Python implementation of Weng-Lin Bayesian ranking, a better, license-free alternative to TrueSkill
This is a port of the amazing openskill.js package.
Installation
pip install openskill
Usage
>>> from openskill import Rating, rate
>>> a1 = Rating()
>>> a1
Rating(mu=25, sigma=8.333333333333334)
>>> a2 = Rating(mu=32.444, sigma=5.123)
>>> a2
Rating(mu=32.444, sigma=5.123)
>>> b1 = Rating(43.381, 2.421)
>>> b1
Rating(mu=43.381, sigma=2.421)
>>> b2 = Rating(mu=25.188, sigma=6.211)
>>> b2
Rating(mu=25.188, sigma=6.211)
If a1 and a2 are on a team, and wins against a team of b1 and b2, send this into rate:
>>> [[x1, x2], [y1, y2]] = rate([[a1, a2], [b1, b2]])
>>> x1, x2, y1, y2
([28.669648436582808, 8.071520788025197], [33.83086971107981, 5.062772998705765], [43.071274808241974, 2.4166900452721256], [23.149503312339064, 6.1378606973362135])
You can also create Rating objects by importing create_rating:
>>> from openskill import create_rating
>>> x1 = create_rating(x1)
>>> x1
Rating(mu=28.669648436582808, sigma=8.071520788025197)
Ranks
When displaying a rating, or sorting a list of ratings, you can use ordinal:
>>> from openskill import ordinal
>>> ordinal(mu=43.07, sigma=2.42)
35.81
By default, this returns mu - 3 * sigma, showing a rating for which there's a 99.7% likelihood the player's true rating is higher, so with early games, a player's ordinal rating will usually go up and could go up even if that player loses.
Artificial Ranks
If your teams are listed in one order but your ranking is in a different order, for convenience you can specify a ranks option, such as:
>>> a1 = b1 = c1 = d1 = Rating()
>>> result = [[a2], [b2], [c2], [d2]] = rate([[a1], [b1], [c1], [d1]], rank=[4, 1, 3, 2])
>>> result
[[[20.96265504062538, 8.083731307186588]], [[27.795084971874736, 8.263160757613477]], [[24.68943500312503, 8.083731307186588]], [[26.552824984374855, 8.179213704945203]]]
It's assumed that the lower ranks are better (wins), while higher ranks are worse (losses). You can provide a score instead, where lower is worse and higher is better. These can just be raw scores from the game, if you want.
Ties should have either equivalent rank or score.
>>> a1 = b1 = c1 = d1 = Rating()
>>> result = [[a2], [b2], [c2], [d2]] = rate([[a1], [b1], [c1], [d1]], score=[37, 19, 37, 42])
>>> result
[[[24.68943500312503, 8.179213704945203]], [[22.826045021875203, 8.179213704945203]], [[24.68943500312503, 8.179213704945203]], [[27.795084971874736, 8.263160757613477]]]
Choosing Models
The default model is PlackettLuce. You can import alternate models from openskill.models like so:
>>> from openskill.models import BradleyTerryFull
>>> a1 = b1 = c1 = d1 = Rating()
>>> rate([[a1], [b1], [c1], [d1]], rank=[4, 1, 3, 2], model=BradleyTerryFull)
[[[17.09430584957905, 7.5012190693964005]], [[32.90569415042095, 7.5012190693964005]], [[22.36476861652635, 7.5012190693964005]], [[27.63523138347365, 7.5012190693964005]]]
Predicting Winners
You can compare two or more teams to get the probabilities of each team winning.
>>> from openskill import predict_win
>>> a1 = Rating()
>>> a2 = Rating(mu=33.564, sigma=1.123)
>>> predictions = predict_win(teams=[[a1], [a2]])
>>> predictions
[0.45110901512761536, 0.5488909848723846]
>>> sum(predictions)
1.0
Available Models
BradleyTerryFull: Full Pairing for Bradley-TerryBradleyTerryPart: Partial Pairing for Bradley-TerryPlackettLuce: Generalized Bradley-TerryThurstoneMostellerFull: Full Pairing for Thurstone-MostellerThurstoneMostellerPart: Partial Pairing for Thurstone-Mosteller
Which Model Do I Want?
- Bradley-Terry rating models follow a logistic distribution over a player's skill, similar to Glicko.
- Thurstone-Mosteller rating models follow a gaussian distribution, similar to TrueSkill. Gaussian CDF/PDF functions differ in implementation from system to system (they're all just chebyshev approximations anyway). The accuracy of this model isn't usually as great either, but tuning this with an alternative gamma function can improve the accuracy if you really want to get into it.
- Full pairing should have more accurate ratings over partial pairing, however in high k games (like a 100+ person marathon race), Bradley-Terry and Thurstone-Mosteller models need to do a calculation of joint probability which involves is a k-1 dimensional integration, which is computationally expensive. Use partial pairing in this case, where players only change based on their neighbors.
- Plackett-Luce (default) is a generalized Bradley-Terry model for k ≥ 3 teams. It scales best.
1.2k Jan 05, 2023
73 Dec 01, 2022
338 Dec 26, 2022
19 Dec 13, 2022
1 Dec 22, 2021
3 Apr 03, 2022
2 Oct 06, 2022
31 Nov 17, 2022
0 Dec 06, 2021
847 Dec 19, 2022
23.6k Jan 03, 2023
264 Nov 08, 2022
183 Jan 03, 2023
325 Dec 23, 2022
10 Apr 04, 2022
675 Jan 08, 2023
1.3k Jan 08, 2023
8 Dec 01, 2022
16 Aug 14, 2022
5k Jan 09, 2023