This is the code of the paper "On the Limitations of Elo: Real-World Games, are Transitive, not Additive" accepted in AISTATS 2023 (https://arxiv.org/abs/2206.12301).

First, run pip install -e . in this folder to install the package and its dependency.

Then one can launch 2 scripts which run in minutes:

• ipython -i examples/plot_elo_breaks.py to produce Figure 2 of the paper "Elo score fails to rank players for a transitive game"
• ipython -i examples/plot_disc_rating_works.py to produce Figure 3 of the paper "Extended Elo score manages to rank players for a transitive game"

Code to produce Figure 6 and 7 is provided, although running these experiments might take time (hours). For Figure 6 (prediction performances on spinning top data) one should go in the expes/spinning_top folder, and successively run:

• main_pred.py
• print_results.py
• figure_pred.py

For Figure 7 one should go to the expes/chess_starcraft folder and successively run:

• main_chess.py
• main_starcraft.py
• figure_chess_starcraft.py

If you found this code useful, please cite

@article{bertrand2022limitations, title={On the Limitations of Elo: Real-World Games, are Transitive, not Additive}, author={Bertrand, Quentin and Czarnecki, Wojciech Marian and Gidel, Gauthier}, journal={arXiv preprint arXiv:2206.12301}, year={2022} }