Accepted Manuscript

Robust Goal Programming for Multi-objective Optimization of

Data-driven Problems: a Use Case for the United States
Transportation Command’s Liner Rate Setting Problem

Robert W. Hanks , Brian J. Lunday , Jeffery D. Weir

PII: S0305-0483(17)30687-4
DOI: https://doi.org/10.1016/j.omega.2018.10.013
Reference: OME 1983

To appear in: Omega

Received date: 18 July 2017

Accepted date: 18 October 2018

Please cite this article as: Robert W. Hanks , Brian J. Lunday , Jeffery D. Weir , Robust Goal
Programming for Multi-objective Optimization of Data-driven Problems: a Use Case for the
United States Transportation Command’s Liner Rate Setting Problem, Omega (2018), doi:
https://doi.org/10.1016/j.omega.2018.10.013

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Highlights
 Demonstrates the applicability of robust goal programming and specifically, the robust goal
programming model that utilizes cardinality-constrained robustness via 𝐿𝐿2-norm uncertainty
sets.
 Offers a new methodology that leverages robust goal programming applied in concert with a
mapping methodology for future transportation rate-related problems, setting robust shipping
rates that address uncertainty in the cost-related data, while seeking to meet various goals.

T
IP
CR
US
AN
M
ED
PT
CE
AC

Preprint submitted to Omega– July 2017

 ACCEPTED MANUSCRIPT

Robust Goal Programming for Multi-objective Optimization of Data-

driven Problems: a Use Case for the United States Transportation
Command’s Liner Rate Setting Problem

Robert W. Hanksa, Brian J. Lundaya, Jeffery D. Weira

T
IP
Abstract

Robust goal programming (RGP) is a recently developed, powerful new optimization modeling technique

CR
that conjoins two widely accepted operations research disciplines: robust optimization (RO) and goal
programming (GP). In lieu of applying a probability distribution over possible outcomes, an approach

US
considered by stochastic programming, RO utilizes uncertainty sets to account for data uncertainty. This
characteristic of RO is an important attribute because identifying such a probability distribution is
challenging, at best. Given this RO context, RGP additionally incorporates GP, traditionally a
AN
deterministic procedure, to address optimization problems having multiple objectives. As such, RGP has
potential to help address a wide array of data-driven applications, ranging from financial management to
engineering design.
M

As a motivating use case for the utility of an RGP approach, this paper demonstrates the
applicability of RGP by way of the data-driven United States Transportation Command
ED

(USTRANSCOM) liner rate setting problem. USTRANSCOM is responsible for the technical direction
and supervision of over $7 billion [1] of annual passenger, cargo, mobility, and personal property
movements in support of the Department of Defense (DoD). Transporting people and material with both
PT

organic and contracted assets, USTRANSCOM supports DoD organizations and agencies on a
reimbursable basis, annually setting and charging rates for air and liner (i.e., sea) transport for their
CE

customers and reimbursing the transportation providers accordingly. The Cost Recovery Branch within
TCJ8, the Financial Management and Program Analysis staff organization for USTRANSCOM, annually
AC

sets liner shipping rates specific to each combination of origin, destination, commodity type, booking
terms, and container size for the upcoming fiscal year (FY). As a government entity, USTRANSCOM
seeks to neither make a profit nor operate at a loss in any given FY. The current rate setting methodology
a
R. Hanks (corresponding author), B. Lunday, J. Weir
Air Force Institute of Technology, Department of Operational Sciences, 2950 Hobson Way, Wright-Patterson AFB,
OH 45433, United States
Email addresses: robert.hanks.5@us.af.mil (R. Hanks), brian.lunday@afit.edu (B. Lunday), jeffery.weir@afit.edu (J.
Weir)

2
 ACCEPTED MANUSCRIPT

assumes existing data is deterministic, resulting in process inaccuracies that contribute to unexpected
surpluses or deficits each FY. Moreover, the current method fails to consider an additional
USTRANSCOM objective: meeting customer’s expectations that liner rates will change annually in
accordance with industry-specific inflation. Considering the different goals and inherent parametric
variance, the use case herein incorporates a decision maker’s risk preference regarding parametric
variability via a priori analysis to inform RGP techniques and improve the USTRANSCOM liner rate
setting process.

T
Keywords: Robust optimization; Goal programming; Robust goal programming; Decision making under

IP
risk

CR
___________________________________________________________________________________

1 Robust goal programming

US
Robust goal programming (RGP) is a recently developed sub-discipline within optimization that
combines uncertainty modeling via robust optimization (RO) with the widely applied multi-objective
AN
optimization technique goal programming (GP). As an alternative to RO, stochastic programming (SP),
utilizes probability distributions to consider uncertainty in data [2]; however, the increasing demand to
solve data-driven problems, wherein optimization problems are increasingly complex in terms of size and
M

dimensionality [3], the task of identifying a probability distribution over parametric data to use SP is often
unreasonable [4]. As a result, RO has ascended as a modeling technique for data-driven problems having
ED

either uncertain or highly variable parameters, and the utility of RGP is increasing correspondingly.
Building upon RGP modeling developments [5] as well as work by Hanks et al. [6] that informs a method
for parameterizing such RGP models for a given decision makers (DM) risk preference (i.e., the degree of
PT

risk aversion) within the context of data-driven problems, the research herein motivates the use of RGP
via a financial use case.
CE

1.1 Literature review

RGP is a relatively new area in the field of RO that was first introduced by Kuchta [5] to combine
AC

cardinality-constrained robustness and interval-based uncertainty with popular GP techniques. When

compared to Soyster’s [7] strict robustness, cardinality-constrained robustness [4,8] is a less conservative
RO model predicated on the idea that not all coefficients subject to parametric uncertainty will realize
their worst-case deviational values. More explicitly, consider a robustness measure , for every
constraint of a linear program, that regulates the required level of conservatism and takes
on values in the interval , | |-, wherein is the set of coefficients that are vulnerable to
parametric uncertainty in the interval [ ] (i.e., interval-based uncertainty set), and

3
 ACCEPTED MANUSCRIPT

designates the maximum allowable deviation of parameter . Bertsimas and Sim [4,8] justified
that it is not likely that every take on uncertainty. In the event is not integer-valued, up to
⌊ ⌋ of the coefficients may deviate by , and exactly one coefficient will deviate by at most
( ⌊ ⌋) . Conversely, when the -values are discrete (i.e. integer-valued), every coefficient that
does deviate from their nominal value may do so at a maximum allowable value.
Applying parametric analysis for varying values of in combination with an a posteriori

T
assessment of a DM’s risk preference, Hanks et al. [9] extended Kuchta’s [5] RGP model via cardinality-
constrained robustness and norm-based uncertainty sets. The model that applies the 𝐿 -norm to account

IP
for data uncertainty is presented below [9], and an interested reader is referred to that work for alternative

CR
RGP models that utilize either strict robustness or ellipsoidal uncertainty sets.

∑ (1)

US
AN
∑ (2)
M

√ ∑( ) ( ) | | ⌊ ⌋ (3)
ED

(4)

(5)
PT

The objective function (1) minimizes the sum of positive deviations from goals , and Constraint (2)
calculates the positive and negative deviations (i.e., and ) from each goal, where the total induced
CE

variability is represented by for each goal. Constraint (3) applies 𝐿 -norm uncertainty sets to the
possible sets of coefficients subject to parametric uncertainty (i.e., | | ⌊ ⌋ ), where
AC

the least upper bound (i.e., infimum) penalty, is calculated for every combination of

√∑ ( ) ( ) , and wherein represents the remainder value when is not integer-

valued (i.e., ⌊ ⌋). Constraint (4) characterizes additional constraints on the decision space that
are neither related to the DM’s goals nor considered with respect to possible parametric uncertainty.
Finally, Constraint (5) ensures all decision variables are non-negative.

4
 ACCEPTED MANUSCRIPT

Although RGP is a relatively new modeling concept, its potential for practical use to address
data-driven problems is extensive. To date, data-driven applications of RGP in the literature include
portfolio selection [10,11], the capital budgeting problem [12], supply chain network design [13], supply
chain electric power [14], and intermodal routing problems [15]. RGP has also been discussed in other
forums regarding its potential use for multi-criteria linear programming [16] as well as compromise
programming [17].
There have been a few persistent shortcomings related to the application of RGP models in the

T
literature to date. First, the RGP models apply parametric analysis to consider possible combinations of

IP
-values, which can be computationally expensive when the number of decision variables or the number
of goal-related constraints are high. Second, the risk preference of the DM is not taken into consideration

CR
until after-the-fact (i.e., a posteriori). Consequently, the DM must decide which solution set is best
among many alternative optimal solutions, each of which represents a different risk attitude; selecting in

US
this manner is a daunting-if-not-impossible task for a DM, and the presentation of the myriad of solution
options can potentially bias the identification of their true risk preference. Lastly, when
current RGP methodologies are intractable because they will yield an infinite number of combinations of
the
AN
parametric variations.
To further develop the practical nature of RGP, Hanks et al. [6] account for a DM’s risk
preference via a priori analysis using a variety of rigorous risk elicitation methods that can then be
M

mapped to an RGP risk parameter (i.e., ). More specifically, the authors present three theorems that
enable a mathematical mapping of risk attitude to RGP model parameters by utilizing piecewise
ED

differential equation models. Hanks et al. [6] develop mappings for several risk elicitation techniques:
the bisection, Holt-Laury (HL), and Eckel-Grossman (EG) methods, as well as the Bomb Risk Elicitation
Task (BRET). The mapping used when eliciting risk via the bisection, HL, and EG methods is presented
PT

in Equation (6) whereas the mapping applied in conjunction with the BRET is offered in Equation (7).

( )
CE

( | ) { ( )
(6)
AC

( )
(7)
( | ) { ( )

By eliciting the DM’s risk a priori, Equations (6) or (7), as appropriate, one can generate a
corresponding risk parameter as a function of the quantitative risk attitude test measurement , wherein
denotes the assumed risk neutral point in the RGP realm and is the -value corresponding to a risk
neutral attitude. In doing so, a robust solution is identified using the existing RGP methods that will not

5
 ACCEPTED MANUSCRIPT

only provide the DM with a single solution, the solution set will also correspond to the DM’s measured
risk preference. As an aside, it is important to note that, within the RGP context, risk represents the
willingness of a DM to ignore the variability in data used to inform the parameters corresponding to their
given objectives. To that end, a DM’s risk preference can be specific to, and parameterized for, each of
the DM’s objectives. Moreover, either two-sided risk (i.e., net deviation from a goal) or one-sided risk
(i.e., semi-variance) can be considered for a given goal.
It should be noted that the aforementioned formulation described (1)-(5) as well as Equations (6) and

T
(7) are previously presented in Hanks et al. [9] and Hanks et al. [6], respectively, and have not been

IP
previously applied to any realistically-sized problem instance. In the following sections, these RGP-oriented
formulations and methods are applied to the USTRANSCOM liner rate setting problem as a use case to

CR
motivate the application of RGP to other problems.

1.2 Research Contributions

US
This paper demonstrates the applicability of the RGP methods in the literature, specifically the cardinality
constrained via 𝐿 -norm uncertainty sets model as set forth by Hanks et al. [9]. Moreover, via a
AN
demonstrated use case, this work illustrates the detailed application of RGP as a modeling methodology
to solve a transportation shipping rate-setting problem, wherein the solution identifies robust shipping
rates that address uncertainty in the cost-related data while seeking to meet two goals of the shipping
M

contract managing agency. By demonstrating the application and positive outcomes of RGP for the
agency, this work seeks to motivate the consideration and application of RGP to a yet wider array of data-
ED

driven optimization problems having both parametric variability and competing objectives.
This paper is organized as follows. Section 2 summarizes the USTRANSCOM liner rate setting
problem and provides a brief synopsis of liner rate setting literature. Section 3 presents a
PT

USTRANSCOM-specific RGP formulation, given historic liner cost-data and selected sources of
uncertainty. Considering a range of DM risk attitudes, Section 4 applies the model for setting FY16 rates,
CE

comparing the outcomes to currently utilized rate setting methods, and Section 5 concludes the work and
identifies possible extensions for future research.
AC

2 A use case for RGP: the USTRANSCOM liner rate setting problem

USTRANSCOM is responsible for the technical direction and supervision for all DoD passenger,
cargo, mobility, and personal property movements, wherein they provide global mobility solutions during
both peacetime and wartime environments [18]. As a mechanism to support reimbursable financial
transactions for shipping, USTRANSCOM maintains a Transportation Working Capital Fund (TWCF) to
service payments. For transportation services provided, USTRANSCOM pays civilian shippers from the

6
 ACCEPTED MANUSCRIPT

TWCF accordingly to rates for each shipment that may vary by shipper, services provided, and the
specific contract. The TWCF is reimbursed by rates charged by USTRANSCOM to supported DoD
customers. However, to simplify the DoD customers’ budget planning processes, improve financial
predictability, and operate in a manner that is legally impartial with respect to shipping providers,
USTRANSCOM does not charge DoD with the exact contracted rate for each shipment; instead, its Cost
Recovery Branch (hereafter referred to as TCJ8, the abbreviation for its parent organization) seeks to
recoup expenses in aggregate for a given fiscal year (FY) by setting and leveraging a set of shipping rates

T
specific to origin, destination, commodity type, booking terms, and container size, but agnostic to shipper

IP
and specific contract. Over the past ten years [19], each of these rates has been calculated annually using
the weighted average, inflation-adjusted historical cost over recent FYs, adjusted to recoup overhead costs

CR
in proportion to the projected shipping workload for the upcoming FY.
If the rates set and imposed by TCJ8 are not accurate with respect to the actual shipping costs

US
paid to the transportation provider in the following FY, plus overhead recoupment, a surplus or deficit
results. As such, maintaining a positive balance in the TWCF that carries over between FYs serves to
hedge against the risk of not being able to pay shippers if USTRANSCOM recoups insufficient costs
AN
from DoD customers and operates at a deficit in any given year. However, it is also not desirable to have
a large TWCF balance either. As such, USTRANSCOM seeks to provide the shipment of DoD goods
and services while incurring neither a profit nor a deficit. Whereas operating at a deficit may reduce the
M

balance of the TWCF to a level that does not support reimbursable operations, operating at a profit would
be an irresponsible use of taxpayer resources. Beyond the goal of operating at cost, USTRANSCOM also
ED

seeks to manage customer expectations. Unconcerned with USTRANSCOM’s internal processes to

recoup overhead costs and manage the TWCF balance, DoD customers expect their rates to change from
year-to-year in accordance with industry-specific inflation rates. Should USTRANSCOM’s processes set
PT

rates that deviate notably from customer expectations, or if USTRANSCOM accumulates a high TWCF
balance, DoD customers may seek to better meet their expectations or save costs, respectively, by
CE

utilizing third party logistics (3PL’s) directly for their shipment needs, albeit while simultaneously
incurring a contracting burden. Such an outcome would entail a failure of USTRANSCOM to provide
AC

effective support to its customers.

Rate setting is a complex process that takes a notable level of detailed analysis to complete.
However, most of the literature related to rate setting is proprietary because 3PLs seek a competitive
advantage in the market and, consequently, useful published information is limited. However, a few
works pertaining to liner rate setting methodologies, policies, and theoretical inferences do influence this
study of the USTRANSCOM liner rate setting problem. Shneerson [20] extensively discussed liner rates
and made conclusions that liner freight rates can be explained via pricing and demand. The author

7
 ACCEPTED MANUSCRIPT

identified the most important factors influencing liner rates to be the ratio of volume to weight (i.e.,
stowage factor) and the unit value of the commodities being shipped. At the time, it was common
practice to label cost as the most influential factor when setting liner rates, but Shneerson showed that the
stowage factor is the most influential element. However, in a later study done by Brooks and Button [21],
it is demonstrated that stowage factors can sometimes be a misleading variable when determining liner
rates. The authors identified that the type of customer (e.g., freight forwarder, consignee, or shipper) can
vary the requirement related to liner rate setting. Moreover, the authors discussed extensively how the

T
aggregation of rates can yield inaccurate rates, as the direction of a shipping route directly influences the

IP
cost of transport. Also closely related to this study is work by Skinner et al. [22], wherein the authors
provided econometric modeling results showing the amount of variability in rate setting can be attributed

CR
as follows: 75-80% to distance (or OD pair), 8-10% to geography, and 1-3% to volume. Beyond these
works, there are several recent studies on shipping costs and rate setting [23,24,25], but none are

2.1 Current USTRANSCOM liner rate setting process US

sufficiently related to inform the specific problem considered herein.
AN
Any DoD liner shipment of goods or services has fixed costs (e.g., maintenance, government employee
wages) and variable costs (e.g., tonnage, fuel based on distance traveled). The fixed (or overhead) costs
are accounted for via a net, relative adjustment to every inflation-adjusted average cost (i.e., the
M

accumulated operating result (AOR)), whereas the variable costs are calculated by considering 2-1/2
years-worth of data to calculate a weighted average cost. These variable costs depend on the distance
ED

between the origin and destination (i.e., OD pair), commodity type, booking terms, and container size.
TCJ8 is responsible for setting transportation rates for their DoD customers, and the rates must be
set prior to each FY. Unique to the liner rate setting problem for USTRANSCOM is the following
PT

characteristic. Although the rates USTRANSCOM pays for transportation services may vary over the set
of military or contracted civilian providers, even across civilian providers operating under different multi-
CE

year contracts, USTRANSCOM is mandated to set and charge a single shipping rate to DoD customers
that does not vary by the servicing provider. To wit, USTRANSCOM cannot legally charge different
rates that would exhibit bias among providers. A benefit of this structure is greater budget predictability
AC

for DoD customers, albeit with greater budget uncertainty for USTRANSCOM, reinforcing the need for
the TWCF. The primary goal when setting these rates is to obtain a net operating result (NOR) of zero
profit, where the NOR is simply the rates charged to DoD customers by USTRANSCOM minus the costs
spent by USTRANSCOM to reimburse government or contracted transportation agencies variable costs as
well as fixed costs to administer the process (i.e., overhead costs). In other words, obtaining a $0 NOR
will result in a steady TWCF balance.

8
 ACCEPTED MANUSCRIPT

TCJ8 configures liner rates using a 2-1/2-year, weighted average from the previous FY’s costs.
Each weighted average cost is then adjusted by a refresh rate and an AOR factor to estimate the next FY’s
rate. The refresh rate considers yearly inflation that is predetermined and written into a contract based on
evaluation of the most utilized carrier’s historical data. The AOR factor considers the amount of surplus
or deficit (i.e., NOR) for a two-year period and offsets that same amount when setting future rates to
recover overhead costs. The AOR factor adjusts rates to both recoup overhead costs and offset the net
gains (or losses) made over the previous two years of operations. In general, the liner rate being

T
calculated [26,27] is a base rate that accounts for most of the shipping, whereas the AOR factor accounts

IP
for both fixed personnel overhead as well as geographic aggregation of demands to reduce the amount of
rates within a given FY.

CR
A widely-used and accepted method for considering overhead costs is activity-based costing (ABC)
[28]. In fact, TCJ8 has utilized ABC in the past when considering overhead costs for selected accessorial

US
charges such as special port handling [29]. Yet, ABC for the USTRANSCOM liner rate setting problem
is difficult to implement given the tensions between (1) the need to provide rates to DoD customers that
are constant for a given FY and do not vary by the shipping provider and, (2) the significant number of
AN
combinations of activities to cost when considering rates in the aggregate and under uncertain conditions
when dealing with multiple shipping providers. To date, TCJ8 has not yet identified an acceptable
compromise between these tensions. Therefore, for the purposes of this study, we do not implement
M

ABC, but an interested reader is referred to the seminal work by Kaplan and Cooper [28] for an overview
on the method, or to more recent, related application studies by Kaplan et al. [30], Tsai et al. [31], and
ED

Oker and Adiguzel [32].

There exist two concerns of note with the current USTRANSCOM liner rate setting process. First,
when considering 2-1/2 years of historical cost data to set each rate, TCJ8 uses a point estimate for the
PT

expected cost, ignoring variability in the historical data. Identifying probability distributions over
intervals of uncertainty for historical cost data is not practical [4,8]; thus, accounting for the variability
CE

inherent in the rates via uncertainty sets renders RO preferable to SP. Second, the current rate setting
process seeks to manage the TWCF balance, but it does not address any concern with meeting customer
AC

expectations (i.e., adjusting rates annually in accordance with industry specific inflation). Because
USTRANSCOM has multiple goals and associated target values, GP is an appropriate multi-objective
approach. Combining the needs to both account for data uncertainty using well-known RO techniques
and address more than one rate setting objective using GP, the USTRANSCOM liner rate setting problem
is well suited for RGP.

9
 ACCEPTED MANUSCRIPT

3 Current and RGP-based solution methodologies

The USTRANSCOM liner rate setting problem requires elaboration within the context of an optimization
modeling framework. Section 3.1 presents the notation utilized in the USTRANSCOM liner rate setting
problem including various sets, parameters, and decision variables. The current methodology for setting
yearly liner rates (as implemented by TCJ8) is explained in Section 3.2. In Section 3.3, important model
assumptions are presented, especially with regard to parametric variability within the model. The section

T
concludes by setting forth a variant of the RGP model described in (1)-(5) as it pertains to the
USTRANSCOM liner rate setting problem.

IP
The number of liner rates under consideration in this process is notable. Depending on the ability

CR
of historical cost records, liner rates are specific to each combination of 59 possible origination zones (o),
59 possible destination zones (d), 10 types of commodity classification categories (c), nine distinct
booking terms that stipulate different combinations of accessorial services at the origin or destination (b),

US
and three different categories of container sizes (i.e., breakbulk, a container less than 40 feet, and a
container greater than or equal to 40 feet) (s). Considering all possible combinations of the
AN
aforementioned zones, categories, terms, and sizes, TCJ8 seeks to leverage data to inform 939,870 liner
rates annually. Hereafter, we refer to a specific combination of these options as an odcbs-combination,
but we refrain from indexing all of the following, relevant notation and calculation specific to a given rate
M

with ―odcbs‖, as the result would be cumbersome to read. Fortunately, the liner rate setting problem is
separable by odcbs-combination, as will be discussed in Section 3.3.1, so the notational convenience does
ED

not detract from the precision of the ensuing exposition.

3.1 Notation
PT

Prior to setting forth the various formulations we examine, it is necessary to introduce the associated sets,
parameters, decision variables, and functions. The notation is defined as follows:
CE

Sets

 * + set of all USTRANSCOM liner rate setting goals, indexed as .

 * +: set of all FY’s, indexed as
AC

Parameters

 : number of observations/shipments for a given odcbs-combination during FY .

 : the amount of measurement tons (mtons) for each observation/shipment for a
given odcbs-combination during FY .

10
 ACCEPTED MANUSCRIPT

 : cost per mton for each observation/shipment (does not consider refresh rates)
corresponding to shipments for a given odcbs-combination during FY .
 : weighted average, refresh-rate adjusted cost to ship an mton corresponding to shipments
for a given odcbs-combination during FY .
 : refresh rate adjustment for any given rate set in FY .
 : AOR/overhead adjustment for any given rate set in FY
 * + for all shipments for a given odcbs-combination.

T
: weight attributed to each goal,
 * + for all shipments for a given odcbs-combination

IP
: number of products for each goal,
subject to deviation (uncertainty), where , - (due to the separable nature of each odcbs-

CR
combination specific rate).
 : the allowable amount of parametric uncertainty observed in the refresh rate during FY .
 : the allowable amount of parametric uncertainty observed in the AOR factor during FY


.
US
: the allowable amount of parametric uncertainty witnessed in corresponding to a
AN
shipments for a given odcbs-combination during FY .
 : the amount of parametric uncertainty observed in during FY

 : the amount of parametric uncertainty observed in during FY .

M

 : the combined amount of parametric uncertainty observed in the coefficient pertaining to goal
* + corresponding to a shipments for a given odcbs-combination during FY .
ED

Decision Variables

 : rate ($/metric ton) for DoD customers for a given odcbs-combination during FY .
PT

 : penalty associated to a shipment for a given odcbs-combination during FY pertaining to

each goal .
CE

 : amount of deviation above a target value for corresponding to a shipment for a given
odcbs-combination during FY pertaining to each goal .
 : amount of deviation below a target value for corresponding to a shipment for a given
AC

odcbs-combination during FY pertaining to each goal .

 : amount of deviation above a target value for corresponding to a shipment for a given
odcbs-combination during FY pertaining to each goal .
 : amount of deviation below a target value for corresponding to a shipment for a given
odcbs-combination during FY pertaining to each goal .

11
 ACCEPTED MANUSCRIPT

Functions

 ∑ ( )( ) ∑ ( )

∑ , wherein represents the total inflation-adjusted costs for all shipments

for a given odcbs-combination over FY’s and .

 ∑ ∑ ∑ , wherein calculates the mtons shipped

for a given odcbs-combination over FY’s and .

T
 . / ∑ ( )( ) ( ) ∑ ( ) ( )

IP
∑ ( ), wherein . / represents the total inflation-adjusted inverse cost for all

CR
shipments f for a given odcbs-combination over FY’s and .

3.2 Calculation method for the current rate setting methodology

US
As a means of comparison, it is important to understand how TCJ8 currently calculates yearly liner rates
for USTRANSCOM. To include more data and construct a better point estimate, TCJ8 takes into
AN
consideration the cost of every movement over the previous two years as well as the first six months of
the current FY, and it inflation-adjusts each movement into current year dollars. It should be reiterated
that the normalization of the temporal, historical data to a common FY is conducted to compute specific,
M

USTRANSCOM parameters. Whereas the actual normalization process may vary with a particular
application, data normalization for temporal data is a common technique which can be paired with the
ED

RGP methodology for any number of relevant problems. Using these inflation-adjusted shipment costs,
TCJ8 calculates the weighted average cost (WAC) of each odcbs-combination (for which there exists cost
record data) over the past 2-1/2 years, as per Equation (8).
PT

(8)
CE

The WAC is then adjusted by both the refresh rate and the AOR factor to set the upcoming FY’s liner
rate, as per Equation (9).
AC

( )( ) (9)

3.3 Applying RGP to the liner rate setting problem

The RGP formulation presented in this section is a modified version of the RGP model described above in
(1)-(5) and as set forth by Hanks et al. [9]. However, it is appropriate to first review the relevant RGP
modeling assumptions, followed by discussing the parametric variability observed in the USTRANSCOM

12
 ACCEPTED MANUSCRIPT

liner rate setting problem and, finally, presenting an RGP model to assist TCJ8 in their liner rate setting
endeavors.

3.3.1 RGP-specific modeling assumptions

In applying the RGP methodology to the USTRANSCOM rate setting problem, several assumptions are
required. First, we assume that all -combinations are subject to parametric uncertainty, which is

T
verifiably correct when examining cost data. Second, the rate setting process is separable by -
combination. This assumption is valid because, if each rate is robust and contributes to neither a deficit

IP
nor a surplus to the TWCF, then the aggregation of rates will share the same attributes. Thus , -

CR
for our modeling implementation. In turn, the rate setting process being separable alleviates the need for
the remainder value as described in Section 1.1. It is further assumed that corresponds to a
risk-averse DM, whereas and respectively denote a risk-neutral and risk-seeking DM.

US
This assumption relates higher levels of risk to a willingness to ignore the variability present in the data-
driven modeling parameters. As a result, direct implementation of a risk-elicitation method as described
AN
in Section 1.1, is not required for this demonstration. We further assume the DM’s risk parameter is the
same for every -combination during FY and, in this study, that each of USTRANSCOM’s
two goals are both equally important and well scaled, allowing for a cardinality weighting of deviations
M

from goals within the RGP objective function. The refresh rate and AOR parametric uncertainties each
are assumed to be normally distributed, denoted herein as ( ) and ( ),
ED

respectively. Finally, for the data examined, we assume the current NOR is equal to zero, as a more
detailed result is not known for the historical data used in this study.
PT

3.3.2 RGP goal constraints

The first goal for the USTRANSCOM liner rate setting problem is to minimize deviation from $0 NOR.
CE

Alternatively stated, TCJ8 seeks to set liner rates such that next FY’s rate is as close to the previous FY’s
cost-adjusted WAC-based rate as possible, per Equation (9). However, the RGP model described in (1)-
(5) demands that parametric uncertainty be considered only as a cost coefficient (i.e., the constraints are
AC

transformed into data-driven parameters within the RGP format, resulting in coefficients for decision
variables). Thus, because the uncertainty is present in the parameters , , and , the NOR
goal constraint seen in Equation (9) is transformed as follows:

( )( )( ) (10)

13
 ACCEPTED MANUSCRIPT

Conducting this transformation does require more work when calculating certain parameters, specifically
. Calculating the and then taking the reciprocal is merely an approximation, whereas the

calculation set forth in Equation (11) is more appropriate, where each is considered for a 2-1/2-year

period to calculate . / .

. / (11)

T
IP
Similarly, the second goal of interest to USTRANSCOM and TCJ8 is to set liner rates that are

CR
robust from year-to-year, setting the rate as close to the cost adjusted rate from last year.

( ) (12)

US
To ensure the parameter subject to uncertainty is represented in the coefficient (i.e., the constraint is
transformed into a data-driven parameter), Equation (12) is rewritten as:
AN
( ) (13)
M

3.3.3 Addressing parametric variability within the RGP model

In addition to these goal constraints, there are penalty constraints that must be considered to account for
ED

possible parametric variability. Within the context of this use case, if the data-driven parameters are
included in the denominator of a simple fraction, a weighted standard deviation is utilized to calculate the
parametric variability. If the data-driven parameter is a part of a more complicated functional form, a
PT

Monte Carlo simulation is leveraged to identify the mean and standard deviation for the function. While
these methods are utilized to parameterize this USTRANSCOM use case, we recognized that alternative
CE

methods to parameterize an RGP instance may be considered.

It is assumed that there are three sources of variability present in the USTRANSCOM liner rate
setting problem. The first parameter subject to uncertainty is the refresh rate such that ( ),
AC

where is the average refresh rate over the past three years and is the corresponding standard
deviation over the same three years. Given these values, a Monte Carlo simulation is applied for 10,000
replications to compute values for , from which the average and standard deviation are computed.

The second parameter that is considered to model uncertainty is the AOR factor. The parametric
variability witnessed in the AOR factor is calculated in the exact same fashion as the refresh rate, wherein
( ), is the average AOR during the previous three years, and is the

14
 ACCEPTED MANUSCRIPT

analogous standard deviation over the same three years. A Monte Carlo simulation of 10,000 replications
is conducted to identify the mean and standard deviation for of values for .

Finally, the third factor contributing to variability in the RGP model is considered in the

calculation. Unlike the refresh rate and AOR factor, is dependent on each odcbs-combination.

Because the calculation considers a weighted average, the weighted standard deviation is

T
computed to account for the variability witnessed in , per Equation (14).

IP
∑ ( )

CR
√ (14)
( )∑

)( US
Although each source of uncertainty is important, the combined uncertainty comprised in the

coefficient of the NOR goal (i.e., ( ). /) merits additional discussion. To identify

AN
the combined effect of the various uncertainties, the variance of multiple random variables is considered
using the formula described in Equation (15).

( ) , - ( , -) ( ) ( ) ( ), ( )- ( ), ( )- (15)
M

However, Equation (15) only considers the variance of two random variables, but we are interested in the
ED

combined variance of three random variables. As a result, first calculate V (( )( )) using

Equation (15) and define the following:

PT

( )( ) (16)
CE

(17)
AC

Because the ( ) (( )( )), three variables are condensed into two wherein we can

apply Equation (15) to get the parametric variability coinciding with the coefficient of the NOR goal, per
Equation (18).

√ (( )( )( )) (18)

15
 ACCEPTED MANUSCRIPT

Addressing the parametric variability witnessed in the second goals’ coefficient is a much simpler
process. Because the second goal depends entirely on the allowable variability of the refresh rate, the
parametric variability pertaining to the coefficient of the yearly, robust rates’ goal is outlined in Equation
(19) below.

√ ( ) (19)

T
Of note, is equivalent to .

IP
CR
3.3.4 USTRANSCOM liner rate setting RGP model formulation

The RGP method that enforces cardinality-constrained robustness via 𝐿 -norm uncertainty sets described
in (1)-(5) considers one-sided variability. For the USTRANSCOM liner rate setting problem, two-way

USTRANSCOM liner rate setting problem and the fact that

US
deviation must be taken into consideration. Furthermore, because of the separable nature of the
, -, the RGP model discussed in (1)-
AN
(5) is actually equivalent to the cardinality-constrained RGP model that uses 𝐿 -norm uncertainty sets [9].
This equivalency is demonstrated below, along with the final RGP model that is applied to solve the
USTRANSCOM liner rate setting problem for each odcbs-combination.
M

( ) ( ) (20)
ED

( )( )( ) (21)
PT

( ) (22)

√( ) (23)
CE

√( ) (24)
AC

(25)
(26)

When referencing Constraints (23) and (24), it is evident that the square root of a non-negative squared

term .√( ) / is the term itself ( ) and is therefore equivalent to the cardinality-constrained model that

uses 𝐿 -norm uncertainty sets. Moreover, the absolute value operator in the norm is not necessary, as the
elements of the argument are each constrained to be non-negative. Furthermore, it should be noted that

16
 ACCEPTED MANUSCRIPT

Constraint (25) is included for additional side-constraints in the general-sense; however, the
USTRANSCOM liner rate setting problem does not currently require side constraints. Finally, because
the USTRANSCOM liner rate setting problem considers both positive and negative deviation from goals,
and because we assume , the below RGP formulation is adopted as a specific formulation to
account for two-way deviation.

(27)

T
IP
( )( )( ) (28)

CR
( )( )( ) (29)

( )
US (30)
AN
( ) (31)

(32)
M

(33)

(34)
ED

4 Testing, results, and analysis for the USTRANSCOM liner rate setting problem
PT

The RGP model in (27)-(34) is applied to solve the USTRANSCOM liner rate setting problem for FY16
using FY12-FY15 cost data and evaluated using FY16 cost data. For comparison, the WAC method is
CE

applied and evaluated for FY16 using the same data. Before any mention of the results, however, caveats
are offered regarding the shortcomings of the FY12-FY16 cost data. Of note, all testing is performed
utilizing Microsoft Visual Basic and Applications [33] to manage the data while invoking the commercial
AC

solver LINGO, version 11.0 [34] to solve the associated RGP formulation instances. Although specific
run times are not reported due to the batching of rate calculations, the entire solution procedure took
approximately 12 hours of run time on an Intel(R) Xeon CPU E3-540 1225 v3 having a 3.3 GHz
processor and 32 GB of RAM.
There are some notable caveats to the quantitative results of any of the methods considered in this
study, despite a notably illustrative demonstration of the RGP model’s application vis-à-vis the WAC

17
 ACCEPTED MANUSCRIPT

method. While conducting this study, there were substantial issues regarding the data used by TCJ8, as
well as the impact of the AOR factor on the overall rate setting process.
There is substantial consideration related to the Five V’s of the cost record data [35]: volume
(amount of data), velocity (speed at which data is generated), variety (various types of data), value
(possible insights created from data), and finally, veracity (accuracy of data) [36]. In this regard,
challenges discovered in the USTRANSCOM data were two-fold: (1) the veracity of the data, and (2) the
sparsity of the data. It was noticed that there were certain shipments of data that had negative-valued

T
mtons or even zero-valued mtons, but still had a cost associated. Table 1 reports the number of shipments

IP
reported within each FY that had negative-valued or zero-valued mtons, but an associated cost.

CR
Table 1: Veracity Issues in TCJ8 Data

# of shipments with # of shipments with

FY12 0 2
FY13 0 0
FY14
FY15
FY16
0
729
588
US 8
18
8
AN
Even though the percentage of these particular shipments is low relative to the total number of shipments
within a given FY, including this type of data in the calculation of the next FY’s rate for any odcbs-
combination can detrimentally affect the outcomes (i.e., the NOR, robust rates) of the rate setting process.
M

Because of this, the overall precision of the USTRANSCOM data is problematic.

Of equal if not greater importance is the issue of data sparsity. To address this concern, TCJ8
ED

considers 2-1/2 years-worth of refresh-rate adjusted data instead of only current year data to reduce the
likelihood of a rate not having any cost records to inform it. Even with this technique, the amount of cost
data is still insufficient to produce accurate estimates for every liner rate in the next FY. For example, in
PT

calculating the FY16 rates using the WAC and RGP methods, there were 2880 distinct odcbs-
combinations. When considering each observed -combination receipt in FY16, there were 1,641
CE

individual combinations. Yet, of these 1,641 rates, only 1,133 were in common with the 2,880 estimated
rates set forth by the WAC and RGP methods (see Table 2). Consequently, 30.96% of the odcbs-
AC

combinations and 13.68% of the total mtons shipped in FY16 are not being taken into consideration for
the NOR calculation because there is no basis of comparison (i.e., there were 508 observed odcbs-
combinations in FY16 that are not present in FY15). Ultimately, the lack of collective rates year-to-year
is a direct result of sparsity seen in the data, wherein the only way to address this issue is added processes
(e.g., consult a subject matter expert) to the TCJ8 rate setting methodology.

18
 ACCEPTED MANUSCRIPT

Table 2: FY16 RGP Rate versus FY16 Actual Rate Data Comparison
Actual FY16 Data Captured in FY16 RGP Data % Captured in FY16 RGP Data
odcbs-combinations mtons odcbs-combinations mtons odcbs-combinations mtons
1.6K 2.4M 1.1K 2.0M 69% 86.3%

In this regard, according to TCJ8, the FY16 direct costs recovered by liner rates was $599.8 million,
whereas the non-direct (overhead) cost for liner rates was $62.1 million. However, because the RGP
method accounts for approximately 86.32% of the total mtons shipped via liner in FY16, we assume

T
86.32% of the total FY16 liner costs and adjust accordingly, per Table 3. Applying this adjusted price for

IP
this study is essential in reporting a true NOR comparison of the three rate setting methodologies.

Table 3: FY16 Liner Cost Data

CR
Actual FY16 Cost Data Adjusted FY16 Cost Data
Direct Non-Direct Direct Non-Direct
$599.8M $62.1M $517.7M $53.6M
Total $661.9M

US $571.4M

One final topic regarding this study concerns the AOR factor. As it is currently understood, the
AOR factor is dependent upon the total amount of shipments over the course of a FY, wherein the AOR
AN
factor acts a cost recoupment factor in the hopes to garner a $0 NOR by distributing overhead costs
proportionally to all odcbs-combinations within an FY. At this time, the actual AOR calculation is not
fully understood, which is concerning due to its high variability over recent FY’s (as seen in Table 4) and
M

large impact on the rate setting process.

Table 4: Recent AOR Factors

ED

FY08 FY09 FY10 FY11 FY12 FY13 FY14 FY15

AOR Factor 20% -7% 12.2% 5.5% 8% 8% 22.3% 38.3%
PT

In an effort to limit the AOR factor’s impact, it is highly recommended that it only be applied to the
yearly odcbs-combinations that are historically reliable, consistently priced, and high in cost. For all
CE

other odcbs-combinations (i.e., lack consistency, low cost, etc.), the rates should be inflation-adjusted
using the refresh rate alone. With this subtle change, the overhead costs will continue to be recovered and
the AOR factor will have a far less influence on the variability in rates, and ultimately the NOR.
AC

The NOR calculation considers the total revenue generated for each rate setting methodology and
subtracts the adjusted total cost described in Table 3. The results are presented in Table 5Table 6
regarding the NOR for varying levels of risk preference, . More specifically, Table 5 compares the
NOR of all rate setting methods for parametric variability up to one standard deviation. In an attempt to
capture the impact on the NOR when more variability is induced, Table 6 demonstrates the effect on the
NOR for parametric variability up to three standard deviations.

19
 ACCEPTED MANUSCRIPT

Table 5: NOR Comparison for Varying Risk Preferences (one-standard deviation)

Rate Setting Methodology
Risk Preference RGP WAC*
Total NOR $14.4M $21.3M
Total NOR $17.9M $21.3M
Total NOR $21.1M $21.3M
*The WAC method is independent of a DM’s risk-preference.

Table 6: RGP Method NOR Comparison for Varying Risk Preference and Allowable Variability

T
Allowable Parametric Variability

IP
Risk Preference
Total NOR $14.4M $14.4M $14.4M
Total NOR $17.9M $17.8M $17.8M

CR
Total NOR $21.1M $21.1M $21.1M

Table 5 illustrates that each of the three rate setting methods easily recover the adjusted FY16

US
total costs, wherein the RGP method is the closest to achieving the goal of $0 NOR. In fact, if a DM is
risk-averse, the RGP model provides USTRANSCOM with roughly $6 million less profit than the current
AN
method. The RGP method also outperforms the current rate setting method when a DM is risk-neutral,
but the difference in NOR is decreased dramatically. Finally, if a DM assumes there is no parametric
variability (i.e., the DM is risk-seeking), the RGP method generates roughly $200K less in revenue when
M

compared to the WAC method. Therefore, the NOR gets worse as a DM’s risk preference tends towards
more risk-seeking behavior because of the significant variability observed in the actual FY15 and FY16
rates. Similar results are documented in a theoretical study by Hanks et al. [6], wherein a DM who
ED

assumes a large amount of parametric variability (i.e., a risk-averse DM) is rewarded with a lower NOR
because the observed parametric variability is in fact significant. Similarly, the risk-seeking DM predicts
PT

that there will be no variability in the data, but this assumption is clearly not accurate and, as a result, the
NOR is pointedly worse than the NOR realized by the risk-averse DM. Comparable results ensue when
the allowed parametric variability increases by one or two standard deviations, as depicted in Table 6. As
CE

a result, because of the complexity observed in the USTRANSCOM liner rate setting problem, parametric
variability is a certainty – and when RGP is implemented, a risk-averse attitude should be put into effect
AC

to reduce the NOR.

While considering the second goal of setting robust rates annually, the difference between the
rates set by each methodology and the inflation-adjusted rate (i.e., ( ) ) is evaluated. In
particular, Figure 1-Figure 2 illustrate the average and standard deviation of the rate difference,
respectively, whereas

20
 ACCEPTED MANUSCRIPT

Table 7 offers a summary of the minimum, maximum, and range of the rate difference for each rate
setting methodology.

$15

$10

$5

$0

T
k=1 k=2 k=3
-$5

IP
-$10

CR
RGP WAC

Figure 1: Average of the Rate Difference for Varying Methods and Risk Preferences

$700
$600
US
AN
$500
$400
$300
M

$200
$100
ED

$0
k=1 k=2 k=3
RGP WAC
PT

Figure 2: Standard Deviation of the Rate Difference for Varying Methods and Risk Preferences
CE

Table 7: Minimum, Maximum, and Range of the Rate Difference for Varying Methods and Risk Preferences
Rate Setting Methodology
Risk Preference RGP WAC*
Minimum -$1.5K -$3.1K
AC

Maximum $1 $17.7K
Range $1.5K $20.9K
Minimum -$0.8K -$3.1K
Maximum $1 $17.7K
Range $0.8K $20.9K
Minimum $0 -$3.1K
Maximum $0 $17.7K
Range $0 $20.9K
*The WAC methods is independent of a DM’s risk-preference

21
 ACCEPTED MANUSCRIPT

Figure 1 and Figure 2 confirm that the RGP method yields a lower average of the rate difference
as well as a considerably lower standard deviation of the rate difference as compared to results
corresponding to the WAC. Moreover,

Table 7 reports that the RGP method has a lower minimum, maximum, and range of the rate difference
when compared to the other rate setting methodologies. As the DM’s preference towards risk transitions
from risk-aversion to risk-seeking, all metrics (i.e., average, standard deviation, minimum, etc.) decrease.

T
This confirms intuition because, as a DM becomes less risk-averse, they assume less parametric

IP
variability and robust rates are more easily obtained.

CR
5 Conclusion and recommendations
In a data-driven environment, multi-objective optimization problems are much more complex and
combinatoric in nature [3,37,38] , making the utilization of RO, and specifically RGP, much more

US
enthralling than previous methods. This paper demonstrates the applicability of RGP to real-world, data-
driven problems with uncertain and fluctuating parameters – problems such as the USTRANSCOM liner
rate setting problem.
AN
USTRANSCOM’s mission is to ―provide full-spectrum, global mobility solutions and enabling
capabilities to their customers in peace and war‖ [18]. TCJ8 plays a significant role in supporting this
M

mission by setting liner rates annually, which can impact contractual negotiations, fiscal planning, and
overall readiness. Currently, the rate setting methodology being utilized by TCJ8 to set their yearly liner
rates is not as effective as it could be due to underlying assumptions about the deterministic nature of
ED

historical average cost data. The current method also considers only one of USTRANSCOM’s two goals
for line rate setting. This study demonstrates the use of RGP for the USTRANSCOM liner rate setting
PT

problem as a means to consider parametric variability while seeking to attain two rate-related goals.
Although the RGP model demonstrates that it can reduce the NOR and produce robust shipping rates, its
effectiveness is limited by the veracity and sparsity of the data, as well as the overall importance of the
CE

AOR factor.
We propose several recommendations to improve the liner rate setting methodology utilized by
AC

TCJ8. First, TCJ8 should take active measures to ensure their cost record data is complete and accurate.
Second, a modified process should leverage more than 2-1/2 years of historical, inflation-adjusted data to
address the inability of the current process to set each specific rate. Third, TCJ8 should conduct
additional research regarding the AOR factor and its impact on the rate setting process. Fourth, TCJ8
should investigate how unequal weighting of USTRANSCOM goals effects the RGP model’s solution.
Finally, TCJ8 should try and automate the overall rate setting process [19].

22
 ACCEPTED MANUSCRIPT

As a follow-on study, we suggest TCJ8 recoup their overhead costs differently in an attempt to
reduce the impact of the AOR factor on the TWCF balance. More specifically, we recommend recouping
overhead costs by marginally increasing the inflation-adjusted costs (via an AOR factor) on the highest
volume of shipping combinations that have low uncertainty in the workload forecasts, while using only
inflation-adjusted costs to calculate rates for all other shipping combinations. In doing so, the AOR
factor’s influence on any one rate will not be substantial, and the variability in the NOR incurred resulting
from uncertainty in parts of the liner forecast will be reduced.

T
Moreover, as a future study we suggest that TCJ8 consider recouping their overhead costs by

IP
evaluating a weighting schema for the various TCJ8 goals in regards to the RGP methodology as applied
to the USTRANSCOM liner rate setting problem.

CR
While the results of this paper demonstrate the applicability of RGP as a use-case for
USTRANSCOM, the potential for RGP utilized in other forums and various problem sets is without

US
question. It would be interesting to see RGP applied to a problem set have more than two goals. Rather
than assuming a DMs risk attitude, it would be intriguing to apply RGP to a data-driven problem wherein
the DMs risk preference is elicited a priori. Finally, it would be fascinating to observe RGP applied in
AN
different application areas – such as architecture or sports.

Disclaimer
M

The views expressed in this article are those of the authors and do not reflect the official policy or
position of the United States Air Force, the Department of Defense, or the United States Government.
ED

Acknowledgments

The authors are grateful to the Deputy Editor and three reviewers for their constructive comments that
PT

have improved not only the substance of this paper, but also its overall presentation. The authors also
gratefully thank TCJ8, specifically Mr. Tim Landvogt and Ms. Jasolyn Evans, for supplying the liner data
CE

and providing insights regarding the liner rate setting process used by the TCJ8 Cost Recovery Branch.

References
AC

[1] Sparkman BT. USTRANSCOM workload forecasting. in: a presentation at the 85th Symposium,
Military Operations Research Society, 2017.
[2] Ben-Tal A, El Ghaoui L, Nemirovski A. Robust optimization. Princeton, NJ: Princeton University
Press; 2009.
[3] Pflüger D, Peherstorfer B, Bungartz H. Spatially adaptive sparse grids for high-dimensional data-
driven problems. Journal of Complexity 2010;26:508–22.
[4] Bertsimas D, Sim M. Robust discrete optimization and network flows. Mathematical
Programming 2003;98:49–71.

23
 ACCEPTED MANUSCRIPT

[5] Kuchta D. Robust goal programming. Control and Cybernetics 2004;33:501–10.

[6] Hanks RW, Weir JD, Lunday BJ, Robbins MJ. Mapping a decision maker’s risk preference to
robust optimization parameters. Technical Paper, Department of Operational Research, Air Force
Institute of Technology, Wright-Patterson Air Force Base, Oh 2017.
[7] Soyster AL. Technical note—convex programming with set-inclusive constraints and applications
to inexact linear programming. Operations Research 1973;21:1154–7.
[8] Bertsimas D, Sim M. The price of robustness. Operations Research 2004;52:35–53.
[9] Hanks RW, Weir JD, Lunday BJ. Robust goal programming using different robustness echelons

T
via norm-based and ellipsoidal uncertainty sets. European Journal of Operational Research
2017;262:636–46.

IP
[10] Ghahtarani A, Najafi AA. Robust goal programming for multi-objective portfolio selection
problem. Economic Modelling 2013;33:588–92.

CR
[11] Nik MSA, KhajezadehDeafuli M. Conflict between goal programming and robust goal
programming in optimal investment portfolio selection. International Journal of Scientific Study
2017;5:720–5.
[12]

[13]
US
Koosha H. A robust goal programming model for the capital budgeting problem. Journal of the
Operational Research Society 2018;69:1105–13.
Taha R, Abdallah K, Sadek Y, El-Kharbotly A, Afia N. Multi-objective supply chain network
AN
design under demand uncertainty using robust goal programming approach. in: Proceedings of the
2015 International Conference on Operations Excellence and Service Engineeing, Orlando, FL:
2015.
[14] Qu B, Lu L. Optimization model for electric power supply chain operation based on system
M

engineering. System Science, Engineering Design and Manufacturing Information - 2011

International Conference, vol. 2, 2011, p. 77–80.
[15] Ayudhya N, Suksawat W. Robust goal programming approach to an intermodal routing decision
ED

problem. Engineering and Applied Science Research 2016;43:130–2.

[16] Kuchta D. A concept of a robust solution of a multicriterial linear programming problem. Central
Euorpean Journal of Operations Research 2011;19:605–13.
PT

[17] Rezaie MH, Ghahtarani A, Najafi AA. Ideal and anti-ideal compromise programming for robust
bi-objective portfolio selection problem. Economic Computation and Economic Cybernetics
Studies and Research 2015;49:307–21.
CE

[18] USTRANSCOM. United States Transportation Command 2017. https://www.ustranscom.mil/

(accessed June 11, 2017).
[19] MRC LLC. TWCF rate study executive summary. Joint Distribution Process Analysis Center,
AC

USTRANSCOM: 2015.
[20] Shneerson D. The structure of liner freight rates: a comparative route study. Journal of Transport
Economics and Policy 1976:52–67.
[21] Brooks MR, Button KJ. The determinants of shipping rates: a north atlantic case study. Transport
Logistics 1996;1:21–30.
[22] Skinner W, Maserang D, Caplice C. The future of transportation rates and fuel surcharges – the
shipper’s view. in: a presentation at the Annual Global Conference of the Council of Supply Chain
Management Professionals, Chicago, IL: 2009.

24
 ACCEPTED MANUSCRIPT

[23] Joo S-J, Min H, Smith C. Benchmarking freight rates and procuring cost-attractive transportation
services. International Journal of Logistical Management 2017;28:194–205.
[24] Forkenbrock DJ. Comparison of external costs of rail and truck freight transportation.
Transportation Research Part A: Policy and Practice 2001;35:321–37.
[25] Hummel D. Transportation costs and international trade over time. Journal of Economic
Perspectives 2007;21:131–54.
[26] Military Surface Deployment and Distribution Command. FY 2018 direct booking program billing
rates & guide 2018.

T
https://www.ustranscom.mil/dbw/docs/FY18_Direct_Booking_Rates_Setting_Guide.pdf (accessed
June 4, 2017).

IP
[27] Military Surface Deployment and Distribution Command. FY 2018 liner ocean transportation
program stabilized breakbulk / dry cargo and container billing rates 2018.

CR
https://www.ustranscom.mil/dbw/docs/FY17_Liner_Breakbulk_and_Container_Rates_and_Guida
nce.pdf (accessed June 4, 2017).
[28] Kaplan RS, Cooper R. Cost and effect. Boston: Harvard Buisness School Press; 1998.
[29]

[30]
US
Military Surface Deployment and Distribution Command. FY15 port handling billing rates 2015.
https://www.ustranscom.mil/dbw/docs/FY15_Port_Handling_Rates_and_Definitions.pdf
(accessed June 7, 2017).
Kaplan RS, Witkowski M, Abbott M, Guzman AB, Higgins LD, Meara JG, et al. Using time-
AN
driven activity-based costing to identify value improvement opportunities in healthcare. Journal of
Healthcare Management 2014;59:399–412.
[31] Tsai WH, Yang CH, Chang JC, Lee HL. An Activity-Based Costing decision model for life cycle
M

assessment in green building projects. European Journal of Operational Research 2014;238:607–

19.
[32] Oker F, Adiguzel H. Time-driven Activity-Based Costing: An implementation in a manufacturing
ED

company. Journal of Corporate Accounting and Finance 2016;27:39–56.

[33] Albright SC. VBA for modelers: developing decision support systems. Cengage Learning; 2013.
[34] LINDO Systems Inc. Powerful LINGO solvers 2016.
PT

http://www.lindo.com/index.php?view=article&catid=4%3Alingo&id=13%3Apowerful-lingo-
solvers&option=com_content&Itemid=3 (accessed July 24, 2016).
[35] Kitchin R, McArdle G. What makes big data, big data? Exploring the ontological characteristics of
CE

26 datasets. Big Data & Society 2016;3:1–10.

[36] Marr B. Big data: The 5 V’s everyone must know. LinkedIn Pulse 6 2015.
[37] Wang Z, Glynn PW, Ye Y. Likelihood robust optimization for data-driven problems.
AC

Computational Management Science 2016;13:241–61.

[38] Delage E, Ye Y. Distributionally robust optimization under moment uncertainty with application
to data-driven problems. Operations Research 2010;58:595–612.

25