alphazero-othello / README.md
sr5434's picture
Update README.md
9ab4297 verified
metadata
license: mit

AlphaZero trained to play Othello using Jax and PGX. I used a TPU v4-8 provided by the TensorFlow Research Cloud to build this. Currently, we only have a checkpoint for steps 13270 and 15154, but we will have better models soon. Model evaluations:

Step Win % vs PGX baseline Draw % vs baseline Lose % vs baseline
13270 ~46.8% 6.25% ~46.8%
15154 62.5% 0% 37.5%
17039 81.25% 3.125% 15.625%
22190 87.5% 0% 12.5%