# TrueSkill

> https://www.microsoft.com/en-us/research/project/trueskill-ranking-system

https://www.microsoft.com/en-us/research/project/trueskill-ranking-system

A <Bayesian> <skill rating system> developed by <Microsoft research>.

Summary:

- <ELO> uses <point estimation>, producing a single skill value, whereas TrueSkill uses <interval estimation>, giving a range of skill. The skill estimate is represented by the mean, and its certainty by the [standard deviation](https://wiki.g15e.com/pages/Standard%20deviation.txt). The standard deviation narrows as more games are played. (In other words, playing more games increases certainty in the skill estimate.)
- An obvious result (the expected winner wins) has a small effect on ratings; an unexpected result (the expected loser wins) has a large effect.
- Rating calculations can vary with a game's characteristics:
	- Beta: A value that varies with how strongly skill correlates with winning and losing in the game.
	- Dynamics Factor/Tau: The amount of "noise" added to keep the standard deviation from converging to zero as games are played.
	- Draw Probability: The probability of a draw. A draw in a game with a high draw probability is an "obvious result." Conversely, a weak team beating a strong team despite a high draw probability is a "very surprising result." In ELO, drawn games cannot affect ratings. But a draw itself is important information, especially when draws are unlikely.
- Calculations also account for practical situations:
	- Partial Play: Players who influenced the outcome only slightly because they joined midway, left midway, took a break, and so on.
	- Partial Update: Cases where we want this game's result to affect ratings only partially, for reasons such as an "unstable network."
- Unlike ELO, it can calculate ratings for each participant in team matches with various team configurations.
- It can predict how good a team configuration is (that is, how evenly matched the game will be): Match Quality.