Professional Documents
Culture Documents
Abstract: Decision making for managing risks to critical infrastructure systems requires accounting for (1) the uncertain behavior of
disruptive events; and (2) the interdependent nature of such systems that lead to large-scale inoperability. This paper integrates a dynamic
Downloaded from ascelibrary.org by New York University on 05/12/15. Copyright ASCE. For personal use only; all rights reserved.
risk-based interdependency model, the dynamic inoperability input-output model, with a multiobjective decision tree to analyze preparedness
decisions. The use of a dynamic model allows for resilience and recovery decisions to be incorporated in the decision-making framework,
and uncertainty is accounted for using probability distributions. The multiobjective inoperability decision tree is applied to the study of
transportation infrastructure disruptions, namely closures of an inland waterway and an inland waterway port. A data-driven multiregional
study of the Port of Catoosa in Oklahoma, along the Mississippi River Navigation System, is discussed and suggests careful consideration
when investing larger amounts toward port security. DOI: 10.1061/(ASCE)IS.1943-555X.0000171. © 2013 American Society of Civil
Engineers.
Author keywords: Decision tree; Disruptive event; Infrastructure preparedness; Multiobjective decision; Dynamic inoperability
input-output model (DIIM).
Distribution of Disruptive Events X̄ð1 − X̄Þ
β̂ ¼ ð1 − X̄Þ −1 ð13Þ
s
Discussed in more detail in the methodological development
section, the states of nature describe the magnitude of a disruptive
event at the inland port, where such an event may be rare in nature Dynamic Multiobjective Inoperability Decision Tree
with large impacts. Originally introduced through the idea of the
scale invariance (Richardson 1948), which is an inverse power scal- Extending a static version of the multiobjective inoperability deci-
ing between independent and dependent variables, the power-law sion tree (MOIDT) (Santos et al. 2008), this section describes a
Downloaded from ascelibrary.org by New York University on 05/12/15. Copyright ASCE. For personal use only; all rights reserved.
distribution has been used to model different social, political, and dynamic and stochastic MOIDT to model the (1) investment in in-
financial patterns in history. frastructure preparedness, and (2) the interdependent benefits of
Studies aimed at determining the probability distribution of such investments (or adverse impacts of a lack of investment).
the severity of a terrorist attack agree that, according to empirical The use of the DIIM considers the cumulative effect of the
data on worldwide terrorist events from 1968 to 2008, the proba- total economic loss, a risk measure of interest to decision makers.
bility of a terrorist event claiming x deaths follows a power-law Also, the problem is solved here by taking into account the uncer-
distribution (Clauset and Youngal, unpublished data, 2005; tainty in the parameters of the DIIM as well as in the states of nature
Johnson et al. 2006; Clauset et al. 2007): in the decision tree. As a consequence, the parameters of the beta
distribution for each entry in the demand perturbation matrix, c ,
PðxÞ ¼ Cx−λ ð8Þ
depend on the state of nature as well as on the point estimates of
This model, which is provided in Eq. (8) for scale parameter λ each entry.
and normalizing constant C, has been extended to a more general
form to describe other covariates such as the size and experience of Infrastructure Preparedness Decision Problem
the terrorist organization and the type of weapons used in the attack
(Clauset and Gleditsch 2009; Clauset and Wiegel 2009). Power-law An influence diagram is a useful graphical tool that helps to visu-
relationships are also used to model the severity of natural disasters alize the decision problem in terms of the decision, uncertainties,
such as tornadoes (Malamud and Turcotte 2012), other severe objectives, and the influential relationship each has on the others.
weather conditions (Dessai and Walter 2000), and large earth- For the infrastructure preparedness decision problem in Fig. 2, the
quakes (Mega et al. 2003), among others. decision to make, represented by a rectangle, is a certain amount
Statistical methods such as the least square methods, the of dollars invested in preparedness activities. This decision will
maximum likelihood function, or an extended combination of these directly affect one of the objective functions, represented by a dia-
with other statistical tests are used to estimate the parameters of this mond shape, which is the cost. It will also indirectly affect the other
distribution (Clauset et al. 2009): λ, C, and minimum value of the objective function, which is the expected total economic loss as it
random variable xmin . will impact two uncertainties, represented by a circular shape:
(1) the probability of a disruptive event occurring (in case of a man-
made attack), and (2) the severity of the disruptive event. Hence, the
Distribution of Demand Perturbation problem presented is a multiobjective decision problem in which
The manner in which the disruptive event affects changes in the decision maker is interested in minimizing both the cost of pre-
demand is itself represented with a probability distribution. As in paredness and the expected total economic loss in case of a disrup-
Santos et al. (2008) and Santos (2008), the beta distribution is used tive event. Both objectives are competing (as increased investment
here to describe uncertainty in c . An advantage of using this in preparedness would expectedly lead to fewer losses). And while
distribution, whose probability density function is shown in both are measured in dollar terms, the objectives are noncommen-
Eq. (9), is that its support is on [0,1]. This is beneficial in describing surable in that investment budgets would likely come primarily
a parameter such as ci , the minimum and maximum are control-
lable, and the range of values of its parameters determine several
shapes for the probability distribution:
Γðα þ βÞ α−1
fðxÞ ¼ x ð1 − xÞβ−1 ð9Þ
ΓðαÞΓðβÞ
αβ
σ2 ¼ ð11Þ
ðα þ βÞ2 ðα þ β þ 1Þ
Fig. 2. Influence diagram describing the infrastructure preparedness
The parameters of this distribution are estimated using the
investment decision-making process
method of moments. Given the values of the sample mean, X̄,
has a negative exponential relationship with the factor of influence, the recovery process depends on the case study and will be further
θI , computed in Eq. (14). This relationship ensures that as the developed in the illustration in case study section.
amount of investment is increased, the reduction in demand after The decision is then based on the aggregate economic loss
a disruptive event is exponentially decreased. These relationships computed at the time, T, when all industries are fully recovered,
are represented in vector form in Eq. (16), where X̄ increases as and since the decision tree is a one-period, more precisely one-year
the proportion of the maximum demand perturbation increases, de- decision tree, there is one investment to be made.
pending on the severity of the event, and it decreases exponentially
as the factor influence (or equivalently the amount invested) Simulation of the Decision Tree
increases:
Alluded to previously, stochastic decision trees require the use of
X̄ ¼ γe−θI cmax ð16Þ Monte Carlo simulation in the folding back process. For each alter-
native emanating from the first decision node, a fixed cost repre-
Likewise, standard deviation s is computed by assuming that the senting the amount invested is assumed, and θI for an investment of
maximum demand perturbation lies three standard deviations I can be computed. The following steps guide the simulation for a
above the mean, shown in vector form in Eq. (17). No further sufficiently large number of iterations:
restrictions or assumptions are applied to sample variance, which 1. Generate a random variable from the power law distribution
results in larger variations when the sample mean is smaller: within the bounds of the minimum and maximum severity,
guided by Eq. (8).
1
s ¼ ðcmax − X̄Þ ð17Þ 2. Given the severity, factor of influence, θI , and the elicited
3 values cmax and qmax ð0Þ, calculate the sample mean, X̄, and
Similar to the sample mean, the initial inoperability is a function standard deviation, s, of the demand perturbation from
of both the investment strategy, θI , and the severity of the event, γ, Eqs. (16) and (17), the parameters of the DIIM from Eqs. (18)
as shown in Eq. (18), where qmax ð0Þ represents the vector of initial and (19), and the probability, pI , of a disruptive event occur-
inoperability under the most extreme circumstances: ring from Eq. (15).
3. Using the method of moments, compute estimates for the para-
qð0Þ ¼ γe−θI qmax ð0Þ ð18Þ meters of the beta distribution from Eqs. (12) and (13) and
draw random variables from this distribution for each industry
The values of qmax ð0Þ could either be determined by experts or element in the demand perturbation vector. It is assumed that
estimated depending on the system considered in the application. the number n 0 of initially perturbed industries can represent a
Resilience is often defined from two aspects: (1) reducing the subset of all industries, n 0 ≤ n.
impact of a disruptive event, and (2) improving the speed with 4. Compute the inoperability vector at each point in time
which recovery occurs (Henry and Ramirez-Marquez 2012; Zobel from Eq. (3). P
2011). The effect of an investment in improving the first aspect is 5. Compute the total economic loss, QðTÞ ¼ xT τt¼1 qðtÞ.
found in Eq. (18), where the initial inoperability experienced after a 6. Compute the expected value of the total economic loss as the
disruptive event is lessened with an investment in infrastructure product of the loss and the probability p of a disruptive
preparedness. The second aspect is addressed in Eq. (19), where event, EQðTÞ ¼ QðTÞp.
preparedness investments can also decrease the time to full recov- Repeating the above steps for a number of iterations, N, for one
ery of industry i, T i . Like the previous relationships, recovery particular investment cost results in a distribution for EQ under this
time is reduced according to e−θI T i , an exponentially decreasing specific investment. The mean values of the EQ distributions for
function of the factor of influence, θI . As recovery commences, each investment could be compared to determine the appropriate
inoperability decreases over time, consequently impacting total investment level. Further, the conditional expected value (upper-
economic losses across all sectors over time, QðtÞ ¼ xðtÞT qðtÞ: tail value) provides an idea of how each investment performs in
h i extreme conditions (Asbeck and Haimes 1984).
ln q ðeqi−θð0ÞI T Þ
i
ki ¼ −θ i ð19Þ
ðe I T i Þð1 − aii Þ Illustrative Example: Inland Waterway Port
Investment
The amount invested in port security and system hardening, in
general having an exponential impact on the effect of a disruption, The methodology developed in this paper is illustrated with a
is a more realistic representation of the real life behavior of im- decision problem regarding inland waterway port security and resil-
provements in the risk management of infrastructure systems. ience. This data-driven case study involves the Port of Catoosa in
ceeding in its mission of ensuring “a secure and resilient Nation mum perturbation and initial inoperability represent a one-year clo-
with the capabilities required across the whole community to pre- sure of the port. Since this case study is motivated by the amount of
vent, protect against, mitigate, respond to, and recover from the grant from PSGP allocated to the group of ports to which the Port of
threats and hazards that pose the greatest risk” (DHS 2011). Catoosa belongs, the maximum amount of investment is deter-
Among this set of grants is the Port Security Grant Program mined accordingly. It is also assumed that all the industries require
(PSGP), whose purpose is “ : : : to support increased port-wide risk the same time, T ¼ T i , to fully recover, where T depends on the
management; enhanced domain awareness; training and exercises; duration of the closure of the port. In this particular example we
expansion of port recovery and resiliency capabilities; and further consider κ ¼ 0.5, and assume that if the port was closed for D days,
capabilities to prevent, detect, respond to, and recover from the industries relying on the port as well as the interdependent in-
attacks : : : and : : : assist ports in obtaining the resources required dustries will all recover in T ¼ κD days. This assumes that while
to support the NPG’s associated mission areas and core the port was closed, the products that were supposed to be exported
capabilities.” were held at the port and those that were supposed to be imported
According to the grant program overview report (DHS 2012), were held at their original port, and after the port reopened, the port
the ports in the United States were categorized into three different needs T days to ship all the products held and hence the industries
groups, and DHS divided the grant of the PSGP among the three would be fully recovered. This would mostly affect the manner
groups. However, the ports within each category were to compete in which the demand perturbation is decreasing in the recovery
for the funding available in their group. According to that report, period, and instead of adding c ðtÞ, in Eq. (6), we are adding
the group in which the Tulsa Port of Catoosa is placed contains a the change in the demand perturbation, Δc , which considers that
total of 22 ports and received a grant of $4,875,000. The approach once the recovery starts, the commodity flows through the port at a
provided in this paper could provide a helpful means to determine rate of 1=κ; in this case the flow is twice as much as it usually is to
the appropriate investment amount allocated to each port based on make up for the loss and the time during which the commodity had
the port’s size, its mix of annual commodity flows, and its location, to wait at the port. During the recovery, the daily demand pertur-
among other factors impacting the probability of a disruptive event bation is subtracted from the initial demand perturbation and
occurring as well as its consequences in terms of economic losses. updated accordingly every day until Δc ðTÞ ¼ 0. Note that while
the port was closed for D days, the recovery process does not
start until the port reopens. This suggests that the port and associ-
Assumptions ated industries are not functioning for D days, after which the
The probability of a (rare) event causing a disruption of an extreme recovery process commences and spans T ¼ κD days before the
nature at the port is assumed to be similar to the probability of a port and industries are fully recovered. However, since the DIIM
terrorist attack since this type of event would incur large impacts models the recovery process, the model is implemented for T ¼ κD
and rarely occurs. This assumption is merely for illustration pur- days and not the full period of time required to fully recover,
poses; a sensitivity analysis will follow to address scenarios with D þ κD:
different probabilities. Bun (2006) developed a mathematical
1 1
model to measure the probability of a nuclear terrorist attack, find- Δc ð1Þ ¼ c ð0Þ − c ð0Þ Δc ðtÞ ¼ Δc ðt − 1Þ − c ð0Þ
ing that, with a set of plausible parameters, the probability of a nu- T T
clear terrorist attack in the next 10 years is 0.29. Extending from ð20Þ
this result, the probability associated with a terrorist event occurring
in a one-year period is assumed to be 0.29=10 ¼ 0.029, which Finally, we assume that the demand perturbation comes from the
corresponds to one period of a decision tree. This computation loss in exports and is computed as the ratio of exports of industry i
assumes independence between probabilities of occurrence of a to the total production output of that industry at a specific time, t, as
terrorist attack over the years and generalizes the likelihood of a shown in Eq. (21):
nuclear terrorist attack to any broadly defined disruptive event.
Therefore, a baseline probability for a disruptive event occurring ei ðtÞ
ci ðtÞ ¼ ð21Þ
in one year is assumed to be 0.02 for this particular application. xi ðtÞ
Given that the result used is specific to a nuclear attack, it might
seem unrealistic to model any type of disruptive event occurring at Similarly, the inoperability is the result of imports not reaching
the Port of Catoosa. However, since determining the appropriate the port and causing a shortage in the material needed to produce in
estimate for the probability of a disruptive event lies beyond the the industries relying on such commodities. It is then the ratio of
scope of this research, a baseline probability is set to be slightly imports to the total production output of industry i, provided in
less than the given result and followed up by a sensitivity analysis Eq. (22) (MacKenzie et al. 2012a, Pant et al. 2011):
If both the numerator and the denominator in Eqs. (21) and (22)
are yearly estimates, the above metrics can serve as the maximum
demand perturbation and the maximum inoperability because the
worst-case scenario is considered to be a full year of port closure
and they are used to compute the initial demand perturbation and
inoperability. The elements of the sample mean and initial inoper-
ability vectors in Eqs. (16) and (18) are then expressed in Eqs. (23)
and (24), where ei ðtÞ, mi ðtÞ, and xi ðtÞ, respectively, represent the
exports, imports, and total production output for industry i at time t
of one year:
Downloaded from ascelibrary.org by New York University on 05/12/15. Copyright ASCE. For personal use only; all rights reserved.
−θI ei ðtÞ
X̄ i ¼ γe ð23Þ
xi ðtÞ
mi ðtÞ
qi ð0Þ ¼ γe−θI ð24Þ Fig. 4. Frequency distribution of the expected total economic losses
xi ðtÞ
across all regional industries
The input data described above rely on either (1) parameter
assumptions that could be altered depending on the application,
or (2) port-specific data. Several data sources from the Tulsa Port measured in dollars, they are not necessarily commensurate as
of Catoosa (2009), the U.S. Army Corps of Engineers (2010), and the preparedness investment would likely come from a port author-
the Bureau of Transportation Statistics (2009) were used by ity or DHS funding program while economic losses would be
MacKenzie et al. (2012a) to derive estimates of the dollar amount shared across multiple industries. The resulting Pareto-optimal
of commodities flowing through the Port of Catoosa. These frontier, plotting expected total economic loss, EQðTÞ, computed
estimates are inputs to Eqs. (21) and (22). The interdependencies once the industries are fully recovered, versus investment is shown
between the critical infrastructure and the rest of industries are in Fig. 5. Note that, as Fig. 4 suggests, there exists a probability
expressed by two matrices in the DIIM model: (1) the A matrix distribution for expected total economic loss for each realization
describes how inoperability propagates among industries when the of investment. Shown in Fig. 5 are the mean, M, of these distribu-
infrastructure is disrupted, and (2) the K matrix governs the recov- tions, as well as conditional means CM 0.01 and CM 0.05 . Although
ery of infrastructure and industry sectors. The entries in the A the Pareto-optimal frontiers in Fig. 5 are not completely smooth, it
matrix are computed using data from the Bureau of Economic is concluded that none of the investment strategies is dominant due
Analysis, and the formula for the entries of the K matrix is to the overall shape of the curve. Additional iterations would likely
discussed the DIIM background section. The idea of expressing result in smoother curves.
initial inoperability and demand perturbation as a fraction of the Here, the conditional expected total economic loss values are
worst-case scenario has been used in a different study on port upper-tail, or extreme, values associated with higher consequence,
disruptions (Pant et al. 2011). The length of closure of the port but lower probability, events (Asbeck and Haimes 1984). The
is chosen to be anywhere between two weeks and two months, conditional expected value is found with Eq. (25) and indicator
a reasonable assumption in the case of any disaster [e.g., see Pant function in Eq. (26):
et al. (2011)]. And the time to full recovery, controlled by κ, is ran-
domly chosen. These parameters can be altered according to each
case and largely depend on the decision makers and the type of risk
they seek to mitigate.
1 if EQðiÞ > β
δðiÞ ¼ ð26Þ
0 otherwise
Conclusion
24(3), 697–713. U.S. Dept. of Transportation. (2011). America’s marine highway, U.S. De-
Santos, J. R., and Haimes, Y. Y. (2004b). “Modeling the demand reduction partment of Transportation, Maritime Administration, Washington, DC.
input-output (I-O) inoperability due to terrorism of interconnected infra- Zobel, C. W. (2011). “Representing perceived tradeoffs in defining disaster
structures.” Risk Anal., 24(6), 1437–1451. resilience.” Decis. Support Syst., 50(2), 394–403.