CORAL (Communicative Representation for Adaptive Reinforcement Learning)

Official implementation of the CORAL framework.

Setup

This code was developed with Python 3.10.18

If you have the correct python version you need to install the dependencies via

pip install -r requirements.txt

Usage

Pre-training (Stage 1)

Train Information Agent and Control Agent jointly:

bash train.sh

Or manually:

python CORAL.py --mode pretrain \
  --envs Navix-Empty-8x8-v0,Navix-GoToDoor-8x8-v0,Navix-SimpleCrossingS9N3-v0 \
  --total_timesteps 50000000 \
  --ia_path ./models/ia/ \
  --message_dim 32 \
  --num_envs 128

Deployment (Stage 2)

Deploy frozen Information Agent with new Control Agent on unseen environments:

bash deploy_example.sh

Or manually:

python CORAL.py --mode deploy --deploy_env Navix-DoorKey-8x8-v0 --ia_path ./models/ia/ --total_timesteps 10000000