Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Decision Support Systems 50 (2011) 394–403

Contents lists available at ScienceDirect

Decision Support Systems


j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / d s s

Representing perceived tradeoffs in defining disaster resilience


Christopher W. Zobel ⁎
Department of Business Information Technology, Virginia Polytechnic Institute and State University, Blacksburg, VA 24061-0235, United States

a r t i c l e i n f o a b s t r a c t

Article history: Two of the primary measures that characterize the concept of disaster resilience are the initial impact of a
Received 9 April 2010 disaster event and the subsequent time to recovery. This paper presents a new analytic approach to
Received in revised form 1 October 2010 representing the relationship between these two characteristics by extending a multi-dimensional approach
Accepted 3 October 2010
for predicting resilience into a technique for fitting the resilience function to the preferences and priorities of a
Available online 30 October 2010
given decision maker. This allows for a more accurate representation of the perceived value of different
Keywords:
resilience scenarios to that individual, and thus makes the concept more relevant in the context of strategic
Adjusted resilience decision making.
Optimization Published by Elsevier B.V.
Visualization
Tradeoffs
Decision support
Preference modeling

1. Introduction comparing such resilience, Bruneau et al. [4] introduced the concept of
the resilience triangle, which incorporates measures of both the
As public awareness continues to grow of the long-term implica- robustness against initial loss due to a disaster and the rapidity of the
tions of disasters such as the 2010 Haiti Earthquake, the concept of recovery process. This initial concept was extended by Zobel [21], who
disaster resilience is becoming more widely recognized as an defined the related measure of predicted resilience and then presented
important means of characterizing the ability of individuals, commu- a multi-dimensional approach for visually comparing resilience across
nities, and organizations to recover from the impacts of such events. different disaster scenarios.
Resilience is commonly defined as “the act of rebounding or springing In order to improve the utility of resilience as a comparative
back” [12] from a disaster, and a resilient organization often is measure, the following discussion extends these previous efforts by
described as one which is able to quickly return to normal (or even introducing an approach for explicitly incorporating a decision
improved) operations after such an event has occurred [14,15]. This maker's perceptions of the relative value of different resilience
ability to recover, however, can also be improved by efforts to mitigate scenarios into a new, more representative, measure of adjusted
against and prepare for the initial impact of a disaster, and therefore resilience. Because different individuals may view the same resilience
there also exists support for a broader definition of resilience which scenario from a variety of different perspectives, such a measure,
incorporates both pre-event and post-event activities [10,17,19]. which can be fine-tuned to accurately represent a specific perspective,
This broader definition of resilience allows it to be used for will be far more useful for a decision maker than would be a measure
strategic decision making, in the sense that it provides a means for that represents a single fixed interpretation of resilience. By fitting the
assessing the relative risk of alternative scenarios. For example, adjusted resilience function to the given decision maker's perspective
whether planning a new hospital in an earthquake zone or expanding over several different scenarios, one can help to make resilience much
a supply chain into a politically unstable region, the extent to which more consistent and meaningful for that individual and thus help to
alternative options support both initial resistance to and recovery support more effective decision making.
from a possible disaster can have an impact on their long-term The following discussion begins with an overview of previous
effectiveness. It thus can be important to be able to measure the work on resilience and on the resilience triangle in particular,
relative amount of resilience associated with these factors for the including a look at the predicted resilience measure. It then motivates
different alternatives, in order to better inform the process of and derives a formulation for adjusted resilience and characterizes the
choosing between them. As a first step towards measuring and bounds on its defining parameters. After a brief discussion about
capturing decision makers' preferences, an optimization model is
presented for fitting the adjusted resilience function to the stated
⁎ Tel.: +1 540 231 1856. preferences of a given decision maker, and an example is provided of
E-mail address: czobel@vt.edu. its implementation. The paper concludes with a general discussion of

0167-9236/$ – see front matter. Published by Elsevier B.V.


doi:10.1016/j.dss.2010.10.001
C.W. Zobel / Decision Support Systems 50 (2011) 394–403 395

the technique in the context of its ability to support more effective


decision making in disaster operations management.

2. Background

2.1. Disaster resilience

Different aspects of the concept of resilience are currently being


studied from a number of viewpoints within the academic research
Fig. 1. The original resilience triangle (adapted from [4]).
community [7,10,11,13,18]. In general, many such studies tend to
focus on the resilience of either the physical (technological) aspects of
a system [3,6] or the social (human) aspects of a system [7,9], but they then derived a simple quantitative measure for the loss of
there is also significant discussion about studying combined human– resilience in a system:
environment interactions (socio-ecological systems) [8,19]. Given the
importance of recognizing the larger context within which these t1
R = ∫t ½100−Q ðt Þdt ð1Þ
factors exist [1], even the articles that focus on a specific aspect of 0

resilience will typically discuss the significance of other factors (such


as legal or organizational issues) that can affect resilience overall. The where Q(t) represents the quality of the system's infrastructure at a
resulting diversity of perspectives represented in these different given time t. The vertical axis in Fig. 1 effectively represents the level
approaches to the subject has led to a large number of different of robustness of the given system, and the horizontal axis represents
working definitions that have been developed for the concept of the associated rapidity of recovery [3].
disaster resilience [20]. By generalizing the concept of infrastructure beyond that of just a
Even with differences in opinion about the scope of the concept physical system, Bruneau et al. [4] also use Eq. (1) to discuss what they
and about terminology, however, there is ongoing interest in refer to as the technical, organizational, social, and economic
developing quantitative techniques for measuring disaster resilience, dimensions of resilience. Thus, for example, robustness can be
in order to better support planning and decision making [8]. To be measured with respect to social, organizational, or economic
useful for this purpose, such techniques must either be focused “functionality,” as well as with respect to physical functionality. As
specifically on a particular aspect of resilience (such as physical defined by Eq. (1), the resilience measure may have very different
resilience), or else they must be general enough to support applica- values within each of these dimensions, depending on the nature and
tion to both the physical and social aspects of resilience, as well as impact of a given disaster. Chang and Shinozuka [5] extend this idea
application to other important “environmental” aspects of the concept by establishing the notion of predefined performance standards for
(such as economic resilience or political resilience). Because of its both robustness and rapidity, against which actual system perfor-
simplicity, the resilience triangle of Bruneau et al. [4], provides a strong mance can be compared. They then define resilience as the probability
basis for developing quantitative measures that can be applied to that these standards will be met, in each of the technical, organiza-
resilience in this more general context, as well as in a more focused tional, social, and economic dimensions.
fashion. It is due to this combination of simplicity and general appli- Bruneau and Reinhorn [3] further demonstrate that redundancy
cability that the resilience triangle was chosen as the basis for the can be visualized by using multiple simultaneous resilience triangles.
development of the adjusted resilience measure discussed in the later They also offer a more precise quantification of the quality function, Q
part. (t), in terms of both loss and performance standards, and they discuss
some of the probabilistic aspects of assessing resilience. Cimellaro,
2.2. Resilience triangle Reinhorn, and Bruneau [6] provide a further extension to these ideas
by explicitly defining the area beneath the curve for Q(t) as a direct
Bruneau et al. [4], define disaster resilience as being characterized measure of resilience itself. This allows for a reduction in the size of a
by the extent to which the following factors are present in either a resilience triangle to be reflected as an increase in the corresponding
physical or a social system: resilience of the system that it represents.
(1) Robustness—the strength of a system, or its ability to resist the
2.3. Multi-dimensionality and non-linearity
impact of a disaster event, in terms of the amount of damage or
loss of functionality that results because of the event.
Although there are advantages to using a single calculated value to
(2) Rapidity—the rate or speed at which a system is able to recover
define disaster resilience, it is also important to recognize the
to an acceptable level of functionality, after the occurrence of a
potential issues associated with doing so. In particular, if resilience
disaster event.
is considered to be a function only of the area of the resilience triangle,
(3) Resourcefulness—the level of capability for dynamically
then very different combinations of initial loss and recovery time can
responding to a disaster event, by identifying and implement-
correspond to exactly the same resilience value. Thus, for example, a
ing solutions to improve rapidity and/or robustness.
public building which suffers only slight damage in an earthquake, but
(4) Redundancy—the extent to which components of the system
which has a long recovery time due to scarcity of resources, may have
are substitutable, and therefore able to be replaced or
exactly the same measured amount of resilience as a privately-owned
augmented when functionality has been lost or reduced.
building with significant initial damage but a much quicker recovery
The last two factors, resourcefulness and redundancy, are time (See Fig. 2). In reality, however, these two different scenarios
generally considered to be the “means” by which disaster resilience may not be equivalent for a given decision maker, and he or she may
can be improved, with the corresponding “ends” being measured by actually prefer to be faced with one situation over the other (and
the impact of these improvements on the factors of robustness and might thus implicitly consider that situation to have greater
rapidity [4]. resilience).
In order to explicitly represent resilience as a combination of these In order to better capture and represent such preferences, Zobel
first two factors, Bruneau et al. [4] introduced the concept of the [21] introduced a multi-dimensional representation of resilience that
resilience triangle, as illustrated in Fig. 1. Based upon this concept, incorporates both the resilience value and its two defining
396 C.W. Zobel / Decision Support Systems 50 (2011) 394–403

Fig. 2. Two triangles representing the same resilience for different situations.

characteristics of robustness and rapidity. This approach, which we This larger encompassing area is established by specifying a
will discuss in more detail in the later part, supports visually parameter T* that serves as a strict upper bound on the set of possible
comparing different scenarios with respect to all three of these values for T (assuming that beyond a certain time limit, a system has
attributes at the same time, and it thus provides a more comprehen- an effective resilience of 0) [21]. This allows various instances of
sive means of assessing and differentiating between their relative predicted resilience to be compared on the same relative scale.
levels of resilience. It is important to note, however, that Zobel's [21] The actual predicted resilience function for the ordered pair (X, T)
approach does not provide a means for explicitly representing a is then given by [21]:
particular decision maker's preferences between scenarios. Bruneau
and Reinhorn [2] have suggested addressing this need by treating T*− XT XT
RðX; T Þ = 2
= 1− X∈½0; 1; T∈½0; T* : ð2Þ
resilience as a nonlinear measure that could be calibrated to fit T* 2T*
different situational preferences. By doing so, one would allow the
resilience measure to clearly and directly indicate the relative value of By construction, the minimum possible value for R(X, T) under this
different situations to a particular decision maker, and thus help to formulation (i.e., X = 1 and T = T*) is 0.5, and the corresponding
capture the varying perceptions and interpretations of the actual maximum possible value (for X = 0 or T = 0) is 1.0. This is simply a
disaster resilience inherent in a given system. result of basing the function on the area of the resilience triangle, and
With this in mind, the primary focus of the discussion in the later it is straightforward to rescale R(X, T) to the interval [0, 1], based upon
part is to propose an approach for capturing the non-linearity of the preferences of a given decision maker. We shall examine this in
decision makers' preferences that also explicitly incorporates the more detail in Section 5.1.
multi-dimensional nature of resilience into its representation. We The relationship between X and T can now be written as:
begin the discussion by defining the notion of predicted resilience, as
introduced in Zobel [21], and provide an overview of the multi- XT = ðX−0ÞðT−0Þ = M; where M = ð2 T*Þð1–RÞ; ð3Þ
dimensional approach for visualizing the relative predicted resilience
of different systems. We then develop a new dynamic formulation for which is the equation of a rectangular hyperbola centered at (0,0) and
resilience that allows us to fit a resilience function to the preferences with asymptotes lying on the X and T axes. Thus for a fixed resilience
of a given decision maker, and we discuss an approach for optimizing value, R, (and a fixed T*) the set of all possible combinations of X and T
the accuracy of this fit. describes a hyperbola. By systematically varying R, we therefore may
generate a series of equilateral hyperbolas, all of the same shape, in
3. Predicted disaster resilience the upper right quadrant of the plane (see Fig. 4), with larger values of
R corresponding to curves closer to the origin.
Zobel [21] defines the predicted resilience for a given system to be a Observations which lie on the lower right portion of each
function of the predicted amount of initial loss and the associated resilience curve represent scenarios for which the initial disaster-
recovery time for a future disaster event, where both values are related loss is very high but the recovery time is small. Such a situation
estimated based on the information available before the event may occur, for example, when an affected facility is located in an area
actually occurs. The initial loss value, X, is measured as a percentage with a very strong infrastructure and quick access to resources in
of the total available functionality of the system (in the context of any support of rebuilding. In contrast, observations which lie on the upper
of the four dimensions of resilience), and the recovery time, T, is
measured in a relevant time unit, such as weeks or months. Although
each of these measures could be represented probabilistically [3], the
predicted resilience function assumes that they are deterministic
estimates of the expected future behavior of the system [21]. With
respect to the four factors used to characterize disaster resilience,
recovery time (T) represents a measure of rapidity, and 1 − X provides
a measure of robustness, as in Ref. [6]. Predicted resilience is thus
considered a function of these two factors.
The calculation of a value for predicted resilience is based on
approximating the area under the quality curve represented by Fig. 1.
Because the area above that curve may be approximated by
calculating the area of the corresponding resilience triangle, the
predicted resilience is generated by simply subtracting the area of that
triangle from a fixed larger area within which it is contained, and then
representing the result as a percentage of that larger area (see Fig. 3). Fig. 3. The predicted resilience triangle as a proportion of T*.
C.W. Zobel / Decision Support Systems 50 (2011) 394–403 397

values than those that are actually produced by Eq. (2). As with the
previous example, simply calculating a value for resilience is not
sufficient if it does not match the needs of the decision maker and
does not therefore support effective decision making.
In order to address this issue, and thus to make the concept of
predicted resilience more applicable to actual decision makers, we
introduce an approach to analytically representing a given decision
maker's adjusted resilience function. This offers a means for
calibrating the predicted resilience to fit a decision maker's situational
preferences, as suggested by Bruneau and Reinhorn [2], and thus it
ultimately provides more value for resilience as an implementable
analytic measure.

4.1. Representing preferences

The basis for our development of a function for adjusted resilience


is the notion of threshold values for both X and T which identify the
upper limits on the preferred set of values for each parameter. This
idea was introduced by Chang and Shinozuka [5], in the context of
defining a region of acceptable resilience within which all values
satisfy both constraints. We may extend this notion of a single
preferred region, however, to recognize the additional regions which
can be identified as “preferred” with respect to only one of the two
Fig. 4. Predicted resilience curves.
parameters, as well as the region in which both parameters exceed
their desired levels. Our approach to incorporating decision maker
preferences is based upon these four regions and upon the relative
left portion of each curve represent facilities which suffer a small perception of the resilience values within them, as provided by a given
initial loss but are faced with a relatively long recovery time. An decision maker. Fig. 5 thus illustrates the subdivision of the
example of such a case might be a well-constructed facility that is underlying resilience domain, given specific threshold values of X̂
located in a relatively remote area, and for which fewer resources are and T̂.
immediately available for restoring functionality in a timely manner. As an example of how different preferences may be represented in
Because both types of facility may have the same calculated resilience, this context, we may compare the relative perception of observations
however, this multi-dimensional representation offers an approach in region 1 (desirable X and T) with the perception of observations in
for easily differentiating between the very different situations that regions 2 (desirable T only) and 3 (desirable X only). If it is very
they represent. important to a decision maker that values of T not exceed the
threshold T̂, then we might expect resilience to decrease more rapidly
4. Incorporating decision maker preferences (with respect to increases in T) in region 3 than is indicated by the

As the previous discussion indicates, the ability to visually


compare the relative resilience of multiple facilities with respect to
both robustness and rapidity, as well as with respect to the predicted
resilience value itself, can provide a decision maker with a more
complete understanding of the impact of a potential disaster event
upon their organization's assets. Furthermore, as suggested by Zobel
[21], the individual preferences of that decision maker can be
represented, to a certain extent, by allowing them to identify and
characterize particular regions of interest within the overall resilience
space.
It is important to recognize, however, that the visualization
approach discussed above does not actually provide the ability for a
decision maker to capture and represent their perception of the actual
resilience value for a given facility. For example, Eq. (2) implies that
for a given disaster event, a facility with a 10% loss of functionality and
a recovery time of 20 months (assuming a maximum acceptable
recovery time of T* = 100 months) would actually have a predicted
resilience of 99%. This value may seem artificially high to some
decision makers. From the standpoint of interpreting such results and
then using them to drive policy decisions, artificially high values may
ultimately have a negative impact on efforts to actually mitigate
against the impact of a potential future disaster.
A decision maker may also wish to reduce (or increase) the
differentiation between the given resilience values within a particular
region of interest. For example, given a preference for smaller initial
losses over smaller recovery times, a decision maker may feel that two
facilities with the same predicted initial loss and a difference of only
one month in recovery time should have much more similar resilience Fig. 5. Resilience regions defined by X̂ and T̂.
398 C.W. Zobel / Decision Support Systems 50 (2011) 394–403

original predicted resilience function. In contrast, if there is a relative the formulation includes a number of adjustable parameters that can
indifference between values of X larger than X̂, for a particular T, then help fit the function to a particular set of preferences. Following the
the relative decrease in actual resilience in region 2 (as X increases) development of the generalized model formulation, we briefly discuss
may actually be perceived to be less than that provided by the the issue of actually collecting these preferences, and then introduce
predicted resilience. Fig. 6 provides an example of the adjusted an approach for determining values for the parameters that can
resilience curves that might correspond to these two scenarios. provide an optimal fit for a given decision maker.

4.2. Assumptions 5. Adjusted resilience function

Several important assumptions are implicit in the resilience profile The adjusted resilience function defined in the later part
reflected in Fig. 6. The first assumption is that there are only four incorporates three different parameters that can be used to adjust
regions of interest, as described above, which represent the more or the original predicted resilience to better fit a decision maker's
less preferred ranges of values for both X and T. Although these ranges preferences. The first parameter, α, is used to adjust the overall slope
could themselves be further sub-divided to represent more levels of of the resilience function, so that its value may be either reduced or
preference, this initial partitioning concisely represents the opportu- increased as necessary. The other two parameters, γ and δ, serve a
nity for characterizing different preferences, and thus best supports similar function, but only for the regions corresponding to “less
the initial development of the underlying model, as given in the later acceptable” values for either X or T. In combination, these three
part. parameters can provide significant flexibility for fitting the adjusted
The second implicit assumption is that, rather than there being a resilience function to the specific preferences of a given decision
discrete drop in resilience immediately above the threshold values, maker or organization.
the adjusted resilience remains a continuous function of X and T
throughout the domain (with the option to suddenly increase the rate 5.1. Adjusting resilience on the entire domain
at which resilience decreases beyond these threshold values). Thus
incremental changes in X or T correspond to incremental changes in As discussed previously, one of the potential issues with the
adjusted resilience. original predicted resilience function is that it may generate values
The third assumption inherent in Fig. 6 is that adjusted resilience that are generally scaled higher (or lower) than a decision maker
maintains, within each region, the basic hyperbolic nature of the might otherwise prefer. Because resilience can be only defined on the
original predicted resilience function. This ensures not only that interval [0,1], however, the amount and type of change that can be
increasing values of X or T will always correspond to decreasing applied is restricted. In particular, one cannot simply add a fixed
resilience (and thus that the marginal slope of the resilience function amount to the resilience of each observation because this will lead to a
is always decreasing) but also that the slope in a region will always be number of values greater than one. We thus introduce a parameter, α,
constant for any fixed value of X or T. This constant marginal rate of that is used to adjust the original resilience value by a proportion of
change is a consequence of the triangular nature of the resilience the amount by which it differs from the maximum value of 1, and
function, and it directly reflects the underlying relationship between which therefore keeps the relative impact of this adjustment
X and T. consistent for all observations in the domain.
With these assumptions in mind, we can explicitly define a The following equation allows for adjusting the predicted
formulation for the adjusted resilience function. Because different resilience defined in Eq. (2), and also maintains the symmetric
decision makers will likely have different preferences in each region, hyperbolic relationship between X and T.

       
XT XT XT−αXT ð1−αÞXT
Rα ðX; T Þ = 1− +α = 1− = 1− :
2T* 2T* 2T* 2T*
ð4Þ

In order for Eq. (4) to represent resilience, as it has been defined,


not only must all Rα(X, T) be contained within the interval [0,1] but
also the choice of α must support the assumptions stated in
Section 4.2. Section 6 will therefore be used to derive bounds on the
set of possible values for α, as well as those for each of the other
parameters discussed below.
It is important to note that setting α = 0 in Eq. (4) gives the
original predicted resilience function of Eq. (2), and setting α = − 1
effectively rescales resilience to the interval [0, 1]. Because Rα(X, T) is
symmetric, any such changes in the value of α will have exactly the
same impact on both X and T. In order to allow for adjustments
associated with only one of these two variables, we must therefore
introduce additional parameters, γ and δ. The following discussion
describes how these parameters are subsequently incorporated into
the new resilience function.

5.2. Adjusting resilience outside each “preferred” region individually

Let X̂ and T̂ represent a decision maker's chosen threshold values


Fig. 6. Example of adjusted resilience curves. for X and T respectively. The specification of theses threshold values
C.W. Zobel / Decision Support Systems 50 (2011) 394–403 399

provides the opportunity to consider adjusting resilience outside of 5.3. Adjusting resilience outside both “preferred” regions
the “preferred” region for each variable. Because we are assuming that
the perceived resilience function is continuous over the entire If both X and T meet or exceed their respective thresholds at the
domain, however, any changes to the formulation of the resilience same time, we encounter a situation in which each of α, γ, and δ may
function as it transitions from a “more preferred” to a “less preferred” be simultaneously incorporated into the adjusted resilience function.
region must reflect this continuity. One approach to this is to restrict The new function given in Eq. (7) reflects the overall combination of
the adjustment of the “less preferred” resilience values to a proportion these different adjustments:
of the possible amount of available change, as we did in Section 5.1.  
However, rather than being bounded above by a maximum resilience Region 4 : X≥ X̂; T ≥ T̂ :
value of 1, observations in the “less preferred” regions are instead  
ð1−αÞðXT Þ
bounded above by the resilience values that are achieved at the Rαγδ ðX; T Þ = 1−
2T*
boundaries with the “preferred” regions. 0 0  11
  ð1−αÞ X̂T
The additional adjustment of the resilience function thus proceeds ð1−α ÞðXT Þ
+ δ@ −@ AA
as follows: Let γ be the parameter used to specify the amount of 2T* 2T*
change associated with T, and let δ be the corresponding parameter 0 0  11
used to specify change in the resilience function with respect to X.   ð1−αÞ X T̂
@ ð1−αÞðXT Þ @ AA
Given the current adjusted resilience function, as defined in Eq. (4), +γ −
2T* 2T*
we may further adjust it for values of T ≥ T̂ and X≤ X̂ by incorporating
0     1
γ in the following way: ð1−αÞ XT + δT X− X̂ + γX T− T̂
= 1−@ A
2T*
 
Region 3 : X≤ X̂; T ≥ T̂ : ð7Þ
 
ð1−αÞðXT Þ
Rαγ ðX; T Þ = 1− As with Eqs. (5) and (6), the resilience function given in Eq. (7)
2T*
0 0  11 transitions smoothly and continuously to the corresponding
  func-
  ð1−αÞ X T̂
ð1−α ÞðXT Þ tions in the adjoining regions since Rαγδ X̂; T = Rαγ X̂; T and
+ γ@ −@ AA    
2T* 2T* Rαγδ X; T̂ = Rαδ X; T̂ .
  1 ð5Þ
0
ð1−αÞ XT + γX T− T̂ 5.4. Overall adjusted resilience
= 1−@ A
2T*
0   1 By combining Eqs. (4)–(7), we may therefore define the overall
ð1−αÞX T + γ T− T̂ adjusted resilience function, for a given choice of α, γ and δ, as
= 1− @ A
2T* follows:

8  
> ð1−αÞXT
>
    >Rα ðX; T Þ = 1−
>
> 2T*
for X ≤ X̂; T ≤ T̂
>
>   1
>
> 0
It can easily be seen that Rαγ X; T̂ = Rα X; T̂ for all X≤ X̂, >
>
> ð1−αÞX T + γ T− T̂
>
regardless of the value of γ. This ensures that the resilience varies
>
>
>Rαγ ðX; T Þ = 1−@ A for X ≤ X̂; T N T̂
>
> 2T*
>
<
continuously across T for any X in the preferred region. Furthermore, 0   1
Rðα;γ;δÞ ðX; T Þ = ð1−αÞT X + δ X− X̂
>
>
since (1 − α)γ in Eq. (5) is being multiplied by a non-negative value >
>
>Rαδ ðX; T Þ = 1−@ A for X N X̂; T ≤ T̂
>
> 2T*
(because T≥ T̂, and X, T ≥ 0), the sign of (1 − α)γ determines whether >
>
>
> 0      1
>
>
or not the slope in Region 3 is being increased or decreased over that >
> ð1−αÞ XT + δT X− X̂ + γX T− T̂
>
> @ A for X N X̂; T N T̂
>
:Rαγδ ðX; T Þ = 1−
>
associated with Eq. (4). 2T*

In much the same way, we may also adjust Eq. (4) for values of ð8Þ
X≥ X̂ and T≤ T̂ by incorporating δ as follows:

where X ∈ [0, 1] and T ∈ [0, T*]. Each combination of α, γ and δ


 
represents a different type of adjustment to the original predicted
Region 2 : X≥ X̂; T≤ T̂ :
  resilience function, and thus each such combination represents a
ð1−αÞðXT Þ different overall resilience profile.
Rαδ ðX; T Þ = 1−
2T*
0 0  11
  ð1−αÞ X̂T 6. Establishing bounds on the parameter values
ð1−α ÞðXT Þ
+ δ@ −@ AA
2T* 2T*
ð6Þ Given the range of possible resilience values and the assumptions
0   1 presented in Section 4.2, the adjusted resilience function given above
ð1−αÞ XT + δT X− X̂
= 1−@ A must satisfy the following for all X ∈ [0, X *] and T ∈ [0, T *]:
2T*
0   1 Rðα;γ;δÞ ðX; T Þ ≥ 0 ð9Þ
ð1−αÞT X + δ X− X̂
= 1−@ A
2T* Rðα;γ;δÞ ðX; T + ΔT Þ ≤ Rðα;γ;δÞ ðX; T Þ ∀ΔT ∈½0; T*−T  ð10Þ

    Rðα;γ;δÞ ðX + ΔX ; T Þ ≤ Rðα;γ;δÞ ðX; T Þ ∀ΔX ∈½0; X*−X  ð11Þ


Then, as above, Rαδ X̂; T = Rα X̂; T for all T≤ T̂, and the
combined function is continuous across X for any given T in the These necessary conditions allow us to easily identify bounds on
preferred region. parameters α, γ and δ:
400 C.W. Zobel / Decision Support Systems 50 (2011) 394–403

6.1. Constraint 1: negative slopes 6.2. Constraint 2: R(X*,T*)N = 0

Eqs. (10) and (11) indicate that the partial derivatives of R must be By definition, R(0,0) = 1 and all predicted resilience values must
negative in each of the four quadrants. We may thus derive the fall on the interval [0,1]. Because the marginal slope of the predicted
following bounds on the values of the parameters α, γ and δ: resilience function is negative everywhere in the state space,
however, this implies that the smallest possible predicted resilience
Region 1 : X≤ X̂; T≤ T̂ : must be achieved at R(X*,T*). We may therefore ensure that all values
    of R are properly constrained by requiring that R(X*,T*)N = 0. This
dR ð1−αÞT ð1−αÞT
=− ≤0⇒ ≥0⇒ð1−αÞ≥ 0⇒α≤1 requirement provides the following relationship and establishes our
dX 2T* 2T*
final bounds on α, γ and δ:
ð12Þ
0     1
This same upper bound on the value of ∝ is reached by similarly ð1−αÞ X*T* + δT X*− X̂ + γX* T*− T̂
analyzing dR/dT. Because α takes on a single value across all four RðX*; T* Þ = 1−@ A≥ 0
2T*
quadrants, Eq. (12) holds for all X ∈ [0, X *] and T ∈ [0, T *].

Region 2 : X≥ X̂; T≤ T̂: 2T*


  ⇒ ≥ X*T* ð1 + δ + γÞ−γX* T̂−δ X̂T*
dR ð1−αÞT ð1 + δÞ ð1−αÞ
= ≤0⇒ð1−αÞT ð1 + δÞ≥0⇒δ≥  1
dT 2T* ð18Þ
ð13Þ
As in Eqs. (15) and (16), this derived relationship provides
This lower bound on the value of δ is independent of the values of conditional bounds on each of the parameters. In particular, it
the other parameters, and given this result, dR/dT will always also be provides a lower bound on the value of α, and an upper bound on
less than or equal to zero within Region 2. Since δ must have the same the values of γ and δ, each relative to the values of each of the other
value in Regions 2 and 4, Eq. (1) holds in both quadrants. parameters.

7. Fitting decision makers' preferences


Region 3 : X≤ X̂; T≥ T̂:
 
dR ð1−αÞX ð1 + γÞ Given the adjusted resilience function and the associated bounds
= ≤0⇒ð1−αÞX ð1 + γÞ≥0⇒γ≥  1
dT 2T* on each of the component parameters, we may now consider the task
ð14Þ of finding the set of parameter values that most accurately represents
the implicit perceptions that a particular decision maker may have
about the “true” disaster resilience of a given system. In order to
As above, this lower bound on γ is independent of the values of the
accomplish this task, it is necessary to first capture and quantify a
other parameters. It also leads to a result of dR/dX ≤ 0 within Region 3,
representative set of such “resilience perceptions” which can be used
and it applies to both Region 3 and Region 4.
as the basis for deriving the most appropriate parameter values. There
are a number of techniques in the literature which could be useful for
Region 4 : X≥ X̂; T≥
0T̂ :  1 this purpose, including the analytic hierarchy process (AHP) [16]
dR ð1−αÞ T ð1 + δ + γÞ−γ T̂ which provides a widely used approach for establishing preferences
=− @ A≤ 0
dT 2T* between different alternatives.
 1 Although it could be effective, however, a highly structured
0
ð1−αÞ T ð1 + δ + γÞ−γ T̂ approach such as the AHP is not necessarily required to gather a
ð15Þ
⇒@ A≥ 0 decision maker's perceptions. For example, one alternative approach
2T*
would be to have the decision maker directly specify what he or she
 
⇒ð1−αÞ T ð1 + δ + γÞ−γ T̂ ≥ 0 perceives to be the relative resilience associated with a number of
different hypothetical scenarios, based on experience and intuition.
⇒T ð1 + δ + γÞ≥γ T̂ These perceived values could then be compared and validated against
a baseline of calculated predicted resilience values for each individual
scenario. Similarly, the decision maker could instead specify their
Since (1 + δ) is non-negative and T N T̂, this relationship is true for
 i perception of the “true” resilience of scenario by starting with the
all T∈ T̂; T* , as long as (1 + δ + γ) ≥ 0. Given (1 + δ + γ) b 0 however,
 i baseline values and indicating the relative amount by which these
the relationship will hold for all T∈ T̂; T* only if it is true for T *. The values would need to change in order to reflect their preferences.
following thus provides a conditional lower bound on γ, given δ: For the sake of the following discussion we simply assume that a
reasonable set of such values has somehow been determined, and that
it can be used as the basis for choosing parameters to fit the adjusted
⇒T* ð1 + δ + γÞ ≥ γ T̂ ð16Þ
resilience function to a decision maker's preferences. A more in-depth
look at the advantages and disadvantages of different approaches for
Similarly, from analyzing dR/dT, we get a conditional lower bound achieving these specified resilience values is an important future
on δ, given γ: research area from the standpoint of establishing alternative options
for implementation.

…⇒X* ð1 + δ + γÞ ≥ δ X̂ ð17Þ
7.1. Optimization model

Both of these lower bounds will apply within any quadrant that Let S be a finite set of scenarios (observations) which have been
utilizes the associated parameter within its portion of the adjusted chosen as the basis for collecting perceived resilience values from a
resilience function. decision maker. For each (xi, ti) ∈ S, we may then let R(*, *, *)(xiti) be the
C.W. Zobel / Decision Support Systems 50 (2011) 394–403 401

corresponding perceived resilience value that the decision maker has This initial set, S, of representative scenarios can be chosen in such
specified for that scenario, and define RS* = {R(*, *, *)(xiti) : (xiti) ∈ S}. a way as to leverage the fact that, by definition, the adjusted resilience
Given RS* and a specification for T̂ and for X̂, we wish to find values function is linear within each of the regions defined by X̂ and T̂. In
for α, γ and δ that will “best” fit the adjusted resilience function particular, the adjusted slope of the function within Regions 2 and 3
R(α, γ, δ)(x, t) to the decision maker's perceived resilience structure on (due to the combination of α and δ or α and γ) can be determined
S. Depending on the number of different scenarios in S and the extent from just two adjusted observations in each case, whereas the
to which RS* accurately represents the decision maker's preferences, adjusted slope in Region 1 (due to just α) can be determined by a
we may then conclude that R(α, γ, δ)(x, t) represents the overall single adjusted observation from that region. As a simple illustration
adjusted resilience function for that particular decision maker. of the approach, therefore, our example uses a total of four
The following optimization model provides a simple approach for observations: one observation on the threshold between Region 1
optimizing the fit between R(*, *, *)(x, t) and R(α, γ, δ)(x, t) on S, based on and Regions 2 and 3, and one more within each of Regions 2 and 3.
the definition of the adjusted resilience function and the derived In general, using such a limited number of observations to define
bounds on the individual parameter values: the decision maker's adjusted resilience function would typically be
appropriate only in the context of an initial attempt at codifying the
 2 individual's perceptions. Not only might it be difficult for an individual
ðP1Þ min f ðα; γ; δÞ = ∑i∈S R ; ;  ðxi ; ti Þ−Rðα;γ;δÞ ðxi ; ti Þ to accurately assess the resilience for each observation, but also he or
s.t. ***
she might be inconsistent in assessing the relative differences
α≤1 between the observations. As discussed previously, assessing a larger
number of observations in an iterative fashion can allow for these
δ ≥−1
relative differences to be more precisely balanced and refined.
γ ≥−1 With this in mind, the set of scenarios given in Table 1 were chosen
as the basis for generating an initial adjusted resilience function for
T* ð1 + δ + γÞ−γ T̂ ≥ 0 the warehouse construction manager. Each one includes the calcu-
lated predicted resilience value as well as the manager's perceptions
X* ð1 + δ + γÞ−δ X̂ ≥ 0 of the “true” resilience value.
2T*  
Note that the observations in Table 1 have been chosen so that the
− X*T* ð1 + δ + γÞ−γX* T̂−δ X̂T* ≥ 0 □
ð1−αÞ original predicted resilience value is the same for observations 1 and
3, and also for observations 2 and 4, since each set of points lies on a
The solution to (P1) is the set of parameters, (α, γ, and δ), that single resilience curve. Additionally, observations 1 and 2 share the
minimizes the squared error in the objective function. These same value for T and observations 3 and 4 share the same value for X.
parameters then serve to define the adjusted resilience function for This provides the opportunity for a decision maker to assess the
the individual whose preferences are reflected by RS* . relative resilience of a given observation with respect to at least
Deriving this adjusted resilience function allows us to generalize two other points, and thus helps with the overall consistency of the
the behavior exhibited by RS* , so that we may provide the decision assessments.
maker with values for the relative resilience of new scenarios that are According to the perceived resilience values in Table 1, the
consistent with his or her preferences. The extent to which these new manager believes that the actual resilience embodied in an outcome
values can actually match the decision maker's “true” preferences will of 30% loss of functionality and 15 weeks of recovery time is (or
depend, however, on the number of observations in RS* , as well as on should be) less than that provided by the predicted resilience
the accuracy of each observation in RS* . For this reason, it is important function. The same is true for the combination of 15% loss and
to consider the specification of a particular adjusted resilience 30 weeks of recovery time. It is important to note that these two
function as part of an iterative process of refinement. As the set of scenarios have had, and must have, their resilience adjusted by the
analyzed scenarios, S, is expanded or updated, (P1) can be re-solved same amount (if it is adjusted at all). This is because both scenarios fall
and the adjusted resilience function updated so that it reflects the within what has been identified as the “acceptable” region for both
most recent assessment of the decision maker's resilience threshold values (corresponding to “small enough” values for both
perceptions. variables together), and thus they are only affected by changes to the
α parameter. If this identical amount of adjustment is not consistent
7.2. Example with the decision maker's perceptions of their relative value, then it
indicates the need to revisit the specification of X̂ and T̂. This could be
The following example illustrates the effect of applying the accomplished by extending R(α, γ, δ)(X, T) to be a function also of the
optimization model to a set of decision maker-specified resilience parameters X̂ and T̂. If these parameters are also allowed to vary
values. within the optimization then the resulting problem can be used to
Consider an organization that is planning the construction of a new further clarify the notion of the “preferred region” that they serve to
warehouse facility in an area susceptible to potential hurricane define, and thus to more accurately match the decision maker's
damage, and suppose that the manager in charge of the planning overall resilience perceptions.
process has specified that T̂ = 30 weeks and X̂ = 30% are the upper The perceived resilience values in Table 1 also indicate that in
limits on what would be considered manageable levels of recovery comparing scenarios 2 and 4, the manager prefers a longer recovery
time and infrastructure loss, respectively. Furthermore, suppose that
T* = 100 weeks represents an overall upper limit on the projected Table 1
amount of recovery time that might actually be undertaken, due to Resilience scenarios for deriving an adjusted resilience function.
any storm that might occur.
Observations Predicted Perceived
Given these assumptions, the initial predicted resilience function
resilience resilience
(Eq. (2)) may be formulated and presented to the manager as a Number X T
preliminary indication of the resilience of the facility due to different 1 0.3 15 0.9775 0.95
combinations of initial loss and recovery time. By using this function 2 0.5 15 0.9625 0.90
3 0.15 30 0.9775 0.95
as a baseline, the manager can then specify the extent to which his or
4 0.15 50 0.9625 0.92
her perceptions differ for several representative scenarios.
402 C.W. Zobel / Decision Support Systems 50 (2011) 394–403

time and less initial damage over a shorter recovery time but more Table 3
loss of functionality, perhaps due to the potential for continuing to Comparative resilience results.

utilize part of the facility during repairs. There is also a much larger Predicted Perceived resilience Adjusted
difference between the perceived resilience values for scenarios 1 and resilience (specified by decision maker) resilience
2 than there is between the original predicted resilience values, 0.9775 0.95 0.9455
implying that the manager feels that even with a relatively short 0.9625 0.90 0.9037
recovery time, 50% loss of functionality is more significant than 0.9775 0.95 0.9455
0.9625 0.92 0.9237
initially indicated.
Once the perceived resilience values were specified, the optimi-
zation routine was used to identify the set of parameters that provide
the closest fit between these values and the adjusted resilience are available both for continuing operations at a reasonable level and
function. The routine was implemented in Microsoft Excel, using the for recovering in the desired time frame from the physical damage
Risk Solver Platform, version 9.0. that will have occurred.
Table 2 provides the resulting parameter values that define the Similarly, given the existing infrastructure at the potential site, if
adjusted resilience function. As can then be seen in Table 3, the the organization wished to achieve full recovery within 6 months,
adjusted resilience is much closer to the stated perceived resilience then they could estimate the maximum amount of infrastructure loss
than it is to the original predicted resilience, thus more accurately that would still allow them to attain a target of at least 90% resilience.
reflecting the decision maker's true preferences. The two sets of From an infrastructure investment standpoint, the ability to balance
values are not identical because the parameters, and thus the adjusted the two factors would also allow them to consider the “true” relative
resilience function, are restricted by the upper and lower bounds on value of attempting to decrease the recovery time to five months, in
their values, given the chosen values for X̂, T̂, and T*. terms of the increased amount of associated “acceptable” loss that this
Fig. 7 illustrates the difference between the adjusted resilience would then allow.
function given by Table 2, and the original predicted resilience Because it is well recognized that disaster resilience is associated
function. Although further refinement of the adjusted resilience could with more than just physical infrastructure, we must also consider the
consequently be based on either of these functions, one would expect extent to which the proposed measure can be used to capture and
quicker convergence to a truly representative set of parameters if it represent the multi-dimensional nature of resilience. As an example
were the most recent adjusted resilience function that was used as of this in the literature, Chang and Shinozuka [5] calculate the
the basis for updating the representation of the decision maker's resilience associated with each of the concept's four dimensions
preferences. (technical, organizational, social, and economic) in turn, and then
simultaneously present the results in the context of these multiple
8. Conclusions dimensions. A similar approach could easily be taken with adjusted
resilience, so that a value for the adjusted resilience is generated for
This research effort extends the concept of disaster resilience by each of the dimensions in turn, and with the combined set of values
providing an approach for visually and analytically representing the (and the corresponding set of resilience curves) providing a broader,
perceptions of an individual decision maker of the “true” resilience multi-dimensional view of the resilience of the entire system.
value associated with the predicted initial loss (X) and associated time Since the process of gathering a decision maker's perceptions
to recovery (T) for a system within any disaster scenario. The of resilient behavior, and then generating and visualizing their as-
discussion specifically builds upon the notion of the resilience sociated adjusted resilience function(s), will tend to be iterative in
triangle, as introduced by Bruneau et al. [4], and the complementary
measure of predicted resilience, as defined by Zobel [21], and it
provides for a more complete understanding of the potential
significance of a given disaster event. This, in turn, allows for
improved support for strategic disaster planning and mitigation.
There are a number of situations in which the ability to establish an
accurate representation of the relative value of different disaster
scenarios can provide such decision making support. For example, in
analyzing a new construction project for a distribution center a
decision maker may determine that he or she wishes the new facility
to have at least as much overall resilience as an already existing
center, given a potential disaster scenario. Provided with the
opportunity to construct the new facility in such a way as to keep
initial losses to 10%, the decision maker would then be able to identify
the associated recovery time that would need to be achieved in order
to then reach the desired level of resilience. Changes could then be
made in the organization's overall plan in order to provide support for
such activities as pre-positioning additional emergency stock in
nearby facilities or establishing pre-determined contracts with
suppliers or construction companies, in order to ensure that resources

Table 2
Optimal parameter values.

Parameter Value

α −1.4203
γ −0.3995
δ 0.1514
Fig. 7. Comparison of predicted (red) and adjusted (blue) resilience curves.
C.W. Zobel / Decision Support Systems 50 (2011) 394–403 403

nature, computer-based decision support is a necessary component [10] T. McDaniels, S.E. Chang, D. Cole, J. Mikawoz, H. Longstaff, Fostering resilience to
extreme events within infrastructure systems: characterizing decision contexts
of incorporating the concept of adjusted resilience into the planning for mitigation and adaptation, Global Environmental Change 18 (2) (2008)
operations of an organization. Data management, optimization, and 310–318.
visualization capabilities are all critical for implementing the concept [11] D. Mendonça, Decision support for improvisation in response to extreme events:
learning from the response to the 2001 World Trade Center attack, Decision
and for supporting a flexible decision making process. Despite some Support Systems 43 (3) (2007) 952–967.
complexity in its modeling and representation, however, the actual [12] Available at http://dictionary.oed.com/.
concept of adjusted resilience is relatively simple and straightforward. [13] M. Randles, D. Lamb, E. Odat, A. Taleb-Bendiab, Distributed redundancy and
robustness in complex systems, Journal of Computer and System Sciences (in
As such, we feel that it has the opportunity to play an important role
press), doi:10.1016/j.jcss.2010.01.008.
in improving the ability of individuals and organizations to assess [14] A. Rose, Defining and measuring economic resilience to disasters, Disaster
potential disaster resilience, and thus to make decisions that can help Prevention and Management 13 (4) (2004) 307–314.
[15] A. Rose, G. Adosu, S.Y. Liao, Business interruption impacts on the electric power
improve their protection against and recovery from the occurrence of
system resilience to a total blackout of a terrorist attack of Los Angeles: customer
a disaster event. resilience to a total blackout, Risk Analysis 27 (3) (2007) 513–531.
[16] T.L. Saaty, How to make a decision—the analytic hierarchy process, European
Journal of Operational Research 48 (1) (1990) 9–26.
[17] M. Shinozuka, S.E. Chang, T.-C. Cheng, M. Feng, T.D. O'Rourke, M.A. Saadeghvaziri,
References X. Dong, X. Jin, Y. Wang, P. Shi, Resilience of integrated power and water systems,
in: MCEER (Ed.), MCEER Research Progress and Accomplishments: 2003–2004,
[1] R. Bea, I. Mitroff, D. Farber, H. Foster, K. Roberts, A new approach to risk: the 2004, pp. 65–86, Buffalo, NY.
implications of E3, Risk Management 11 (1) (2009) 30–43. [18] D.E. Snediker, A.T. Murray, T.C. Matisziw, Decision support for network disruption
[2] M. Bruneau, A. Reinhorn, Seismic resilience of communities-conceptualization mitigation, Decision Support Systems 44 (2008) 954–969.
and operationalization, Proceedings, International Workshop on Performance- [19] K. Tierney, M. Bruneau, Conceptualizing and measuring resilience: a key to
based Seismic-Design, Bled, Slovenia, June 28–July 1, 2004, Bled, Slovenia. disaster loss reduction, TR News, 2007, pp. 14–17.
[3] M. Bruneau, A. Reinhorn, Exploring the concept of seismic resilience for acute care [20] H. Zhou, J. Wang, J. Wan, H. Jia, Resilience to natural hazards: a geographic
facilities, Earthquake Spectra 23 (2007) 41. perspective, Natural Hazards 53 (1) (2010) 21–41.
[4] M. Bruneau, S.E. Chang, R.T. Eguchi, G.C. Lee, T.D. O'Rourke, A.M. Reinhorn, M. [21] C.W. Zobel, Comparative visualization of predicted disaster resilience, Proceed-
Shinozuka, K. Tierney, W.A. Wallace, D. von Winterfeldt, A framework to ings of the 7th International ISCRAM Conference., 2010, Seattle, WA.
quantitatively assess and enhance the seismic resilience of communities,
Earthquake Spectra 19 (4) (2003) 733–752. Christopher W. Zobel is an associate professor of Business Information Technology
[5] S.E. Chang, M. Shinozuka, Measuring improvements in the disaster resilience of at Virginia Tech. He received the Ph.D in Systems Engineering from the University of
communities, Earthquake Spectra 20 (3) (2004) 739–755. Virginia, an M.S. in Mathematics from the University of North Carolina at Chapel Hill,
[6] G. Cimellaro, A. Reinhorn, M. Bruneau, Seismic resilience of a hospital system, and a B.A. in Mathematics from Colgate University. His primary research interests
Structure and Infrastructure Engineering 6 (1) (2010) 127–144. are in the area of intelligent decision support systems and knowledge management
[7] S. Cutter, L. Barnes, M. Berry, C. Burton, E. Evans, E. Tate, J. Webb, A place-based for multi-organizational networks, particularly in the realm of disaster operations
model for understanding community resilience to natural disasters, Global management. Dr. Zobel has published articles in Decision Sciences, the International
Environmental Change 18 (4) (2008) 598–606. Journal of Production Research, and the European Journal of Operational Research,
[8] R. Klein, R. Nicholls, F. Thomalla, Resilience to natural hazards: how useful is this among others, and he is a member of the Decision Sciences Institute (DSI), the
concept? Global Environmental Change Part B: Environmental Hazards 5 (1–2) Institute for Operations Research and the Management Sciences (INFORMS), and the
(2003) 35–45. International Association for the Study of Information Systems for Crisis Response
[9] S. Manyena, The concept of resilience revisited, Disasters 30 (4) (2006) 434–450. and Management (ISCRAM).

You might also like