This repo implements the state-of-the-art MARL algorithms for networked system control, with observability and communication of each agent limited to its neighborhood. For fair comparison, all algorithms are applied to A2C agents, classified into two groups: IA2C contains non-communicative policies which utilize neighborhood information only, whereas MA2C contains communicative policies with certain communication protocols.
Available IA2C algorithms:
- PolicyInferring: Lowe, Ryan, et al. "Multi-agent actor-critic for mixed cooperative-competitive environments." Advances in Neural Information Processing Systems, 2017.
- FingerPrint: Foerster, Jakob, et al. "Stabilising experience replay for deep multi-agent reinforcement learning." arXiv preprint arXiv:1702.08887, 2017.
- ConsensusUpdate: Zhang, Kaiqing, et al. "Fully decentralized multi-agent reinforcement learning with networked agents." arXiv preprint arXiv:1802.08757, 2018.
Available MA2C algorithms:
- DIAL: Foerster, Jakob, et al. "Learning to communicate with deep multi-agent reinforcement learning." Advances in Neural Information Processing Systems. 2016.
- CommNet: Sukhbaatar, Sainbayar, et al. "Learning multiagent communication with backpropagation." Advances in Neural Information Processing Systems, 2016.
- NeurComm: Inspired from Gilmer, Justin, et al. "Neural message passing for quantum chemistry." arXiv preprint arXiv:1704.01212, 2017.
Available NMARL scenarios:
- ATSC Grid: Adaptive traffic signal control in a synthetic traffic grid.
- ATSC Monaco: Adaptive traffic signal control in a real-world traffic network from Monaco city.
- CACC Catch-up: Cooperative adaptive cruise control for catching up the leadinig vehicle.
- CACC Slow-down: Cooperative adaptive cruise control for following the leading vehicle to slow down.
- Python3 == 3.5
- Tensorflow == 1.12.0
- SUMO >= 1.1.0
First define all hyperparameters (including algorithm and DNN structure) in a config file under [config_dir] (examples), and create the base directory of each experiement [base_dir]. For ATSC Grid, please call build_file.py to generate SUMO network files before training.
- To train a new agent, run
python3 main.py --base-dir [base_dir] train --config-dir [config_dir]
Training config/data and the trained model will be output to [base_dir]/data and [base_dir]/model, respectively.
- To access tensorboard during training, run
tensorboard --logdir=[base_dir]/log
- To evaluate a trained agent, run
python3 main.py --base-dir [base_dir] evaluate --evaluation-seeds [seeds]
Evaluation data will be output to [base_dir]/eva_data. Make sure evaluation seeds are different from those used in training.
- To visualize the agent behavior in ATSC scenarios, run
python3 main.py --base-dir [base_dir] evaluate --evaluation-seeds [seed] --demo
It is recommended to use only one evaluation seed for the demo run. This will launch the SUMO GUI, and view.xml can be applied to visualize queue length and intersectin delay in edge color and thickness.
The paper results are based on an out-of-date SUMO version 0.32.0. We have re-run the ATSC experiments with SUMO 1.2.0 using the master code, and provided the following training plots as reference. The paper conclusions remain the same.
| Grid | Monaco |
|---|---|
 |
 |
The pytorch impelmention is also avaliable at branch pytorch.
For more implementation details and underlying reasonings, please check our paper Multi-agent Reinforcement Learning for Networked System Control.
@inproceedings{
chu2020multiagent,
title={Multi-agent Reinforcement Learning for Networked System Control},
author={Tianshu Chu and Sandeep Chinchali and Sachin Katti},
booktitle={International Conference on Learning Representations},
year={2020},
url={https://openreview.net/forum?id=Syx7A3NFvH}
}