rstt 0.6.4

Ranking Simulation Testing Tool

RSTT

Simulation Framework for Tournament and Ranking in Competition

• :warning: ALPHA version. Package still under construction. Do not hesitate to suggest features addition
• :bulb: Design for simulation based research
• :minidisc: Production of large synthetic dataset
• :computer: Automated simulation workflow
• :page_with_curl: Document your model by referercing class sources
• :chart_with_upwards_trend: Enhance Analysis by comparing trained models to simulation models.
• :wrench: Design and integrate your own custom components
• :question: Support and advise on Discord

Installation

The package is available on PyPi. To install, run

pip install rstt

User Documentation is available on readthedocs.

Description

The package provides everything needed to simulate competition and generate synthetic match dataset. It contains ranking implementation (such as Elo and Glicko ...), popular tournament format (Single elimination bracket, round robin, ...), many versus many game mode with automated outcome (score/result) generation methods. Additionaly different player model are available, including time varing strenght.

RSTT is a framework, letting user developp and intergrate with ease their own models to test.

Getting Started

Code Example

from rstt import Player, BTRanking, LogSolver, BasicElo
from rstt import SingleEliminationBracket

# some player
population = Player.create(nb=16)

# a ranking to infer player's skills.
elo = BasicElo(name='Elo Ranking', players=population)

# display the ranking to the standard output
elo.plot()

# create a competition - the solver param specify how match outcome are generated
tournament = SingleEliminationBracket(name='RSTT World Cup 2024', seeding=elo, solver=LogSolver())

# register player, unranked partcipants get assigned lower seeds.
tournament.registration(population)

# play the tournament - the magic happens!
tournament.run()

# update ranking based on games played
elo.update(games=tournament.games())

# display the updated ranking
elo.plot()

# The LogSolver implies a 'Consensus' Ranking based on 'the real level' of players.
truth = BTRanking(name='Consensus Ranking', players=population)
truth.plot()

Simulation Based Research

RSTT is meant for science and simulation based research in the context of competition. Whenever possible code is based on peer reviewed publication and cite the sources.

The following papers can be good read to start a journey in the field:

• Anu Maria [1], covers steps to follow and pitfalls to avoid in simulation based research.
• D. Aldous [2] presents base models in the context of sport competition and introduce research questiions. Several classes and features of RSTT are implemetion from this paper.
• S. Tang & Cie [3] Is a recent example of reseach. It uses synthetic dataset to provide insight about observations in real game data set.

Tutorial

Simulation based research should not be code dependant, rather model dependant. Thus I propose as tutrial a reproduction of result from A Krifa & Cie [4], which was originaly performed in R. It is a great example of the rstt features as it uses ranking, solver and scheduler models.

Package Concept

The rstt package is build on 5 fundamental abstraction:

• Player: who participate in games and are items in rankings. Different models are available including ones with 'time varying skills'
• Match: which represent more the notion of an encounter than a game title with rules. It contains players grouped in teams to which a Score (the outcome) is assigned once.
• Solver: Protocol with a solve(Match) that assign a score to a game instance. Usually implements probabilistic model based on player level.
• Scheduler: Automated game generator protocol. Matchmaking and Competition are scheduler, the package includes standards like elimination bracket and round robin variations..
• Ranking: Implmeneted as a tuple (standing, rating system, inference method, observer) that estimate a skill value (or point) for player.

Regarding ranking's component.

• Standing: is an hybrid container that implement a triplet relationship between (rank: int, player: Player, point: float) and behave like a List[Player ], Dict[Player, rank] and Dict[rank, Player]
• RatingSystem: store data computed by ranking for player
• Inferer: in charge of statistical inference and provide a rate() method.
• Observer: manage the workflow from the observation that triggers the update of a ranking to the new computed ratings of players.

Community

How to cite

If you use RSTT, consider linking back to this repo!

Source

[1] Anu Maria. (1997). Introduction to modeling and simulation. In Proceedings of the 29th conference on Winter simulation (WSC '97). IEEE Computer Society, USA, 7–13. https://doi.org/10.1145/268437.268440

[2] Aldous, D. (2017). Elo ratings and the sports model: A neglected topic in applied probability? Statistical Science, 32(4):616–629, 2017.

[3] Tang, S., Wang, Y., & Jin, C. (2025). Is Elo Rating Reliable? A Study Under Model Misspecification. arXiv preprint arXiv:2502.10985.

[4] Adrien Krifa, Florian Spinelli, Stéphane Junca. On the convergence of the Elo rating system for a Bernoulli model and round-robin tournaments. [Research Report] Université Côte D’Azur. 2021. ⟨hal-03286065⟩.