Notice

Our article and code have been updated and the new code is avaliable at https://github.com/LiangZhang1996/DataLight.

Introduction

Official code for article Data Might be Enough: Bridge Real-World Traffic Signal Control Using Offline Reinforcement Learning

Usage

First, download the memory from our google drive, memory . Or you can prepare your own memory with our provided codes.

1. run run_offline.py, then get the well-trained models;
2. run run_test.py to test the models on each dataset;
3. run run_cycle.py to get the DataLight-Cycle model;
4. run summary.py to get the performance of each model.

Different configurations

1. Change the offline data at line 39 of run_offline.py to use different offline data.
2. Change the parameters at line 46 of run_offline.py to use different amounts of offline data.
3. Refer to DataLight/generate_offline_data/ for generating offline data.

Reference

For baseline methods, refer to Advanced_XLight and DynamicLight.

License

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