Skip to content

Latest commit

 

History

History
30 lines (22 loc) · 1.05 KB

STATE.md

File metadata and controls

30 lines (22 loc) · 1.05 KB

What already works

  • Breakout egui user-driven game:
cd src && cargo run -p breakout-game --release

  • Defining and using a model in Python with Tensorflow/Keras (see src/q-learning/python_model)

  • Using a SavedModel (defined and saved in Python) in Rust for inference and learning. We are able to train our integration test model ql_model_ballgame_3x3x4_5_512 in Rust so that it is able to solve the ballgame task.
    => Unfortunately I discovered, that its apparently not yet supported by Tensorflow to restore the saved model weights directly using Rust code. See readme.md and integration tests in src/ql-in-rust

  • Integration test:

git clone git@github.com:bitmagier/q-learning-breakout.git
cd q-learning-breakout
cd tf_model
python3 create_ql_model_ballgame_3x3x4_5_512.py
cd ..

cargo test --test learn_ballgame --release

What's next

  • Doing the Q-learning in Python using a Rust library to feed it with the agent environment data (State, Action, Reward) (Pyo3 seems to be the way to call Rust from Python)