Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: The low probability – high consequence nature of hazardous materials (hazmat) incidents dictate a risk-
Received 16 December 2019 averse route planning approach. However, preparing routing plans for multiple hazmat shipments
Revised 28 November 2020 between various origin–destination pairs also raises the question of risk-equity, and not just minimiza-
Accepted 7 January 2021
tion of hazmat risk. Hence, the objective is to plan an equitable routing plan for different rail hazmat ship-
Available online 12 January 2021
ments while not only ensuring the safety of citizens but also precluding certain population zones from
being subjected to intolerable levels of hazmat risk. To this end, we propose an analytical framework that
Keywords:
makes use of a conditional value-at-risk (CVaR) measure of risk to generate minimum risk shipment
Risk equity
Conditional value-at-risk
routes while promoting risk-equity in both the arcs and the yards of the railroad network. While the com-
Railroad transportation mercial solver, CPLEX, lacks the ability to generate integer solutions for even small problem instances, a
Hazardous materials Lagrangian relaxation method aimed at being maximized using the Subgradient optimization algorithm
Integer optimization is applied to provide a lower bound. The proposed framework is finally used to study several problem
instances using the realistic infrastructure of a railroad operator, and to conclude that risk-equity can
be achieved by re-routing, and that the design of train services along with the trade-off between yard-
risk and arc-risk determine the number of re-routing and the optimal value of CVaR.
Ó 2021 Elsevier Ltd. All rights reserved.
https://doi.org/10.1016/j.cor.2021.105222
0305-0548/Ó 2021 Elsevier Ltd. All rights reserved.
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
assessment methodologies that would both ensure risk-averse rout- is referred to as traditional risk. This expected consequence
ing, and incorporate the risk tolerance of a decision maker to gener- approach, originally developed to evaluate risk from highway ship-
ate multiple routes between a given origin–destination pair. To close ments and from fixed facilities, has been appropriately adapted to
this gap, Hosseini and Verma (2017) made the first attempt to study rail hazmat shipments in Bubbico et al. (2004a), Bubbico
develop a value-at-risk (VaR) assessment methodology to facilitate et al. (2004b), Verma (2011), Bagheri et al. (2014), and Cheng
risk-averse routing of rail hazmat shipments. It is pertinent that et al. (2017). Lack of data and/or limitations associated with the
though VaR has a simple interpretation for hazmat shipments, viz., expected consequence approach has led to the development of
how many people are exposed to hazmat risk given a certain confi- alternative risk measures such as incident probability that neglects
dence level, it is critiqued for ignoring the tail of the distribution, the incident consequences (Abkowitz et al., 1992), and population
i.e., overlooks catastrophic events. To overcome the indicated short- exposure that focuses only on the consequences (ReVelle et al.,
comings, Hosseini and Verma (2018) subsequently developed a con- 1991; Verma and Verter, 2007). Most recently, Hosseini and
ditional value-at-risk (CVaR) measure that quantifies the risk that Verma (2017), Hosseini and Verma (2018) have proposed risk
would be encountered in the unfavorable tail of the distribution to assessment methodologies that make use of value-at-risk and
avoid extreme events. Through numerical experiments, they conditional-value-at-risk measures to incorporate the catastrophic
demonstrate that the latter methodology not only incorporates the nature of rail hazmat shipments.
characteristics of railroad accidents and the risk tolerance of the Risk management: As indicated earlier, over the past four dec-
decision maker to generate risk-averse routing of shipments ades, the railroad industry has spent considerable effort in reduc-
between a single pair of origin–destination yards, but also under- ing the frequency of tank car accidents as well as the likelihood
scores its superiority over both the expected consequence and the of releases in case of an accident. However, the more recent risk
VaR measures. management initiatives have focused on tank-car design to mini-
In the current paper, we study the routing of multiple rail haz- mize the likelihood of hazmat release (Barkan, 2008); strategic
mat shipments of different volumes between multiple pairs of ori- placement to reduce the derailment probability of hazmat railcars
gin–destination yards, i.e., a multi-commodity flow problem. It is (Thompson, 1992; Verma, 2011; Cheng et al., 2017); routing of rail-
important that the ensuing routing plan could result in the over- cars (Glickman, 1983; Glickman et al., 2007; Verma, 2009; Verma
loading of certain rail-tracks and/or railyards, which in turn will et al., 2011) and, incorporating the risk preference of the decision
increase the corresponding hazmat risk. Thus, to prevent certain makers to generate dissimilar routes (Hosseini and Verma, 2017,
populated zones from being subjected to intolerable hazmat risk, 2018).
we promote equity in the spatial distribution of hazmat risk
throughout the railroad network. Hence, the proposed analytical 2.2. Catastrophic avoidance in hazmat transportation
methodology incorporates physical infrastructure of railroad trans-
portation system that entails using a limited number of train ser- The expected consequence approach is risk-neutral and thus
vices to move shipments and undergoing transfer operations at unable to capture the public posture against hazmat shipments.
intermediate yards, and then generates minimum CVaR risk routes Abkowitz et al. (1992) made the first effort to overcome this limi-
for rail hazmat shipments while ensuring equitable distribution of tation by proposing a perceived risk model for highway hazmat
hazmat risk in the given railroad network. Finally, the proposed shipments, where risk-averseness was incorporated via a percep-
methodology was used to study and analyze several problem tion factor. However, the value of the perception factor is difficult
instances generated using the realistic infrastructure of a railroad to both understand and quantify, which motivated Erkut and
operator, and to gain insights. Through numerical experiments Ingolfsson (2000) to analyze three catastrophic avoidance models
we illustrate the importance of train service design in ensuring risk for highway shipments -i.e., maximum risk (MM); mean–variance
equity; demonstrate the impact of risk-equity constraints on opti- (MV); and, disutility (DU). More recently, there have been efforts to
mal CVaR values, on optimal routing of hazmat shipments, and the generate dissimilar routes based on the risk preference of the deci-
trade-off between yard-risk and arc-risk; and, make use of Lagran- sion maker. To that end, Kang et al. (2014a), Kang et al. (2014b)
gian lower bound to validate the high-quality solutions resulting proposed value-at-risk model for highway shipments, while
from the proposed methodology. Hosseini and Verma (2017) proposed an equivalent methodology
The rest of the paper is organized as follows. Section 2 reviews for rail hazmat shipments. Finally, to overcome the limitations
the relevant literature, followed by the problem description and associated with value-at-risk approach, conditional value-at-risk
assumptions in Section 3. The proposed analytical methodology, (CVaR) measures were developed for highway hazmat shipments
i.e., mathematical program and the solution technique, is devel- (Toumazis and Kwon, 2013, 2016; Faghih-Roohi et al., 2016), and
oped in Section 4. Solution and analyses of the realistic size case for rail hazmat shipments (Hosseini and Verma, 2018).
study are discussed in Section 5. Finally, conclusions, contributions
and directions for future research are outlined in Section 6. 2.3. Risk equity in hazmat transportation
2. Literature review To the best of our knowledge, Fang et al. (2017) made the only
attempt to capture risk-equity on a specific rail service leg, which
For expositional reasons and also to position this work in the was accomplished by altering the train speed when routing and
context of the existing literature, the relevant literature is orga- scheduling rail hazmat shipments. However, most of the risk-
nized under three themes: first, risk assessment and management equity work has been done in the highway domain. Risk equity
initiatives for rail hazmat shipments; second, peer reviewed has been defined as the largest difference in the risk level among
engagements dealing with catastrophic avoidance (i.e., high conse- a set of individuals (Keeney, 1980). Gopalan et al. (1990) made
quence); and third, consideration of risk equity in hazmat the first effort to generate an equitable set of routes for road haz-
transportation. mat shipments such that the difference in hazmat risk in various
zones is less than the threshold. Current and Ratick (1995) pro-
2.1. Risk assessment and management of rail hazmat shipments posed a multi-objective model that makes both facility-siting and
hazmat routing decision, and impose risk equity by minimizing
Risk assessment: Risk is most commonly defined as the product the maximum allowable risk exposure for each zone. Subse-
of the probability and the consequence of an undesirable event, and quently, Carotenuto et al. (2007) proposed a model that generates
2
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
minimal risk paths for road hazmat shipments, and limits the risk Y : Set of yards in the network; indexed by i; j; k
on each link as a mechanism to enforce risk equity. Bianco et al. A : Set of ðundirectedÞ arcs in the network; indexed by ði; jÞ
(2009) proposed a bi-level network design model to capture the and ðk; jÞ
interaction between the local regulator seeking to minimize haz- S : Set of train services in the network; indexed
mat risk over its jurisdiction and the regional regulator aiming to by s ðand=or s0 Þ
minimize total transport risk while assuring risk-equity. Kang Ys 2 Y : Set of yards in trainservice fsg; indexed by is ; js ; ks
et al. (2014b) used a value-at-risk approach to model equitable As 2 A : Set of ðdirectedÞ service legs in trainservice fsg;
routing of road hazmat shipments, and proposed a relaxation indexed by ðis ; js Þ and ðks ; js Þ
heuristic to generate an efficient solution. Romero et al. (2016) V : Set of shipments in the network; indexed by v
developed a model to analyze the facility location and routing N ðv Þ : Number of hazmat railcars in shipment v
problem for hazmat, where the objective is to minimize total Oðv Þ : Origin of shipment v
canister-miles and transportation accident risk. The possibility of Dðv Þ : Destination of shipment v
considering equity in the sites selected and transportation routes
recommended is also included. Equity is expressed in terms of The above notations are then used to define the following
the Gini coefficient and the population is represented as the house- binary decision variables for routing the shipments through the
holds within a certain distance of the transportation facilities railroad network using the available train services:
selected. More recently, Taslimi et al. (2017) proposed a bi-level 8
model where the hazmat carriers select minimum-cost routes in < 1; if shipment v is carried using arc ði; jÞ
>
the network, while the regulatory authority at the upper level xvis js ¼ of train service fsg ðservice leg ðis ; js ÞÞ
>
:
seeks to consider risk-equity by minimizing the maximum trans- 0; otherwise
port risk in a zone. The interaction between the two levels is
achieved through finding the best locations for hazmat response 8 0
teams and making additional road segments available. < 1; if k is a transferring yard between train services fs g
>
xvks0 ks ¼ and fsg for shipment v
To sum, to the best of our knowledge, there is an absence of >
:
risk-averse methodology for equitable routing of rail hazmat ship- 0; otherwise
ments. Considering the demonstrated advantages of the condi-
tional value-at-risk methodology over other catastrophic Next, we can build the routing constraints ðX 2 wÞ based on the
avoidance techniques, we make the first attempt in that direction. decision variables as follows:
It is important to mention that the proposed analytical approach is X X
xvis js xvjs is
distinct from anything present in the literature in the following js js
ways: first, characteristics of railroad accidents -i.e., sequence of 8
> 1; if is ¼ Oðv Þ
events resulting in release from railcar, derailment and conditional >
>
>
>
probabilities of release; and the possibility of multiple sources of < 0; if is –Oðv Þ or Dðv Þ
¼ 8v ; 8is
release, need to be incorporated; second, unlike highway ship- >
>
> ð for any non transferring yard is for shipment v Þ
ments, one has to work with pre-defined trains services to move >
>
:
shipments, which may involve transfer operations at intermediate 1; if is ¼ Dðv Þ
rail yards; third, risk equity is ensured across the entire railroad
network, i.e., along each rail-track and in each rail-yard. X X
xvks js xvjs0 ks0 ¼ 0;
js js0
3. Problem description
for any transferring yard ks0 s for shipment v 8v ; 8ks0 s
In an effort to formulate the optimization program, we first Shipment v traverses through route Oðv Þ Dðv Þ using S out of
introduce the appropriate concepts and notations pertinent to rail- S available train services in the network, which consists of a set of
S
road transportation system in Section 3.1, which is followed by a transferring yards Yv ¼ s2S Ys and a set of service legs
S
high-level overview of the CVaR methodology proposed in A ¼ s2S As , i.e. totally nv ¼ jYv [ Av j items. It is pertinent to
v
Hosseini and Verma (2018), and the discussion about incorporating note that the routing constraints possess a multi-commodity flow
risk equity, respectively, in the next two subsections. structure. We define the following parameters for the route
Oðv Þ Dðv Þ of shipment v :
3.1. Railroad transportation system
pvk : Incident probability at transferring yard k resulting from
In a railroad transportation system, the physical infrastructure
transporting Nðv Þ hazmat railcars
comprises of rail yards and tracks. Any two yards are connected
cvk : Consequence at transferring yard k resulting from transport-
by tracks, which are called the service legs of a train traveling non-
ing Nðv Þ hazmat railcars
stop between them. A sequence of service legs and intermediate
pvij : Incident probability in arc ði; jÞ resulting from transporting
yards constitutes an itinerary available to a railcar for its journey.
We focus on the tactical planning problem of a railroad company Nðv Þ hazmat railcars
that regularly transports regular and hazmat freight between dif- cvij : Consequence in arc ði; jÞ resulting from transporting Nðv Þ
ferent pairs of origin–destination yards and needs to ensure that hazmat railcars
demand is satisfied. Hence, the objective is to route rail hazmat ship-
ments using a conditional value-at-risk methodology such that haz- To capture the characteristics of railroad accidents, incident
mat risk is not only minimized but also equitably distributed in the probability in the yards and arcs should take into consideration
railroad network. the probability that a train meets with an accident and the condi-
Sets, indices and parameters tional probabilities that a hazmat railcar derails and releases its
contents. We make use of a decile-based approach to compute
Network G ¼ ðY; A; SÞ these probabilities (i.e., the length of a train is divided into ten
equal parts):
3
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
X 10
17:13 106 for transferring yards. We then use average values
pvij ¼ P Aij yvr P Dr jAij P HjDr ; Aij P RjH; Dr ; Aij
of conditional probabilities across the ten deciles, hence pvk and
r¼1
pvij do not depend on the train configuration any more, although
they are still dependent upon the number of hazmat railcars
X
10
pvk ¼ PðAk Þ yvr ðPðDr jAk Þ PðHjDr ; Ak Þ PðRjH; Dr ; Ak ÞÞ Nðv Þ: pvij ¼ arc ði; jÞ0 s length ðmileÞ 7:35 1011 N ðv Þ and
r¼1
pvk ¼ 6:42 1010 N ðv Þ. On the other hand, given a hazmat inci-
dent, ArcGIS (ESRI, 2007) was used to estimate consequences cv k
where P Aij (or PðAk Þ) is the probability that a train meets with an
and cvij as the population exposure due to the release from Nðv Þ
accident on arc ði; jÞ (or at yard k); P Dr jAij (or PðDr jAk Þ) is the prob-
hazmat railcars traversing transferring yard k and arc ði; jÞ, respec-
ability of derailment of a railcar in the rth decile of the train given
tively. Readers are encouraged to refer to Hosseini and Verma
the accident on arc ði; jÞ (or at yard k); P HjDr ; Aij (or PðHjDr ; Ak Þ) (2017) for complete details. It should be evident that both the inci-
is the probability that a hazmat railcar derailed in the r th decile of dent probability and the consequence will increase with the total
the train given the accident on arc ði; jÞ (or at yard k); number of hazmat railcars.
P RjH; Dr ; Aij (or PðRjH; Dr ; Ak Þ) is the probability of release from a
hazmat railcar derailed in the rth decile of the train given the acci- 3.2. CVaR for rail hazmat shipments
dent on arc ði; jÞ (or at yard k). Therefore, pvk and pvij depend on yvr ,
the numbers of hazmat railcars in shipment v placed at decile r of Let C vðtÞ denote the tth smallest value in the set
each train (i.e., train configuration), which in total equal Nðv Þ: n o
P10 v cvk [ cvij : k 2 Yv &ði; jÞ 2 Av , PvðtÞ be the corresponding incident
r¼1 yr ¼ N ðv Þ; where yr 8v .
v train length
5
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
v s;s0
Subject to
!
XX
Nðv Þxvis js pij cij dij 8ði; jÞ ð10Þ X2w
v s
This Lagrangian function separates into jV j distinct shortest
where pk (or pij ) and ck (or cij ) are the incident probability and con- path problems; it will become more evident if we make the follow-
sequence, respectively resulting from transporting one hazmat rail- ing modifications
car at transferring yard k (or in arc ði; jÞ). The above constraints 8
> v v v
< pk ck C ðrÞ þ Nðv Þu p c ;
demonstrate the expected consequences resulting from the total if cvk > C vðrÞ 8k; 8v
wvk ¼ 1a k k k
hazmat shipments that pass through each transferring yard and ð11Þ
>
:
arc in their journey from origin to destination. Hence, the Lagran- Nðv Þuk pk ck ; otherwise
gian function LðuÞ will be
8
LðuÞ ¼ min >
> v v v
< pij cij C ðrÞ
r
0 0 þ Nðv Þuij pij cij ; if cvij > C vðrÞ 8ði; jÞ; 8v
wvij ¼ 1a ð12Þ
X 1 X X >
>
@C v þ @ pvk cvk C vðrÞ xvks0 ks :
ðr Þ Nðv Þuij pij cij ; otherwise
v 1a s;s0 k;cvk >C ðvr Þ
! where wvk , wvij 0. This way the Lagrangian function will be
X X
þ pvij cvij C ðvrÞ xvis j !
X v
XX XX
wvk xvk 0 ks þ wvij xvis j
s
s ði;jÞ;cvij >C ðvrÞ LðuÞ ¼ min C ðr Þ þ
r s s
v s;s0 k s ði;jÞ
! X X
X XX u k dk uij dij
þ Nðv Þxvks0 ks uk pk ck k ði;jÞ
k v s;s0
! Subject to
X XX X X
þ Nðv Þ xvis j uij pij cij u k dk uij dij
ði;jÞ v s
s
k ði;jÞ
X2w
which can be restated as follows
Subject to
6
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
Step 1:
Step 2:
Step 3:
X v X X
C ðrÞ þ f ðv Þ
r
LðuÞ ¼ min uk dk uij dij to increase for some fixed number of iterations. Lub is an upper
r
v k ði;jÞ bound on the optimal value LðuÞ, which can be the value of any pri-
mal feasible solution.
where
As alluded, Lagrangian relaxation method has been introduced
!
XX XX to supply us with a lower bound on the optimal solution. It enables
f ðv Þ ¼ min wvk xvk 0 ks þ wvij xvis j
r
s s
ð13Þ us to evaluate the maximum distance from the optimality of the
X2w
s;s0 k s ði;jÞ
solutions generated by the heuristic to be described in the next
which is a shortest path problem, as it represents finding the short- subsection.
est path for shipment v in the network where the costs (weights) of
yards and arcs are denoted by wvk and wvij , respectively. Finally, we 4.2.2. Determining k-minimal CVaR paths
propose an algorithm to solve LðuÞ, whose details are outlined in In this section, we propose a greedy heuristic algorithm to
Fig. 1. determine k-minimal CVaR paths while the equity constraints are
Subgradient search algorithm to determine the Lagrangian multi- satisfied. This algorithm is developed based upon Yen’s k-
pliers The Lagrangian dual problem, max fLðuÞju 0g, is in general shortest path algorithm (Yen, 1971). Yen’s procedure is a classic
difficult to solve exactly, and may be approximately solved by algorithm for ranking the k-shortest simple paths among the set
means of the well-known and widely-used subgradient technique of all paths between O-D pairs. It is an iterative algorithm that at
(Held et al., 1974). A vector c is a subgradient of LðuÞ at u if and only each iteration selects the shortest path from the candidate set of
if LðuÞ LðuÞ þ ðu uÞT c. Let CðuÞ denote the set of optimal solu- paths, which is then updated and deviates from the paths selected
tions for the Lagrangian function at u, if x 2 CðuÞ, then it can be in the previous iterations. Our proposed algorithm not only pro-
shown (Martin, 1999) that the following vectors c are subgradients vides a good initial primal feasible solution to be used in the Sub-
of LðuÞ at u gradient algorithm, but also can be regarded as a stand-alone
! solution algorithm, as will be shown in the next section. Other
XX approaches to finding a set of spatially dissimilar paths in hazmat
ck ¼ dk þ N ðv Þxvks0 ks pk ck 8k
v s;s0
route planning can be found in Dell’Olmo et al. (2005), Caramia
et al. (2010), and Bronfman et al. (2015).
! The main idea in this algorithm is that we start with the ship-
XX
cij ¼ dij þ N ðv Þxvis j s pij cij 8ði; jÞ ment which has the highest number of hazmat railcars and do its
v s optimal routing in such a way that the CVaR risk associated with
the generated route is minimized. We repeat this process for the
Next, we utilize them to develop the algorithm depicted in
next highest number of hazmat railcar shipment, but then check
Fig. 2.
if the created minimum CVaR route violates any of the risk equity
To calculate the Lagrangian function for the first time, we ini-
constraints for the arcs and/or transferring yards or not. If so, we
tialized the Lagrangian multipliers to zero
remove the corresponding arcs and/or transferring yards from
(uk ¼ 0; 8k and uij ¼ 0; 8ði; jÞ). Then the multipliers get updated
the network and do a rerouting to find the next minimal CVaR route
by the Subgradient optimization algorithm (Fig. 2) with the objec-
for this shipment. This process is repeated as many times as
tive of maximizing the Lagrangian function (LðuÞ). Through the
required until a minimal CVaR route is found for the shipment.
above algorithm, the sequence of dual solutions fuq g approaches
Likewise, this procedure is iterated until all hazmat shipments in
an optimal solution fu g, however the value of L uqþ1 is not nec-
the network are routed. It is worthwhile mentioning that when a
essarily greater than Lðu Þ. In fact, according to the fundamental
q
yard is removed from the network as a transferring yard, it may
theoretical result (Held et al., 1974), the sequencefLðuq Þg con-
still be used in the routing of the next shipments as a transship-
verges to Lðu Þ if the sequence tq converges to zero ment yard.
P1
and q¼0 t q ¼ 1. It is common to determine t q by a formula such as To implement the proposed heuristic algorithm, we first rewrite
the objective function of problem (P), (4), by using (8) as follows:
hq Lub Lðuq Þ
tq ¼ X X v
kcq k2 min
r
CVaRra v ¼ min
r
C ðr Þ þ f r v
v v
where hq is a positive scalar between 0 and 2 (Martin, 1999). We
determine hq by setting h0 = 2 and halving hq whenever LðuÞ fails where
7
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
!
XX
v XX
v 5. Computational analysis
f r
v ¼ min
X2w
wk xvk 0 ks þ
s
wij xvis j s
ð14Þ
s;s0 k s ði;jÞ
In this section, the proposed analytical approach is applied to a
problem instance generated using the realistic infrastructure of a
Class I railroad operator, in the mid-west United States, that was
where
introduced in Verma et al. (2011). In addition, insights about various
pertinent drivers are developed to help managers make better deci-
8 sions. To that end, we first outline the specifics of the case study,
< pvk cvk C vðrÞ
wvk ¼ 1a
; if cvk > C vðrÞ 8k; 8v ð15Þ
which is then solved and analyzed to generate insights in the follow-
: ing subsections.
0; otherwise
5.1. Problem setting
8
>
< pvij cvij CvðrÞ Fig. 4 depicts the infrastructure of a Class I railroad operator and
wvij ¼ 1a
; if cvij > C vðrÞ 8ði; jÞ; 8v ð16Þ was recreated using a geographical information system. Each of the
>
: 25 yards in the network is both a demand and a supply point for
0; otherwise
the others. However, we considered 560 different hazmat ship-
Next, we develop the steps as outlined in Fig. 3 to obtain the k- ments (o-d pairs) in the computational study, and where number
minimal CVaR paths as the equity considerations are taken into of hazmat railcars to be shipped between a given o-d pair, N(v),
account. ranges between [1,15]. We also assume a total of 31 different train
8
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
Table 4
(a) Number of shipments rerouted by the k-minimal CVaR paths algorithm to ensure risk equity for transporting 560 shipments (b) k-minimal CVaR paths algorithm gap from the
Lagrangian lower bound.
dij &dk 105 Subgradient Optimization Algorithm k-minimal CVaR Paths Algorithm
P
v CVaRa ðv Þ
Computational max LðuÞ Number of Shipments Re-routed (RR = re- Computational r Gap from the
Time (20 (Lagrangian lower routing) Time Lagrangian lower
iterations) bound) bound
1st 2nd 3rd 4th 5th 6th 7th
RR RR RR RR RR RR RR
235 & 1 21 min, 38 s 151,221.05990 6 0 0 0 0 0 0 1 min, 3s 151,337 0.08%
230 & 1 20 min, 19 s 151,221.05551 9 0 0 0 0 0 0 1 min, 3s 151,404 0.12%
225 & 1 20 min, 44 s 151,221.05524 14 3 0 0 0 0 0 1 min, 3s 151,589 0.24%
220 & 1 20 min, 05 s 151,221.06661 20 4 0 0 0 0 0 1 min, 2s 151,823 0.40%
215 & 1 20 min, 58 s 151,221.06563 26 3 0 0 0 0 0 1 min, 5s 152,025 0.53%
210 & 1 21 min, 18 s 151,221.06385 33 6 0 0 0 0 0 1 min, 6s 152,250 0.68%
205 & 1 20 min, 03 s 151,221.07832 42 11 0 0 0 0 0 1 min, 12 s 152,658 0.95%
200 & 1 22 min, 6s 151,221.07973 49 10 0 0 0 0 0 1 min, 9s 153,174 1.29%
195 & 1 20 min, 19 s 151,221.08388 55 14 4 0 0 0 0 1 min, 14 s 153,448 1.47%
190 & 1 18 min, 31 s 151,221.10173 64 20 8 0 0 0 0 1 min, 11 s 154,162 1.94%
187 & 1 21 min, 50 s 151,221.08231 68 21 8 0 0 0 0 1 min, 21 s 154,645 2.26%
1 & 19 20 min, 52 s 151,221.07527 4 0 0 0 0 0 0 0 min, 58 s 151,417 0.13%
1 & 18 18 min, 51 s 151,221.12554 5 0 0 0 0 0 0 1 min, 0s 151,454 0.15%
1 & 17 22 min, 33 s 151,221.10585 29 5 0 0 0 0 0 1 min, 7s 151,670 0.30%
1 & 16 20 min, 52 s 151,221.11382 48 9 0 0 0 0 0 1 min, 11 s 151,953 0.48%
1 & 15 20 min, 23 s 151,221.10554 59 17 0 0 0 0 0 1 min, 3s 152,551 0.88%
1 & 14 18 min, 37 s 151,221.13063 71 25 1 1 0 0 0 1 min, 12 s 153,414 1.45%
1 & 13 19 min, 5s 151,221.13670 88 34 2 2 1 0 0 1 min, 10 s 154,327 2.05%
1 & 12 19 min, 20 s 151,221.14796 107 55 14 8 6 0 0 1 min, 23 s 157,092 3.88%
1 & 11 21 min, 13 s 151,221.14074 129 72 31 17 12 2 0 1 min, 24 s 159,564 5.52%
1 & 10 22 min, 41 s 151,221.14410 153 93 52 32 27 13 2 1 min, 36 s 163,606 8.19%
235 & 19 21 min, 29 s 151,221.09926 9 0 0 0 0 0 0 1 min, 5s 151,523 0.20%
230 & 18 21 min, 19 s 151,221.03638 13 0 0 0 0 0 0 1 min, 14 s 151,611 0.26%
225 & 17 20 min, 44 s 151,220.95954 40 11 4 1 0 0 0 1 min, 12 s 151,971 0.50%
220 & 16 20 min, 40 s 151,220.95171 65 29 11 3 0 0 0 1 min, 20 s 152,710 0.98%
minimal CVaR path algorithm (Fig. 3). This exercise enabled us to 5.3. Risk equity analyses
discern the importance of train service design in ensuring risk
equity, and motivated the re-design of some pre-defined train In this subsection, we analyze how imposing risk equity con-
services. For instance, relying on the train services as identified straints on transferring yards and arcs affect optimal CVaR routing
in Verma et al. (2011), all shipments from yards at Cadillac and of the rail hazmat shipments throughout the network. To do so, we
Grand Rapids (i.e., yards 16 and 17) to other yards are forced to initially solve problem (P) without enforcing risk equity con-
go through service leg Cadillac-Lansing, i.e., 16 ? 18 (Fig. 5). Since straints -i.e., ignoring constraints (9) and (10), where the objective
a total of 346 hazmat railcars are leaving these two yards, is just to route all the hazmat shipments such that the total CVaR
and hence the hazmat risk load on arc (16,18), i.e., the LHS of risk is minimized. The resulting solution was decoded, and it turns
equation (10), is always greater than or equal to out that the maximum hazmat risk was registered on the service
ð346Þ 161 0:735 1010 1940 ¼ 0:007943, where 161 is the leg Chicago-Portage -i.e., arc (2,6), and at the transferring yard at
length of arc (16,18) in miles, and 1940 shows the population Chicago with respective risk values of 242 105 and 21 105 .
exposure around this arc. Subsequently, (P) was solved three more times: first, with just
Therefore, it is not possible to reduce the risk load on this over- the risk equity constraints on the arcs, the details of which are dis-
loaded arc (just by rerouting the shipments) unless new service cussed in Section 5.3.1; second, with just the risk equity constraints
legs are provided. In fact, when arc (16,18) is removed from the on the yards, and the resulting insights are provided in Sec-
network by the k-minimal CVaR paths algorithm, no alternative tion 5.3.2; and third, with constraints on both yards and arcs, and
service legs are provided by the train services to be used for rout- the related details are discussed in Section 5.3.3. The k-minimal
ing shipments like (O = 17, D = 14) and (O = 16, D = 14), hence Dijk- CVaR paths algorithm is then applied to each of these three scenar-
stra’s shortest path algorithm becomes infeasible too soon and the ios while dij and dk values are set to gradually decrease from the
algorithm stops. We resolved this issue by re-designing the two of maximum risk load numbers found above (when there are no risk
current train services in the network, namely, train service {Chi- equity constraints in the problem). Reducing d values continues
cago To Indianapolis} and {Indianapolis To Cadillac}, and changed until the algorithm reports an error resulting from the infeasibility
them from Chicago -> Portage -> Lafayette -> Indianapolis and Indi- of Dijkstra’s shortest path algorithm since no alternative route can
anapolis -> Fort Wayne -> Grand Rapids -> Cadillac to Grand Rapids - be found after all overloaded arcs and/or transferring yards are
> Portage -> Lafayette -> Indianapolis and Indianapolis -> Fort Wayne removed from the network.
-> Grand Rapids -> Lansing, respectively. This way, two new service
legs Grand Rapids -> Portage and Grand Rapids -> Lansing will be
5.3.1. Risk equity constraints only on arcs
added to the network (see Fig. 5), which help to prevent overload-
In the first scenario, as the max risk allowed on the arcs declines
ing on arc (16,18) and consequently enhance risk equity in the
gradually, the transferring yard risks increase. Fig. 6 depicts how
whole network.
making the dij values tighter causes the risk load at the riskiest
10
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
met on all the arc of the network. Fig. 7a depicts the original route
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
where hazmat risk on arc (3,5), connecting Washington to Granite
k-minimal CVaR Paths Algorithm
1538.3
1948.8
1538.3
1948.8
1538.3
1948.8
944.9
1035.6
3805.4
944.9
1035.6
3805.4
944.9
1035.6
3805.4
the output of the model proposed in Hosseini and Verma (2018).
However, the methodology proposed in this paper seeks to impose
P
16 s
14 s
10 s
10 s
10 s
9s
9s
9s
9s
9s
9s
9s
(Fig. 7c). Finally, following removal of four arcs and 3rd re-
routings, the shipments reach their destination without violating
max LðuÞ (Lagrangian
Subgradient Optimization Algorithm
the threshold hazmat risk imposed on all the arcs in the network
(Fig. 7d). For visual impact, we have used dotted line scheme to
lower bound)
1538.3
1948.8
1538.3
1948.8
944.9
1035.6
3805.4
944.9
1035.6
3805.4
944.9
1035.6
3805.4
vices in the network. In addition, the CVaR values are indicated for
each routing scheme, and it is intuitive to predict that consequent
CVaR values increase, given removal of overloaded arcs from the
Computational
(20 iterations)
network.
Time
s
s
s
s
s
s
s
s
s
19
19
21
23
29
19
21
23
29
18
21
22
28
20
20
26.1%
37.8%
52.1%
52.0%
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
0.0%
Gap
mat risk loadings at the top arcs of the network. Clearly tighter
threshold values at the yards are increasing hazmat risk at some
Best Lower
of the most used arcs in the network, and the top eight are depicted
1538.3
1948.8
1136.1
1282.7
1822.9
1136.1
1282.7
1822.9
944.9
1035.6
3805.4
944.9
1035.6
944.9
1035.6
Bound
shipments, and the process is similar to that for arcs. For instance,
944.9
1035.6
3805.4
944.9
1035.6
2062.3
3805.4
944.9
1035.6
2062.3
3805.4
1538.3
1948.8
1538.3
1538.3
the transferring yards that violate the risk threshold value are
removed in the subsequent re-routing, and are depicted using a
CPLEX vs k-minimal CVaR paths algorithm solutions for very small problem instances.
overloaded transferring yard will not be used for routing new ship-
of 6 h
of 6 h
of 6 h
of 6 h
of 6 h
of 6 h
Limit
Limit
Limit
Time
1s
5s
1s
4s
the original route, but then used in the second rerouting as a trans-
Number of
shipment yard.
1617
1328
1725
2123
4112
1303
2098
4087
1220
2015
1700
4004
824
932
907
CPLEX MIP Solver
14,078
14,078
2818
4223
5632
2818
4223
5632
2818
4223
5632
7039
7039
7039
10
10
2
3
4
5
2
3
4
5
1&16
values are not violated at the yards and arcs in the network.
Table 5
Table 6
CPLEX vs k-minimal CVaR paths algorithm solutions for larger problem instances.
dij &dk 105 Number CPLEX MIP Solver (time limit of 6 h) Subgradient Optimization k-minimal CVaR Paths Algorithm
of Algorithm (20 iterations)
Shipments P
v CVaRa ðv Þ
Number Number of Best Best Gap Computational max LðuÞ Computational r Gap from
of Constraints Integer Lower Time (Lagrangian Time the
Variables Solution Bound lower Lagrangian
bound) lower
bound
220 & 16 15 21,123 6121 7202.7 2169.3 69.9% 34 s 7058.0 11 s 7058.0 0.0%
20 28,154 8094 9922.0 1328.0 86.6% 44 s 9640.0 12 s 9772.1 1.4%
44 61,946 17,513 None None N/A 1 min, 50 s 18,953.3 18 s 19,318.2 1.9%
147 206,925 58,260 None None N/A 5 min, 6 s 47,431.0 22 s 47,653.2 0.5%
287 403,971 113,395 None None N/A 11 min, 10 s s 80,252.7 37 s 80,617.7 0.4%
560 788,272 221,024 None None N/A 20 min, 40 s 151,220.9 1 min, 22 s 152,710.4 1.0%
Fig. 6. Hazmat risk at transferring yards as risk equity threshold becomes tighter on arcs.
12
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
Fig. 7. k-minimal CVaR paths determination for the 468th shipment by imposing risk equity constraint (only) on arcs: dij ¼ 187 105 .
Fig. 8. Hazmat risk at top eight arcs as risk equity threshold becomes tighter at the yards.
13
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
Fig. 9. k-minimal CVaR paths determination for the 547th shipment by imposing risk equity constraint (only) on transferring yards: dk ¼ 10 105 .
14
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
Fig. 10. k-minimal CVaR paths determination for the 497th shipment by imposing risk equity constraint on arcs and transferring yards: dij ¼ 220 105 and dk ¼ 16 105 ,
respectively.
15
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
Fig. 11. Increase in the total optimal CVaR as a result of imposing risk equity constraints.
P
In Table 4, we compare the solution values v CVaRra ðv Þ (i.e. the
less than 20 s with a 0.0% gap from the corresponding Lagrangian
lower bounds. For larger problem instances, the number of deci- values of total CVaR for all the shipments in the network) gained by
sion variables and constraints increase exponentially: for the case the k-minimal CVaR paths algorithm for all the different scenarios
with 560 hazmat shipments there are more than 788,000 variables with the Lagrangian lower bound. It lists the values of the relative
and 221,000 constraints. This coupled with the complexity of gap (in percentage) of the solution values with respect to the lower
objective function and constraints forbid CPLEX to output an inte- P
v CVaRa ðv Þ LB =LB %. Comparing to the lower
r
bound LB, i.e.
ger solution and lower bound even after a time limit of 6 h in most
of the cases, while our algorithm could find better solutions in less bound, we conclude that the quality of the solution values is high
tan 90 s (see Table 6). with an average relative gap of 1.40% and a maximum value of
We attempted to solve the Lagrangian dual problem 8.19%. The computational time requirements of both methods for
max fLðuÞju 0g to achieve an effective lower bound on the opti- all the scenarios are also reported in Table 4. A maximum of
mal solution values for the scenarios studied in Section 5.3. To 100-second computational time validates the quickness of the pro-
P
v CVaRa ðv Þ, which are gained by
do so, we plug the values of r posed solution methodology as well.
the k-minimal CVaR paths algorithm (Fig. 3) and used to build
the above discussions in Sections 5.2 and 5.3, into the Subgradient 6. Conclusion
algorithm (Fig. 2), where at each step, they are used as Lub , an upper
bound on the optimal value LðuÞ, in the step size t q calculation. At In this paper, we have studied equitable routing of rail hazmat
each iteration q of the Subgradient algorithm, we first calculate shipments between multiple pairs of origin–destination yards. To
Lðuq Þ using the Lagrangian function algorithm (Fig. 1), which is this end, we propose an analytical methodology that finds the best
then utilized to compute the subgradients cq . Consequently, we route for each hazmat shipment such that the sum of the resulting
update Lagrangian multipliers uq by means of the step size CVaR-based hazmat risk along the routes is minimized while
hq ðLub Lðuq ÞÞ adhering to the maximum threshold risk imposed on rail-arcs
tq ¼ kcq k2
.
and transferring yards in the given railroad network. A k-
To find good initial values for variables u (at iteration q ¼ 0), we minimal CVaR paths algorithm was developed and experimentally
set them (all the multipliers u0k , 8k and u0ij , 8ði; jÞ) to random values evaluated on the realistic infrastructure of a Class I railroad opera-
in the ranges ½0; 1; ½0; 1 10; ½0; 1 102 ; ½0; 1 103 ; tor in the United States, whose efficiency and effectiveness was
½0; 1 10 ; and ½0; 1 105 . It turns out that initiating the Subgradi-
4 demonstrated by the lower bound provided by a Lagrangian
ent algorithm with random variables u in ½0; 1 will ultimately lead to method, and by the (near-)optimal solutions found by CPLEX for
the best (maximum) LðuÞ. An example is provided in Fig. 12 for the very small problem instances. It was also demonstrated that the
inability of CPLEX to generate integer solutions for even small
case in which we impose the maximum risk of 235 105 and
problem instances’ mathematical model stresses the complexity
19 105 on the arcs and yards, respectively. While using larger ini- of solving the problem and the necessity of developing a special-
tial u values, Subgradient algorithm gradually increases LðuÞ, by ized solution methodology.
using random values in ½0; 1 as initial u, LðuÞ fails to increase after This paper has a three-fold contribution: first, this is the first
some number of iterations. This observation shows the algorithm effort that illustrates how to incorporate characteristics of railroad
has already found the maximum/best lower bound and is not able accidents in developing a CVaR-based risk methodology with risk
to find a better one. Recall that we aim to maximize the Lagrangian equity consideration; second, through numerical experiments we
function, max fLðuÞju 0g, to find the best lower bound on the opti- demonstrate the impact of risk-equity constraints on optimal CVaR
mal solution (as the main problem is minimization). values, on optimal routing of hazmat shipments, and the trade-off
16
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
Fig. 12. Changes in LðuÞ by iterating Subgradient algorithm with different initial values of u for risk equity dij ¼ 235 105 and dk ¼ 19 105 .
between yard-risk and arc-risk; and third, demonstrate how the Conceptualization, Formal analysis, Investigation, Writing - review
design of train services can influence the risk equity in the whole & editing, Supervision, Funding acquisition.
network, and how beneficial it could be to invest in a revised
design of them. Acknowledgement
There are a number of areas of future research. First, given the sig-
nificance of hazmat freight, CVaR risk and risk-equity could be con- This research has been supported by a grant from the Social
sidered explicitly when designing train services. Second, integrate Sciences and Humanities Research Council of Canada (435-1178-
routing and scheduling to facilitate safer routing of shipments, and 2015). The authors would like to thank two anonymous referees
yard transfer operations considering other ongoing operations at a whose feedback improve the paper significantly.
given yard. Third, develop robust formulations to tide over the
imprecision associated with accident probability and consequence
data, and devise routing strategies that are less sensitive to changes Appendix A. Supplementary data
in arc/yard attributes or data inaccuracy. Fourth, investigate the
trade-off between risk-equity and travel time (or cost). Supplementary data to this article can be found online at
https://doi.org/10.1016/j.cor.2021.105222.
17
S. Davod Hosseini and M. Verma Computers and Operations Research 129 (2021) 105222
Abkowitz, M., Lepofsky, M., Cheng, P., 1992. Selecting criteria for designating Hosseini, S.D., Verma, M., 2018. Conditional value-at-risk (CVaR) methodology to
hazardous materials highway routes. Transp. Res. Rec. 1333. optimal train configuration and routing of rail hazmat shipments. Transp. Res.
Bagheri, M., 2009. Risk-based model for effective marshaling of dangerous goods Part B: Methodol. 110, 79–103.
railways cars Ph.D. Dissertation. University of Waterloo. Kang, Y., Batta, R., Kwon, C., 2014a. Value-at-risk model for hazardous material
Bagheri, M., Verma, M., Verter, V., 2014. Transport mode selection for toxic gases: transportation. Ann. Oper. Res. 222 (1), 361–387.
rail or road? Risk Anal. 134 (1), 168–186. Kang, Y., Batta, R., Kwon, C., 2014b. Generalized route planning model for hazardous
Barkan, C.P., 2008. Improving the design of higher-capacity railway tank cars for material transportation with VaR and equity considerations. Comput. Oper. Res.
hazardous materials transport: optimizing the trade-off between weight and 43, 237–247.
safety. J. Hazard. Mater. 160, 122–134. Keeney, R.L., 1980. Equity and public risk. Oper. Res. 28(3-Part-I), 527–534.
Bianco, L., Caramia, M., Giordani, S., 2009. A bilevel flow model for hazmat Martin, R.K., 1999. Large Scale Linear and Integer Optimization: A United Approach,
transportation network design. Transp. Res. Part C: Emerg. Technol. 17 (2), 175– Springer, New York, pp. 393–415.
196. Revelle, C., Cohon, J., Shobrys, D., 1991. Simultaneous siting and routing in the
Bronfman, A., Marianov, V., Paredes-Belmar, G., Lüer-Villagra, A., 2015. The disposal of hazardous wastes. Transp. Sci. 25 (2), 138–145.
maximin HAZMAT routing problem. Eur. J. Oper. Res. 241, 15–27. Romero, N., Nozick, L.K., Xu, N., 2016. Hazmat facility location and routing analysis
Bubbico, R., Di Cave, S., Mazzarotta, B., 2004a. Risk analysis for road and rail with explicit consideration of equity using the Gini coefficient. Transp. Res. Part
transport of hazardous materials: a GIS approach. J. Loss Prev. Process Ind. 17 E: Logist. Transp. Rev. 89, 165–181.
(6), 483–488. Taslimi, M., Batta, R., Kwon, C., 2017. A comprehensive modeling framework for
Bubbico, R., Di Cave, S., Mazzarotta, B., 2004b. Risk analysis for road and rail hazmat network design, hazmat response team location, and equity of risk.
transport of hazardous materials: a simplified approach. J. Loss Prev. Process Comput. Oper. Res. 79, 119–130.
Ind. 17 (6), 477–482. Thompson, R., 1992. Hazardous materials car placement in a train consist. US
CAPP Crude oil: forecast, market and transportation. Can Assoc Pet Prod 2014. Department of Transportation, Federal Railroad Administration, Office of
Available at: http://www.capp.ca. Research and Development.
Caramia, M., Giordani, S., Iovanella, A., 2010. On the selection of k routes in multi- Toumazis, I., Kwon, C., 2013. Routing hazardous materials on time dependent
objective hazmat route planning. J. Manage. Math. 21, 239–251. networks using conditional value-at-risk. Transp. Res. Part C: Emerg. Technol.
Carotenuto, P., Giordani, S., Ricciardelli, S., 2007. Finding minimum and equitable 37, 73–92.
risk routes for hazmat shipments. Comput. Oper. Res. 34 (5), 1304–1327. Toumazis, I., Kwon, C., 2016. Worst-case conditional value-at-risk minimization for
Cheng, J., Verma, M., Verter, V., 2017. Impact of train makeup on hazmat risk in a hazardous materials transportation. Transp. Sci. 50 (4), 1174–1187.
transport corridor. J. Transp. Saf. Security 9, 167–194. TRANSPORT CANADA, 2016. Transportation in Canada 2016: Statistical Addendum.
Current, J., Ratick, S., 1995. A model to assess risk, equity and efficiency in facility Accessed via request.
location and transportation of Hazardous materials. Location Sci. 3 (3), 187–201. U.S. Department of Transportation,(2015. Hazardous materials shipments by
Dell’Olmo, P., Gentili, M., Scozzari, A., 2005. On finding dissimilar pareto-optimal transportation mode. Bureau of Transportation Statistics Available at: https://
paths. Eur. J. Oper. Res. 162 (1), 70–82. www.census.gov/econ/cfs/2012/ec12tcf-us-hm.pdf.
Erkut, E., Ingolfsson, A., 2000. Catastrophic avoidance models for hazardous Vaezi, A., Verma, M., 2017. An analytics approach to dis-aggregate national freight
materials route planning. Transp. Sci. 34 (2), 165–179. data to estimate hazmat traffic on rail-links and at rail-yards in Canada. J. Rail
ESRI, 2007. ArcView GIS 9.1. ESRI Inc., 380 New York Street, Redlands, CA, USA. Transp. Plann. Manage. 7, 291–307.
Faghih-Roohi, S., Ong, Y.-S., Asian, S., Zhang, A.N., 2016. Dynamic conditional value- Vaezi, A., Verma, M., 2018. Railroad transportation of crude oil in Canada:
at-risk model for routing and scheduling of hazardous material transportation developing long-term forecasts, and evaluating the impact of proposed
networks. Ann. Oper. Res. 247 (2), 715–734. pipeline projects. J. Transp. Geogr. 69, 98–111.
Fang, K., Ke, G.Y., Verma, M., 2017. A routing and scheduling approach to rail Verma, Manish, 2009. A cost and expected consequence approach to planning and
transportation of hazardous materials with demand due dates. Eur. J. Oper. Res. managing railroad transportation of hazardous materials. Transportation
261 (1), 154–168. research part D: transport and environment 14 (5), 300-308.
Glickman, T.S., Erkut, E., Zschocke, M.S., 2007. The cost and risk impacts of rerouting Verma, M., 2011. Railroad transportation of dangerous goods: A conditional
railroad shipments of hazardous materials. Accid. Anal. Prev. 39, 1015–1025. exposure approach to minimize transport risk. Transp. Res. Part C: Emerg.
Glickman, T.S., 1983. Rerouting railroad shipments of hazardous materials to avoid Technol. 19, 790–802.
populated areas. Accid. Anal. Prev. 15, 329–335. Verma, M., Verter, V., 2007. Railroad transportation of dangerous goods: population
Gopalan, R., Kolluri, K.S., Batta, R., Karwan, M.H., 1990. Modeling equity of risk in the exposure to airborne toxins. Comput. Oper. Res. 34, 1287–1303.
transportation of hazardous materials. Oper. Res. 38 (6), 961–973. Verma, M., Verter, V., Gendreau, M., 2011. A tactical planning model for railroad
Held, M., Wolfe, P., Crowder, H.P., 1974. Validation of subgradient optimization. transportation of dangerous goods. Transp. Sci. 45 (2), 163–174.
Math. Program. 6, 62–88. Verma, M., Verter, V., Zufferey, N., 2012. A bi-objective model for planning and
Hosseini, S.D., Verma, M., 2017. A Value-at-risk (VaR) approach to routing rail managing rail-truck intermodal transportation of hazardous materials. Transp.
hazmat shipments. Transp. Res. Part D: Transp. Environ. 54, 191–211. Res. Part E: Logist. Transp. Rev. 48, 132–149.
Yen, J.Y., 1971. Finding the k shortest loop-less paths in a network. Manage. Sci. 17
(11), 712–716.
18