Joint-Local Grounded Action Transformation for Sim-to-Real Transfer in Multi-Agent Traffic Control

🚦 Overview

This repository builds upon the UGAT codebase (forked from UGAT) to support our proposed Joint-Local Grounded Action Transformation (JL-GAT) framework. Our approach focuses on improving sim-to-real transfer in multi-agent traffic signal control by allowing flexible control over centralized, decentralized, and hybrid grounding configurations.

🔧 How to Run JL-GAT

To run the JL-GAT experiment, first follow the instructions section in UGAT to download the necessary resources (CityFlow & SUMO), then use the following command:

python run_s2r.py --network cityflow1x3 --agent presslight

You can swap out the network name with:

• cityflow1x3 for the 1×3 network
• cityflow4x4 for the 4×4 network

This will initiate training and evaluation under the specified environment.

⚙️ Configuring JL-GAT

To customize the GAT configuration:

Open the config file:

JL-GAT/configs/tsc/base.yml

Modify the following parameters:

• gattype: Choose from:
  • "centralized"
  • "decentralized"
  • "jlgat"
• gat: Set to false to run without GAT.
• prob_grounding: Set a value (e.g., 0.2) to control the probability of applying grounding during training.