Neurosymbolic Transformers for Multi-Agent Communication

Dependencies

Python 3

OpenAI Gym 0.11.0

PyTorch

Deep graph library (https://www.dgl.ai/)

Video illustration for the various tasks:

Please find a short video at tasks_video.mp4 to get a better understanding of the tasks used in this paper.

Pre-trained models:

You can find the pre-trained models for all the tasks in main/results/ folder.

random_cross task -> main/results/model_random_cross/

random_grid task -> main/results/model_random_grid/

unlabeled_goals task -> main/results/model_unlabeled/

two_groups_cross in Figure 1 -> main/results/model_2groups_cross/

Generating plots:

  #To generate plots for Figures 2 and 6
  # Random cross
  python3 -m main.plot_results --result_dir=model_random_cross --env=FormationTorch-v0
  # Random grid
  python3 -m main.plot_results --result_dir=model_random_grid --env=FormationTorch-v0
  # Random cross
  python3 -m main.plot_results --result_dir=model_unlabeled --env=UnlabeledGoals-v0

The above cmds should create plots at main/results/RESULT_DIR/plots/

Simulating pre-trained models:

   python3 -m main.replay  --result_dir=RESULT_DIR --baseline=BASELINE

RESULT_DIR can be model_random_cross, model_random_grid, model_unlabeled, model_2groups_cross.

BASELINE can be tf-full, hard, dist, prog, prog-retrained, dt, dt-retrained, det, det-retrained.

Note that prog-retrained is the version corresponding to our full approach.

Training models from scratch

To train models from scratch, run the following scripts. (Note that this part takes several hours and requires GPUs)

  # Random cross
  bash run_random_cross.sh
  # Random grid
  bash run_random_grid.sh
  # Random cross
  bash run_unlabeled_goals.sh