摘要
After an earthquake, the sudden influx of casualties from multiple zones overwhelms medical resources such as ambulances and hospital beds. This post-disaster mass casualty incident (MCI) is highly unpredictable, with conditions continuously evolving over time, making it urgent to develop real-time and flexible ambulance dispatch and casualty distribution strategies to maximize the expected patient survival rate. To address this, this study establishes a Markov Decision Process (MDP) model to make real-time decisions about ambulance dispatch, ambulance redeployment, and hospital selection based on the current state of the MCI system, optimizing resource allocation. Due to the high-dimensional and uncountable state space of this MDP model, and the inability to accurately estimate its transition probabilities, we propose a novel simulation-based approximate dynamic programming solution method, integrating a deep-neural-network-based approximate policy iteration (DNN-API) algorithm with a hybrid agent-based and discrete event simulation to efficiently generate near-optimal decisions. In collaboration with the National Science and Technology Center for Disaster Reduction in Taiwan, we conduct computational experiments using a potential earthquake scenario in Taipei City to assess the efficiency and effectiveness of our proposed simulation-based approximate dynamic programming algorithm compared to the least square time difference-based ADP method and current practice heuristics.