wwxfromtju / mackrl

Code for Multi-Agent Common Knowledge Reinforcement Learning (NeurIPS 2019)

Fast MVC-style deep MARL framework

Includes implementations of algorithms:

• MACKRL
• CENTRAL-V

Installation instructions

Build the Dockerfile:

$ ./build.sh

StarCraft II

Set up StarCraft II. Download this specific version here: (SC2.3.16.1)[http://blzdistsc2-a.akamaihd.net/Linux/SC2.3.16.1.zip] move to

coma/3rdparty/StarCraftII

(and unzip of course using password iagreetotheeula)

and then copy

/src/envs/starcraft2/maps

to the

3rdparty/StarCraftII/Maps/Melee maps folder (which you will have to create first).

Run an experiment

Run one of the EXPERIMENTs from the folder src/config/experiments on a specific GPU using some special PARAMETERS:

cd fastmarl/src
../run.sh <GPU> python3 main.py --exp_name=<EXPERIMENT> with <PARAMETERS>

Keep an eye on your docker containers, they will be named <USER>_fastmarl_GPU_<GPU>_<RANDOM>:

docker ps

If you do not want them anymore, kill a container named NAME with

docker kill <NAME>
docker rm <NAME>

If you want to get rid of all your containers, execute

fastmarl/kill.sh

Run SC2 baselines

Run

exp_scripts/coma_baselines/run.sh <Number of runs per scenario, e.g. 5>

Results are automatically logged to both tensorboard

tensorboard --logdir=./results/tb_logs

and, if MongoDB has been set up (see main.py for config details), a database will be created.

If MongoDB is not available, then Sacred will produce output files under

./results/sacred