You are on page 1of 14

Managing Capacity and Flow at Theme Parks

Author(s): Reza H. Ahmadi

Reviewed work(s):
Source: Operations Research, Vol. 45, No. 1 (Jan. - Feb., 1997), pp. 1-13
Published by: INFORMS
Stable URL: .
Accessed: 10/10/2012 19:01

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Operations Research.
Universityof Califomia at Los Angeles, Los Angeles, California
(ReceivedFebruary1995;revisionreceivedSeptember1995;acceptedApril 1996)

The growthof serviceindustriesand their impacton the U.S. economyhave attractedconsiderableattentionin recentyears.While
some servicesectors,most notablyairlineand telecommunication industries,have been in the forefrontof model development,the
industryis ratherfragmented,and similarrigoris lackingin most other sectors.
Thispaperdescribesan applicationof a model-basedapproachto some of the short-termridecapacityandvisitorflowissuesfaced
by the Six FlagsMagicMountain(SFMM),a majornationalthemepark.Specifically,we considerdailyoperationsat the themepark
and focus on the generationand evaluationof alternativestrategiesfor managingride capacitiesand visitorflow. Managementof
demandinvolvestwo aspects:(a) understandingcustomerpreferencesas revealedby routingbehavior,and (b) usingthe model to
evaluatethe implicationsof changesin transition-behavior.
A crucialcomponentof the studyrelatesto the empiricaldata collected.Besidesverifyingthe validityof the models,these data
provideseveralinsightsfor developingschemes to managethe day-to-dayoperationsof the park.The SFMMmanagementwas
activelyinvolvedin variousphasesof this studyand as a resulthas been introducingthe proposedmodelsin a phasedmanner.

The theme park studied here, Six Flags Magic Moun- the waiting time is not necessarily a desirable objective.
tain (SFMM) located in southern California,provides For example Larson (1987) argues that for fast food cus-
a day-long total entertainmentpackage. Theme parks pos- tomers, satisfactionin a single-queue system may be higher
sess several interesting characteristicsthat influence both than in a multi-queue system, even though customerswait
analysisand managementof their operations.First, the ser- longer in a single-queue system. Also, in a theme park,
vice package is not homogeneous-the experienceincludes waiting may contribute to the experience; this notion was
thrill rides, shows, arcade games, and food and beverages. verified by the result of a customer satisfaction survey at
Second, customerpreferencesare not uniform,and the mar- the park. Although excessive waiting times are quite unde-
ket could be segmented into several groups. Third, the sirable, low waiting times tend to have a negative impact as
park attendance level fluctuates significantly,according to well. Another customer surveycommissionedby the theme
the season, day of the week, and time of day. Fourth, park suggested that beyond a threshold level of average
customer perceptions (e.g., about delays and queues) play number of rides, further rides provided little improvement
an important role in evaluations of the park's operations. in customer satisfaction. These are interesting issues and
Even more important,the interdependence between these are subjects for further research.
measures is not obvious and requires insights into custom- In this paper, we describe an application of a model-
ers' needs and expectations. based approach to some of the short-term operational is-
For example, the correlation between capacity utiliza- sues faced by the SFMM. Specifically,we consider daily
tion and waiting time is well recognized, and the tradeofs operations at the park and focus on optimally setting a
between capacity, operating costs, and waiting times have ride's nominal capacity,analyzingand managingthe park's
been addressed in a wide range of applications.However, visitors transition patterns, and developing models to sug-
the impact of waiting times on customer satisfactionis not gest routing tours. A crucial component of our study re-
very clear. Traditionallywaiting has been viewed as a neg- lates to the empirical data collected. In addition to
ative measure, and studies in the operations management verifying the validity of the models developed, these data
literature typically assume a monotonic relationship be- provide insights for improvingthe day-to-dayoperations of
tween waiting times and customer satisfaction. However, the park.
there is some evidence to suggest that this may not be The remainder of the paper is organized as follows. In
universallytrue in all instances in service industries.While the next section we describe the park and discuss different
few customers tolerate or desire long waits, it appears that customer classes and their service experiences at the park.
in some situations customer experience and perception of Section 2 provides a mathematical model for managing
service is enhanced by some waiting, and thus minimizing ride capacity,a model for generatingthe desired transition

Subject classifications: Service. Design. Capacity.

Area of review: OR PRACrICE.

Operations Research s e a r c 0 064X64 X /45410001 $05.00

Vol. 45, No. 1, January-February 1997 1 0 1997 INFORMS

pattern, for influencing future customers' movements in seated alone, thus occupying a whole unit by themselves,
the park, and a model for providing a routing sequence and families and small groups may do the same thing. As a
based on specific customers preferences for the rides. result, the ride's observed capacity may vary significantly
Also, we develop a neural network model to estimate the from the nominal capacity,and for some rides it may be as
observed capacityof the rides for any given nominal capac- low as 60% of the nominal capacity. For example, the
ity. In Section 3 we describe the results of our data collec- observed capacity at Log Jammer, operating with 32 logs,
tion and related analysis. Detailed analysis of the ride has typicallybeen around 1,200, whereas the nominal ca-
capacity model, its potential for improving park opera- pacity is 1920 rides per hour for weekends.
tional performance, and other related managerial consid- The variety of special shows offered is the second class
erations are addressed in Section 4. In Section 5 we focus of attractionat the theme park. From an operationalview-
on the implementation aspects of influencing customers point these can be viewed as batch processes with the cycle
transition patterns in the park. We characterizetransition time determined by the show's characteristics.The batch
patterns that lead to improvedpark performance,and pro- capacityessentiallyis fixed, and short-termoperational de-
pose several policies to influence visitors' behavior while cisions concern the number of shows and the correspond-
they are touring the park. In Section 6 we conclude our ing schedule.
study and discuss issues to be investigated in future work. The complexity in managing the operations of the park
is primarily due to the wide variation in the customers'
arrival pattern and their entertainment preferences. The
customers exhibit a wide variation in their preferences,
The primaryattractionof the SFMM theme park lies in its and their perceptions of service can be classifiedinto three
thrilling roller coaster rides, which have catchy names like main groups: (1) younger visitors, especially teenagers, (2)
Viper, Colossus, and Ninja. The rides are complemented family visitors, and (3) senior citizens. The primaryattrac-
by a variety of special shows such as the Dolphin act, the tion for younger visitors lies in thrill rides, and teenagers
U.S. high diving team, and the Batman stunt show. A wide appear to be less sensitive to long waits. In contrast,senior
selection of arcades, gift shops, and eating establishments citizens are influenced by waiting times and tend to plan
completes the entertainmentservicesprovidedby the park. the sequence of rides in order to reduce their waiting
From an operations perspective, the rides offered at the times. Families often are constrainedby height limitations
park can be classified broadly into three categories-(1) that exclude certain rides. Family groups tend to have a
group rides, (2) continuous rides, and (3) individualrides. lower tolerance for long waits than teenagers. At an aggre-
Colossus and Flashback are examples of rides in which gate level, the behavior of the three customer groups may
customers are grouped together for a roller coaster ride. be characterized by their transition behavior within the
In contrast to the group rides, Metro and Orient Express park, tolerance to waiting, and thresholdlevel for the num-
are examples of continuous rides, and the ride pace is well ber of rides per visit.
regulated. Buccaneer and Granny Grand Prix are exam-
ples of individualrides in which the pace is less controlled.
Effective management of the rides requires clear under-
standing of ride capacity. Generally, a ride's nominal ca-
pacity is determined by the number of operating units
(cars, boats, trains, etc.), the number of seats per operating In this section we provide a set of models for managingthe
unit, its trip time, and loading and unloading time. For ride capacity and demand in the short-termoperations at
example, Jet stream could be operating with either 20, 25, the park. As mentioned earlier, capacity changes are ef-
28, or 32 boats. The ride cycle time is estimated to be fected primarilyby changingthe number of operatingunits
seven minutes, with 4.5 minutes for trip time and the re- on each ride. However, management of ride demand in-
mainder for loading and unloading customers. Each boat volves two aspects: (1) understanding customers' prefer-
can accommodate five passengers; consequently the ride ence through their transition behavior, and (2) using the
nominal capacity could vary from 850 to 1360 customers model to evaluate the implicationsof changes in transition
per hour. Typically,the cycle time for the rides is constant, behavior. Thus we do not attempt to characterize the
and the capacity is adjusted by altering the number of trade-offsbetween customer satisfactionand waiting times.
carts. Park management changes the ride capacity based Instead, we assume that a desired average number of rides
on the park attendance level and queue lengths at different is targeted(based on such trade-offs),and that the model is
rides. used to determine the optimal ride nominal capacity and
A ride's observed capacity may differ from its nominal to propose strategies to achieve this target. For the first
capacity. For the continuous and individual rides, the ob- level of analysis we aggregate the customer classes into
served and nominal capacities are primarilythe same. But one class. In addition,we focus on the rides and ignore the
for the group rides, the observed capacity is a function of shows and other attractions offered by the park. Figure 1
how the visitors are "grouped"and "loaded"on the oper- provides a schematic of the network of rides in the park,
ating units of the ride. Individualcustomersmay wish to be where many small and adjacent rides have been grouped







_t g;)- i_
/SUS ~~~~JET.


0> - / /\ MmB. METROB


T t ~~~~~~~7
> _
/ _ \
/ \ I N. NINJA

/ ~~~~~~~~~~~~~~~~~~~OEA. ORIENT EXPRESS A


/ / \ _/ ~~~~~~~~~~~~~~~~~~~~~PSY.
REV. / /~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~FA ECO




F / \ =/ 8 ~~~~~~~~~~~~~~~~~~~~~~~~TUR.
9 / \ 6

\ | W~~~~~~~~~~~~~~~~~~~~~~~~ZR.


Figure 1. Network ofrid .PAT

Figure 1. Network of rides.

together. Next, we describe the general characteristicsof also a decision variable and the model simultaneouslysets
each model. the ride's capacity and defines the transition within the
(I) A neural network model, Ride Capacity Model park. Based on the "optimum"transitionmatrixwe gener-
(RCM), is constructedto determine observed ride capacity ate several policies to influence demands for the rides.
for any group ride, where "customergrouping"is critical. (IV) Finally, we construct two mathematical models to
In general, the neural network uses the historical arrivals design touring plans of the park to avoid congestion. The
and departures and the ride's nominal capacity to deter- Ride Selection Problem (RSP) decides on the set of rides
mine the customer groupingpattern and learn the dynam- with high customer utility. The Ride Visiting Problem
ics of grouping. Based on these observations, the model (RVP) sequences the set of rides to be visited, given the
estimates the ride's observed capacity. anticipatedwaiting times for the rides during the day. Fig-
(II) Visitors' arrivalpatterns at the rides, the transition ure 2 shows the relationshipof these models.
patterns within the park, and the rides' observed capacity
are used to determine the optimal nominal capacity level 2.1. Ride Capacity Model
for the rides. The objective of the Capacity Management In this section, we describe the neural networkmodel used
Model (CMM) is to maximize the park service level sub- to approximatea ride's observed capacity. The output of
ject to park operatingbudget, customer thresholdvalue for
the desired number of rides, and maximum tolerable TransitionPattern OptimalCapacity
queue lengths at the rides.
(III) Existing transition patterns within the park often RCM Observed Capacity RSP Set of Rides
result in various levels of demand for different rides. This ExpectedWaitingTimes RVP OptimalTosrs

demand variation manifests itself in large fluctuations in

the queue lengths for some rides at different hours. To ~PM OptimalCapacity
influence the park transitionpattern, we develop the Flow
Pattern Model (FPM), in which the transition pattern is Figure 2. Overall configurationof models used.

the RCM is used in both the capacity and flow pattern In our model we have set C = f(D - Y) = (D - Y)3, which
models. The neural network is used to approximate the through our experimentationwas found to be quite suit-
group rides, such as Colossus, Jet Stream, Log Jammer, able. Equation (4) determines how to change the weights
Metro, Ninja, Psyclone, and Revolution. For the other cat- along the connections in the network, where the partial
egories of rides, the "observed capacity"was provided by derivativesare easily computed.After trainingthe network
the management. with sample data we freeze the weights, and then for any
We first discuss how the ride throughput(i.e., the num- input vector we can get the predicted output algebraically.
ber of customers served at each ride) is approximated.In The neural network used in our implementation con-
our model, the ride throughput is estimated by the mini- sisted of 14 nodes in the input, 13 in the hidden layer, and
mum of the observed ride capacity and the number of one node in the output layer. The nodes in the input layer
customers waiting in the ride queues, specifically: correspond to the number of vehicles (one node for each
capacity level), ride trip time, ride loading and unloading
Sit = Min(Qit, Cig) , (1) time, seats per unit, type of ride (batch, discrete or contin-
uous), number of waiting lines, popularityof the ride, and
where Si, is the throughput of ride i at time period t and a bias node. The nodes in the input and output layer are
Qit is the ride queue length. The ride observed capacity, fully connected. The additional node in the input layer
Cig,is defined by the neural network approximationof the corresponds to the bias node, which is generally included
operating characteristicsof the rides. Approximation (1) in backpropogationnetworks. All the available rides data
was motivated by the observationthat the number of rides were used to train and to stop training the network, to
taken was mainly based on how customers were grouped avoid over-training.For further discussion on how to train
to get on the rides. But in instances when the queue length neural network see Masson and Wang.
is smaller than the existing number of operating vehicle
units available at the ride, the impact of customer group- 2.2. Capacity Management Model
ing is negligible. The quality of approximation (1) is de- The capacity management model (CMM) determines the
scribed in Section 3. capacity levels for the rides in the park during different
We have constructed a simple neural network for each time periods. The model utilizes the transition probabili-
type of ride. The networksconsist of one input, one hidden ties as the input parametersof the model. Notation used in
layer, and one output layer which has only one node at the the formulations is given in Table I. The CMM can be
output layer. For further information on neural networks formallypresented as follows:
the reader is referred to Masson and Wang (1990) and
Fort (1988). The amplification(2) and transformation(3) n\
functions of the networkmay be describedby the following
Max Min wisit)
Qit =Qit - 1 + I it- sit 1 1 (i, t), (5)
A n
U E WAXA (2) Iit = iOtPOPji+ k Sj- 1 V (i, k), t E Tk , (6)
A=1 j=1

(1 + e -(3) Sit =Min( Qit, E CigYigk) V (i, k), t E Tk , (7)

where wk is the weight at the output node via the connec-
E Yigki=l V(i,k), (8)
tions from the hidden layer node A, X. is the input from g=1
hidden layer processing element A, and Y is the actual n G K
output of the network. The sigmoid nonlinear transforma- E , x
i=1 g=1 k=1
ITkI AigYigk < B, (9)
tion function is given by (3), where a is a measure of noise
in the system. The sigmoid function is continuous and n T T

monotone and is used in many applications.The objective E E Sit -- (1 + y)TV E Iot, (10)
i=1t=1 t=1
of the model is to train the network such that the error in
the output is minimized. Let E = D - Y represent the Qit QXit Vfi, t),(1
error at the output node duringthe trainingsession, where Yigk E (0, 1) V (i, g, k) , (12)
D is the desired output (the observed capacity). The data
Qit, Iit, sit 0 U(, t) . (13)
set for trainingthe networkswas collectedby the data collec-
tion team. We set the change in the weights to be propor- The CMM maximizesthe minimumweighted numberof
tional to the negative of the derivativeof the cost function, rides given in any time period. This measure, as discussed
C, with respect to the connection weights, such that AwA= further in Section 3, is a more relevant measure of park
- aC/law. From the chain rule we get: performance and provides uniformityin deliveryof service
rather than the total number of rides given throughoutthe
AwA = -aC/aWl = -a ciaY * aY/aU * aU/awl . (4) day. The objective is achieved by changing the capacityof

Table I
i, j: index of the rides, ij = 1 ... , n
g: index for capacity level, g 1, .. G
t: index of time periods, t = 1,..., T
k: index of transition patterns, k = 1, ... K

1: index of time periods in Tour Design Model, I = 1, ..., L

F set of pairs of rides that are far apart
Tk a set definingthe time periods in transitionpatternk
Qi, length of the queue at ride i at the beginningof time period t
Poik = probabilityof customersgoing to ride i upon their arrivalduringtransitionpatternk
PR = probabilityof customersgoing to ride j from ride i duringtransitionpatternk
F1it= numberof customersgoing to ride j from ride i duringtime period t
it = numberof customersarrivingat-ridei at the beginningof time period t
Iot = numberof customersarrivingat the parkat time period t
Sit= numberof customersservedby ride i duringtime period t
Yigk - if capacitylevel g is used at ride i duringthe transitionpatternk, 0 otherwise
Ril expected waitingtime for ride i duringtime interval1 given the attendancelevel
Zl= 1 if ride i is visited in time interval1, 0 otherwise
QXit= maximumacceptablequeue length,definedby management
qxit minimumdesirablequeue length, definedby management
Gig= actualcapacityof ride i operatingat level g, computedthroughneuralnetworkmodel
Aig = cost of operatingride i at capacitylevel g
Wv customerpreferenceassociatedwith ride i
B - budgetaryoperatinglimit
TV = visitors'ride thresholdvalue
y = a safety factorfor ride thresholdvalue, 1 > y 0

the rides at different time intervals. Constraint (5) com- The results of this model were used as the first step in
putes the queue length at each ride for every time period. influencing the transition patterns around the park. Next
In constraint(6) we capture the movement of park visitors we discuss the flow pattern model.
by the number of customers that arrive at ride i in time
period t. Visitors either go directly to ride i upon their 2.3. Flow Pattern Model
arrivalat the park or join ride i from other rides based on In the Flow Pattern Model (FPM), in addition to identify-
the transition matrix. Constraint (7) determines the num- ing the capacity level of each ride, we focus on capturing
ber of customers receiving service at ride i and time t. The the desired transition probabilities and movement of visi-
ride capacity level is determined by constraint (8). Con- tors in the park. The FPM also seeks to find the optimum
straints (9), (10), and (11) define, respectively, the bound distributionof customer arrivalsat the park.
on the operating budget, the total rides given, and the T n
maximumtolerable queue lengths for the rides. Constraint Max Mn | WiSit (RSP)
(10) is motivated by our empirical analysis,which indicates
that customers have a threshold value for the number of s.t.
rides they take during their visit to the park. Constraints
Qit =Qit - 1 + IOtPOik Sit - 1
(9) and (11) are managerial inputs and significantlyinflu-
ence the service package delivered to the park visitors.
+ >Fj1t V(i, k)jt ETk, (15)
The CMM is a mixed integer linear program of moder- j=1
ate size and was used in the first phase of our implemen- n
tation. Details of model verification and implementation Sit = >Fit V (i, t), (16)
are discussed in Section 4. A variation of CMM in which j=1

arrivalprobabilitiesare also decision variablesmay be con- Pijk O

0= Fijt = O V (i, j, k), t E=Tk (17)
structed by adding constraint (14). qxit Qit ffi Qxit V (i t) , (18)
n (7), (8), (9), (10), (12), (14),
EPoik=1 V(k). (14)
:> v t) (1 9)
i=l Fsj, Q!t, Iit n
(i j.

The FPM maximizes the minimumweighted number of 100

rides given in any time period. Constraint (15) identifies 80 10:OOAM - 11:00AM
the queue length for various rides, at the beginningof each
time period. Constraint (16) captures the conservation of w 60-
flow at each ride and in each time period. Constraint(17)
ensures that flows for which the existing transition proba-
bilities are zero are forced to be zero, i.e., visitors do not 20 -
travel along those links. This constraint also avoids having LEGEND
1. Metro
the visitors travel a greater distance than required to get to 2. OrientExpress
3. 3. Sky
Sky Tower
0 glN MmJ UP
1 2 3 4 5 6 7 8 9 1011121314151617181920212223242526272829
a ride and limits their behavior to their existing movement 4. Buccaneer

5. CircusWheel 100
in the park. Constraint (18) imposes, a lower and upper 6. GrandCarousel
7. GrannyGranPrix
bound on the queue lengths for the rides. This constraintis 8. Jet Stream 803:00PM - 4:00PM
designed to influence the visitors' perception of the rides. 9. JollyRoger
10. Log Jammer
Influencing the time-dependent transition patterns is a 11. Scrambler
12. SpinOut
challenging and difficulttask. Therefore, we used the out- 13. Ninja
14. Reactor
put of the FPM model to obtain an average time indepen- 15. RoaringRapids
? 40

dent transition matrix. This matrix can be obtained from 16. Sandblasters
17. Subway 20

the optimal flows along the links of the park network (Fijt) 18. Swashbuckler
19. SwissTwist
and the optimum number of customers receiving service at 20. TidalWave 1 2 3 4 s 6 7 8 9 1011121314151617181920212223242526272829
21. Turbo
different rides (Si,). The optimum transition probabilities, 22. Colossus 100

23. Flashback
that generate the optimum flow according to FPM, are 24. Freefall
given by the following equation: 25. Goldrusher
26. Psyclone
80 - 9:00PM- 10:OOPM
27. Revolution
K F. 28. Viper W 60
T 1
P-= IT E zJt V(i,]). 29. Z-Force

k=1 tETk Sit 40

The optimum transition matrix provides a guideline for 20

park managers to seek measures to decrease or increase
the flows of visitors across different links in the park. In 12 3 4 5 6 7 8 9 101112131415161718192021223242526272829
Section 5 we describe the implementationof this model in RIDENUMBER

Figure 3. Ride waiting times.
2.4. Tour Design Model
Tour design is concerned with developing alternate touring L
plans of the park, and keeping visits as much as possible Zil = 1 V(i), (21)
within the timeframes given and time spent waiting in the 1=1
lines or traveling from one ride to another. These plans Z51 + Z{l = 1, 8 and u GEF V (1), (22)
are designed to avoid congestions on days with moderate
or heavy attendance. On lighter days these plans will save ZilEe(0, 1) V(i,l). (23)
time, but will not be as essential to successfulpark visits as The objective function maximizes the weighted number
on crowded days. Customers experience distinct waiting of rides visited, a surrogatefor customer satisfaction.Con-
times at rides during different time intervals of the day. straint (20) decides on the set of rides that could be visited
This informationis essential in designing a good tour (see within each time interval. Based on the forecast of the
Figure 3 for variations in rides' waiting times). We solve attendance level, the expected waiting time for each ride
the tour problem in two stages. In the first stage we deter- (Ril) is computed from the CMM. In constraint (21) we
mine which rides are to be visited in each time interval. define the preferred rides to be visited. With constraint
This problem is referred to as the Ride Selection Problem (22) we impose a limit of one trip between rides that are
(RSP). In the second stage we provide a sequence for the quite far apart because their inclusion in the same time
rides to be visited, called the Ride Visiting Problem interval would force visitors to travel across the park in
(RVP). The RSP is formulated as follows: every time interval. This constraint would limit the travel
n L time imposed for the rides selected.
Max E > WiZil (FPM) The RSP is similar in structure to the multiple-choice
t=l 1=l
knapsackproblem, and the solution procedureprovidedby
s.t. Gavish and Pirkul (1991) was modified to incorporatecon-
straint (22), where a set of knapsack subproblemshad to
be solved. This approachgenerates many feasible solutions
and effectively solves this class of problems. Solutions to
the RSP define the set of rides along with the time inter- Step 4. The shortest path from the source node S to the
vals in which they have to be visited. terminal node T determines the tour for visiting the rides.
To complete the touring plan we need to'find the order- Step 2 of the procedure finds a Hamiltonian path for the
ing of the rides that minimizesthe visitors'travel time. The rides in each time interval.The open tours in this step are
problem of finding the ride visiting order (RVP) can be improved by the implementation of a simple P-OPT pro-
modeled as a variation of the TravelingSalesman Problem cedure. Step 3 joins the L Hamiltonianpaths optimally.In
with group precedence. The rides in each period must be our implementation,we solve the RVP problem for each
completed prior to the rides in the next time period.. feasible solution found for the RSP. This iterative ap-
Crucial to the implementationof our proposed solution proach enables us to find the grouping of the rides that
procedure is the computation time needed to obtain a provides maximum utility, based on the preference for
good solution. Consequently,we focus on generating fast each ride and the total number of rides. Implementation
solutions. Spacefilling curves are a natural tool to effi- of the tour design model is discussed in Section 5.
ciently solve combinatorial optimization problems in
Euclidean space (see Bartholdi and Platzman 1988). Heu-
ristics based on this technique are quite fast in execution 3. DATA COLLECTIONAND ANALYSES
and are particularlysuited for environmentswhere time or In this section we describe the results of our data col-
computing resources are limited. An interesting feature of lection and related analysisthat were undertakenas a part
this approachis that clusters of points in the original space of the study reportedin this paper. The primarypurpose of
maintain their closeness in the reduced space. Combinato- the empirical study was to obtain the requisite data to
rial problems are much easier to solve in the reduced implement the models presented in Section 2. In addition,
space. We use the following procedure to solve RVP: this phase of our study was used to develop some insights
into the operations of the theme park. This additional
Procedure to Solve RVP:
data, some of which was qualitative, played a key role in
Step 1. Let the spacefillingcurve 0 map the unit interval establishing credibilitywith the management of the park
onto the unit square. For the set of rides selected in the and facilitatedimplementationefforts.Managerialimplica-
time interval 1, say N1, compute the inverse image of rides tions of our empirical study are discussed later in this
in N1 under 0-1. Let 0-1(N1) denote the corresponding section.
points. To generate a data base to support our study we relied
Step 2. For each 1 E L, sort 0-1 (N1)'s according to on direct observations as well as routine information col-
their positions on the unit interval and identify the first lected by the park over an extended period of time. For
and last ride to be visited in each time interval.Let a1 and the sake of brevity, we do not present the details of this
f31denote these rides. data. Instead we provide illustrativeexamples and describe
the results of the study. As the reader will appreciate,
Step 3. Constructa layered network G = (N, E) with the some of the raw data is proprietaryinformationand is not
following characteristics available in the public domain.
(I) Layers:The number of layers in the network is L + 2.
Layers 0 and L + 1 are the source (S) and the terminal
3.1. Primary Data
node (T).
(II) Nodes: There are two nodes in each layer 1, except the Determining the transitionpatternwithin the park was the
source and terminal, referred to as a1 and I3l. focus of the primarydata collection in this study. Specifi-
(III) Arcs: Directed arcs in the networkfully connect every cally, a survey was administrated to 6,101 customers to
adjacent layer. Node S is connected to all nodes in layer 1. obtain information related to their movements within the
Node T is connected to all nodes in layer L. park. Essentially, a questionnaire elicited three pieces of
(IV) Costs: Let g(x) denote the cost of arc x in the net- informationfrom each respondent individualor group: (a)
work. The cost of each arc is computed as follows: the preceding ride visited by the customer, (b) the subse-
g(S -> al3fl]) = time to get to node al[31] quent ride proposed to be visited, and (c) the age range of
g(a1[f31]-> al+[f31+1]) = total time needed to go from the group. The data on the subsequent ride to be visited
node a1[fl3]to node 013+1[a1+1] to node 01+i [a,+1] and was found to be unreliable since the customers did not
visit all the rides designated in time interval 1 + 1, alwaysfollow their plans and hence (b) was not used in the
starting at node f3l+1[al+1] and ending at node remainderof the analysis.The data on preceding ride visit
at1+JP[1+1]. was used to constructthe transitionmatrixPijk describedin
= total time needed to go from
g(a1[31] -> f31+1[a1+1]) Section 2. Since the survey was administered throughout
node 0a1[R31]to node ay1+1[131+F
] and visit all the rides the entire day, we were able to identify the time-
designated in time interval 1 + 1, starting at node dependent transitions.Our results indicate that it is possi-
and ending at node fl31+1[oa1?1].
?a1+1[f31?1] ble to distinguish three different time-of-day-dependent
g(alL[fL] -> T) = travel time from ride aL[L] to the transitionpatterns--one each in the morning,alfternoon,
entrance. andevening.The morningtransitionpatterncoveredthree

time periods, from 10 a.m. to 1 p.m. The afternoon tran- queue lengths at the rides. An important aspect of
sition pattern occurred until 7 p.m., and the evening customer-relateddata is the distributionof customer arriv-
transition covered the remaining time intervals until als at the park. Although the park attendance level varies
10 p.m. An aggregatetransitionmatrixof the entire day was significantlythroughout the year and shows considerable
obtained by averagingthe three transitionmatrices. Based seasonality, the percentage of total arrivals at different
on the survey data, we were able to identify some domi- hours of the day is quite predictable. An empirical distri-
nant flow patterns with implicationsfor the park manage- bution was constructed (see Law and Kelton 1982) based
ment. These findings are discussed later in Section 3.3. on this data set and was used in the capacity and flow
pattern models in Sections 4 and 5. The distribution of
3.2. Secondary Data visitors' arrival times along with the distribution of the
The extensive data base of the park provided supplemen- number of visitors in the park at any given hour was used
tary information for our study. The data in this data base, for short-termwork force scheduling of the park.
which was developed over a period of 20 months, between
January 1990 and August 1992, may be broadly classified 3.3. Preliminary Data Analyses and Their
into two groups: ride or operations-relatedand customer- Implications for the Theme Park
related data.
In this section we describe our initial analyseswith empir-
(a) Ride or operations-related data. For each of the major ical data and discuss briefly the implications for the park
rides and shows, the park management collects the follow- management.For ease of exposition, the materialhas been
ing informationon an hourlybasis: (1) ride nominal capac- classified into two categories.
ity, (2) hourly throughput, (3) wait times, (4) queue
lengths. These data were used to define the performance (a) Transitionmatrix,performancefunctions, and in-park
functions for each ride. The empirical data were used to flows. One of the objectives of the data analysis was to
validate modeling efforts to capture the dynamicsof rides. validate the transition matrix and the performance of the
In our implementation we focused on two measures- transition functions used in the formulationof the models
throughput and queue lengths. The expressions for these described in Section 2. This was achieved by observingthe
two measures, given by Equation (1), are rather straight- queue lengths and waiting times and comparingthem with
forward and were found to be acceptable for our study. the results provided by the approximationfunctions and
We used the above data to assess the popularity of each through the neural network model. The minimum, aver-
ride. The index for capturing customer preferences was age, and maximum average waiting time differenceswere
used to define the weights Wi in the objective function of found to be 0, 6, and 14 minutes, respectively.While the
the models describedin Section 2. The index was also used sample of three days used in this experimentis too small to
to develop an ABC classification of rides similar to permit statisticallysignificantconclusions, the results were
schemes commonly found in inventory control. Based on encouraging and the park management was satisfied with
customer preferences and ride capacity, we classified the the credibility of the proposed model of the functional
rides into four groups. Class A, with four rides (Colossus, equations.
Viper, Ninja, and Goldrusher), accounted for 20% of the In addition, our preliminaryanalysisidentified dominant
total rides. As discussed later, this qualitative insight was flow paths and gave some insights for developing imple-
useful in designing implementation schemes to influence mentation schemes to influence customer flows. We found
transition patterns in the park as well as in developing that:
routing schemes for group visitors. Also, based on this (i) Visitors to the park typically maintain a clockwise
classificationof the rides and customer transitionbehavior, (60%) or counterclockwise(nearly 40%) direction for vis-
the park was able to evaluate the impact of closing a spe- iting the various rides/shows. This behavior was more
cific ride, relocating a ride, or adding new ride. prominent in the mornings and tended to weaken in the
We were also providedwith informationabout the variable afternoon. The later behavior is not very surprising.Since,
cost of operating the rides at different capacity levels. the most popular rides are farther along the clockwise
These cost functions include the staffing, electricity, and path, and for some customers the afternoon trip is not the
maintenance charges. In general, three differentcost func- firstone and hence they are less willingto go aroundthe park
tions describe the operations costs of the rides. For a sub- to visit the less attractiverides/shows.
set of the rides, the operating cost is independent of the (ii) Visitors tend to remain in the same vicinity. Typi-
capacitylevel, and for other rides the operating cost varies cally more than 50% tended to stay in the same area as the
with every capacity level.
previous ride.
(b) Customer-relateddata. The following data collected (iii) Within a vicinity, customers tend to visit A and
by the park formed a majorinput into our analysisand pro- B category rides first rather than the less popular C and D
vided the necessaryinformationto operationalizethe models category rides.
of Section 2: daily park attendance, hourly arrivaland de- (iv) A very small fraction of customers are willing to
parture counts, number of hours park was open, and choose a ride twice in a row.
(v) There was a substantialvariation (between rides and efficiency of all theme park locations. Our preliminary
time of day) in the wait times. For example, the average analysis indicated the deficiencies of this performance
wait times for the more popular rides such as Viper varied measure. Theme parks operate in different environments
between 15 and 20 minutes in the evenings (after 7:00 with differentcustomer demographics,various park sched-
p.m.) to between 40 and 55 minutes in the peak time ules, and different lengths of open season. Also, distinct
(12:00 noon-4:00 p.m.). Charts in Figure 3 show the aver- customer classes may have different ride threshold values;
age wait time for the rides on Saturdays,at three different thus, merely counting the number of rides given is seldom
times of the day. The wait times for children's rides are a sole measure of customer satisfaction. The problem of
small and keep decreasing until the park closes. The wait service measurement is further complicated by noticing
times for Freefall, Ninja, Flashback,Goldrusher, and Psy- that many service systems do not experience a homoge-
clone have their peak at mid-day. Park visitors wait the neous input and thereby may not have a single crucial
longest in the afternoon, between 2 p.m. to 6 p.m., before measure by which to evaluate their performance.Some of
they can experience this group of rides. But the queues our findings in this respect can briefly be described as
and wait times for Revolution, Viper, and Roaring Rapids follows:
are much higher in the morning than at other times during (i) Customers tend to spend more time in the park-
the day. about 30 minutes-for every additional hour that the park
(vi) Customers do not move on from A and B rides to is open.
the other A and B rides more often than to other kinds of (ii) The average number of hours spent in the park
rides. However, certain pairs of rides show high traffic increases by 1.4 hours for every increase of one in the
flows. For example 77% of visitors leave Turbo for Sub- logarithm of the total number of guests.
way, 73% leave Orient Express for Sky Tower, 62% go (iii) The rides/person/hourdecreases with the logarithm
from Grand Carousel to Log Jammer, and 53% of visitors of the total number of guests.
at Swashbucklergo to Colossus. Other significant traffic (iv) Average number of rides per person per hour is a
flow can be derived from the aggregate transition pattern. better measure of customer satisfactionthan avertagenum-
In no case did a significanttrafficflow occur between two ber of rides per person per day, and consequently a better
C and D rides in different areas of the park. Also, more measure of efficient park operation. This measure moti-
people prefer to ride on Viper immediately after the park vated the objective function for the models in Section 2.
opens than any other ride. Finally, few people are willing These findings served as a basis for achieving improved
to ride on the same A and B ride twice in a row, indicating operational performance at the park. The models con-
that only a small proportion (less than 6.8%) of the cus- structed in Section 2 were also influenced by these analy-
tomers would repeat a long line after enduring the first ses. The ride threshold value and average ride per person
one. per hour were used by the park managementto constructa
(vii) As discussed in Section 1 and indicated by the park new trackingsystem for customer satisfaction.The analysis
management, operating profits are positively correlated in this section has led to the search for a more definitive
with the duration of visitors' in-park stay. But the average measure of service performance, to include other dimen-
in-park stay does not continuously increase with the num- sions of the service package delivered by the park.
ber of rides given per person per day. Further analysis of
average in-park stay indicates that there is a threshold 4. MANAGINGRIDES CAPACITY
value for average number of rides of about 12 rides per
person per day. Although this ride threshold value could Poor management of theme park capacity may result in
not be reached on many crowded days, on days when it considerable undesirable customer waiting or under-
could be reached and surpassed, customers chose not to utilization of the available capacity. Capacitymanagement
stay any longer in the park, indicatingthat the extrabenefit at the theme park is exacerbated not only by low utiliza-
of an additional ride is almost nil, except in customer per- tion of some rides but also by lengthy queues for other
rides, during distinct hours of the day. Customers also
ception of the park. Also, an increase in the total number
of guests attending the park tends to increase the length of have varying degrees of tolerance for waiting for thrill
in-park stay, since it will take customers longer to reach rides during different hours of the day-adding further to
their ride threshold or, otherwise put, to "get their mon- the complexity of park management. In this section we
discuss different aspects of ride capacity management at
ey's worth." When expecting high customer attendance,
the park. These analyses include discussion of various pol-
the park management can also increase the durationof the
icies and their implicationsfor improvingtheme park per-
park's open hours to provide a higher level of service to
formance. First we verify the quality of the models
(b) Parkperformancemeasure. The Operations Depart-
ment within the theme park corporation uses "numberof 4.1. Model Verification and Potentials
rides per person per day," to measure park performance. In Section 3 we established the validity of the approxima-
This is the industrystandardfor determiningthe operating tion functions and verified the quality of the transition

220000 240000

C 200000 220000

l 1d80000 to 200000
1000 E 180000
10 1400000 0
040000 g ACTUAL
. 120000 Estimated
E = 140000--
z 100000 120000
120000 - t ~ Ad-Hoc Adjustment

80000 *e? --- - CMMResults

0 5 10 15 20 25
Time Periods 80000
0 5 10 15 20 25
Figure 4. Model's prediction vs. actual performance. Time Periods

Figure 5. Expected improvementfrom CMM.

functions used in our mathematicalmodels. To verify the
CMM's potential contribution empirically, we extracted
policies. In the static policy we attempt to obtain the opti-
hourly informationon actual capacitylevels and number of
mal-but constant-ride capacity throughout the day. In
rides given throughout the period from February22, 1993
this version of the model, constraint (8) is replaced with
to August 8, 1993. The attendance level for these days
ranged from 9,261 to 35,646 and the cumulativenumber of
rides given in one day (excluding the shows) ranged from G K

101,518 to 206,724. We simulated the park environmentby E E Yigk= 1 V (iM (24)

g=1 k=1
using the same arrivalpatterns, park operating hours, and
capacity levels used during the day for each ride. Figure 4 The obvious advantage of this policy is its stable work-
compares the actual service provided versus the model force requirementfor operatingthe rides. For the dynamic
predictions. The average error was +0.79% and the maxi- policy, K, the number of time intervals in which the ride
mum error was +2.75%. Although the model overesti- capacity could change (in constraint (9)) was set to three
mates the park performance,the magnitude of percentage to correspond to the changes in the number of customer
of error is well within the acceptable range. This step of transition patterns. Although this policy requires variable
model verification provided further justification for using work-force sizes, the schedules could be easily accommo-
three transitionmatricesthroughoutthe day to capture the dated within the existing park operating performance
movement of customers while in the park. and/or by redeployment of operators among the rides. In
In addition to verifying the model's prediction power, the flexible policy, K was set equal to T (the number of
we also had to establish the degree of improvement that hours the park was open), to evaluate the extent of ride
managementcould expect to obtain from implementingthe capacity changes during the day.
CMM,insteadof the existingad hoc way of changingthe ride We simulated these alternative scenarios for different
capacity level. Additionally, we performed an experiment attendance levels at the park. The attendance level was
in which the operating budget and the maximum queue changed from 10,000 to 34,000 to provide a larger and
lengths at various rides were forced to be similarto the ex- more uniform spectrum of attendance statistics. The em-
periment discussed at the beginning of this section, while pirical distribution of percentage of arrivalswas used to
the capacitywas allowed to change with the changes in the get the estimated hourly arrivals.Furthermore,the operat-
customer transitions intervals. The result shows that, on ing budget, which was set by the park management, was
average, a 6.89% improvement was gained by the CMM, approximately75% above the minimum operating budget
rangingfrom 2.6% to 15.2%over the currentmyopic prac- needed, and the queue lengths were limited by the maxi-
tice. Adjustingfor the model's average overestimatingten- mum queue lengths found in Section 4.1. Figure 6 plots
dency of 0.79%, the improvement amounts to 0.87 the performance of various policies. The results have sev-
additional ride per person per day. To place the CMM's eral significantmanagerialimplications:
improvementin the appropriatecontext, we derived lower
and upper bounds on park performance by setting the 216000
operatingbudget constraint(9) to minimumand maximum @ 2060000

levels possible. The average improvements amounted to o 196000

37% of the gap between the upper and lower bound values *) 1860000

for the test data. Figure 5 depicts the improvements. 1760000

n 166000 FLEXIBLE
E 156000
4.2. Capacity Management Policies: z
146000 v STATIC
Three capacity planning strategies were examined. These 1360000
strategies vary in terms of ride capacity usage and imply 9 14 19 24 29 34
Attendance Level
different work-force scheduling for operating the rides.
The policies are referred to as static, dynamic,and flexible Figure 6. Comparisonof different capacitypolicies.

230000 arrivals in the first two hours. The squared correlation

C 220000 * - coefficientwas 0.896. We developed a regression model to
> 210000 - Attendance
obtain an updated estimate of the cumulative arrivals,
(a 200000 Attendance
:9 190000
which was used to readjust the ride capacities for the re-
'~180000 mainder of the day. Estimates of remaininghourly arrivals
were obtained by using the empirical distribution of per-
E 1600000
centage of arrival during the day. Correspondingly,the
140000 capacitymodel had to be run twice to set the ride capacity
1 3 5 7 9 01 13 15 levels, once prior to the opening of the park and then after
the statistics of the second hour were available.
Figure 7. Service-budgettrade-off curves. Lindo's (1992) industrialoptimizationpackagewas used
to solve all the resulting mixed integer programs in our
(i) With low park attendance-less than 15,000-park experimentations.The average time for solving the prob-
visitors reach their ride threshold values with the park lems examined was 18 minutes (excluding the generation
simply operating at minimum capacity level. of input matrix), and the maximum time was 37 minutes.
(ii) With high levels of attendance-more than 27,000- All the problems were solved on HP Vectra 486 machine
the park needs to operate at maximum capacity level. In with 32-bit processor and 66-MHZ speed.
addition, show schedules have to be increased to achieve
the desired customer service level.
(iii) Maximum benefit from CMM is achieved in the 5. MANAGINGFLOWS IN THE PARK
range of 15,000 to 27,500 customers attending the park.
In Section 3 we described the existing transition patterns
This range covered 42% of the days in our data set.
and concentrated on identifying how visitors move
(iv) The performances of dynamic and flexible policies
throughoutthe park. We identified three dominant transi-
are quite close. The flexible model only results in 0.29%
tion patterns and analyzedtheir implicationsfor the theme
additional improvement. The static model causes 4.85%
lower performance. park. In this section we focus on characterizingthe opti-
mum transition patterns and develop managerial policies
The operating budget in the preceding simulation was
to influence customer behavior in order to improve the
based on the park management'sdecision. To evaluate the
impact of this managerial decision on park performance, park's service level.
we developed a budget-servicetrade-off curve, which pro-
vides an estimate of the expected service level delivered 5.1. Optimum Transition Patterns
for different levels of total operating budget. Figure 7 The FPM developed in Section 2 is used to capture the
shows the budget-service plot. The trade-off curves are desired transitionprobabilitiesand movement of visitors in
shown for three different attendance levels-15,000, the park. The FPM, in addition to identifyingthe optimum
20,000, and 25,000-for which the capacity management transition patterns, also configures the capacity level at
model was most effective. These trade-offcurveswere cru- each ride. To make the results amenable to implementa-
cial in bringingthe park service deliverycloser to customer tion, the following measureswere taken. The queue length
expectation and desired threshold value, since the park of each ride was restricted to 80% of the maximumqueue
managers could evaluate the impact of budget allocation length observed at different attendance levels. This con-
on park performance. straintreflects the visitors'waiting behavior and the theme
park'sdesire to reduce the visitors'waiting times. Also, the
4.3. Implementation Aspects flow patterns in the park were restricted to the routes
The park attendance level varies significantlythroughout taken by the customers, so that they would not have to
the year and shows considerable seasonality.The disparity travel long distances to get to their next rides. The effect of
in forecastand actualattendancelevel complicatesthe imple- this constraint is to limit the managerial measures to ef-
mentationof the CMM.Althoughthe averageforecasterror forts at increasing or decreasing the flow of visitors within
was negligible, the forecasting system currently in place the present customer transition patterns. The budgetary
results in significant forecast errors, either in overesti- constraints and the rides threshold were defined by the
mating or underestimating the actual attendance. This park managers.
observationwas compelling enough to abandon the idea of To understandthe extent of the optimum transitionpat-
setting ride capacity based solely on the park's forecast terns' influence on park performanceand number of rides
of attendance level. Instead we developed a hybrid pol- given, we plot the results of our simulation in which the
icy that uses both the forecast of arrival for the day and attendance level at the park was varied from 10,000 to
the actual information about the visitors' arrival up to 34,000. We also report the results obtained from the re-
11 a.m. vised version of the CMM, in which the arrivingtransition
The hybridpolicy was motivated by the observationthat probabilitiesare decided in addition to the ride capacities.
total park attendance is highly correlated with the total These results are comparedwith the results obtained from

attractions are viewed as less attractive and popular than
370000 Current

El Arrival
the thrill rides offered by the park. The longest wait for a
320000 - Optimal
show usuallydoes not exceed the durationof one complete
performance. In addition, our interviews indicated that
270000 many visitors defer their plans to attend the shows until
later in the afternoon, hoping to use the early part of the
E 220000
z day for the most attractive rides. This observation moti-
170000 vates the scheduling of the shows with higher frequencies
in the afternoons from 2 p.m. to 6 p.m. Colossus, Ninja,
120000 Flashback, Goldrusher, and Psyclone experience their
8000 13000 18000 23000 28000 33000 38000
Attendance Level
longest queues during this time interval.
Shows attract visitors away from the crowded rides, in
Figure 8. Expected improvementsfrom flow management. effect reducing park congestion levels. To appraise the
likely impact of the shows and generate plausibleschedules,
the CMM, which would indicate the additional improve- we modifiedthe transitionprobabilitiesaccordinglyto accom-
ment expected over and above what the CMM model modate the shows.Given the proposedshow schedulesin any
would provide. In all the models the budgetarylimits were time period, the functionalequationsprovidenew estimates
the same and time-of-the-daytransitionpatternswere sim- of the queue lengthsat majorrides.Developingan optimiza-
ilar to the transition intervals identified in Section 3. Fig- tion model that yields the optimumschedule and frequency
ure 8 depicts the results obtained. On average, a 17% of the showsis part of our ongoingresearch.
improvement over the optimum results obtained from the Additionally, the communicationdevices (TV monitors,
CMM was obtained. However, 57% improvementcould be signs, etc.) in the park could be used favorablyto alter the
expected by alternatingthe overallflow patterns.The aggre- existing transition patterns. TV monitors at each location
gate transition flow is obtained over different attendance could provide informationabout waiting time, show sched-
levels ranging from 15,000 to 25,000. Comparing the cur- ules, and indicate rides with shorter queue lengths. Signs
rent and the desired transition patterns identified the posted at the waiting line areas could indicate the approx-
paths along which the flow has to be adjusted to achieve imate waiting time to deter customerswho have low toler-
the preferred behavior. Some of the significantflows to be ances for experiencing long waiting times. This practice
increased are: Spin-out > Roaring Rapids, Swashbuckler points toward a different queue management:rather than
> Scrambler,Roaring Rapids > Revolution, and Orient hiding the actual queue length, which makes the wait more
Express > Spin-out. The major flows to be decreased are: uncertain,reduce the visitor's anxietyby providingreliable
Z-force > Reactor, Orient Express > Sky Tower, Grand informationabout expected waiting time, thereby reducing
Prix > Flashback,and Psyclone z> Jet Stream. stress while waiting (Maister 1984). Lastly, moving attrac-
With the current transitionpatterns, the customers'ride tions such as magicians and hypnotists could also direct
threshold value could be reached when the attendance customers toward less congested parts of the park.
level is less than 15,000. With the optimum arrival pat-
terns, the ride threshold values could be attained with up 5.3. Designing Planning Tours
to 19,000 customers arrivingat the park. Given the opti- Until recently, the theme park provided few suggestions
mum movement within and upon arrival at the park, the for visitors to better plan their tour of the park. These
ride threshold value was attained at all the various atten- suggestions basically indicated the peak hours for water
dance levels. With a relaxed operating budget and the new rides, coasters, restaurants, shows, and gift shops. These
arrivalpatterns, the ride threshold could be reached up to recommendations fell short of a comprehensive touring
the 25,000 level, indicating the need to influence the flow plan. The tour design models (RSP and RVP) have stimu-
patterns within the park. lated the park management to seek ways to operationalize
of these procedures for all park visitors. These models are
5.2. Policies to Influence Transition Patterns
now in use for generating touring schedules for guided
Several policies were designed to induce visitor behavior group tours, which are provided by the park.
toward the optimum patterns. In this section we present The touring problem was used to create various alterna-
one of these options in detail and address its likely impact tive tours for different categories of visitors, where toler-
the park performance. ance to waiting time, preference for rides, and height
The scheduling (number and timing) of the shows and limitations, as well as the length of the tour, were taken
other entertainment in the park is crucial to park perfor- into account.
mance. Although they create additional dimensions to the
service package offered by the park, these activities are
designed to improve the operational efficiencyof the park 6. CONCLUSIONS AND FURTHER RESEARCH
and alleviate the loads on the rides with long queue We have described a preliminaryapplication of a model-
lengths during the day. Generally, shows and theatrical based approach to managing the capacity and flow at
theme parks. The management of the park was actively in data collection. The paper has also benefited signifi-
involved in various phases of this study and provided con- cantly from the valuable comments provided by the editor,
siderable guidelines for shaping the direction of the the associate editor, and the two anonymousreferees. This
research. research was partially supported by grants from UCLA
An interestingresearch aspect of theme parks and other AcademicSenate and Centerfor TechnologyManagement.
service systems in which customers experience long queues
lies in understandingdifferent aspects of customer percep-
tions of the waiting times. It would be particularlyinterest- REFERENCES
ing to explore how the capacity and flow management
problems should be modeled, such that visitors'perception BARTHOLDI, J. J. AND L. K. PLATZMAN.1988.HeuristicsBased
on SpacefillingCurves for CombinatorialProblems in
of acceptable waiting times as well as the throughput of
EuclideanSpace.Mgmt.Sci. 34, 3, 291-305.
the rides is incorporated.Such analysisrequires additional
FORT,J. C. 1988. Solvinga CombinatorialProblemVia Self-
data on visitor perceptions of ride waiting times in order to
OrganizingProcess:An Applicationof the KohonenAl-
improve customer satisfaction with the service delivery gorithm to the TravelingSalesmanProblem.Biological
system. Cybemetics,33-40.
Finally, the analysis presented in this paper has used a GAVISH, B., ANDH. PIRKUL. 1991. Algorithmfor the Multi-
single customer class and we did not incorporate the de- Resource GeneralizedAssignmentProblem.Mgmt.Sci.
mographyof the customers in the park transitionpatterns. 37, 6, 695-713.
Our questionnaire included the age categories of the cus- LARSON, C. R. 1987. Perspectiveon Queues: Social Justice
tomers, but further data regarding the distributionof dif- and the Psychologyof Queues.Opns.Res. 35, 6, 895-905.
ferent customer classes at different times of the year need LAW,A. W. ANDW. D. KELTON. 1982. SimulationModeling
to be collected to facilitate extension of our models. andAnalysis.McGrawHill.
LINDO.1992.Lindo:Optimization ModelingLanguage.LINDO
System,Inc., Chicago,IL.
MAISTER,D. H. 1984.ThePsychologyof Waitingin Lines.HBS
The author is grateful to Professors Devanath Tirupati, Note 9-684-064.
Rakesh Sarin, and Donald Erlenkotter for their valuable MASSON, E. ANDY. WANG. 1990.Introductionto Computation
comments, Ted London, Jane Thompson, Adekunle and Learningin ArtificialNeural Network.EJOR, 47,
Shoyombo,and Joan Forster-Yamauchifor their assistance 1-28.