search

sotetsuk / pgx

♟️ Vectorized RL game environments in JAX
http://sotets.uk/pgx/
Apache License 2.0
372 stars 23 forks source link

AlphaZero training on standard chess #1163

Closed janetwise closed 6 months ago

janetwise commented 6 months ago

Does your AlphaZero example code support standard chess training besides the current 5 games? I tried to remove the baseline model eval as there isn't any baseline model for chess but the training didn't work.

The Gardner chess training did work. Thank you!

sotetsuk commented 6 months ago

What do you mean by "didn't work?" Any runtime error?

janetwise commented 6 months ago

What do you mean by "didn't work?" Any runtime error?

Thank you for your reply! I tried to run under example/alphazero: python train.py env_id="chess" seed=0.

As the chess game doesn't come with the baseline model (there are 5 baseline models not including 8x8 chess) to do eval with, I then commented out the related eval code and tried to get around the baseline eval part in train.py. Then it had some errors in the transformer code... Didn't really spend lot of time on it. What would you suggest update the code to make it work with chess?

Are you able to run/test your alphaZero training on 8x8 chess? Or do you plan to provide a baseline model for chess so the alphazero training can run?

sotetsuk commented 6 months ago

Then it had some errors in the transformer code...

We do not use transformer. So I'm not sure what kind of errors you have.

Or do you plan to provide a baseline model for chess so the alphazero training can run?

We have a plan to provide powerful opponents for chess, shogi and Go. For shogi and Go, we already have powerful opponents in private and they will be public in a few months. For chess, we are preparing opponent models now. We hope we can provide it in 2024.

We do not provide the support for examples. Closed.