elonmj/west-africa-traffic-rl

Multi-class ARZ traffic modeling + DQN signal control for motorcycle–car mixed traffic in West African urban corridors (data-scarce deployment pathway) # Multi-Class ARZ + RL Traffic Signal Control (West Africa) A reproducible Python implementation of a **multi-class macroscopic traffic model** (motorcycles + cars) coupled with **Deep Reinforcement Learning (DQN)** for adaptive signal control on a real 2×2 grid network extracted from OpenStreetMap. This repository is the codebase associated with the paper: > **A Multi-Class Macroscopic Traffic Model with Reinforcement Learning Signal Control for Heterogeneous Urban Networks in West Africa** > > Submitted to the *International Journal of Intelligent Transportation Systems Research* (Springer Nature). ## Why this project matters Most RL traffic-signal papers assume data-rich environments (detectors, calibrated microsimulators, expensive tooling). This implementation demonstrates a practical pathway for **data-scarce cities**: - model heterogeneous traffic (motos + cars) with a multi-class ARZ formulation, - extract real urban topology from OpenStreetMap (Quartier Ganhi, Cotonou, Benin), - train an adaptive controller on a standard CPU, - evaluate across multiple demand regimes, - generate publication-ready figures from the same pipeline. ## Key results **Network case study** (2×2 grid, Quartier Ganhi, Cotonou): | Scenario | Improvement vs fixed timing | |------------|----------------------------| | Light | **+83.8%** | | Moderate | **+55.7%** | | Heavy | **+34.8%** | | Saturated | **+29.0%** | | **Overall**| **+45.7%** | - Training budget: **60,000 steps** (~53 min on a standard CPU) - Hardware: standard CPU (no GPU required) ## Repository structure ```text ├─ params.py # Physical parameters, scenarios, DQN config ├─ solver.py # Multi-class ARZ finite-volume solver (LxF) ├─ environment.py # Single-link Gymnasium environment + baseline ├─ train.py # Single-link training pipeline ├─ network_params.py …