Recurrent MADDPG for Partially Observable and Limited Communication Settings

Code for this paper

Presented at ICML 2019 RL for Real Life Workshop

Known dependencies

Before starting, please make sure you've satisifed the following reqs:

baselines

Please also:

cd maddpg && pip3 install -e . && cd ..
cd multiagent-particle-envs && pip3 install -e .

Try it out

In order to run the algorithm you'll need to be in the rmaddpg/maddpg/experiments directory. In this directory, there is a train.py file which will allow you, among other things, try out the recurrent critic, recurrent actor, or both the recurrent actor-critic (R-MADDPG) model.

cd maddpg/experiments
python3 train.py --scenario simple_ --seed 1 --critic-lstm --actor-lstm