whr

Whole History Rating

CI	status
conda.recipe
pip builds
cibuildwheel

Description

A Python interface incorporating a C++ implementation of the Whole History Rating algorithm proposed by Rémi Coulom.

The implementation is based on the Ruby code of GoShrine.

Installation

To install it from PyPI:

pip install whr

To install it from source code:

git clone [e-mail]:wind23/whole_history_rating.git
pip install ./whole_history_rating

To build this package from the source code, you will need a recent version of Python 3 installed, along with setuptools>=42 and pybind11>=2.10.0. Furthermore, depending on your operating system, you may also require the installation of the appropriate C++ build environment. If you are uncertain about the required dependencies, you can begin by attempting pip install and follow the instructions provided by your system to install the necessary components.

If you encounter compatibility issues while using the latest version, you can also try the older version implemented purely in Python:

pip install whr==1.0.1

Usage

Here is an easy example about how to use the package:

In [1]: import whr
   ...: import math
   ...:
   ...: base = whr.Base(config={"w2": 30})
   ...: base.create_game("Alice", "Carol", "D", 0)  # Alice and Carol had a draw on Day 0
   ...: base.create_game("Bob", "Dave", "B", 10)  # Bob won Dave on Day 10
   ...: base.create_game("Dave", "Alice", "W", 30)  # Dave lost to Alice on Day 30
   ...: base.create_game("Bob", "Carol", "W", 60)  # Bob lost to Carol on Day 60
   ...:
   ...: base.iterate(50)  # iterate for 50 rounds

In [2]: print(base.ratings_for_player("Alice"))
   ...: print(base.ratings_for_player("Bob"))
   ...: print(base.ratings_for_player("Carol"))
   ...: print(base.ratings_for_player("Dave"))
[[0, 78.50976252870765, 185.55230942797314], [30, 79.47183295485291, 187.12327376311526]]
[[10, -15.262552175731392, 180.95086989932025], [60, -18.086030877782818, 183.0820052639819]]
[[0, 103.91877749030998, 180.55812567296852], [60, 107.30695193277168, 183.1250043094528]]
[[10, -176.67739359273045, 201.15282077913983], [30, -177.3187738768273, 202.03179750776144]]

In [3]: print(base.get_ordered_ratings())
[('Carol', [[0, 103.91877749030998, 180.55812567296852], [60, 107.30695193277168, 183.1250043094528]]), ('Alice', [[0, 78.50976252870765, 185.55230942797314], [30, 79.47183295485291, 187.12327376311526]]), ('Bob', [[10, -15.262552175731392, 180.95086989932025], [60, -18.086030877782818, 183.0820052639819]]), ('Dave', [[10, -176.67739359273045, 201.15282077913983], [30, -177.3187738768273, 202.03179750776144]])]

In [4]: evaluate = whr.Evaluate(base)
   ...: test_games = [
   ...:     ["Alice", "Bob", "B", 0],
   ...:     ["Bob", "Carol", "W", 20],
   ...:     ["Dave", "Bob", "D", 50],
   ...:     ["Alice", "Dave", "B", 70],
   ...: ]
   ...: log_likelihood = evaluate.evaluate_ave_log_likelihood_games(test_games)

In [5]: print("Likelihood: ", math.exp(log_likelihood))
Likelihood:  0.6274093351974668

To learn more about the detailed usage, please refer to the docstrings of whr.Base and whr.Evaluate.

Running Tests

To run the test suite:

python tests/test_whr.py

Or using pytest:

pytest tests/test_whr.py -v

API Reference

whr.Base

Main class for computing Whole History Ratings.

Constructor:

• whr.Base(w2=300, virtual_games=2): Initialize the rating system
  • w2: Variance parameter controlling rating volatility over time
  • virtual_games: Number of virtual draws added to first day for regularization

Methods:

• create_game(black, white, winner, time_step, handicap=0): Add a single game

  • black: Name of the black player
  • white: Name of the white player
  • winner: "B" (black wins), "W" (white wins), or "D" (draw)
  • time_step: Integer representing the time period (e.g., day number)
  • handicap: Optional handicap value (default 0)

• create_games(games): Add multiple games at once

  • games: List of game records, each in format [black, white, winner, time_step, handicap]

• iterate(count): Run Newton's method iterations

  • count: Number of iterations to perform (typically 50-100)

• iterate_until_converge(verbose=True): Iterate until convergence

  • Returns the number of iterations performed

• ratings_for_player(name): Get rating history for a player

  • Returns list of [time_step, rating, uncertainty] for each time period

• get_ordered_ratings(): Get all players' ratings ordered by final rating

• log_likelihood(): Get the log-likelihood of the current model

whr.Evaluate

Class for evaluating prediction accuracy on test data.

Constructor:

• whr.Evaluate(base): Initialize evaluator with a fitted WHR model

Methods:

• get_rating(name, time_step, ignore_null_players=True): Get a player's rating at a specific time
• evaluate_ave_log_likelihood_games(games, ignore_null_players=True): Compute average log-likelihood on test games

References

Rémi Coulom. Whole-history rating: A Bayesian rating system for players of time-varying strength. In International Conference on Computers and Games. 2008.