USM

Code for Underexplored Subspace Mining for Agents in Sparse-Reward Cooperative Games

This codebase is built on top of the PyMARL framework for multi-agent reinforcement learning algorithms.

Dependencies

• Docker
• NVIDIA-Docker (if you want to use GPUs)

Setup instructions

Build the Dockerfile using

cd docker
./build.sh

Set up StarCraft II.

./install_sc2.sh

Run an experiment in PWE

cd pwe
python main.py --config=qmix_uesm --env-config=<ENV> with env_args.map_name=<ENV> 

Possible ENVs are:

• pushbox
• rooms
• hh_island

Run an experiment in SMAC

cd smac
python main.py --config=qmix_uesm --env-config=sc2_sparse with env_args.map_name=<MAP>

Possible MAPs are:

• 3m
• 2m_vs_1z
• 3s_vs_5z
• 25m