GoudaGame's picture
Update the model card
2a782af verified
|
Raw History Blame Contribute Delete
1.07 kB
metadata
license: mit
tags:
  - reinforcement-learning
  - card-games
  - estimation

Self-taught

Plays Estimation with one neural network for every decision, bids and cards alike. It learned from nothing by playing itself, rewarded for where it finished at the table: first whole games, then the five-round games the leaderboard plays, then a mix of five-round, ten-round and whole games while being told how many rounds are left, so one network plays every length. The file also holds a small second network that learned, from the same games, where a seat is likely to finish given the scores so far; the look-ahead bots use it to judge a round by what it does to the game. The file is a NumPy .npz of the network's weights. Used by the Self-taught, Self-taught + look-ahead and Self-taught + guesses bots.

Rated on the Estimation Arena leaderboard. Code and how to run it: estimation-arena. Running it currently needs the rules engine, which is private.