1. Introduction

Official code for article <DynamicLight: Dynamically Tuning Traffic Signal Duration with DRL>.

The article is shown at: https://doi.org/10.48550/arXiv.2211.01025.

2. Requirements

python3.6,tensorflow=2.4, cityflow, pandas, numpy

cityflow needs a linux environment, and we run the code on Manjaro Linux.

3. Quick start

3.1 Base experiemnts

For DynamicLight, run:

python run_dynamic.py

For DynamicLight-Lite, run:

python run_dynamic_lite.py

For tranfer, configure and run run_test.py

3.2 Extended experiemnts

3.2.1 Different state representations

For number of vehicles, traffic movement pressure, and queue length, you should change the used feature name in run_dynamic1.py and run it.
For number of vehicles under segmented road, just run run_dynamic.py.

3.2.2 Different neural networks

As described in the article,

For Network 1, run run_dynamic.py
For Network 2, change the model name as DynM in run_dynamic_multi.py and run it
For Network 3, change the model name as DynMA in run_dynamic_multi.py and run it

3.2.3 Different duration action spaces

You should configure the action space in run_dynamic.py and run it.

3.2.4 Different phase control methods

For M-QL, run run_dynamic.py
For Efficient-MP, run run_dynamic_mp.py
For M-SV with queue length as the reward, configure the reward and run run_dynamic_value.py
For M-SV with presure as the reward, configure the reward and run run_dynamic_value.py

3.2.5 Differnt feature fusion methods

Configure the model name as <Dyn01, Dyn02,Dyn03,Dyn05> in run_dynamic.py and run it. You can refer to the DynamicLight/models for more details.

3.2.6 DynamicLight-Cycle

Direct train: run run_dynamic_cycle.py
Under transfer: first run run_dynamic.py to well train DynamicLight; next configure and run run_test.py to transfer it.

4. Baseline methods

For the baseline methods, refer to Efficient-XLight and Advanced-XLight

yuanzhi0515/DynamicLight