PARL/examples/QMIX

• QMIX based on PARL and PaddlePaddle2.0
• • QMIX
  • StarCraft2 Environment
• Benchmark Results
• How to Use
• • Dependencies
  • Start Training

QMIX based on PARL and PaddlePaddle2.0

We reproduce the QMIX based on PARL and PaddlePaddle>=2.0.0, reaching the same level of indicators as the paper in StarCraft2 benchmarks.

QMIX

QMIX is a value-based multi-agent reinforcement learning algorithm.
Learn more about QMIX from: QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning

StarCraft2 Environment

Paper: The StarCraft Multi-Agent Challenge
Github Repositories: smac

Benchmark Results

• We trained our model in 5 different scenarios: "3m", "8m", "2s_3z", "3s_5z" and "1c_3s_5z".
• The difficulty in all scenarios are set to be "7" (very difficult).
• We trained our model 3 times for each scenario.

How to Use

Dependencies

• python3.6+

• parl>=2.0.0

• smac
• PaddlePaddle>=2.0.0

Start Training

1. Modify the config in qmix_config.py.
2. Start training:

   python train.py
   

3. View the training process with tensorboard:

   tensorboard --logdir ./