PyTorch implementation for our ICLR 2019 paper M^3RL: Mind-aware Multi-agent Management Reinforcement Learning.

• Python 3.6
• Numpy >= 1.15.2
• PyTorch 0.3.1
• termcolor >= 1.1.0

The environments Collection_v1, Collection_v0, and Collection_v2 correspond to the S1, S2, and S3 settings in the paper respectively. The multiple bonus levels settings can be run in Collection_v3 (each agent has 1 skill) and Collection_v4 (each worker has 3 skills).

Crafting_v0 and Crafting_v1 stand for the standard setting and the multiple bonus levels setting respectively in the paper.

Please refer to examples for how to run training and testing in different settings.

See LICENSE for additional details.