Python implementation and usage

EbTech / Elo-MMR

Skill estimation systems for multiplayer competitions

https://arxiv.org/abs/2101.00400

MIT License

173 stars 21 forks source link

Python implementation and usage #28

Open kb-open opened 5 months ago

kb-open commented 5 months ago

Thanks for the implementation. But how do we use this library in Python?

EbTech commented 5 months ago

This library would probably have to be packaged with https://github.com/PyO3/pyo3

It would be great if someone wants to help.

kb-open commented 5 months ago

I have a few questions:

Can Elo-MMR (theoretically at least) provide probability that a 2-player will be draw (like TrueSkill provides)? Basically, I'm looking for a metric for quality / fairness of a 2-player match.
I see that there exists a Python implementation here. May I know whether you still would like one more Python implementation for this project of yours?

EbTech commented 5 months ago

Ties are tricky to model consistently across different games, and I think I recall them being somewhat messy in TrueSkill as well. You could try to do similar math for Elo-MMR; I'm not sure how it would turn out. We obtained simpler formulas by taking the limit as the likelihood of ties goes to zero, but in doing so we gave up on predicting ties.

Making Python wrappers for the Rust implementation would be different from making a pure Python implementation, so your second point depends on which you prefer.

BeniaminC commented 3 months ago

I open sourced the Python implementation. It is particularly slow, but I plan to release the CuPy, Numba, and standard Numpy implementation soon(tm). You can take a look and make adjustments for yourself. It is preparing for Python 3.13, which will release the GIL for threads. It has both individual updates (standard) and team updates which utilizes an aggregation method (sum of ratings by default). https://github.com/BeniaminC/Open-ELO/blob/main/src/openelo/systems/elommr.py