conda env create -n emergent_communication -f environment.yml
conda activate emergent_communication
python setup.py develop
If you are on Windows, use the env_windows.yml file instead.
To try training agents on the finder environment using DQN:
python train\train_dqn.py --env finder --save test_dqn
Each experiment done in the paper has its own separate branch:
- solo_agent: only the listener is trained, receiving perfect messages from a hard-coded speaker
- easy_speaker: a simple mlp speaker is added, taking the perfect message as input
- easy_speaker_bias: a positive signalling bias is used to help train the speaker
- speaker_bias: the speaker now receives only the visual input
Code for environments is taken from https://github.com/eugenevinitsky/sequential_social_dilemma_games