Metamodel optimization of a complex, rural-urban emergency medical services system

Complex simulation systems, such as those involving emergency medical services (EMS), are often too computationally demanding to be used in optimization problems. Metamodeling is an attractive alternative, in which a sample of system configurations is evaluated using simulation, and a fast predictive model is developed as a surrogate for the slow simulator. Though the metamodeling literature is extensive, there has been little exploration of how much data is required to construct metamodels that can be used to solve optimization problems effectively, particularly in the context of a complicated rural-urban EMS system environment. In this work, the EMS system in northern St. Louis County, Minnesota is studied, with the goal of discovering station configurations with improved response times. The underlying physical system is complex, with 12 stations spread across both rural and urban areas and a fairly large geographic footprint. A decade of call data is used to develop and validate a stochastic discrete event simulator (DES) for this system, and then the simulator and raw data are used to select realistic station configurations to train the metamodel. Results are first given for just a single station within the system, and then increasingly complex settings are examined culminating with consideration of all 12 stations. Overall, though the metamodeling approach was effective for simpler cases, it requires a tremendous amount of data for larger settings. Specifically for the St. Louis County example, improved configurations were found for the one- and two-station cases, but the amount of data required to produce effective metamodels for the five- and twelve-station versions of the system was computationally infeasible given current DES and optimization heuristic implementations.

Automated summary

Metamodeling requires a tremendous amount of data for optimizing complex EMS systems.