iKrystyan/kigali-ems-pipeline-optimization

This is the multi-agent reinforcement learning system that aims to optimize the emergency services pipeline in rwanda. # Kigali EMS — Multi-Agent Reinforcement Learning for Emergency Dispatch A research-grade simulation platform that applies **Deep Q-Networks (DQN)** and **Multi-Agent Reinforcement Learning (MARL)** to optimise emergency medical service dispatch and traffic signal control across Kigali, Rwanda. The system is built on a full digital twin of the Kigali road network derived from OpenStreetMap, simulated with SUMO (Simulation of Urban MObility), and deployed through a live FastAPI backend with a React dashboard. This project was developed as an academic capstone in partial fulfilment of a Bachelor of Science degree in Software Engineering. ## Table of Contents - System Overview - Key Features - Repository Structure - Prerequisites - Local Setup - Running the Notebooks - Training the Models - Evaluation - Live Server & Dashboard - Architecture - Model Details - Benchmark Results - Configuration Reference - Acknowledgements ## System Overview Emergency response time is one of the strongest predictors of patient survival in trauma and cardiac events. In a rapidly urbanising city like Kigali, a growing population, dense road network, and a limited ambulance fleet make optimal dispatch a genuinely hard combinatorial problem. Static rule-based protocols — dispatch the nearest idle unit — ignore real-time hospital congestion, dynamic traffic, and incident severity. This platform addresses that gap by training two cooperative AI agents: - **Dispatch DQN** — a Deep Q-Network that decides *which ambulance* to send to each incoming incident, learning to minimise the total time-to-care (drive time + emergency department wait time). - **Traffic MARL** — a parameter-shared Multi-Agent Q-Learning controller that manages traffic signal phases citywide, prioritising green waves for approaching emergency vehicles. Both agents are trained offline in a headless SUMO simulation of Kigali and evaluated against three heuristic baselines: random dispatch, nearest-idle dispatch, and se …