/QDPP

Multi-Agent Determinantal Q-Learning

Primary LanguageJupyter Notebook

What is Q-DPP

Q-DPP is a novel function approximator for cooperative multi-agent reinforcement learning problems based on determinantal point process (DPP). Q-DPP promotes agents to acquire diverse behavioral models; this allows a natural factorization for the joint Q-functions with no need for a priori structural constraints (e.g. the monotonicity constraint in QMIX), requirements often posed by centralized-training-decentralized-exeuction methods. A bonus of Q-DPP is that it offers an unified view of current major solvers including VDN, QMIX, and QTRAN.

https://github.com/koulanurag/ma-gym

Multi-Agent Determinantal Q-Learning https://arxiv.org/abs/2006.01482

How to run Q-DPP

  • Setup the environment:
conda create -n dpp python=3.5
conda activate dpp
pip install -r requirements.txt
conda install matplotlib
conda develop add ./ma-gym
  • Set up StarCraft II and SMAC:
cd pymarl
bash install_sc2.sh
  • Run quick examples:
cd pymarl
python src/main.py --config=qdpp --env-config=grid
python src/main.py --config=qmix --env-config=grid  with env_args.game_name=Spread-v0
  • Run all experiments in parallel:
cd pymarl/scripts
bash head_run.sh

The Performance of Q-DPP

We evaluate Q-DPP on five different games against popular baseline models, and present the executable commands for reproducibility and corresponding experimental performance.

Game State
Multi-Step Matrix Game Discrete
Blocker Game Discrete
Coordinated Navigation Discrete
Predator-Prey (2 vs 1 & 4 vs 1) Discrete
StarCraft II (2m_vs_1z) Continuous
  • Stochastic Normal-form Game
cd pymarl
python src/main.py --config=qdpp_nmatrix --env-config=nmatrix_idx with embedding_init=normal

matrix

  • Blocker Game, Coordinated Navigation, and Predator-Prey
cd pymarl
python src/main.py --config=qdpp_prey --env-config=grid_prey5
python src/main.py --config=qdpp_prey --env-config=grid_prey7
python src/main.py --config=qdpp_spread --env-config=grid_spread
python src/main.py --config=qdpp_blocker --env-config=grid

Blocker Game:

blocker

Coordinated Navigation:

navi

Predator-Prey:

pp

  • Initial results on StarCraft II (details of Deep Q-DPP can be found in Deep_QPP.pdf.)
cd pymarl
python src/main.py --config=qdpp --env-config=sc2 with env_args.map_name=2m_vs_1z

2m_vs_1z

Q&A:

  • How can I find codes related to Q-DPP model? / Why do I only find ma-gym and pymarl packages?

We have added Q-DPP into pymarl. The implementation is in ./pymarl/src/ with qdpp in their file names, including ./pymarl/src/controllers/qdpp_controller.py, ./pymarl/src/learners/qdppq_learner.py and ./pymarl/src/modules/mixers/qdpp.py.

  • Why do my local codes fail to run?

Please check your environment, especially ma-gym package. Since we modified some parts of the package, please make sure that you installed ma-gym with the one that we provided. If you are testing on StarCraft environment, you can execute the following command in your terminal to check if SC2 is installed:

echo $SC2PATH

A valid directory is expected. Empirically, the provided codes should work well with PyTorch v1.4.0. Due to NVIDIA driver issues, it is still possible to see incompatible datatype errors raised. We recommend to check if data are copied to correct devices via commands like .cuda() or .cpu().