These examples illustrate how to implement a couple reinforcement learning algorithms to play Atari games. This uses the OpenAI Gym through its Python api so the gym Python package has to be installed. The Python library is executed in Rust using rust-cpython.
The policy gradient example uses the CartPole-v0 environment. It can be
run using the following command.
cargo run --example reinforcement-learning --features=python pgThe A2C example can use any Atari game supported in the OpenAI gym. Training can be launched by running:
cargo run --example reinforcement-learning --features=python a2cThis produces some weight files in the current directory.
If you get some Python import error when using conda, this can
be related to rust-cpython not being able to use the Python
runtime as a shared library. rust-cpython uses
Py_ENABLE_SHARED to determine the link mode which does not
seem to work properly with conda as reported in this
issue.
