Proceedings of the 2012 Winter Simulation Conference C. Laroque, J. Himmelspach, R. Pasupathy, O. Rose, and A. M. Uhrmacher, eds. EXPLORING BOUNDS ON AMBULANCE DEPLOYMENT POLICY PERFORMANCE Eric Cao Ni Susan R. Hunter Shane G. Henderson Huseyin Topaloglu School of Operations Research and Information Engineering Cornell University Ithaca, NY 14853, U.S.A. ABSTRACT Ambulance deployment involves controlling a fleet of ambulances, often in real time, in an attempt to keep response times small. Simulation has been used to devise redeployment policies, and bounds have been obtained from a combination of comparison methods for queues (coupling) and simulation. These techniques yield varying results on two realistic examples. In an attempt to understand the varying results, we explore the performance of the policies and bounds on artificial models. 1 INTRODUCTION Ambulance deployment is the practice of positioning ambulances around a city, perhaps using real-time information. Positions are chosen to attempt to keep response times – the time from when a call is received until an ambulance arrives at the scene of the call – small. Response times are usually summarized by the fraction of calls with response times under some time threshold that is typically taken to be 8 or 9 minutes. We describe a call as “late” if its response time exceeds the time threshold. An important practical goal, then, is to choose a deployment policy that minimizes the fraction of calls that are late. Recently, methods have been proposed for computing bounds on what can be achieved with any deployment policy, whether the policy uses real-time information or not (Maxwell et al. 2012). Such bounds can be used to 1. help identify when a given policy is close to optimal, and 2. help determine whether deployment strategies have the potential to improve performance to within some target, or to firmly establish that some other approach, such as increasing resources, is needed. The computational results in Maxwell et al. (2012) showed that the difference in performance between an implementable policy and the bound (henceforth called the gap) was small (about 2%) in one realistic example of a city, and large (about 10%) in another. A gap of 2% indicates that the given policy is near optimal, but the gap of 10% suggests that either the policy or the bound, or both, can be greatly improved. We want to understand why the gap varies in this way. In particular, what characteristics of a city will lead to small gaps, and in cases where the gap can be expected to be large, are there other bounds that can be derived? In this paper we provide a simulation study that attempts to shed light on this question. We generate a number of artificial “cities” that have varying characteristics that we hypothesize might have an impact on the gap. We then use a simulation study to compute the gaps for each artificial city and look for important factors. The factors we consider are 978-1-4673-4780-8/12/$31.00 ©2012 IEEE 497 978-1-4673-4782-2/12/$31.00 ©2012 IEEE