1. Introduction

Official code for article DataLight: Offline Data-Driven Traffic Signal Control

2. Requirements

python=3.6, tensorflow=2.4, cityflow

3. Quick start

3.1 Prepare offline dataset

Refer to ./memory/readme.md to prepare the offline dataset.

random10.pkl, medium10.pkl, and expert10.pkl are the default offline datsets.

3.2 Model training

3.2.1 Base model

Configure the used memory at line 28 and run run_DataLight.py

3.2.2 Run baselines

For online RL baselines, utilize AttentionLight

For offline RL baselines:

• For BC, configure the used memory and run run_BC.py
• For BCQ, configure the used memory and run run_BCQ.py
• For CQL, configure the used memory and run run_CQL.py
• For DTL, configure the used memory and run run_DT.py

3.3 Model testing

Configure the save directory and run run_test.py, run_test_BC.py, and run_test_DT.py to test the model on other datasets.

3.4 Ablation study and case study

• For low data scenario, configure the data percent at line 41 and run run_data_perccent.py
• For state study, configure the state representations at line 60 and run run_ablation1.py
• For reward study, configure the reward at line 61 and run run_DataLight.py
• For network w/o sequential encoding, configure the state line 60 and run run_ablation1.py
• For training with cyclical offline dataset, configure the used memory as cycle20.pkl and run run_DataLight.py

License

This project is licensed under the GNU General Public License version 3 (GPLv3) - see the LICENSE file for details.