Optimizing Ambulance Dispatch and Relocation

Abstract

This thesis presents a high-fidelity simulator-based framework for modeling, optimiz- ing, and learning real-time ambulance dispatch and relocation strategies. We begin by formulating and solving three static optimization models to identify optimal am- bulance placements across a real road network in Princeton, NJ. These models are solved as mixed-integer linear programs using Gurobi and rely on a synthetic, spatially heterogeneous and realistic demand generator. To evaluate real-time operational policies, we build a discrete-event ambulance simulator that handles call arrivals with time limits, dynamic dispatching, and patient transport. We test a greedy baseline strategy and show that even with ample fleet size, calls can time out due to poor relocation logic. This motivates the use of reinforcement learning (RL) for dynamic decision-making. We develop two OpenAI Gym-compatible environments and train Proximal Pol- icy Optimization (PPO) agents to minimize response delay and maximize patient coverage. Our environments incorporate static travel-time routing on road graph us- ing OpenStreetMaps (OSM), stochastic on-scene and hospital service durations, and priority-weighted call handling. While our RL agents match or outperform baselines in low-fleet settings, they underperform in high-fleet settings, highlighting the impor- tance of reward shaping, evaluation fidelity, and simulation accuracy. Crucially, RL agents execute decisions in under 0.3ms post-training, offering real-time applicability that static optimization cannot match. We conclude with discussion on how EMS resource allocation is impacted by privatized healthcare structures in the United States. We outline future work explor- ing social-welfare-maximizing incentives, adversarial multi-agent RL between compet- ing ambulance providers, and simulation-grounded regulatory strategies for equitable emergency care.