SMARTS is a multi-agent reinforcement learning (MARL) framework designed for real-time transit scheduling and deployment. Each transit route is managed by an agent that learns optimal dispatch decisions under varying demand, traffic, and network conditions.
This repository provides:
-
Simulation environment for transit networks
-
SMARTS MARL training pipeline
-
Graph Neural Network + Transformer encoder–decoder models
-
Customized Proximal Policy Optimization (PPO) implementation
-
Scripts for reproducing all experiments and figures in the paper
Synchronous Multi-Agent RL: Each route is controlled by an agent with shared parameters but individualized rewards.
Graph Neural Network (GNN) Backbone: Learns spatial structure from the transit network.
Transformer Encoder–Decoder: Captures global dependencies and dynamic interactions.
Custom PPO Training: Designed for multi-route coordination and stable learning.
Generalizable: Works across different network topologies without fine-tuning.
Real-time Scheduling: Reduces waiting time and increases fleet utilization compared to baseline dispatch strategies.
git clone https://github.com/yourusername/SMARTS.git
cd SMARTS
# (Optional) Create a virtual environment
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
# Install dependencies
pip install -r requirements.txt