How based-watermelon/barrel-ly-learning works
This repository combines a playable Donkey Kong environment with imitation-learning autoplay and an actor-critic model. The main runtime loop gathers game state, obtains actions from the behavior-cloning policy in autoplay mode, and advances and renders the game. The documented training workflow uses gameplay demonstrations to train that policy; actor-critic is a separate model component. The README and tree establish these capabilities, but the training notebook and RL experiment source were not sampled, so their internal wiring is not inferred.
Main entry points: lfdautoplay.py Donkey Kong game · bc_policy.pt Trained BC model · lfdmodel.ipynb BC training · actor_critic.py Actor-critic model
Components 10
Game runtime
- Donkey Kong game
environment/lfdautoplay.py - State features
environment/feature_extraction.py
Imitation learning
- BC training
model/lfdmodel.ipynb - Trained BC model
model/bc_policy.pt - BC policy
environment/lfdautoplay.py
Demonstration data
- Gameplay dataset
- CSV merger
environment/csvmerger.py
Reinforcement learning
- Actor-critic model
model/actor_critic.py - RL experiment
environment/rl_test.py
Other components
- Player
Connections 6
- Player → Donkey Kong game: plays
- Donkey Kong game → State features: extracts state
- Gameplay dataset → BC training: trains from
- BC training → Trained BC model: produces
- Donkey Kong game → Trained BC model: loads weights
- Actor-critic model → Trained BC model: loads BC weights