search

breakwa11 / GoAIRatings

Estimate Go AI ratings by real games
MIT License
138 stars 27 forks source link

Rank vs ELO #14

Open kris-computer-go opened 5 years ago

kris-computer-go commented 5 years ago

Can you explain please how you calculate ELO?

I have thousands of games between LZ / ELF / PhoenixGo / MiniGo but I don't know how to calculate ELO from this.

About rank it seems to me that you assume 1 kyu / dan difference for every 200 points, right?

breakwa11 commented 5 years ago

https://en.wikipedia.org/wiki/Bradley%E2%80%93Terry_model

Right

amj commented 5 years ago

@kris-computer-go i'm not the author but i'm fairly certain the code used to compute the ratings is the same as the code used here: https://github.com/tensorflow/minigo/blob/master/ratings/ratings.py#L190

I scraped the .md files and converted them into a list of numbers, used the same fixed point of zen7 1D=1800, and the numbers lined up exactly. So i'm assuming @breakwa11 is using the same constants and the choix module's implementation of bradley terry.

Just a guess though! :)