Skip to content

Scala realization of Glicko2 rating system with ability to tune by passing different parameters

License

Notifications You must be signed in to change notification settings

andriykuba/scala-glicko2

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

16 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Scala implementation of the Glicko-2 system

Build Status Maven Central

The Glicko rating system and Glicko-2 rating system are methods for assessing a player's strength in games of skill, such as chess and go.

Glicko2 on Wikipedia

The Glicko-2 rating system is intended to estimate the skill of player performance in head-to-head competiotion in a variety of games.

Describtion of Glicko-2 system

For more details, please, look at the Mark Glickman site and the example of the Glicko-2 system

Install

Add the library in built.sbt

libraryDependencies += "com.github.andriykuba" % "scala-glicko2" % "1.0.0" 

Usage

Import classes that is used in a calculation.

import com.github.andriykuba.scala.glicko2.scala.Glicko2
import com.github.andriykuba.scala.glicko2.scala.Glicko2.{Player, Win, Loss, Draw}

Update player's rating on the base of a game series

The Glicko-2 system suppose that rating is calculated not after every game, but after some period of games. The more games in the period - the better. Mark Glickman suppose at least 10 games per player in a rating period.

val player = Player(1500, 200, 0.06)
val games = List(
  Win(Player(1400, 30, 0.06)),
  Loss(Player(1550, 100, 0.06)),
  Loss(Player(1700, 300, 0.06)))
  
val updatedPlayer = Glicko2.update(player, games)

The updatedPlayer will be Player(1464.05,151.52,0.059996)

Update player's rating if player did not play a game in this rating period

Only deviation is changed

val player = Player(1500, 200, 0.06)
val updatedPlayer = Glicko2.update(player)

The updatedPlayer will be Player(1500,200.27,0.06)

Creation new user with the default parameters

Create user with the rating parameters for the Parameters object.

val player = Glicko2.defaultPlayer()

The player will be Player(1500,350,0.06)

The 95% confidence interval

It's more informative to show user rating as interval with some level of confidence. Player.ratingLow and Player.ratingHight returns the borders of 95% rating interval. Look more in the example paper

Tuning Parameters

The default game parameters will be used if none are present in the implicit context.

case class Parameters(
  
  // Game constants.
  tau: Double = 0.5, 
  epsilon: Double = 0.000001,
  scale: Double = 173.7178,
  
  // Default values for the new player.
  defaultRating: Double = 1500,
  defaultDeviation: Double = 350,
  defaultVolatility: Double = 0.06,
  
  // Scale of the result.
  ratingScale: Int = 2,
  deviationScale: Int = 2,
  sigmaScale: Int = 6)

Scale is the number of digits after decimal point. If you want to receive rating without decimal part, then set ratingScale to 0.

Create new implicit Parameters object if you want to change some parameters.

implicit val parameters = Glicko2.Parameters().copy(
    defaultDeviation = 100, 
    defaultVolatility = 0.05)
    
val player = Glicko2.defaultPlayer()    

The player will be Player(1500,100,0.05) in this case

Precision

Test data for the Glicko-2 system verification usually are taken form the Example of the Glicko-2 system

Unfortunately that calculation is just example with a very low precision. More accurate calculation, that programs trivially do, gives little different result.

Almost all of the part of calculation use rounding, so we need to remember about precision digits. This is especially actual for the programming with floating points.

In the paper:

r` = −0.2069(173.7178) + 1500 = 1464.06  

Without rounding:

r` = −0.2069(173.7178) + 1500 = 1464.05778718

It rounded to 1464.06 that is correct, but, in program, we can receive result like this

r` = -0.20694097869815928 * (173.7178) + 1500 = 1464.05066845070891

This result rounded to 1464.05.

The calculation in paper has classic error of rounding in the middle of calculation.

µ` = 0 + (0.8722 * 0.8722) × [0.9955(1 − 0.639) + 0.9531(0 − 0.432) + 0.7242(0 − 0.303)] 
   = 0 + 0.7607(−0.272) 
   = −0.2069

Let's do the same calculation with saving of all available middle precision

µ` = 0 + 0.76073284 * (0.9955 * 0.361) + (0.9531 * (-0.432)) + (0.7242 * (-0.303)) 
   = 0.76073284 * (0.3593755 - 0.4117392 - 0.2194326) 
   = 0.76073284 * (-0.2717963) 
   = -0.20676437120049

r` = -0.20676437120049 * (173.7178) + 1500 = 1464.08134831666752

That is rounded to 1464.08! Far away from the 1464.06.

The reason to so different results is losing of precision. We use rounding in the calculation so we need to remember about precision. In the paper µ` has precision 3 (multiplication of 0.361, -0.432 ... ), so result of

−0.2069 * (173.7178) = -35.94221282 

has also precision 3, as well as the result of

-0.20676437120049 * (173.7178) = -35.91865168333248

It means that we can strictly believe only in one digit after decimal point. Default rating, 1500, is the exact number, i.e. it has infinite precision: 1500.000000 ... Then

r` = -35.94221282 + 1500.000000 
   = 1464.05778718 
   = 1464.1 

So all our calculations is correct, but we can not write 1464.06, we must write 1464.1. Unfortunately, numbers in µ` calculation, like 0.9955, 0.361 was rounded as well.

Usually, if you round in the middle of calculation, then rounded numbers must be computed with two more significant figures than the very final result. By this rule, final result of the calculation that use µ` would have precision 1 and the very strict answer is 1460.

We see now, that the 1464.06 is not correct answer. Let's find the correct answer with the precision of 6 (2 digits after decimal point in our case)

µ` = 0.76073284 * (0.995498 * 0.360532) + (0.953149 * (-0.431842)) + (0.724235 * (-0.302841)) 
   = 0.76073284 * (0.358908884936 - 0.411609770458 - 0.219328051635) 
   = 0.76073284 * -0.272028937157 
   = -0.20694134592563

r` = -0.20694134592563 * 173.7178 + 1500 
   = -35.94939534323941 + 1500 
   = 1464.05060465676059 
   = 1464.05

The precision of -35.94939534323941 was 6 with 4 digits after decimal point. The 1500 is exact number, so addition of -35.94939534323941 and 1500 will give precision 4 before decimal point and precision 4 after decimal point, like 1464.0506. It is the final answer so we will round off two last significant number to receive safe result.

The correct answer is 1464.05.

The same with RD` calculation, fortunately it matches a correct value.

σ` is just truncated. The correct answer is 0.0599958431496. It's definitely not 0.05999. It would be 0.06000, If we want to round it to 6 digits after decimal point. It looks like better to use one more digit to represent this number, just to catch changes.

σ` = 0.059996

About

Scala realization of Glicko2 rating system with ability to tune by passing different parameters

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages