xTL

xTL: Reducing Communication Overhead with XAI-Guided, Semantic-Aware DRL for Urban Traffic Light Control

1. Overview
2. Features
3. Installation
4. Usage
5. Configuration
6. Scenarios
7. API Reference
8. Contact

xTL is an open-source framework that leverages eXplainable AI (XAI) to distil lightweight, semantic traffic features from camera images and feed them into a Deep Reinforcement Learning (DRL) agent for traffic light control. By replacing raw image inputs with the location of the last vehicle in each first approaching platoon, xTL:

• Reduces communication overhead by over 90%, to as little as 20 bytes per step
• Cuts DRL training time by up to 21% compared to IntelliLight
• Maintains or improves traffic efficiency (travel time and delay metrics)

Tech stack: Python 3.8+, SUMO for traffic simulation, Ray RLlib for DRL, SHAP for XAI.

• Human-in-the-loop XAI guidance: Generate SHAP saliency maps on a trained DRL policy, then involve AI and traffic-engineering experts to select the most informative traffic features.
• Semantic communication: Transmit only distilled features (vehicle‐platoon locations + signal phase), not full camera frames.
• Plug-and-play DRL: Off-the-shelf Proximal Policy Optimisation (PPO) agent with a compact CNN+LSTM architecture.
• Reproducible evaluation: Scripts to run “simple” and “complex” intersection scenarios in SUMO, matching our published experiments.
• Extensible: Easily add new semantic features or swap DRL algorithms.

1. Prerequisites

   • Python 3.8 or later
   • SUMO (add $SUMO_HOME/bin to your PATH, refer to Configuring Path Settings)

     export SUMO_HOME="/usr/share/sumo/" # on Ubuntu

   • Git

2. Clone the repo

      git clone https://github.com/Guojyjy/xTL.git
      cd xTL

3. (Optional) Create a Conda environment (recommended)

     conda create -n xtl python=3.8
     conda activate xtl

4. Install Python dependencies

     pip install -r requirements.txt

• Train xTL (W/O image for state space) on the complex intersection

    python train.py xTLimage.ini
    python train.py xTL.ini

• Evaluate a pretrained model on the complex intersection for main traffic metrics (travel time, delay, etc.)

    python evaluation/getResults.py \
      --scen xTLimage_3turnRed3 \
      --avg_num 10 \
      --output-dir output \
      --no_safety # optional, to disable safety checks of time-to-collision

• Generate SHAP saliency maps from a trained imaged-based DRL model

    python evaluate_SHAP.py xTLimage.ini

• Environment variables
  • SUMO_HOME: path to your SUMO installation
• Config file
  • xTLimage.ini: config file for training and evaluation on imaged-based xTL.
  • xTL.ini: config file for training and evaluation on xTL with efficient semantic features，i.e., vehicle‐platoon locations, and current signal phase.
  • IntelliLight.ini: config file for training and evaluation on IntelliLight (manually designed traffic features + simplified 2D image input)

We provide two SUMO scenarios for training and evaluation:

• Simple scenario (scenario/Junc1_small): a four-way intersection with two green phases (NS/WE), 90 m road length, about 8 min simulation, 600 veh/h (NS), 180 veh/h (WE), 1 second for a simulation step.
• Complex scenario (scenario/Junc1_3turn): a four-way with through & left-turn phases (NS, NS-L, WE, WE-L), nearly 1 hr simulation, 840 veh/h (NS/S), 324 veh/h (WE/E), 5 seconds for a simulation step.

If you use xTL in your research, please cite our paper:

Jiaying Guo and Shen Wang, “xTL: Reducing Communication Overhead with XAI-Guided, Semantic-Aware DRL for Urban Traffic Light Control,” in Proceedings of the IEEE Intelligent Vehicles Symposium (IV), 2025.

Or in BibTeX format:

@inproceedings{guo2025xtl,
  title={xTL: Reducing Communication Overhead with XAI-Guided, Semantic-Aware DRL for Urban Traffic Light Control},
  author={Guo, Jiaying and Wang, Shen},
  booktitle={2025 IEEE Intelligent Vehicles Symposium (IV)},
  year={2025},
  organization={IEEE}
}

A high-resolution version of our conference poster is available for viewing and download: Download the poster (PDF).

Below is a preview image of the poster:

Jiaying Guo (School of Computer Science, University College Dublin)

Email: <jiaying.guo@ucd.ie>