.NET/C# implementation of Weng-Lin Rating, as described at https://www.csie.ntu.edu.tw/~cjlin/papers/online_ranking/online_journal.pdf
OpenSkillSharp is available for installation via NuGet:
# via the dotnet CLI
dotnet add package OpenSkillSharp
# via the Package Manager CLI
Install-Package OpenSkillSharpStart using the models by creating an instance of your desired model.
// Create a model with default parameters
PlackettLuce model = new();
// Or customize to your needs
ThurstoneMostellerPart model = new()
{
Mu = 28.67,
Sigma = 8.07,
...
}"Player ratings" are represented as objects that implement OpenSkillSharp.Domain.Rating.IRating. These objects contain properties which represent a gaussian curve where Mu represents the mean, and Sigma represents the spread or standard deviation. Create these using your model:
PlackettLuce model = new();
IRating a1 = model.Rating();
>> {Rating} {Mu: 25, Sigma: 8.33148112355601}
IRating a2 = model.Rating(mu: 32.444, sigma: 5.123);
>> {Rating} {Mu: 32.444, Sigma: 5.123}
IRating b1 = model.Rating(mu: 43.381, sigma: 2.421);
>> {Rating} {Mu: 43.381, Sigma: 2.421}
IRating b2 = model.Rating(mu: 25.188, sigma: 6.211);
>> {Rating} {Mu: 25.188, Sigma: 6.211}Ratings are updated using the IOpenSkillModel.Rate() method. This method takes a list of teams from a game and produces a new list of teams containing updated ratings. If a1 and a2 play on a team and win against a team with players b1 and b2, you can update their ratings like so:
List<ITeam> result = model.Rate(
[
new Team() { Players = [a1, a2] },
new Team() { Players = [b1, b2] }
]
).ToList();
result[0].Players
>> {Rating} {Mu: 28.669648436582808, Sigma: 8.071520788025197}
>> {Rating} {Mu: 33.83086971107981, Sigma: 5.062772998705765}
result[1].Players
>> {Rating} {Mu: 43.071274808241974, Sigma: 2.4166900452721256}
>> {Rating} {Mu: 23.149503312339064, Sigma: 6.1378606973362135}When displaying a rating or sorting a list of ratings you can use the IRating.Ordinal getter.
PlackettLuce model = new();
IRating player = model.Rating(mu: 43.07, sigma: 2.42);
player.Ordinal
>> 35.81By default, this returns mu - 3 * sigma, reresenting a rating for which there is a 99.7% likelihood that the player's true rating is higher so with early games, a player's ordinal rating will usually rise, and can rise even if that player were to lose.
For a given match of any number of teams, using IOpenSkillModel.PredictWin() will produce a list of relative odds that each of those teams will win.
PlackettLuce model = new();
Team t1 = new() { Players = [model.Rating()] };
Team t2 = new() { Players = [model.Rating(mu: 33.564, sigma: 1.123)] };
List<double> probabilities = model.PredictWin([t1, t2]).ToList();
>> { 0.45110899943132493, 0.5488910005686751 }
probabilities.Sum();
>> 1Similarly to win prediction, using IOpenSkillModel.PredictDraw() will produce a single number representing the relative chance that the given teams will draw their match. The probability here should be treated as relative to other matches, but in reality the odds of an actual legal draw will be impacted by some meta-function based on the rules of the game.
double prediction = model.PredictDraw([t1, t2]);
>> 0.09025530533015186The recommended default model is PlackettLuce, which is a generalized Bradley-Terry model for k >= 3 teams and typically scales the best. That considered, there are other models available with various differences between them.
- 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 and the accuracy of this model isn't typically as great as others but it can be tuned with custom gamma functions if you choose to do so.
- Full pairing models should have more accurate ratings over partial pairing models, however in high k games (for example a 100+ person marathon), Bradley-Terry and Thurstone-Mosteller models need to do a joint probability calculation which involves a computationally expensive k-1 dimensional integration. In the case where players only change based on their neighbors, partial pairing is desirable.
This project is largely ported from the openskill.py package with changes made to bring the code style more in line with idiomatic C# principles. The vast majority of the unit tests and data for them were taken from the python package, as well as the openskill.js package. All of the Weng-Lin models are based off the research from this paper or are derivatives of the algorithms found in it.
- Julia Ibstedt, Elsa Rådahl, Erik Turesson, and Magdalena vande Voorde. Application and further development of trueskill™ ranking in sports. 2019.
- Ruby C. Weng and Chih-Jen Lin. A bayesian approximation method for online ranking. Journal of Machine Learning Research, 12(9):267–300, 2011. URL: http://jmlr.org/papers/v12/weng11a.html.
If you are struggling with any concepts or are looking for more in-depth usage documentation, it is highly recommended to take a look at the official python documentation here, as the projects are incredibly similar in structure and you will find that most examples will apply to this library as well.