We formulate the Spatiotemporal Mean Field Games to value iteration expression and use DDPG and residual networks to solve the system dynamics. We apply this method to the traffic ringroad with 3 different reward functions, obtaining the following results:
Our conference paper "A Hybrid Framework of Reinforcement Learning and Physics-Informed Deep Learning for Spatiotemporal Mean Field Games" has been published in AAMAS 2023. If you find it helpful, please cite it.
@inproceedings{10.5555/3545946.3598748,
author = {Chen, Xu and Liu, Shuo and Di, Xuan},
title = {A Hybrid Framework of Reinforcement Learning and Physics-Informed Deep Learning for Spatiotemporal Mean Field Games},
year = {2023},
isbn = {9781450394321},
publisher = {International Foundation for Autonomous Agents and Multiagent Systems},
address = {Richland, SC},
booktitle = {Proceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems},
pages = {1079–1087},
numpages = {9},
keywords = {mean field games, physics-informed deep learning, reinforcement learning},
location = {London, United Kingdom},
series = {AAMAS '23}
}
python MFG.py # hyperparams set by default