Graph Neural Network to Solve the Dynamic Emergency Vehicles Routing Problems

The response time of emergency vehicles (EVs) is critical for saving lives, and reducing this time has been extensively studied. Various methods have been proposed, including vehicle localization, dispatching, routing, and EVs preemption. However, most methods assume that an EV is stationary until dispatched to an accident, while in reality, EVs need to respond dynamically to accidents. In this paper, we propose a Deep Reinforcement Learning approach for reducing response time to traffic accidents by anticipating their occurrence. The reinforcement learning algorithm used is proximal policy optimization (PPO) combined with a graph attention neural network to learn the dynamic of the environment. We model the problem as a dynamic travelling repairman Problem (DTRP) and use a probabilistic model of traffic accidents based on open-access data, from two metropolitan areas, to train our graph attention neural network. We compare our approach to two heuristic policies and demonstrate its effectiveness in two metropolitan areas. Our contributions include a representation of the routing problem as a Markov decision process (MDP), a graph representation of the probabilistic model, and a graph attention neural network architecture. Our results suggest that our approach is as efficient as previous approaches and can be easily adapted to other environments.