andyljones / boardlaw

Scaling scaling laws with board games.
https://andyljones.com/boardlaw
MIT License
38 stars 7 forks source link

activelo benchmarking #9

Open andyljones opened 3 years ago

andyljones commented 3 years ago

activelo is the lil' tool I wrote to calculate Elos for a group of agents in as few games as possible. Each step, it uses variational methods to fit a full Gaussian to the posterior distribution of Elos, and then suggests the matching that'd most reduce the uncertainty in that distribution.

I think it works pretty great, but I haven't done any exhaustive testing. So maybe it's awful! But if it's good, it might be publishable.

gwern has an excellent summary of the competition.