Reinforcement Learning Techniques for Multi-Agent Path Finding

Featured Algorithm

• CACTUS: Confidence-based Auto- Curriculum for Team Update Stability [1]

Prerequisites

Run these commands

cd instances
mkdir primal_test_envs

Training maps

Are generated for each training run in run_training.py using cactus.env.env_generator.

Test maps

Go to the Google Drive referenced by the PRIMAL Github repository. Download the archive with all PRIMAL test maps [2] and unpack it in instances/primal_test_envs

Running the Code

Training

Run training of all MARL algorithms in the paper with (creates a folder mit results.json and actor.pth for evaluation):

python run_training.py

The command will create a folder output/ with named result folders per MARL algorithm.

Test

Run evaluation with (parameter filename specifies the result folder with actor.pth)

python eval.py <filename> <map_size> <density>

The completion rates are printed on the command line and can be redirected into a text or JSON file for post-processing.

References

[1] T. Phan et al., "Confidence-Based Curriculum Learning for Multi-Agent Path Finding", AAMAS 2024

[2] G. Sartoretti et al., "PRIMAL: Pathfinding via Reinforcement and Imitation Multi-Agent Learning", RA-L 2019