Multi-Agent Car Parking using Reinforcement Learning

https://arxiv.org/abs/2206.13338 (ICUR '22)

Short presentation slides

Directories

• MultiAgentCarParkingEnvironment - Unity project containing the implementation of the MDP (environment)
• ppo - scripts related to training PPO and its analysis
• q-learning - scripts related to training Q-Learning and its analysis
• results - results

Environment notes

• MultiAgentCarParkingEnvironment comes with a neural network model attached to the CarAgents. Remove the neural network attached to the Car prefab before building if intending to train (and also set the appropriate initMode of the EnvironmentManager).
• Scripts/Logger in the scene should be enabled if one requires a logger within play mode.

Required software

• Unity version 2020.3.14f1
• ML-Agents (https://github.com/Unity-Technologies/ml-agents) - explained in subsequent README files
• Python 3
• C#.NET, Visual Studio or Jetbrains Rider
• CSV file reader e.g. Excel, etc.

General note

Development was done on Windows, with the RL training run on Linux. We have not tested the implementations on other OSs.

Demo

Below shows an example trained model with 7 agents and 16 parked cars:

https://youtu.be/xMCpVxDpogA

License and citing

This work is under the CC BY 4.0 license.

Please cite the arXiv paper if used:

@misc{tanner2022multiagent,
      title={Multi-Agent Car Parking using Reinforcement Learning}, 
      author={Omar Tanner},
      year={2022},
      eprint={2206.13338},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}