Sacha Chernyavskiy

Couldn't make it fail after several attempts. Thus, it's time for valgrind and code analysers.

Hello, everyone! As far as I can understand, it might require some modifications to the `TabularPolicy` class to make the first case fail/invoke a warning. The policy class keeps the...

Sure, I can help with adding the test. Will submit a pr shortly. > I suspect we have something set in the exploitability code that assumes uniform random if the...

@jzleibo, yes, I solved it, but manually resolving dependencies (installed with `--no-deps` and fixed further). Here is a minimal reproducible example: ```Bash ROM nvidia/cuda:12.1.0-runtime-ubuntu22.04 ARG PYTHON_VERSION=3.11 RUN apt-get update &&...

@lanctot, when debugging the approach, I found that the "example" section from the tutorial doesn't exist for the latest branch: [link](https://github.com/google-deepmind/open_spiel/blob/master/docs/alpha_zero.md#python) Should we add an example to look at the...

We tried to leave the API *as close to the original code as possible*, perhaps, not pulling everything from `jax` performance package, so the code should work fine. We will...

@lanctot, if it helps, we can leave a note in the readme about the matter and can release a not merged version, in a separate branch, if people want to...

@adamantivm, great! 1) can you share the hack, if possible, and what was the issue in your case? 2) what version of python were you using and setup in general?

I see! You actually need to respawn `tensorflow` each time a new process is allocated. Quite related to [this](https://ai.stackexchange.com/questions/36318/tensorflow-gpu-and-multiprocessing) answer, ngl. Your patch actually seems to be a fantastic one....

will check if the op hasn't got time