rating > 2400 (remains permanently at 10 thereafter)

The Magic of Rating Differences and Scoring Probabilities


Difference in rating D into scoring probability PD Suppose two players are to play a game with relative strengths vi and vj. Difference in rating D = vi vj The function F that relates D to PD should have the following qualities:
PD = 0.5 if D = 0 and PD > 0.5 if vi >vj vi vj! -" for PD = 0, vi vj! +" for PD = 1 F should also be symmetric around 0.5

Normal distribution fits these properties and makes sense if we assume that the fluctuations in performance are purely random and are not caused by changing properties of I and j Scaling factor of c = 200 2

1 P( D) = 800 !
Source: http://www.schaakbond.nl/rating/elopaper.html

D c "#

1 2 " t 2

Challenges
Ranking of computer players Retrospective rankings not accurate How to model draws (likely in chess) Performance is not typically normally distributed (kurtosis and skew) Performance variance varies between players but is generally modeled as a constant Players entering and leaving the rating pool (turnover) leads to loss of points from the system not a strict zero-sum system and system average rating is not stable Subpools of players: deflation and inflation

Other Topics
Current research focused influence of covariate information like age, playing routine, style and proficiency level on the reliability of the rating system (can incorporate these into the Kfactor) Rating Databases - Quick Chess (QC), Regular Tournament, Correspondence Chess, Online Algorithm has motivated and moved into other areas Video Games - Xbox Online Splinter Cell Sports - College Football, Soccer, Volleyball, Tennis, Golf British Chess Federation (BCF) scale ELO=(BCF*8)+600: 1 BCF point = 8 ELO point USCF Conversion
USCF initially aimed for an average club player to have a rating of 1500.

http://math.bu.edu/people/mg/ratings/approx/approx.html WorldChessRating.com

Das könnte Ihnen auch gefallen