A multiperiod set covering location model for dynamic redeployment of ambulances

Computers & Operations Research 35 (2008) 814–826
www.elsevier.com/locate/cor
A multiperiod set covering location model for dynamic
redeployment of ambulances
Hari K. Rajagopalana
, Cem Saydamb,∗
, Jing Xiaoc
aSchool of Business, Francis Marion University, Florence, SC 29501, USA
bBusiness Information Systems and Operations Management Department, The Belk College of Business, The University of North Carolina at
Charlotte, Charlotte, NC 28223, USA
cComputer Science Department, College of Information Technology, The University of North Carolina at Charlotte, Charlotte, NC 28223, USA
Available online 6 June 2006
Abstract
Emergency medical service (EMS) providers continually seek ways to improve system performance particularly the response
time to incidents. The demand for ambulances fluctuate throughout the week, depending on the day of week, and even the time of
day, therefore EMS operators can improve system performance by dynamic relocation/redeployment of ambulances in response to
fluctuating demand patters. The objective of the model is to determine the minimum number of ambulances and their locations for
each time cluster in which significant changes in demand pattern occur while meeting coverage requirement with a predetermined
reliability. The model is further enhanced by calculating ambulance specific busy probabilities and validated by a comprehensive
simulation model. Computational results on experimental data sets and data from an EMS agency are provided.
᭧ 2006 Elsevier Ltd. All rights reserved.
Keywords: Covering location problems; Hypercube model; Tabu search
1. Introduction
The goal of emergency medical services (EMS) is to reduce mortality, disability, and suffering in persons [1,2]. EMS
administrators and managers often face the difficult task of locating a limited number of ambulances in a way that will
yield the best service to a constituent population. Typically, calls originating from a population center are assumed to
be covered if they can be reached within a time threshold. This notion of coverage has been widely accepted and is
written into the EMS Act of 1973, which requires that in urban areas 95 percent of requests be reached in 10 minutes,
and in rural areas, calls should be reached in 30 min or less [3].
The study of location models for ambulance location has undergone two different phases of evolution. During the
first phase mostly deterministic models emerged while probabilistic and more realistic models emerged during the
second. More recently the phenomenal advances in hardware, commercially available solvers (e.g., CPLEX [4]), and
meta-heuristics have enabled researchers to develop increasingly more realistic and sophisticated models [2,5–10].
Earlier advances can be found in reviews by Schilling et al. [11], Owen and Daskin [12], and most recent and projected
developments are in the latest review by Brotcorne et al. [1].
∗ Corresponding author. Tel.: +1 704 687 2047; fax: +1 704 687 6330.
E-mail addresses: hrajagop@uncc.edu (H.K. Rajagopalan), saydam@uncc.edu (C. Saydam), xiao@uncc.edu (J. Xiao).
0305-0548/$ - see front matter ᭧ 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cor.2006.04.003

H.K. Rajagopalan et al. / Computers & Operations Research 35 (2008) 814–826 815
Probabilistic location models recognize that any given ambulance may be busy when it is called. Such uncertainty
can be modeled within the math programming formulation or using a queuing framework. Most models based on a
mathematical programming approach tend to employ simplifying assumptions such as all units operate independently,
yet have the same busy probability. For example, server independence and system-wide server busy probability is a
common assumption used in Daskin’s maximum expected coverage location problem (MEXCLP) [13] and ReVelle
and Hogan’s maximum availability location problem (MALP) [14]. However, not all models employ these simplifying
assumptions. For example, in their second MALP formulation (MALPII), ReVelle and Hogan allow server busy
probabilities to be different in various neighborhoods, sectors of a given region but not location or vehicle specific. Ball
and Lin [3] take a similar approach found in MALP but instead of computing server busy probabilities in their reliability
model, they compute the probability that a given demand point will not find an available server. The probabilistic location
set covering problem (PLSCP) formulated by ReVelle and Hogan [15] minimized the number of servers needed to
guarantee coverage for a city. This model like MALPII uses neighborhood server busy probabilities. Later Marianov
and ReVelle [16] extended PLSCP using the assumption of neighborhood probabilities in MALPII to formulate queuing
probabilistic location set covering problem (Q-PLSCP). They model each neighborhood as a multi-server loss system
and calculate the neighborhood busy probabilities a priori and then use it as an input into the system.
True probabilistic models are based on spatially distributed queuing theory [17] or simulation [18] and they are by
definition descriptive. Typically, they are employed to evaluate the vehicle busy probabilities and other performance
metrics of a given allocation of ambulances. Larson’s hypercube model [14,17,19] represents the most notable milestone
for approaches using a queuing framework. The hypercube model and its various extensions have been found particularly
useful in determining performance of EMS systems [2,17,20–27]. However, hypercube is computationally expensive.
For m servers the number of simultaneous equations to solve would be 2m. For fleet sizes of about 10 or more this
approach would be computationally impossible to solve with the present day computers. To solve this problem Larson
developed an approximation to the hypercube problem [19] which would require solution only m simultaneous nonlinear
equations for m servers. One of the assumptions used in Larson’s approximation is that service times are exponentially
distributed and identical for all servers, independent of the customers they are serving. Jarvis generalized Larson’s
approximation for loss systems (zero queue) by allowing service time distributions to be of a general type and may
depend on both server and customer [28].
The literature of EMS models has been skewed towards the strategic decision making, where EMS managers make
decisions on the location of ambulances or ambulance stations over a longer period of time. These models either focused
on finding out the minimum number of ambulances needed to adequately cover an area or the maximum coverage that
could be obtained given a set of ambulances. Both kinds of models were complimentary and serve different purposes.
Minimization models were used on deciding the size of the ambulance fleet while maximization models were used to
give an estimate of how good the system could perform over a period of time. Since a long term perspective was taken,
the demand in these models was naturally assumed to be the same for all periods, therefore static. However, in reality,
demand is not static but fluctuates throughout the week, day of week, and even hour by hour within a given day [2].
Dynamic redeployment models can aid managers make daily or even hourly plans to better respond to predictable
demand fluctuations by time and space. There have been two earlier papers on relocation in the EMS literature [10,29].
Repede and Bernardo [29] extended MEXCLP for multiple time intervals to capture the temporal variations in demand
and unit busy probabilities, hence, called their model TIMEXCLP. Their application of TIMEXCLP to Louisville,
Kentucky resulted in an increase of coverage while the average response time decreased by 36%. The most recent
and comprehensive dynamic relocation model was developed by Gendreau et al. [10]. The objective of their dynamic
double standard model at time t (DDSMt
) is to maximize backup coverage while minimizing relocation costs. There
are several important considerations incorporated into this model. While the primary objective is to maximize the
proportion of calls covered by at least two vehicles within a distance threshold, the model penalizes repeated relocation
of the same vehicle, long round trips, and long trips. The model’s input parameters are updated each time a call is
received and DDSMt
is solved. To solve this complex model, particularly at short time intervals, Gendreau et al.
developed a fast tabu search heuristic implemented on eight parallel Sun Ultra workstations. Using real data from the
Island of Montreal, their tests showed that the algorithm was able to generate new redeployment strategies for 95% of
all cases. They randomly selected 33 problem instances and solved them using the tabu search. They then compared
these solutions with those obtained using CPLEX and found that the worst case departure from optimality was only
2%. Apart from these studies, there has been very little work done concerning the periodic relocation (redeployment)
of ambulances in an environment where demand and the location and number of available ambulances are changing.

816 H.K. Rajagopalan et al. / Computers & Operations Research 35 (2008) 814–826
Brotcorne et al. predict that current and future advances in this field are likely to be in probabilistic location models,
dynamic redeployment models, and fast heuristics designed to solve generally large scale problem instances [1].
In this paper we formulate a multiperiod model for dynamic demand environments which minimizes the number
of ambulances required while meeting predetermined ambulance availability requirements. We extend Marianov and
ReVelle’s Q-PLSCP [16] for multiple periods and incorporate ambulance specific busy probabilities using Jarvis’
hypercube approximation algorithm [28] which further enhances the model’s realism. We develop a tabu search heuristic
to solve the model, validate the solutions by a comprehensive simulation model and provide computational statistics.
We also apply the model to a city with significant demand variability both temporally and spatially. This article is
organized as follows. Section 2 reviews PLSCP and Q-PLSCP, Section 3 presents the model and Section 4 details the
search algorithm. Section 5 presents the numerical experiments and conclusions and suggestions for future work are
discussed in Section 6.
2. Background
One of key developments in the literature is the PLSCP proposed by ReVelle and Hogan [15]. PLSCP minimizes the
numberofambulanceswhileguaranteeingapredeterminedminimumcoveragerequirementwithacertainreliability.Let
xi be the number of ambulances positioned at node i, aij =1 if ambulance at node i is within time (or distance) threshold
S of node j, 0 otherwise, pk be the busy probability of all servers in area k, be the reliability of guaranteed coverage,
and bk be the minimum number of servers required to cover each node. Using the chance constraints formulated by
Charnes and Cooper [30] ReVelle and Hogan [14] show that the number of ambulances (servers) required can be
computed by
bk =
log(1 − )
log pk
. (1)
The PLSLCP is formulated as
Minimize
n
i=1
xi (2)
Subject to
n
i=1
aij xi bj ∀j (3)
xi 0 and integer ∀i. (4)
Objective function (2) minimizes the number of servers to be deployed. Constraint (3) requires the number of servers
covering a node j to be greater than or equal to bj where bj = bk, j ∈ k. Marianov and ReVelle [16] propose the
Q-PLSCP as an extension to PLSCP. The key difference between Q-PLSCP and PLSCP is in how bk is calculated using
Eq. (1). In essence, PLSCP calculates an estimate of the pk using the arrival rate and service rate in the area around the
node i assuming independence among servers. In Q-PLSCP the pk’s are still calculated for a neighborhood where each
neighborhood is modeled as an M/M/s/0—loss queuing system. This allows Q-PLSCP to model server cooperation to
a degree while avoiding more complex structures based on spatially distributed queuing theory [17,19,28].
3. The model
We formulate the dynamic available coverage location (DACL) model to determine the minimum number of ambu-
lances and their locations for each time cluster in which significant changes in demand pattern occur while meeting
coverage requirements with a predetermined reliability. An important feature of the DACL is the incorporation of the
hypercube model [17] thus relaxing the simplifying assumptions that all servers have the same busy probability and
operate independently. We follow an approach similar to Galvão et al. [31] but to minimize the computational effort we
utilize Jarvis’ hypercube approximation algorithm [28]. An added advantage of Jarvis’ methodology is that it allows
for server specific general service time distributions.
Let t be the index of time intervals from 1 to T, xik,t be 1 if server i is located at node k at time t and mt be the number
of ambulances at time period t, hj,t be the fraction of demand at node j at time interval t, n be the number of nodes in

the system, and ct be the minimum expected coverage requirement at time t. Let pi,t be the busy probability of a server
at node i at time interval t, t be the average system busy probability at time interval t, P0 be the probability of having
all servers free M/M/m/0-loss system, Pm be the probability of having all servers busy in an M/M/m/loss system, and
Q(m, t , j) be the correction (Q) factor for Jarvis’ algorithm which adjusts the probabilities for server cooperation in
the models. Let,
Q(m, t , j) =
m−1
k=j (m − j − 1)!(m − k)(mk)(
k−j
t )P0
(k − j)!(1 − Pm)j
m!(1 − t (1 − Pm))
∀j = 0, 1, . . . , m − 1. (5)
Also let,
yj,t =
1 if node j is covered by at least one server with t reliability at time t,
0 if not,
(6)
aij,t =
1 if node j is within the distance threshold of server at node i during time interval t,
0 if not
(7)
Minimize
T
t=1
n
k=1 i∈k
xik,t (8)
Subject to
⎡
⎣
⎧
⎨
⎩
1 −
mt
i=1
p
n
k=1aij xik,t
i,t Q
⎛
⎝mt , t ,
n
j=1
mt
i=1
ajk,t xij,t − 1
⎞
⎠
⎫
⎬
⎭
− t
⎤
⎦ yj,t 0 ∀j, t, (9)
n
j=1
hj,t yj,t ct ∀t, (10)
yj,t , xik,t = {0, 1}∀i, j, k, t. (11)
Objective (8) minimizes the number of ambulances deployed. Constraint (9) tracks the nodes that are covered with
the required (alpha) reliability. Constraint (10) ensures that total system wide coverage will be greater than ct but in
conjunction with the constraint (9) only the demand nodes that are covered with alpha reliability will be included in
the system wide expected coverage statistic. For an in-depth analysis of availability and expected coverage metrics the
reader is referred to Galvão et al. [5].
When system-wide or neighborhood busy probabilities are used in the models, these can be known or estimated a
priori and thus the models may be solved using integer linear programming solvers and/or meta-heuristics with some
computational effort. However, this model requires that we compute server busy probabilities at run times, therefore we
developed a search algorithm to solve the problem. The main algorithm uses a tabu search [32] meta-heuristic to select
the best sites for servers and Jarvis’ hypercube approximation algorithm to compute server specific busy probabilities
and the resulting coverage.
4. A search algorithm
Location/relocation problems are typically NP-complete problems [25]. The size of the solution space of locating
m response units in n zones is nm. Because of the complex combinatorial nature of these problems, there have been
various successful attempts to identify near-optimal solutions through the use of meta-heuristic search methods [33]
such as; tabu search [10,34,35]; simulated annealing [31,36]; and evolutionary algorithms [8,9,37,38].
We developed a straightforward search algorithm to determine the minimum number of servers to meet the coverage
and availability requirements. The algorithm starts with an estimated fleet size for the first time interval, m1, calls a
reactive tabu search (RTS) algorithm [39], a key heuristic which we describe below, to find the best set of locations
for the given fleet size, mt . In the RTS we utilize Jarvis’ hypercube approximation algorithm to compute server busy
probabilities and evaluate the resulting coverage using Eqs. (9) and (10). If the coverage exceeds the requirements then
we decrement the fleet size otherwise increment the fleet size and run the algorithm. The best solution defined as the
minimum fleet size that meets or exceeds the requirements is updated and stored as necessary.

For the next time interval, the algorithm takes number of servers and their locations found in the previous time
interval as the initial locations for the RTS algorithm. This approach takes advantage of the nature of variations in
demand for EMS calls. Real life EMS call demand patterns do not seem to have sudden and huge spatial shifts. Rather,
the change is gradual and therefore using the best solution from time period t − 1 as the initial solution for time period
t actually improves the performance of search algorithm.
4.1. A reactive tabu search algorithm
To determine the best locations for given fleet size we developed a variation of tabu search known as reactive tabu
search [39] algorithm and tuned it to this problem domain. The major characteristic of a tabu search heuristic is its ability
to escape from local optima by allowing a degradation of the objective. Hence there is a risk of cycling as previously
visited solution may be revisited. To overcome this problem, a short term memory (tabu list) is used to prevent the
algorithm from visiting recently visited solutions. Different tabu list definitions can be found in the literature. In RTS
the tabu list is determined through feedback through the search. Similar to Gendreau et al. [35] the basic operation of
our algorithm is moving an ambulance from node i to node j. All moves (i, j pairs) are stored in the long term memory.
Initial tabu size is set to one. If a move in the long term memory occurs again, the tabu size increases to include that
move and if the same move in the tabu list is not repeated for 2 ∗m iterations, where m is the current number of servers,
it is removed from the tabu list.
In our implementation we use a one dimensional data structure (an array) of size mt ∗ t + t where mt is the number
of response units (servers) and t is the number of time intervals. The array starts with the coverage value and the server
locations for time interval 1 and then continues with the coverage value and location of the servers for all time intervals.
At iteration 0, we estimate the fleet size for each time interval, mt , using system-wide average busy probability, t ,
average service rate, t and t , the arrival rate, using mt = t / t t . The set of values for mt gives the initial starting
point of the algorithm. We then randomly generate server locations and run Jarvis’ algorithm to compute server specific
busy probabilities and the correction factor (Q) to determine which demand nodes are covered with alpha reliability,
Eq. (9). Next, we evaluate if the system wide coverage requirements are met using Eq. (10).
The basic operations in RTS or any other tabu search technique involves moving an ambulance from a node i to a
node j where node j is the best location in the neighborhood. The pair (i, j) become tabu as long as the tabu is effective.
Neighborhood for this study is defined as the eight nodes surrounding the selected node. The first ambulance is selected
for the basic operation and the best node from its neighborhood is selected only if the pair is not on the tabu list. We
consider this as one iteration. We then systematically select the next ambulance and conduct a neighborhood search,
and repeat this process until the last ambulance is selected after which the first ambulance is selected again. The size
of the tabu list changes according to the exploration or exploitation pressure needed.
The stopping rule for this implementation of RTS is 100 iterations which we determined by running a set of sample
problems for a long period of time (1000 iterations and more). Our experiments showed that the incremental gain after
100 iterations was minimal. The set of locations during the 100 iterations which resulted in the maximum available
coverage is stored.
As mentioned above, if the best coverage after 100 iterations with a given number of servers (mt ) is less than the
required coverage, then we increase the number of servers and search for 100 more iterations and continue this process
until the coverage constraint is met. If the coverage for the initial number of servers is more than the required coverage
after tabu search then we reduce the number of servers by one and continue tabu search for 100 iterations until the
coverage drops below the required coverage, in which case the most recent solution that meets the coverage constraint
is the minimum fleet size.
4.2. Look ahead procedure
During our initial experiments we discovered that despite the fact that Jarvis’ hypercube approximation algorithm
is efficient and fast, it still imposes a major computational overhead for this search process when we use it to calculate
server specific busy probabilities for every relocation (move) while conducting the RTS. To minimize the computing
time, we developed and tested a look-ahead procedure (LAP) within the RTS. Instead of using Jarvis’ algorithm to
compute server specific busy probabilities and the resulting coverage, the LAP uses a system wide average busy
probability for 100 iterations which was determined after extensive testing on sample problems. At the end of 100

iterations, we run Jarvis’ algorithm, calculate server specific busy probabilities, Q factor, and using Eq. (9) determine
which nodes are covered by alpha reliability and with Eq. (10) we compute the resulting system wide expected coverage
including only the nodes that meet the reliability requirement. However, we cannot assume that the current fleet size,
m∗
t , and the resulting coverage are optimal, thus we run the search algorithm with the initial fleet size of m∗
t but without
LAP, that is computing the ambulance specific busy probabilities every iteration via Jarvis’ algorithm. Our results show
that the RTS with LAP takes us very close to number of servers required to satisfy the coverage constraint generally
2–3 times faster than the algorithm without LAP. The details of the complete search algorithm are presented in the
Appendix.
5. Computational results
We assume a hypothetical region (city) spanning 1024 square miles (32 × 32). We divide this region into 64- and
256-zones, representing medium and large scale problem sizes. We assume two time intervals, representing the hours
before and during rush traffic. In t = 1, the calls are uniformly distributed across all zones whereas in t = 2, the
call distribution reflects the shifts in call volume due to relocation of population from residential areas of the region to
employment and perhaps education centers (Fig. 1). We randomly generate 10 problems each for 64-zone and 256-zone
configurations, thus test our model on 20 problems. For coverage constraints the distance threshold is set to eight miles,
and expected coverage requirement is set to 95% and the reliability is set to 90% for both time intervals. All tests are
conducted on a 2.4 MHz. Pentium IV Dell PC with 512 Mb of RAM and the algorithms are coded in Java (jdk 1.4).
In our experimental design we use two metrics to compare the models and solution approaches: (a) quality of solution;
and (b) run (CPU) time. Run time is self-explanatory. The quality of solution is measured by the coverage and the
minimum number of servers required. We solved each of the test problems using the search algorithm with and without
LAP. Fig. 2 displays the difference in the solution quality between the RTS and RTS-LAP for each problem which is
on average 0.25% with maximum differences of 3.94% and −1.26%, respectively.
Table 1 summarizes the run-time statistics for the two versions of the search algorithm. As anticipated, the search
algorithm with the look ahead procedure solved the 64 zone problems on average in 0.30 min, which is 2.9 times faster
than the search algorithm without the LAP. For 256 zone problems the algorithm with the LAP solved the problems
on average in 2 min which is 2.1 times faster than the run times for the algorithm without LAP. The LAP procedure
improves the run time for solution without any reduction in solution quality. A closer examination of the data shows
that the significant gains in using LAP occur when the number of servers in the initial solution is very different from
the number of servers in the final solution. If the number of servers in the initial solution is the same as the number of
servers in the final solution, the algorithm with LAP procedure is actually slightly slower.
Next, we applied our model to Mecklenburg County, North Carolina, a region spread over approximately 540 square
miles with a 2004 population of 801,137. The county’s call demand distribution is known to fluctuate significantly
2 6 10 14 18 22 26 30
2
10
18
26
0
5
10
15
20
25
30
35
40
45
CallVolume
X Coord
Y
Coord
Fig. 1. Demand distribution for t = 2.

Fig. 2. Difference between the coverage from RTS and RTS-LAP.
Table 1
CPU statistics in minutes (average, std. dev.)
Problem size Search algorithm
RTS RTS-LAP
64-zones 0.88, 0.63 0.30, 0.24
256-zones 4.20, 3.60 2.00, 0.98
Table 2
Cumulative demand distribution per time interval for Mondays in Mecklenburg County
Interval no. Time interval Number of calls
1 0:00–3:00 635
2 3:00–6:00 467
3 6:00–9:00 990
4 9:00–12:00 1450
5 12:00–15:00 1513
6 15:00–18:00 1515
7 18:00–21:00 1259
8 21:00–24:00 913
by day of the week and time of the day which provides an opportunity for our research to be applied to real-life
data thus increasing the chance of being implemented by interested parties and agencies. In 2004 the county received
a total of 77,292 calls of which 61,630 of them were classiﬁed as emergency triggering ambulance dispatches. For
this experiment we used data from one day of the week (Monday) for the entire year. There were 8,742 ambulance
dispatches in the year 2004 for all Mondays distributed over the entire region. We chose eight, 3-h time intervals for
our model. However, an EMS manager can choose, for example, 24 1-h blocks or even break the day into several time
intervals with different durations. Essentially the user determines the appropriate duration of each time interval which
could be based on experiential know-how or statistical analysis. Table 2 displays the yearly demand for each of the 3-h
time intervals on Mondays.
We organized the county into 2 mile×2 mile square nodes. This gave us a total of 168 demand nodes and we assumed
that the ambulances can be posted in any one of the nodes except in the nodes that make up the boundary. Most of the
boundary nodes are indeed not nearly the 2 × 2 sq. miles but typically a third to at most half the full size of 4 sq. miles.
This can be seen in Fig. 3 which displays the cumulative demand per node for all Mondays.

Fig. 3. Demand distribution for Mondays and city boundaries.
Fig. 4. Illustration of spatial variations in call volume during different time intervals.
Next, we broke down the demand per node for each of the time intervals. As anticipated not only the call volume
varied temporally (Table 2) it also varied spatially. On early hours of Mondays major portion of the demand is from
the center of the county but during the day (9:00 a.m.–6:00 p.m.) the demand spreads to areas surrounding the center.
As an example Fig. 4 illustrates the spatial variation in demand between the lowest and highest call volume periods,
3–6 a.m. and 3–6 p.m., respectively.
The county data shows that on average an ambulance spends 17 min 23% of the time when they do not have
to transport patient(s) to a hospital and remaining 77% of the time it takes an average of 71 min. In the model,

Table 3
Predicted minimum fleet sizes and coverage statistics for Mondays
Time interval No. of amb. Predicted coverage Simulated coverage (100 replications)
Average std. dev. Maximum Minimum
1 17 97.01% 98.22% 0.48% 99.37% 97.01%
2 14 95.29% 97.92% 0.33% 98.72% 97.00%
3 20 95.56% 96.67% 0.31% 97.27% 96.06%
4 24 95.93% 97.35% 0.34% 98.21% 96.41%
5 23 95.84% 95.84% 0.38% 96.76% 94.91%
6 23 95.58% 96.53% 0.46% 97.36% 95.45%
7 22 95.08% 96.67% 0.36% 97.54% 95.63%
8 18 95.40% 95.76% 0.41% 96.50% 94.74%
we used a weighted average of 58.58 min for service time, which is a service rate of 1.02/h. For coverage constraints the
distance threshold was set to six miles, system-wide expected coverage requirement was 95% with a 90% availability
requirement for nodes that are covered. We ran the ISA with the LAP to determine the fleet sizes and locations for the
eight time intervals.
To measure the quality of a solution we developed a comprehensive discrete event driven simulation model also
coded in Java. Given a fleet size and set of server locations, the simulation model provides an unbiased and accurate
coverage statistic, thus serves as a benchmark. We fed the prescribed solutions (fleet size and server locations) to the
simulation model to determine the expected coverage. This approach allows for measuring the accuracy of estimated
(predicted) solutions which in turn reflects upon the model’s ability to replicate the reality.
Using the locations prescribed by the model, then we ran the simulation using the actual call data from Mondays.
We assumed that all calls come from the node centers and servers are located at the node centers. Given the large
volume of data available (8742 observations), instead of using the empirical service time distribution, we randomly
sampled from the actual service times recorded on Mondays in 2004. The simulation is replicated 100 times for each
prescribed set of locations to give us an estimate of the resulting coverage. Table 3 below summarizes these results
which compare the model predicted versus simulated “true” coverage statistics (average, standard deviation, maximum
and minimum coverage). All eight instances of the problem resulted in higher than 95% coverage on average and are
not very different from the predicted coverage. These results directly support the realism and accuracy of Larson’s
hypercube model [17,19] which we embed in our model using Jarvis’ approximation algorithm [28]. These findings
are consistent with the literature where locations generated by earlier models such as MEXCLP [9,20,40] were found
to be quite robust but the expected coverage predictions could be off significantly.
The data from 2004 shows that to cover the peak periods on Mondays that the EMS agency deploys about 30
ambulances. Presently, using peak demand period data, they schedule two 12-h shifts and typically during the shift they
do not to add or remove ambulances. Given our results, the agency is likely to consider alternative ways in scheduling
daily shifts. As shown by our results the EMS agency can plan ahead with minimum fleet requirements in a given day
(e.g., Monday) and utilize the surplus ambulances for non-emergency patient transport or as backups. More interesting
possibility is to stagger the starting times for the fleet to approximately match or exceed the minimum fleet requirements
prescribed by our model. That is, at the beginning of each time interval (minus the time to travel to their posts) varying
numbers of ambulances can start their 12-h shifts. Furthermore, varying shift lengths such as 8-, 10-, and 12-h shifts
could be devised to match the prescribed fleet sizes. These ideas are promising and challenging. Amiri’s study of multi-
hour service system design [41] applied to telecommunication systems provides some ideas that could be adapted in
this domain. For future studies we suggest the development a decision support system which integrates the scheduling
decisions with fleet sizes while meeting or exceeding the target coverage requirements.
6. Conclusions
We formulated a new model for dynamic environments with the objective of determining the minimum num-
ber of ambulances and their locations for each time cluster in which significant changes in demand pattern occur
while meeting coverage availability requirements. Our model extends Marianov and ReVelle’s Q-PLSCP [16] by

incorporating ambulance specific busy probabilities similar to Galvão, Chiyoshi and Morabito [5]. To compute server
specific busy probabilities we opted to utilize Jarvis’ hypercube approximation algorithm which allows general service
time distributions that are more common in real life data. We developed an incremental search algorithm to solve the
model. The algorithm uses an implementation of reactive tabu search heuristic to determine the best set of locations for
a given fleet size and uses Jarvis’ hypercube approximation to compute server specific busy probabilities. To minimize
the computational burden we devised a look ahead procedure which helped to solve the problems 2–3 times faster. We
compared the quality of solutions via a simulation model. This approach helped us measure the realism of the solutions
generated by the model using the ISA-LAP approach. The results showed that the model produces high quality solutions
in reasonable computing times. We demonstrated how our model can be used in a real setting using data from a county
with significant variations in calls for emergency services.
Given our promising results we suggest some future studies. Other meta-heuristics could very well be developed
to solve the model [42]. We suggest a follow up study that compares the performance of several meta-heuristic
implementations using computing times and solution quality. The model can be further enhanced by including penalties
for successive redeployment of the same ambulances and avoiding long relocation trips introduced by Gendreau et al.
[10]. With dynamic redeployment models how often and under what circumstances should the redeployment strategy
needs to be updated remains an interesting research question. Since there are somewhat conflicting objectives, such as
minimizing the fleet size while meeting a predefined hard coverage constraint, we can approach this problem from a
multi-objective perspective and use the concept of Pareto domination criterion to analyze the compromise among these
conflicting objectives and develop trade-off curves.
Acknowledgments
The authors are indebted to the unknown reviewers for their critical review and suggestions which helped improve
the paper.
Appendix
Let,
mt the number of servers at time t
T the number of time periods
t the arrival rate of calls at time t
t the service rate at time t
t the busy probability at time t
Vt the search vector containing the locations of servers and the coverage at
time t
Ct (Vt ) coverage of the search vector Vt
In(Vt , x) initialize vector Vt of size x with randomly generated server locations
RT S(Vt ) reactive tabu search using vector Vt
GetNeighbors(Vt , j) get the eight closest nodes to the node at jth index in vector Vt
N set of nodes selected by Gen(Vt , j)
BestTabuCoverageWithJarvis(Vt (j), N) swap each node in set N into vector Vt index j and return a vector
which has the best coverage calculate using Jarvis algorithm embedded in
Eqs. (16) and (17). Assuming node x was swapped for node y, the (x,y)
pair should be in the tabu list
BestTabuCoverageWithoutJarvis(Vt (j), N) swap each node in set N into vector Vt index j and return a vector which
has the best coverage calculate using average busy probability in Eqs. (7)
and (8). Assuming node x was swapped for node y, the (x,y) pair should
be in the tabu list
LAP =
True if LookAhead Procedure is used,
False if not.

Main algorithm
1. Loop t → 1 to T
2. mt = t / t t
3. If t ← 1
4. Vt ← In(Vt , mt + 1)
5. Else
6. Vt ← Vt−1
7. End If
8. Vt ← RTS(Vt , LAP)
9. If Ct (Vt ) ct
10. Loop While Ct (Vt ) ct
11. mt = mt − 1
12. Vt ← In(Vt , mt + 1)
13. Vt ← RTS(Vt , LAP)
14. End Loop
15. Else
16. Do Loop while Ct (Vt ) ct
17. mt = mt + 1
18. Go To Step 3
19. End Loop
20. End If
21. If (LAP = =True)
22. Go to Step 3
23. End If
24. End Loop
25. Return Vt for t = 1, . . . , T
Reactive tabu search algorithm RTS(Vt , LAP)
1. Best ← CopyOf(Vt )
2. Loop i ← 1 to 100
3. Loop j ← 1 to mt
4. N ← GetNeighbors(Vt , j)
5. V 1
t ← CopyOf(Vt )
6. If (LAP = =False)
7. V 1
t (j) ← BestTabuCoverageWithJarvis(Vt (j), N)
8. Else
9. V 1
t (j) ← BestTabuCoverageWithoutJarvis(Vt (j), N)
10. End If
11. If ({Vt (j), V 1
t (j)} /∈ LongTermMemory)
12. LongTermMemory ← Add({Vt (j), V 1
t (j)})
13. Counter ← Counter + 1
14. Else
15. TabuList ← Add({Vt (j), V 1
t (j)})
16. Counter ← 0
17. End If
18. If Counter = =2 ∗ mt
19. Tabulist ← RemoveOldestElement(Tabulist)
20. End if
21. V (j) ← CopyOf(V 1
t (j))
22. Ct (Vt ) ← Ct (V 1
t )

23. If (Ct (Best) < Ct (Vt ))
24. Best ← CopyOf(Vt )
25. End If
26. End Loop
27. End Loop
28. Return Best
References
[1] Brotcorne L, Laporte G, Semet F. Ambulance location and relocation models. European Journal of Operational Research 2003;147:451–63.
[2] Goldberg JB. Operations research models for the deployment of emergency services vehicles. EMS Management Journal 2004;1(1):20–39.
[3] Ball MO, Lin LF. A reliability model applied to emergency service vehicle location. Operations Research 1993;41:18–36.
[4] ILOG, ILOG Cplex 7.0 reference manual. 2000: ILOG.
[5] Galvao RD, Chiyoshi FY, Morabito R. Towards uniﬁed formulations and extensions of two classical probabilistic location models. Computers
& Operations Research 2005;32(1):15–33.
[6] Karasakal O, Karasakal EK. A maximal covering location model in the presence of partial coverage. Computers & Operations Research
2004;31:1515–26.
[7] Brotcorne L, Laporte G, Semet F. Fast heuristics for large scale covering location problems. Computers & Operations Research 2002;29:
651–65.
[8] Aytug H, Saydam C. Solving large-scale maximum expected covering location problems by genetic algorithms: a comparative study. European
Journal of Operational Research 2002;141:480–94.
[9] Saydam C, Aytug H. Accurate estimation of expected coverage: revisited. Socio-Economic Planning Sciences 2003;37:69–80.
[10] Gendreau M, Laporte G, Semet F. A dynamic model and parallel tabu search heuristic for real time ambulance relocation. Parallel Computing
2001;27:1641–53.
[11] Schilling DA, Jayaraman V, Barkhi R. A review of covering problems in facility location. Location Science 1993;1(1):25–55.
[12] Owen SH, Daskin MS. Strategic facility location: a review. European Journal of Operational Research 1998;111:423–47.
[13] Daskin MS. A maximal expected covering location model: formulation, properties, and heuristic solution. Transportation Science 1983;17:
48–69.
[14] ReVelle C, Hogan K. The maximum availability location problem. Transportation Science 1989;23:192–200.
[15] Revelle C, Hogan K. The maximum reliability location problem and alpha-reliable p-center problems: derivatives of the probabilistic location
set covering problem. Annals of Operations Research 1989;18:155–74.
[16] Marianov V, ReVelle C. The queuing probabilistic location set covering problem and some extensions. Socio-Economic Planning Sciences
1994;28:167–78.
[17] Larson RC. A hypercube queuing model for facility location and redistricting in urban emergency services. Computers & Operations Research
1974;1:67–95.
[18] Zaki AS, Cheng HK, Parker BR. A simulation model for the analysis and management of an emergency service system. Socio-Economic
Planning Sciences 1997;31(2):173–89.
[19] Larson RC. Approximating the performance of urban emergency service systems. Operations Research 1975;23:845–68.
[20] Batta R, Dolan JM, Krishnamurthy NN. The maximal expected covering location problem: revisited. Transportation Science 1989;23:277–87.
[21] Chan Y. Location theory and decision analysis. Cincinnati: South Western College Publishing; 2001.
[22] Daskin MS. Network and discrete location. New York: Wiley; 1995.
[23] Larson RC. Urban operations research. Englewood Cliffs, NJ: Prentice-Hall; 1981.
[24] ReVelle C. Review, extension and prediction in emergency siting models. European Journal of Operational Research 1989;40:58–69.
[25] Saydam C, Repede J, Burwell T. Accurate estimation of expected coverage: a comparative study. Socio-Economic Planning Sciences
1994;28(2):113–20.
[26] Burwell T, Jarvis JP, McKnew MA. Modeling co-located servers and dispatch ties in the hypercube model. Computers & Operations Research
1993;20:113–9.
[27] TakedaRA,WidmerJA,MorabitoR.Analysisofambulancedecentralizationinanurbanmedicalemergencyserviceusingthehypercubequeuing
model, Computers & Operations Research, 2005, in press, corrected proof available online 17 May 2005, doi: 10.1016/j.cor.2005.03.022.
[28] Jarvis JP. Approximating the equilibrium behavior of multi-server loss systems. Management Science 1985;31:235–9.
[29] Repede J, Bernardo J. Developing and validating a decision support system for locating emergency medical vehicles in Louisville, Kentucky.
European Journal of Operational Research 1994;75:567–81.
[30] Charnes A, Cooper W. Chance constrained programming. Management Science 1959;6:73–9.
[31] Galvão RD, Chiyoshi FY, Morabito R. Towards uniﬁed formulations and extensions of two classical probabilistic location models. Computers
& Operations Research 2005;32(1):15–33.
[32] Glover F, Laguna M. Tabu search. Boston, MA: Kluwer; 1997.
[33] Osman IH, Laporte G. Metaheuristics: a bibliography. Annals of Operations Research 1996;63:513–628.
[34] Gendreau M, Guertin F, Potvin JY, Taillard E. Parallel tabu search for real-time vehicle routing and dispatching. Transportation Science
1999;33:381–90.
[35] Gendreau M, Laporte G, Semet F. Solving an ambulance location model by tabu search. Location Science 1997;5(2):75–88.

[36] Chiyoshi FY, Galvao RD. A statistical analysis of simulated annealing applied to the p-median problem. Annals of Operations Research
2000;96:61–74.
[37] Beasley JE, Chu PC. A genetic algorithm for the set covering problem. European Journal of Operational Research 1996;94:392–404.
[38] Jaramillo J, Bhadury J, Batta R. On the use of genetic algorithms to solve location problems. Computers & Operations Research 2002;29:
761–79.
[39] Battiti R, Tecchiolli G. The reactive tabu search. Journal on Computing 1994;6(2).
[40] Chiyoshi FY, Galvao RD, Morabito R. A note on solutions to the maximal expected covering location problem. Computers & Operations
Research 2002;30:87–96.
[41] Amiri A. The multi-hour service system design problem. European Journal of Operational Research 2001;128:625–38.
[42] Rajagopalan HK, Vergara FE, Saydam C, Xiao J. Developing effective meta-heuristics for a probabilistic location model via experimental
design. European Journal of Operational Research, 2005, in press, corrected proof available online 26 January 2006.

A multiperiod set covering location model for dynamic redeployment of ambulances

More Related Content

What's hot (15)

Similar to A multiperiod set covering location model for dynamic redeployment of ambulances (20)

More from Hari Rajagopalan (6)

Recently uploaded (20)

A multiperiod set covering location model for dynamic redeployment of ambulances