The code in this repository supplements the paper "Sequence Compression Speeds Up Credit Assignment in Reinforcement Learning".
The current implementation of algorithms heavily favor clarity and flexibility over computational efficiency.
Run an individual experiment from the root directory:
# For chain and split environment
python3 -m scripts.evaluate_cns
# Or accumulated-charge environment
python3 -m scripts.train_acc_charge
# Or key-to-door environment
python3 -m scripts.train_keytodoor
The settings for environment and algorithm can be modified in the config_defaults dictionary in each training file, or directly through a Weights & Biases (wandb) sweep.
In case you don't want to use Weights & Biases (for logging):
export WANDB_MODE=disabled- gymnasium
- numpy
- torch
- wandb