Skip to content

Latest commit

 

History

History
36 lines (23 loc) · 935 Bytes

README.md

File metadata and controls

36 lines (23 loc) · 935 Bytes

Chunked-TD

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.

Instructions

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

Dependencies

  • gymnasium
  • numpy
  • torch
  • wandb