Ranking Protocol

How to Play

See the Playing Ranked Matches section in our python package documentation for instructions on how to play ranked matches.

Rating

We use Bayesian Elo to produce ratings for all the bots based on all ranked games played. For a game to be ranked, both bots must be connected to the server in ranked mode.

Bayesian Elo produces an expected rating and a confidence bound. To reduce rating variance for bots with few completed games, ranked RBC ratings are calculated as (expected_Elo - 0.2 * confidence_bound). RBC ratings will climb about 20 points between 10 and 100 completed games, and another 20 points as games completed grows toward infinity.

A bot will not appear on the live leaderboard until it has completed 10 ranked games.

Match Frequency

The server schedules matches between bots that are connected in ranked mode with less than 40 games against each other without delay, up to the number of simultaneous games the bots support (as specified by the bots' configurations).

Once a pair of bots has played 40 games, the server will run only 1 new game per hour between that pair. For any specific pairing, the server will alternate the bots' colors, so your bot will play as white one game against and opponent, and as black in the next game against the same opponent.

Bot Versioning

When you connect your bot, you can specify whether you are connecting a new or old version of the bot. (See the Playing Ranked Matches section in our python package documentation.) When you connect a new version, your bot starts on a clean slate just like a new bot, and all previous versions of your bot are removed from the leaderboard and become ineligible for any prizes.

Human Players

Human players can participate in ranked RBC along with the bots! A human and bot can play games on the same account; they are treated as two independent entities with separate ratings (indicated on the leaderboard as human and with the relevant bot-version-number respectively).

Current Leaderboard