European Journal of Operational Research 318 (2024) 520–533

Contents lists available at ScienceDirect

European Journal of Operational Research

journal homepage: www.elsevier.com/locate/eor

Stochastics and statistics

A disaggregated integer L-shaped method for stochastic vehicle routing

problems with monotonic recourse
Lucas Parada a , Robin Legault b,c , Jean-François Côté d ,∗, Michel Gendreau a
a
CIRRELT and Polytechnique Montréal, 2500 chemin de Polytechnique Montréal, Montreal, H3T 1J4, Quebec, Canada
b
Operations Research Center, Muckley Bldg, 1 Amherst St, Cambridge, MA 02142, United States of America
c
Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, MA 02139, United States of America
d
CIRRELT and Université Laval, 2325 Rue de l’Université, Quebec City, G1V 0A6, Quebec, Canada

ARTICLE INFO ABSTRACT

Keywords: This paper proposes a new integer L-shaped method for solving two-stage stochastic integer programs whose
Stochastic programming first-stage solutions can decompose into disjoint components, each one having a monotonic recourse function.
Integer L-shaped method In a minimization problem, the monotonicity property stipulates that the recourse cost of a component must
Monotonic recourse
always be higher or equal to that of any of its subcomponents. The method exploits new types of optimality
Vehicle routing problem
cuts and lower bounding functionals that are valid under this property. The stochastic vehicle routing problem
is particularly well suited to be solved by this approach, as its solutions can be decomposed into a set of
routes. We consider the variant with stochastic demands in which the recourse policy consists of performing a
return trip to the depot whenever a vehicle does not have sufficient capacity to accommodate a newly realized
customer demand. This work shows that this policy can lead to a non-monotonic recourse function, but that the
monotonicity holds when the customer demands are modeled by several commonly used families of probability
distributions. We also present new problem-specific lower bounds on the recourse that strengthen the lower
bounding functionals and significantly speed up the solving process. Computational experiments on instances
from the literature show that the new approach achieves state-of-the-art results.

1. Introduction inequalities bound the recourse for a much broader range of solutions
than optimality cuts. However, many of the most impactful ones are
In two-stage stochastic programs, first-stage decisions are taken problem-specific, which limits their range of application. Other meth-
before the realization of uncertain parameters. Once these parameters ods like the logic-based Benders decomposition of Hooker and Ottosson
are revealed, second-stage decisions are taken to correct the first-stage (2003) try to overcome the problem by providing cuts that include only
solution. The objective is to find a first-stage solution that minimizes a small subset of variables. These cuts are then effective for a large
the cost of the first stage and the expected cost of the second stage. number of solutions. Other enhancements are reported in Rahmaniani,
These problems have typically been solved using Benders decompo- Crainic, Gendreau, and Rei (2017).
sition (Benders, 1962), or equivalently the L-shaped method (Van This paper introduces a new integer L-shaped method, named
Slyke & Wets, 1969) when they have continuous variables and a large the disaggregated integer L-shaped method (DL-shaped method). This
number of scenarios. For solving programs with binary variables in both
method is tailored for a specific class of two-stage stochastic programs
stages, researchers historically relied on the integer L-shaped method
in which, given a first-stage solution, the recourse cost can be expressed
of Laporte and Louveaux (1993). These methods can be regarded as
as a sum of independent recourse functions involving disjoint sets of
branch-and-cut (B&C) algorithms where the expectation function of the
first-stage variables. The DL-shaped method also requires each of the
objective is relaxed and replaced by a variable that is linearly bounded
resulting recourse functions to be monotonic. In the case of a minimiza-
with inequalities. Nowadays, the L-shaped and integer L-shaped meth-
ods are considered to be computationally impractical because of their tion problem, this property means that adding first-stage variables to a
slow convergence. Several techniques have been proposed in the liter- component cannot decrease its recourse cost. Many two-stage stochastic
ature to mitigate this problem. A common approach is to add lower programs have such structure and property. For example, in the bin
bounding functionals (LBFs) (Birge & Wets, 1986) to the model. These packing problem with uncertain weights, it is possible to decompose

∗ Corresponding author.
E-mail address: [email protected] (J.-F. Côté).

https://doi.org/10.1016/j.ejor.2024.05.012
Received 8 February 2024; Accepted 7 May 2024
Available online 9 May 2024
0377-2217/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
L. Parada et al. European Journal of Operational Research 318 (2024) 520–533

the first-stage solution by bin. This problem is found in the allocation information is gradually revealed. In recent decades, the increasing
of surgeries to operating rooms (Denton, Miller, Balasubramanian, & availability of information technologies has simplified the gathering of
Huschka, 2010). Similarly, in job scheduling problems (Elçi & Hooker, real-time information that helps decision-makers re-optimize a priori
2022; Li, Côté, Callegari-Coelho, & Wu, 2022), jobs are assigned to routes on the fly. As a consequence, new paradigms have emerged, such
machines, and their first-stage solution can be decomposed by machine. as the dynamic and online vehicle routing problems, where the param-
In the latter application, the monotonicity of the objective function eters are updated throughout the solution process. A comprehensive
corresponds to the fact that assigning more jobs to a machine cannot survey on the dynamic vehicle routing problem is presented in Pillac,
decrease the expected makespan of its operations. Gendreau, Guéret, and Medaglia (2013).
We propose an implementation of the DL-shaped method for the Within the a priori optimization paradigm, the DTD and optimal re-
vehicle routing problem with stochastic demands (VRPSD). In this stocking (OR) policies are the most common in the literature. Under the
problem, each customer has to be visited exactly once by a fleet of DTD policy, restocking trips are only allowed in case of failure. Studies
capacitated vehicles. This structure makes it possible to decompose the of the VRPSD with this policy include the works of Gendreau, Laporte,
recourse cost by route. Customer demands are modeled by independent and Séguin (1995), Laporte, Louveaux, and Van Hamme (2002), and Ja-
random variables (RVs) with known distributions and are only observed bali, Rei, Gendreau, and Laporte (2014). In the OR policy, vehicles are
when a vehicle arrives at the customer’s location. A failure occurs when allowed to perform preventive returns to the depot to avoid costly fail-
a vehicle arrives at a customer’s location with a remaining capacity ures. Yang, Mathur, and Ballou (2000) have shown that the OR policy
smaller than the customer’s demand. When this happens, a recourse can be calculated by solving a Bellman equation. Relevant works under
action must be implemented to satisfy the current and next customer this policy include Louveaux and Salazar-González (2018), Salavati-
demands. This work applies the detour-to-depot (DTD) recourse policy, Khoshghalb, Gendreau, Jabali, and Rei (2019), Florio et al. (2020),
which performs a back-and-forth trip from the customer to the depot to and Hoogendoorn and Spliet (2023). Additionally, variants of these
restock the vehicle. We show sufficient conditions on the distribution two policies have also been studied. For example, Ak and Erera (2007)
of the demands for this policy to lead to a monotonic recourse function. propose a recourse policy that is based on pairwise sets of a priori
The contributions of this paper are as follows. First, we introduce routes. Also, in Secomandi and Margot (2009), an a priori planned
the DL-shaped method in the context of the VRPSD. Second, we demon- tour is partly re-optimized using a MDP that is limited to a restricted
strate that, although the DTD recourse function is not monotonic in set of states. Florio, Feillet, Poggi, and Vidal (2022) apply another
general, it is for Poisson distributions and some normal, binomial, type of partial re-optimization in which the order of visit of successive
Erlang, and negative binomial distributions when the sum of the ex- customers in a planned route is allowed to be switched.
pected demands on each route respects the vehicle capacity. Third, Several numerical experiments have been performed in the liter-
we present several new lower bounds on the DTD recourse function ature to compare the effectiveness of the proposed policies. Those
that take advantage of its monotonicity and are used by the LBFs. of Louveaux and Salazar-González (2018) show that the OR policy
To our knowledge, we perform the first numerical comparison of the might yield some savings compared to the DTD policy, but they also
different lower bounds on the recourse. Our numerical experiments noticed that the optimal a priori routes are most of the time the
show that our lower bounds greatly improve over the ones found in same as those from the DTD policy. Their results also indicate that
the literature and that our method achieves state-of-the-art results on more instances can be solved when using the DTD policy. Salavati-
existing benchmark instances. This leads us to propose new instances Khoshghalb et al. (2019) computed the cost of the DTD policy from an
that are more challenging than the existing ones from the literature. optimal OR policy solution. An average increase in the optimal value of
The remainder of the paper is organized as follows. Section 2 about 0.63% could be observed. Florio et al. (2020) indicate that the
discusses the literature on stochastic vehicle routing problems. Sec- benefit of using the OR policy over the DTD policy is close to none.
tion 3 presents the mathematical formulation of the problem. Section 4 However, savings of 1.6% on average can be obtained when reducing
studies the monotonicity of the recourse function. Section 5 details the number of possible routes and allowing the capacity to be exceeded.
the implementation of the new method and the new lower bounds They also noted that the benefit of using the OR policy decreases when
for the problem. Section 6 describes the overall approach. Section 7 the demand variability increases. Finally, the partial re-optimization
presents numerical results for our implementation, and lastly, Section 8 policy of Florio et al. (2022) can further reduce costs compared to
concludes the paper. the OR policy by 0.73% on average. There are still open questions
from these observations. Those analyses are limited to relatively few
2. Related literature instances of small sizes. In light of these results, the DTD policy is still
worth studying.
The stochastic vehicle routing problem with random customer de- Exact methods from the literature for the VRPSD divide into two
mands and multiple depots by Tillman (1969) is considered to be the categories, namely integer L-shaped and branch-and-price (B&P) meth-
first studied version of the VRPSD. Nowadays, the state-of-the-art exact ods. The main advantages of the integer L-shaped method are that it
methods can solve instances with more than 100 customers (Florio, can manage solutions that include long routes (40+ nodes per route), it
Hartl, & Minner, 2020). By this measure, since Tillman (1969), the field is compatible with several demand distributions, including continuous
has linearly increased over time in its solution capability. Although this ones, and it can handle many different types of recourse functions.
is considerable progress for an NP-Hard problem like the VRPSD, Gen- Unfortunately, it is generally inefficient for instances requiring more
dreau, Jabali, and Rei (2016) conclude that we are still in the early than three vehicles. Although several authors have devoted important
stages of the field, with room for significant improvement in exact efforts to propose LBFs (Hjorring & Holt, 1999; Hoogendoorn & Spliet,
solution methods. In this section, we present a historical review of the 2023; Jabali et al., 2014; Laporte et al., 2002) that alleviate this
exact methods developed for the VRPSD. limitation, solving instances with a large number of vehicles remains a
The a priori and re-optimization paradigms are the two model- challenge. On the other hand, B&P methods have shown to be efficient
ing paradigms used in the literature. The choice of either of these for instances that require a large number of short routes, but they
paradigms depends on when the stochastic parameters are revealed also have important limitations. They have been almost exclusively
in the solution process. Formalized by Bertsimas, Jaillet, and Odoni implemented for demands following Poisson distributions. Also, for the
(1990), the a priori paradigm relies on the design of routes that cannot OR policy, they require to store entire routes in the labeling algorithm
be modified after they have been fixed. Re-optimization was formalized of the column generation subproblem, which greatly limits the length
by Dror, Laporte, and Trudeau (1989) as a Markov decision process of the routes the method can handle. Finally, for the DTD policy, they
(MDP) and allows for the construction and adjustment of routes as do not scale well when the demands and the capacity of vehicles are

521
L. Parada et al. European Journal of Operational Research 318 (2024) 520–533

high since the size of the pricing subproblem depends on the vehicle exactly once. A route is defined as a sequence of customers starting
capacity. To mitigate this limitation, some methods from the literature and ending at the depot, and a sequence is feasible if the sum of the ex-
divide each instance’s capacity and expected demands by their greatest pected demands respects the vehicle capacity. This condition was added
common divisor to produce subproblems of tractable size in their by Laporte et al. (2002) to avoid routes that systematically fail while
column generation procedures. Although the instances resulting from having others that are underutilized. Let 𝑀 ⊆ N be the set specifying
this transformation are usually good approximations of the original the number of vehicles that can be used in a solution. For unlimited
∑
problems, they are not equivalent in general and can have different fleet problems, this set is given by 𝑀 = {⌈( 𝑖∈𝑆 𝜇𝑖 )∕𝑄⌉, … , |𝑁|}, and
optimal solutions and values. when exactly 𝑚̄ vehicles must be used, 𝑀 = {𝑚}.̄
The idea of disaggregating the recourse, i.e., separating it into Let 𝐸(𝑆) be the edges with both endpoints in 𝑆 and 𝐸(ℎ) the set of
independent disjoint functions, was studied by Wets (1966). They show edges linked with node ℎ. The VRPSD can be formulated as follows. The
that the recourse can be disaggregated trivially when the second-stage variable 𝑥𝑖𝑗 indicates the number of times edge (𝑖, 𝑗) ∈ 𝐸 is traversed.
linear programming matrix can be decomposed into blocks. In such The variables associated with pairs of customers are binary, and those
cases, a variable is added in the first stage for each block. Then, in linking the depot to a customer can take the values 0, 1, or 2. The
each iteration, the second-stage problem of each block is solved, and variable 𝑥0𝑖 equals 2 if a vehicle serves a single customer 𝑖 ∈ 𝐸. A
an optimality cut that contains only first-stage variables of the block is binary variable 𝑧𝑚 decides whether 𝑚 vehicles are used in the solution.
added. Extending this idea to problems whose second-stage matrix is Let (𝑥) denote the expected recourse cost of the first-stage solution
not decomposable is challenging. 𝑥 = (𝑥𝑖𝑗 ). The model is:
The second stage of vehicle routing problems does not present this ∑
block structure. One of the reasons for this is that the uncertain pa- min 𝑐𝑖𝑗 𝑥𝑖𝑗 + (𝑥) (1)
rameters are progressively revealed as the clients are visited. However, (𝑖,𝑗)∈𝐸
∑ ∑
the first-stage solutions of vehicle routing problems have a separable s.t. 𝑥0𝑖 = 2𝑚𝑧𝑚 , (2)
structure, as they form independent routes. Some works have intro- 𝑖∈𝑁 𝑚∈𝑀
∑
duced techniques to exploit this separability. Séguin (1996) proposed to 𝑥𝑖𝑗 = 2, ℎ ∈ 𝑁, (3)
disaggregate the recourse by routes, Côté, Gendreau, and Potvin (2020) (𝑖,𝑗)∈𝐸(ℎ)
⌈∑ ⌉
proposed to bound the recourse for sets of nodes, and Hoogendoorn ∑ 𝑖∈𝑆 𝜇𝑖
and Spliet (2023) for unstructured sets. These methods add variables in 𝑥𝑖𝑗 ≤ |𝑆| − , 𝑆 ⊆ 𝑁, (4)
(𝑖,𝑗)∈𝐸(𝑆)
𝑄
the first stage to bound some parts of the second stage by using linear ∑
inequalities. Numerical experiments show that these valid inequalities 𝑧𝑚 = 1, (5)
help at reducing computation times. The superiority of the DL-shaped 𝑚∈𝑀

method over these previous works comes from its new optimality cuts 𝑥𝑖𝑗 ∈ {0, 1}, (𝑖, 𝑗) ∈ 𝐸(𝑁), (6)
and new LBFs that simultaneously bound the recourse for paths and
𝑥𝑖𝑗 ∈ {0, 1, 2}, (𝑖, 𝑗) ∈ 𝐸(0), (7)
sets of nodes that apply to a broad range of first-stage solutions.
As noted in the introduction, the DL-shaped method requires the 𝑧𝑚 ∈ {0, 1}, 𝑚 ∈ 𝑀. (8)
monotonicity of the recourse function. The topic of monotonicity re-
The objective (1) is to minimize the sum of travel costs plus the
ceived little attention in the VRPSD literature. To the best of our
expected recourse cost. Constraint (2) imposes that 𝑚 routes must be
knowledge, the articles by Dror and Ball (1987) and Kreimer and Dror
connected to the depot if 𝑚 vehicles are used. Constraints (3) ensure
(1990) constitute the only theoretical works on the matter. They both
that customers are visited exactly once. Constraints (4) impose that the
give sufficient conditions for the probability of a failure to increase
expected demand on each route does not exceed the vehicle capacity
at each customer on a route, assuming independent and identically
and that the routes are connected to the depot. Constraint (5) ensures
distributed (i.i.d.) demands. In this paper, we demonstrate that this is
that the number of vehicles that are used in the solution is an element
sufficient for the recourse function to be monotonic under the DTD
of 𝑀. Constraints (6)–(8) define the domain of the decision variables.
policy. We also introduce sufficient conditions for the monotonicity
For the DTD recourse policy, it is possible to separate the calculation
of the recourse function that can be applied in the independent but
of the recourse by route. Let 𝜈 be the set of routes in solution 𝑥𝜈 . Then,
not identically distributed case. We show conditions under which it is
the expected recourse cost is defined as:
respected for Poisson, normal, binomial, Erlang, and negative binomial
∑
demands. In the case the monotonicity property does not hold, we refer (𝑥𝜈 ) = (𝑟), (9)
to Hoogendoorn and Spliet (2023), Jabali et al. (2014), Laporte et al. 𝑟∈𝜈
(2002) for solving stochastic vehicle routing problems, and more gen-
where (𝑟) = min{1 (𝑟), 2 (𝑟)} is the minimum expected recourse cost
erally to Laporte and Louveaux (1993) for other two-stage stochastic
of both orientations of route 𝑟 = (0, 𝑖1 , … , 𝑖𝑡 , 0). The computation of
programs.
both orientations is required because model (1)–(8) does not capture
route orientation. The expected recourse cost is denoted by 1 (𝑟) and
3. Mathematical formulation
2 (𝑟) in the two orientations. Under the DTD policy, a cost of 2𝑐0𝑖
is incurred when the demand of customer 𝑖 cannot be satisfied. The
This section introduces the notation that will be used throughout
expected recourse of route 𝑟, in the first orientation, is thus:
the paper and the mathematical model of the problem. ( 𝑗−1 )
The VRPSD is defined as follows. Let 𝐺 = {𝑁0 , 𝐸} be an undirected ∑
𝑡 ∑∞ ∑ ∑𝑗

graph where 𝑁0 = {0, 1, … , 𝑛} is the set of nodes, with node 0 being 1 (𝑟) = 2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖𝑗 (10)
𝑗=1 𝑙=1 𝑘=1 𝑘=1
the depot, 𝑁 = {1, … , 𝑛} the set of customer nodes, and 𝐸 = {(𝑖, 𝑗) ∶
𝑖, 𝑗 ∈ 𝑁, 𝑖 < 𝑗} the set of edges. Traveling on edge (𝑖, 𝑗) incurs a travel Note that a restocking trip occurs only when a customer’s demand
cost of 𝑐𝑖𝑗 . Each customer 𝑖 ∈ 𝑁 has a nonnegative demand given by exceeds the vehicle’s residual capacity. We exclude the case of going to
the RV 𝜉𝑖 , with an expected value 𝜇𝑖 and a standard deviation 𝜎𝑖 . It the depot and then moving to the next customer when an exact stock-
is assumed that the demand variables are independently distributed, out occurs. If it is assumed that 𝜉𝑖 ≤ 𝑄 for all 𝑖 with probability 1,
that they are divisible so they can be fulfilled in one trip, or multiple then the second summation of (11) can be limited to 𝑗 − 1 instead of
trips in case of failure. A fleet of identical vehicles is available to satisfy going to ∞. If( demands can exceed the ) capacity of the vehicles, the
∑𝑗−1 ∑𝑗
customer requests. Each vehicle has a capacity 𝑄 and must follow a probability P 𝜉𝑖
𝑘=1 𝑘
≤ 𝑙𝑄 ≤ 𝜉
𝑘=1 𝑘𝑖 tends to zero as 𝑙 increases
route that starts and ends at the depot. Each customer must be visited and the summation can be stopped when it reaches a probability close

522
L. Parada et al. European Journal of Operational Research 318 (2024) 520–533

to zero. In our implementation, the summation is stopped when the Proposition 1. Let 𝑆 ⊆ 𝑁 be a set of customers such that, for any pair
probability is lower than 0.0001. In the paper, the expected recourse (𝑎, 𝑏) ∈ 𝑆 × 𝑆, 𝑎 ≠ 𝑏, and for any subset 𝑆̃ ⊆ 𝑆 ⧵ {𝑎, 𝑏}, there exists integers
cost (𝑟) of a route 𝑟 = (0, 𝑝, 0) will equivalently be denoted by (𝑝). ̃ 𝜌𝑎 and 𝜌𝑏 such that the RVs 𝜉,
𝜌, ̃ 𝜉𝑎 and 𝜉𝑏 can respectively be decomposed
It is standard practice in the literature (e.g., Dror et al., 1989; as the sum of 𝜌,̃ 𝜌𝑎 and 𝜌𝑏 i.i.d. RVs 𝜁𝑘 , 𝑘 ∈ N. The following condition is
Laporte et al., 2002) to express the expected recourse cost under the sufficient for the set 𝑆 to respect the monotonicity condition:
DTD policy in terms of the cumulative distribution function (CDF) of 𝑃𝑙𝑄 (𝜌̃ + 𝜌𝑎 + 𝑗) ≥ 𝑃𝑙𝑄 (𝜌̃ + 𝑗) ∀𝑙 ∈ N, ∀𝑗 ∈ {1, … , 𝜌𝑏 }, (13)
the demand, which has the advantage of being more direct to compute (∑ )
𝑠−1 ∑ 𝑠
for usual probability distributions. where 𝑃𝑢 (𝑠) = P 𝑘=1 𝜁𝑘 ≤ 𝑢 < 𝑘=1 𝜁𝑘 denotes the probability that
( ( 𝑗−1 ) ( 𝑗 )) exactly 𝑠 ∈ N RVs 𝜁𝑘 must be drawn for their cumulative sum to first exceed
∑𝑡 ∑
∞ ∑ ∑
1 (𝑟) = 2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 − P 𝜉𝑖𝑘 ≤ 𝑙𝑄 𝑐0𝑖𝑗 (11) a given threshold 𝑢 ≥ 0.
𝑗=1 𝑙=1 𝑘=1 𝑘=1
(∑ ∑ Proof. Given the assumptions of the proposition, each original cus-
𝑗−1
The more intuitive and succinct expression P 𝜉 ≤ 𝑙𝑄 < 𝑗𝑘=1 tomer 𝑖 ∈ 𝑆 can be seen as a group of 𝜌𝑖 smaller customers with i.i.d.
) (∑ 𝑘=1 𝑖𝑘 ) (∑
𝜉𝑖𝑘 , which is equivalent to the difference P 𝑗−1
𝜉 ≤ 𝑙𝑄 −P 𝑗 demands 𝜁𝑘 , 𝑘 ∈ N and sharing the same location.
) 𝑘=1 𝑖𝑘 𝑘=1
From there, the left-hand side of inequality (12) can be rewritten as
𝜉𝑖𝑘 ≤ 𝑙𝑄 in the case of RVs with nonnegative support, is used through- follows.
out the paper for notational simplicity. However, as customary in (𝜌+𝜌
̃ 𝑎 ̃ 𝑎 +𝜌𝑏
𝜌+𝜌
)
( ) ∑ ∑
the literature, we compute the recourse based on (11), including for P 𝜉̃ + 𝜉𝑎 ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑎 + 𝜉𝑏 = P 𝜁𝑘 ≤ 𝑙𝑄 < 𝜁𝑘
instances using normally distributed demands. All the properties of 𝑘=1 𝑘=1
(𝜌+𝜌 )
the paper directly extend to formulation (11) by substituting the re- 𝜌𝑏
∑ ̃ 𝑎 +𝑗−1
∑ ̃ 𝑎 +𝑗
𝜌+𝜌
∑
stocking probabilities in (10) by the corresponding difference of CDFs = P 𝜁𝑘 ≤ 𝑙𝑄 < 𝜁𝑘
everywhere in the proofs. 𝑗=1 𝑘=1 𝑘=1
𝜌𝑏
∑
= 𝑃𝑙𝑄 (𝜌̃ + 𝜌𝑎 + 𝑗)
4. Monotonicity of the recourse function 𝑗=1

The second equality is obtained by decomposing the probability that

Our algorithm relies on new optimality cuts and lower bounds the 𝑙th restocking trip occurs between customers 𝜌+𝜌̃ 𝑎 +1 and 𝜌+𝜌
̃ 𝑎 +𝜌𝑏
that require the recourse function to decrease monotonically when as the sum of the probabilities that it occurs at customer 𝜌+𝜌
̃ 𝑎 +𝑗, for
removing customers from a route. This section formalizes this property 𝑗=1, … , 𝜌𝑏 .
by the so-called monotonicity condition and monotonicity property, Doing the same for the right-hand side, we obtain the sufficient
which respectively apply to sets of customers and instances of the condition for inequality (12).
VRPSD. Based on these definitions, we prove sufficient conditions for ( ) ( )
P 𝜉̃ + 𝜉𝑎 ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑎 + 𝜉𝑏 ≥ P 𝜉̃ ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑏
the recourse function to be monotonic under the DTD policy when the
𝜌𝑏 𝜌𝑏
demands follow independent Poisson, normal, binomial, Erlang, and ∑ ∑
⟺ 𝑃𝑙𝑄 (𝜌̃ + 𝜌𝑎 + 𝑗) ≥ 𝑃𝑙𝑄 (𝜌̃ + 𝑗)
negative binomial distributions. We also provide counterexamples to 𝑗=1 𝑗=1
illustrate that the recourse function is not monotonic in general.
⟸ 𝑃𝑙𝑄 (𝜌̃ + 𝜌𝑎 + 𝑗) ≥ 𝑃𝑙𝑄 (𝜌̃ + 𝑗) ∀𝑗 ∈ {1, … , 𝜌𝑏 } □

Definition 1. Let 𝑝 = (𝑖1 , … , 𝑖𝑡 ) be a path on graph 𝐺. We say that path In the proofs of Propositions 2 to 7, the expected recourse costs are
𝑝′ = (𝑗1 , … , 𝑗𝑡′ ) is a subsequence of 𝑝 if it can be obtained by removing computed for a single path orientation. This is sufficient since 1 (𝑝) ≥
elements of 𝑝, without changing the ordering of the remaining ele- 1 (𝑝′ ) and 2 (𝑝) ≥ 2 (𝑝′ ) implies that (𝑝) ≥ (𝑝′ ). We now prove
ments. Formally, 𝑝′ is thus a subsequence of 𝑝 if {𝑗1 , … , 𝑗𝑡′ } ⊆ {𝑖1 , … , 𝑖𝑡 } that, if the customers of a path respect the monotonicity condition,
and, for any pair (𝑗𝑎′ , 𝑗𝑏′ ) = (𝑖𝑎 , 𝑖𝑏 ), 𝑎′ < 𝑏′ if and only if 𝑎 < 𝑏. removing customers from this path cannot increase its recourse cost.
This is shown in Proposition 2.
Definition 2. Let 𝑝 = (𝑖1 , … , 𝑖𝑡 ) be a path on graph 𝐺. A path 𝑝′ that
can be written as 𝑝′ = (𝑖𝑎 , … , 𝑖𝑏 ) for some 1 ≤ 𝑎 ≤ 𝑏 ≤ 𝑡 is a subpath Proposition 2. For a path 𝑝 = (𝑖1 , … , 𝑖𝑡 ) composed of a set of customers
respecting the monotonicity condition, the expected recourse cost decreases
of 𝑝. In other words, 𝑝′ is a subsequence of 𝑝 that can be obtained by
monotonically when customers are removed from 𝑝, i.e., (𝑝) ≥ (𝑝′ ) for
removing elements only at the beginning and the end of 𝑝.
any subsequence 𝑝′ of 𝑝.
Definition 3. A set of customers 𝑆 ⊆ 𝑁 is said to respect the
Proof.
monotonicity condition if, for any pair (𝑎, 𝑏) ∈ 𝑆 × 𝑆 such that 𝑎 ≠ 𝑏 ( 𝑗−1 )
and for any subset 𝑆̃ ⊆ 𝑆 ⧵ {𝑎, 𝑏}, the following inequality holds for any ∑
𝑡 ∑
∞ ∑ ∑
𝑗
(𝑝) = 2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖𝑗
positive integer 𝑙 ∈ N: 𝑗=1 𝑙=1 𝑘=1 𝑘=1
( ) ( ) ( 𝑗−1 )
P 𝜉̃ + 𝜉𝑎 ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑎 + 𝜉𝑏 ≥ P 𝜉̃ ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑏 , (12) ∑
𝑡 ∑
∞ ∑ ∑
𝑗
=2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖𝑗
∑
where 𝜉̃ = 𝑖∈𝑆̃ 𝜉𝑖 . 𝑗=1
𝑖𝑗 ∈𝑝′
𝑙=1 𝑘=1 𝑘=1

( 𝑗−1 )
∑
𝑡 ∑
∞ ∑ ∑
𝑗
Definition 4. An instance of the VRPSD with the DTD policy is said to +2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖𝑗
∑
have the monotonicity property if any set 𝑆 ⊆ 𝑁 such that 𝑖∈𝑆 𝜇𝑖 ≤ 𝑄 𝑗=1 𝑙=1 𝑘=1 𝑘=1
𝑖𝑗 ∈𝑝⧵𝑝′
respects the monotonicity condition. ( 𝑗−1 )
∑
𝑡 ∑
∞ ∑ ∑
𝑗
The above definition can be applied by enumeration to verify ≥2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖𝑗
𝑗=1 𝑙=1 𝑘=1 𝑘=1
whether an instance respects the monotonicity property, but it might be 𝑖𝑗 ∈𝑝′
unpractical for large instances since the number of comparisons grows
⎛ 𝑗−1 ⎞
exponentially with |𝑁|. In Proposition 1, we introduce a sufficient con- ∑
𝑡 ∑
∞
⎜∑ ∑ 𝑗
⎟
dition which will later be used to prove that the monotonicity property ≥2 P⎜ 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 ⎟ 𝑐0𝑖𝑗
is respected if the demands are modeled by specific distributions.
𝑗=1 𝑙=1 ⎜ ′𝑘=1 𝑘=1 ⎟
𝑖𝑗 ∈𝑝′ ⎝ 𝑖𝑘 ∈𝑝 𝑖𝑘 ∈𝑝′ ⎠

523
L. Parada et al. European Journal of Operational Research 318 (2024) 520–533

( ) ( )
= (𝑝′ ) ⟺ P 𝜉̃ + 𝜉𝑎 ≤ 𝑙𝑄 − P 𝜉̃ + 𝜉𝑎 + 𝜉𝑏 ≤ 𝑙𝑄
( ) ( )
The second inequality comes from the fact that, for any 𝑙≥1 and for ≥ P 𝜉̃ ≤ 𝑙𝑄 − P 𝜉̃ + 𝜉𝑏 ≤ 𝑙𝑄 (21)
any 𝑗 ∈ {1, … , 𝑡} such that 𝑖𝑗 ∈ 𝑝′ , the following inequality holds. ( ) ( ) ( ) ( )
⟺ 𝐻 𝑙𝑄, 𝜆+𝜆̃ 𝑎 − 𝐻 𝑙𝑄, 𝜆+𝜆 ̃ 𝑎 +𝜆𝑏 ≥ 𝐻 𝑙𝑄, 𝜆̃ − 𝐻 𝑙𝑄, 𝜆+𝜆
̃ 𝑏 (22)
( ) ( )
( 𝑗−1 ) ⎛ 𝑗−1 ⎞ 𝛤 𝑙𝑄 + 1, 𝜆+𝜆̃ 𝑎 𝛤 𝑙𝑄 + 1, 𝜆+𝜆̃ 𝑎 +𝜆𝑏
∑ ∑
𝑗
⎜∑ ∑ 𝑗
⎟ ⟺ −
P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 ≥ P⎜ 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 ⎟ (14) (𝑙𝑄)! (𝑙𝑄)!
⎜ 𝑘=1′ 𝑘=1 ⎟ ( ) ( )
𝑘=1 𝑘=1
⎝ 𝑖𝑘 ∈𝑝 𝑖𝑘 ∈𝑝′ ⎠ 𝛤 𝑙𝑄 + 1, 𝜆̃ 𝛤 𝑙𝑄 + 1, 𝜆+𝜆̃ 𝑏
≥ − (23)
To prove this, we first introduce some notation: (𝑙𝑄)! (𝑙𝑄)!
( ) ( )
{ } ̃ 𝑎 − 𝛤 𝑙𝑄 + 1, 𝜆+𝜆
⟺ 𝛤 𝑙𝑄 + 1, 𝜆+𝜆 ̃ 𝑎 +𝜆𝑏
• 𝑆̃ 𝑗 = 𝑖𝑘 ∈ 𝑝′ ∶ 𝑘 ≤ 𝑗−1 is the set of customers that are visited ( ) ( )
before customer 𝑖𝑗 in path 𝑝′ ≥ 𝛤 𝑙𝑄 + 1, 𝜆̃ − 𝛤 𝑙𝑄 + 1, 𝜆+𝜆 ̃ 𝑏 (24)
∑ ∞ ∞
• 𝜉̃𝑗 = 𝑖∈𝑆̃ 𝑗 𝜉𝑖 is the total demand of these customers ⟺ 𝜆𝑙𝑄 𝑒−𝜆 𝑑𝜆 − 𝜆𝑙𝑄 𝑒−𝜆 𝑑𝜆
{ } ∫𝜆+𝜆 ∫𝜆+𝜆
• {𝑎1 , … , 𝑎𝑠 } = 𝑖𝑘 ∈ 𝑝 ∶ 𝑘 ≤ 𝑗−1, 𝑖𝑘 ∉ 𝑝′ is the set of customers ̃ 𝑎 ̃ 𝑎 +𝜆𝑏
∞ ∞
that are visited before customer 𝑖𝑗 in path 𝑝, but are not visited
≥ 𝜆𝑙𝑄 𝑒−𝜆 𝑑𝜆 − 𝜆𝑙𝑄 𝑒−𝜆 𝑑𝜆 (25)
in path 𝑝′ ∫𝜆̃ ∫𝜆+𝜆
̃ 𝑏
̃ 𝑎 +𝜆𝑏
𝜆+𝜆 ̃ 𝑏
𝜆+𝜆
We can now prove inequality (14). ⟺ 𝜆𝑙𝑄 𝑒−𝜆 𝑑𝜆 ≥ 𝜆𝑙𝑄 𝑒−𝜆 𝑑𝜆 (26)
∫𝜆+𝜆
̃ 𝑎 ∫𝜆̃
⎛ 𝑗−1 ⎞
⎜∑ ∑ 𝑗
⎟ ̃ 𝑏
𝜆+𝜆 ( )𝑙𝑄 −(𝜆+𝜆 ) ̃ 𝑏
𝜆+𝜆
P⎜ 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 ⎟ ⟺ 𝜆+𝜆𝑎 𝑒 𝑎 𝑑𝜆 ≥ 𝜆𝑙𝑄 𝑒−𝜆 𝑑𝜆 (27)
⎜ 𝑘=1′ 𝑘=1 ⎟ ∫𝜆̃ ∫𝜆̃
⎝ 𝑖𝑘 ∈𝑝 𝑖𝑘 ∈𝑝′ ⎠ ̃ 𝑏 (( )
( ) 𝜆+𝜆 )𝑙𝑄 −(𝜆+𝜆 )
⟺ 𝜆+𝜆𝑎 𝑒 𝑎 − 𝜆𝑙𝑄 𝑒−𝜆 𝑑𝜆 ≥ 0 (28)
= P 𝜉̃𝑗 ≤ 𝑙𝑄 < 𝜉̃𝑗 + 𝜉𝑖𝑗 (15) ∫𝜆̃
( ) ̃ 𝑏
≤ P 𝜉̃𝑗 + 𝜉𝑎1 ≤ 𝑙𝑄 < 𝜉̃𝑗 + 𝜉𝑎1 + 𝜉𝑖𝑗 (16)
𝜆+𝜆 ( )
⟺ 𝑓𝑙𝑄 (𝜆 + 𝜆𝑎 ) − 𝑓𝑙𝑄 (𝜆) 𝑑𝜆 ≥ 0, (29)
∫𝜆̃
⋮ (17)
( ) where 𝑓𝑙𝑄 (𝜆) ∶= 𝜆𝑙𝑄 𝑒−𝜆 . Since the inequalities 0 ≤ 𝜆 ≤ 𝜆+𝜆𝑎 ≤ 𝑄 ≤ 𝑙𝑄
≤ P 𝜉̃𝑗 + 𝜉𝑎1 + ⋯ + 𝜉𝑎𝑠 ≤ 𝑙𝑄 < 𝜉̃𝑗 + 𝜉𝑎1 + ⋯ + 𝜉𝑎𝑠 + 𝜉𝑖𝑗 (18) ̃ 𝜆+𝜆
hold for any 𝜆 ∈ [𝜆, ̃ 𝑏 ], it is sufficient to show that the function 𝑓𝑙𝑄 (⋅)
( 𝑗−1 ) is non-decreasing on [0, 𝑙𝑄] to conclude that the difference 𝑓𝑙𝑄 (𝜆 + 𝜆𝑎 ) −
∑ ∑𝑗
=P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 (19) 𝑓𝑙𝑄 (𝜆) is nonnegative everywhere on the interval of integration. The
𝑘=1 𝑘=1 derivative of function 𝑓𝑙𝑄 (⋅) is given by:
By hypothesis, the set 𝑆 = {𝑖1 , … , 𝑖𝑗 } respects the monotonicity 𝜕𝑓𝑙𝑄 (𝜆)
= (𝑙𝑄 − 𝜆)𝑒−𝜆 𝜆𝑙𝑄−1 ,
condition. This allows us to apply inequality (12) 𝑠 times for different 𝜕𝜆
pairs (𝑎, 𝑏) ∈ 𝑆 × 𝑆 and subsets 𝑆̃ ⊆ 𝑆 to obtain inequalities (16)– which is indeed nonnegative ∀ 𝜆 ∈ [0, 𝑙𝑄]. Inequality (29) is therefore
(18). At the 𝑞th application of the inequality, (𝑎, 𝑏) = (𝑎𝑞 , 𝑖𝑗 ) and 𝑆̃ = respected. □
𝑆̃ 𝑗 ∪ {𝑎1 , … , 𝑎𝑞−1 }. □
Propositions 3, 4, 5, 6 and 7 show that, under some conditions, Proposition 4. Any set 𝑆 of customers whose demands are given by
the monotonicity property is respected for customer demands following independent normal RVs 𝜉𝑖 ∼  (𝜇𝑖 , 𝜎𝑖2 ) with integer expected values 𝜇𝑖 ∈ N
𝜎2
independent Poisson, normal, binomial, Erlang, and negative binomial and sharing a common coefficient of dispersion 𝐷 = 𝜇𝑖 ≤ 1 ∀𝑖 ∈ 𝑆 with
distributions. We present next the proofs for the cases of Poisson and ∑ 𝑖
total expected value 𝑖∈𝑆 𝜇𝑖 ≤ 𝑄 respects the monotonicity condition.
normal independently distributed customer demands. The remaining
three proofs are given in Appendix A. For the important case of the
Proof. Consider a pair (𝑎, 𝑏) ∈ 𝑆 × 𝑆 such that 𝑎 ≠ 𝑏 and a subset
Poisson distribution, which is covered in Proposition 3, aside from ∑
𝑆̃ ⊆ 𝑆 ⧵ {𝑎, 𝑏} with total expected demand 𝜇̃ = 𝑖∈𝑆̃ 𝜇𝑖 . Using the
the independence of the demands, the only additional assumption we ̃ 𝜉𝑎 and 𝜉𝑏 can respectively be decomposed
notation of Proposition 1, 𝜉,
make is that the total expected demands of the customers on a route
as the sum of 𝜌= ̃ 𝜌𝑎 =𝜇𝑎 and 𝜌𝑏 =𝜇𝑏 i.i.d. RVs 𝜉𝑘 , 𝑘 ∈ N, such that
̃ 𝜇,
must respect the vehicle’s capacity. In Proposition 8, this assumption
𝜁𝑘 ∼  (𝜇=1, 𝜎 2 =𝐷). From there, the sufficient condition (13) can be
is shown to be required for the case of Poisson distributions. We
rewritten as follows.
also show in Proposition 9 that it is not sufficient in general. The
proofs of Propositions 4 to 7 directly rely on the sufficient condition 𝑃𝑙𝑄 (𝜇̃ + 𝜇𝑎 + 𝑗) ≥ 𝑃𝑙𝑄 (𝜇̃ + 𝑗) ∀𝑙 ∈ N, ∀𝑗 ∈ {1, … , 𝜇𝑏 } (30)
of Proposition 1. The assumptions made in these propositions allow
us to decompose the demands into sums of i.i.d. RVs, but may not be √ The coefficient of variation 𝑐 = 𝜎∕𝜇 of the i.i.d. RVs 𝜁𝑘 is equal to
𝐷. Since 𝐷 ≤ 1 by hypothesis, then 𝑐 ≤ 1. In this case, it follows
minimal.
from Kreimer and Dror (1990) that, for any 𝑙 ∈ N, the sequence
⌊𝑙𝑄⌋ ∑
̃ 𝑎 +𝜇𝑏 ≤ 𝑖∈𝑆 𝜇𝑖 ≤
{𝑃𝑙𝑄 (𝑠)}𝑠=1 is monotonically increasing. Since 𝜇+𝜇
Proposition 3. Any set 𝑆 of customers whose demands are given by inde-
∑ 𝑄 and all the expected demands are integer-valued, this allows to
pendent Poisson RVs 𝜉𝑖 ∼ Poisson(𝜆𝑖 ) with total expected value 𝑖∈𝑆 𝜆𝑖 ≤ 𝑄
conclude that (30) holds. □
respects the monotonicity condition.
Proposition 5. Any set 𝑆 of customers whose demands are given by inde-
Proof. Consider a pair (𝑎, 𝑏) ∈ 𝑆 × 𝑆 such that 𝑎 ≠ 𝑏, a subset
∑ pendent binomial RVs 𝜉𝑖 ∼ Bin(𝑛𝑖 , 𝑝) sharing a common success probability 𝑝
𝑆̃ ⊆ 𝑆 ⧵ {𝑎, 𝑏} with total expected demand 𝜆̃ = 𝑖∈𝑆̃ 𝜆𝑖 and a positive ∑
with total expected value 𝑖∈𝑆 𝑛𝑖 𝑝 ≤ 𝑄 respects the monotonicity condition.
integer 𝑙 ∈ N. To show that inequality (12) holds, we use the fact
that the sum of 𝑠 independent Poisson RVs with means 𝜆1 , … , 𝜆𝑠 is
∑ Proof. See Appendix A. □
Poisson distributed with mean 𝑠𝑘=1 𝜆𝑘 . Also, we use the fact that the
CDF of a Poisson RV with mean 𝜆 evaluated at 𝑙𝑄 ∈ N is given by
𝐻(𝑙𝑄, 𝜆) = 𝛤 (𝑙𝑄+1,𝜆)
∞
, where 𝛤 (𝑙𝑄+1, 𝜆) = ∫𝜆 𝑡𝑙𝑄 𝑒−𝑡 𝑑𝑡 is the upper Proposition 6. Any set 𝑆 of customers whose demands are given by
(𝑙𝑄)! independent Erlang RVs 𝜉𝑖 ∼ Erlang(𝑛𝑖 , 𝜆) sharing a common rate parameter
incomplete gamma function. ∑
( ) ( ) 𝜆 with total expected value 𝑖∈𝑆 𝑛𝑖 ∕𝜆 ≤ 𝑄 respects the monotonicity
P 𝜉̃ + 𝜉𝑎 ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑎 + 𝜉𝑏 ≥ P 𝜉̃ ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑏 (20) condition.

524
L. Parada et al. European Journal of Operational Research 318 (2024) 520–533

Proof. See Appendix A. □ master solution and are generated to prohibit infeasible solutions from
being visited. The optimality cuts are supporting linear functions of
Proposition 7. Any set 𝑆 of customers whose demands are given by (𝑥) that are added whenever a feasible solution is found to improve
independent negative binomial RVs 𝜉𝑖 ∼ NB(𝑟𝑖 , 𝑝) sharing a common success the bound on 𝜃. The feasibility cuts are generally problem-specific.
∑
probability 𝑝 with total expected value 𝑖∈𝑆 𝑟𝑖 (1−𝑝)∕𝑝 ≤ 𝑄 respects the In the context of the VRPSD, they typically correspond to subtour
monotonicity condition. elimination constraints and rounded capacity inequalities. Optimality
cuts have a general form regardless of the problem to which the integer
Proof. See Appendix A. □ L-shaped method is applied. Let 𝑥𝜈1 be the set of positive variables in
a feasible solution 𝑥𝜈 at iteration 𝜈 of the method. Then the following
The next propositions present two counterexamples to illustrate that
inequality:
the recourse function is not monotonic in general.
⎛∑ ∑ ⎞
Proposition 8. Let 𝑝 = (𝑖1 , … , 𝑖𝑡 ) be a path of customers whose demands 𝜃 ≥ ((𝑥𝜈 ) − 𝐿) ⎜ 𝑥𝑖 − 𝑥𝑖 − |𝑥𝜈1 | + 1⎟ + 𝐿, (31)
⎜ 𝜈 ⎟
are given by independent Poisson RVs 𝜉𝑖 ∼ Poisson(𝜆𝑖 ) with total expected ⎝𝑖∈𝑥1 𝑖∉𝑥𝜈1 ⎠
∑
value 𝑡𝑖=1 𝜆𝑖 > 𝑄, and 𝑝′ a subpath of 𝑝. Inequality (𝑝) ≥ (𝑝′ ) does not where (𝑥𝜈 ) is the expected recourse cost of solution 𝑥𝜈 , is a valid
hold in general. optimality cut. It was shown by Laporte and Louveaux (1993) that these
optimality cuts will yield an optimal solution, if one exists, in a finite
Proof. Consider a problem with 𝑄 = 10, a path 𝑝 = (1, 2, 3, 4), and number of steps.
𝑝′ = (2, 3, 4). Let the expected demands be 𝜆1 = 1, 𝜆2 = 1, 𝜆3 = 9, The known shortcoming of the integer L-shaped method is that
and 𝜆4 = 2. By setting 𝑐01 = 𝑐03 = 0, 𝑐02 = 2 and 𝑐04 = 1, we obtain its optimality cuts are active for a single solution. The optimality cut
that (𝑝) ≈ min{0.457, 0.478} = 0.457 < 0.478 = min{0.492, 0.478} ≈ reduces to 𝜃 ≥ (𝑥𝜈 ) at solution 𝑥𝜈 , but for any other solution, it does
(𝑝′ ). □ not improve the inequality 𝜃 ≥ 𝐿. Consequently, without appropriate
LBFs to increase the bound on 𝜃, the classical method generally iterates
Proposition 9. Let 𝑝 = (𝑖1 , … , 𝑖𝑡 ) be a path of customers whose total through many first-stage solutions before converging to an optimal
expected demand respects the vehicle capacity and 𝑝′ a subpath of 𝑝. solution.
Inequality (𝑝) ≥ (𝑝′ ) does not hold in general.
5.1. New optimality cuts
Proof. Consider a problem with 𝑄 = 10, a path 𝑝 = (1, 2, 3, 4),
and 𝑝′ = (2, 3, 4). Suppose that the demand of customer 1 is 1 with The main idea of the DL-shaped method is to replace the variable 𝜃
probability 1, the demand of customer 2 is 2 with probability 0.5 and that bounds the recourse function by a sum of variables 𝜃𝑖 , one for each
3 with probability 0.5, the demand of customer 3 is 0 with probability customer 𝑖 ∈ 𝑁. The purpose of this transformation is to express the
0.5 and 7 with probability 0.5, and the demand of customer 4 is 2 with contribution of each customer to the cost of the second stage by the
probability 1. The total expected demand on path 𝑝 is 9, which respects gradual addition of a new type of optimality cuts during the solving
the vehicle
{ capacity. However,
} the expected recourse of path 𝑝 is 𝑄(𝑝) =
{ } process. These cuts are added for feasible paths that are found, and
min 21 𝑐03 + 21 𝑐04 , 𝑐02 , and that of path 𝑝′ is 𝑄(𝑝′ ) = min 𝑐04 , 𝑐02 . By the cut of each path only involves the variables associated with the
′
taking 𝑐03 < 𝑐04 < 𝑐02 , we obtain that 𝑄(𝑝) < 𝑄(𝑝 ). □ arcs of the path and the 𝜃𝑖 variables of the customers of the path. This
makes them active for each path that is active in the current solution,
5. Disaggregated integer L-shaped method whether it is fractional or integer. They allow to bound the recourse
function more tightly and effectively than the traditional optimality
This section presents the DL-shaped method as a variant of the cuts, as those only bound one solution per cut. This gives the following
classical integer L-shaped method from Laporte and Louveaux (1993). objective function.
First, an outline of the classical method is provided, and we motivate ∑ ∑
the need for our method by exposing the known shortcoming of the min 𝑐𝑖𝑗 𝑥𝑖𝑗 + 𝜃𝑖 (32)
classical one. Next, Section 5.1 presents new optimality cuts, from (𝑖,𝑗)∈𝐸 𝑖∈𝑁

which the validity of our method results. Finally, Section 5.2 introduces The DL-shaped method starts by constructing a MP with the new ob-
the new LBFs we use to approximate the recourse cost. jective function, relaxed integrality constraints, and rounded capacity
The integer L-shaped method for two-stage recourse problems with inequalities. In each iteration of the method, when a feasible integer
binary first-stage variables starts by constructing a so-called master solution is found, an optimality cut is added for each route 𝑟 = (0, 𝑝, 0)
problem (MP). The MP is obtained from the original problem by having a positive recourse ((𝑟) > 0), where 𝑝 = (𝑖1 , … , 𝑖𝑡 ). Such a path
relaxing a set of complicating constraints. This set typically includes the 𝑝 is composed of |𝑝| = 𝑡 − 1 edges. For a first-stage solution 𝑥, let us
∑
integrality constraints and any set of constraints that is exponential in denote by 𝑥(𝑝) = 𝑡−1 𝑗=1 𝑥𝑖𝑗 𝑖𝑗+1 the number of these edges that are active.
size, like the classical subtour elimination constraints. In addition, the Also, let 𝑁(𝑟) be the set of customers that are visited in path 𝑝. The
recourse function is replaced by a continuous variable 𝜃 that bounds new optimality cut follows.
it from below, and is initially lower bounded by a finite value 𝐿. In ∑
the case of the VRPSD, the recourse cost is nonnegative and the trivial 𝜃𝑖 ≥ (𝑟) (𝑥(𝑝) − |𝑝| + 1) (33)
𝑖∈𝑁(𝑟)
bound 𝐿 = 0 can thus be used. The initial MP is defined by the non-
relaxed constraints of the original program, the lower bound 𝐿 on 𝜃, We refer to these cuts as path cuts (P-Cuts). Since they do not
and an objective function of the form min 𝑐𝑥 + 𝜃, where 𝑐 is a known include the depot edges, as opposed to the route cuts of Séguin (1996),
cost vector with the same size as the solution vector. The problem is these P-Cuts are more general. In the specific case of a route 𝑟 = (0, 𝑖, 0)
then solved through a branch-and-bound method that explores the set visiting a single customer 𝑖, inequality (33) reduces to 𝜃𝑖 ≥ (𝑟).
of feasible solutions of the original program and adds linear inequalities We now prove that the P-Cuts are valid optimality cuts. By Laporte
on the fly. The first MP that is solved in the method is denoted as and Louveaux (1993), this implies that the DL-shaped method yields an
the root node, and the following MPs are indexed by the iteration optimal solution in a finite number of steps.
at which they appear in the solving process. The constraints that are
added to the successive MPs are either feasibility or optimality cuts. Proposition 10. Eq. (33) defines a valid set of optimality cuts if the
The feasibility cuts are valid inequalities that are violated by the current problem has the monotonicity property.

525
L. Parada et al. European Journal of Operational Research 318 (2024) 520–533

Proof. Let 𝑥𝜆 be a solution satisfying constraints (2)–(8) and let 𝜆 Proof.⌈ ∑Let 𝑥⌉𝜆 be a solution satisfying constraints (2)–(8). If 𝑥𝜆 (𝑆) ≤
∑ 𝑖∈𝑆 𝜇𝑖
be the set of routes it visits. Let 𝜃 ∗ (𝑥𝜆 ) be the minimizer of 𝑖∈𝑁 𝜃𝑖𝜆 |𝑆| − − 1, then the right-hand side of (36) is non-positive and
𝑄
𝜆
over the feasible set specified by the first-stage solution 𝑥 and by the the constraint is thus respected by any feasible⌈nonnegative
∑ ⌉ solution
constraints (33) that are contained in the MP at a given iteration. The 𝑖∈𝑆 𝜇𝑖
(𝑥𝜆 , 𝜃 𝜆 ) to the MP. Otherwise, 𝑥𝜆 (𝑆) = |𝑆| − 𝑄
. This means
MP is maximally constrained if (33) has been added to the model for that solution 𝑥𝜆 uses the smallest number of vehicles allowed by the
each possible route 𝑟. Under this maximal set of optimality cuts, for capacity constraints to visit the customers of 𝑆 and that, in each route
∑
each route 𝑟 ∈ 𝜆 , the sum 𝑖∈𝑁(𝑟) 𝜃𝑖𝜆 is constrained to be larger than 𝑟 ∈ 𝜆 , the customers of 𝑆 are visited consecutively. In this case,
or equal to (̃𝑟) for each route 𝑟̃ = (0, 𝑝,
̃ 0) such that 𝑝̃ is a subpath of 𝑟. ∑
the S-Cut reduces to 𝑖∈𝑆 𝜃𝑖 ≥ 𝐿(𝑆).⌈ ∑ Let {𝑝⌉𝜆1 , … , 𝑝𝜆𝑚 } be the set of
Any other route produces a trivial bound for 𝑟, as (𝑥𝜆 (𝑝) ̃ − |𝑝|
̃ + 1) ≤ 𝜇𝑖
paths forming solution 𝑥𝜆 , where 𝑚 = 𝑖∈𝑆
. To prove the validity
0 if 𝑝̃ is not a subpath of 𝑟. Since it is assumed that the problem 𝑄 ∑
respects the monotonicity property, it follows from Proposition 2 that of constraint (36), we have to demonstrate that 𝐿(𝑆) ≤ 𝑚 𝜆
𝑗=1 (𝑝𝑗 ).
∑ 𝜆
For each 𝑗 ∈ {1, … , 𝑚}, the segment of path 𝑝𝑗 that is composed
the most restrictive constraint for 𝑟 is 𝑖∈𝑁(𝑟) 𝜃𝑖𝜆 ≥ (𝑟). Hence, by
∑ of customers of 𝑆 is a subpath 𝑝𝑗 of 𝑝𝜆𝑗 , and the set {𝑝1 , … , 𝑝𝑚 } of
definition of 𝜃 ∗ (𝑥𝜆 ), 𝑖∈𝑁(𝑟) 𝜃𝑖∗ (𝑥𝜆 ) = (𝑟) for each 𝑟 ∈ 𝜆 and
∑ ∑ ∑ ∑ these subpaths forms a partition of 𝑆. Since it is assumed that the
thus, 𝑖∈𝑁 𝜃𝑖∗ (𝑥𝜆 ) = 𝑟∈𝜆 𝑖∈𝑁(𝑟) 𝜃𝑖∗ (𝑥𝜆 ) = 𝑟∈𝜆 (𝑟) = (𝑥𝜆 ). The
𝜆 ∗ 𝜆 𝜆
objective values of (𝑥 , 𝜃 (𝑥 )) and 𝑥 for their respective problems are problem respects the monotonicity property, Proposition 2 implies that
therefore identical when the optimality cuts associated with solution (𝑝𝜆𝑗 ) ≥ (𝑝𝑗 ) ∀𝑗 ∈ {1, … , 𝑚}. By definition of the lower bound 𝐿(𝑆),
∑ ∑𝑚
𝑥𝜆 have been generated. Otherwise, the objective value of (𝑥𝜆 , 𝜃 ∗ (𝑥𝜆 )) the sequence of inequalities 𝐿(𝑆) ≤ 𝑚 𝑗=1 (𝑝𝑗 ) ≤
𝜆
𝑗=1 (𝑝𝑗 ) allows to
for the MP can only be less than or equal to that of 𝑥𝜆 for problem conclude. □
(1)–(8). □
5.3. Recourse lower bounds
As shown in Laporte and Louveaux (1993), the linear relaxation of
(1)–(8) can be improved by introducing a nontrivial lower bound on This section presents lower bounds on the recourse used by the 𝑆-
the recourse using a precomputed value 𝐿. Cuts for a given set of customers 𝑆 to be served by 𝑚 vehicles. Some
∑
𝜃𝑖 ≥ 𝐿 (34) lower bounds from the literature are first presented.
𝑖∈𝑁 To our knowledge, the first lower bound on the recourse of a given
Unfortunately, such a bound is generally close to zero for problems set 𝑆 is due to Laporte et al. (2002), where they suppose vehicles are
with an unlimited fleet, as using more vehicles reduces the probability only allowed to fail once. In the best case, the failures would happen
of having a failure. In particular, the recourse cost is close to zero for the customers of 𝑆 that are the closest to the depot. Then, the sum
of demands is seen as a continuous quantity that has to be distributed
when each vehicle only serves a single client. Therefore, it can be
among the 𝑚 vehicles. The problem is modeled as a non-linear program
useful to compute a lower bound 𝐿𝑚 on the recourse given that exactly
having for objective the minimization of the expected recourse.
𝑚 ∈ 𝑀 vehicles are being used. 𝐿𝑚 is computed using the method
defined in Section 5.3.3. This leads to the following valid lower-bound A simpler bound, which is based on the assumption that there is
inequalities. one unique large vehicle with capacity 𝑚𝑄, is proposed by Louveaux
∑ ∑ and Salazar-González (2018). In the best case, the 𝑙th failure would
𝜃𝑖 ≥ 𝐿𝑚 𝑧𝑚 (35) happen at the 𝑙th customer closest to the depot. Then, the expected
𝑖∈𝑁 𝑚∈𝑀 recourse cost can be easily calculated by summing the probability of
the 𝑙th failure multiplied by twice the 𝑙th closest distance to the depot
5.2. Lower bounding functionals
among the customers of 𝑆.
In the following, we propose three new bounds, 𝐿1 (𝑆), 𝐿2 (𝑆) and
This section presents a new type of LBFs that extends some ideas
𝐿3 (𝑆), which have the advantage of providing tighter values than those
proposed in Côté et al. (2020). Similarly to the previously introduced
of Laporte et al. (2002) and Louveaux and Salazar-González (2018). All
optimality cuts, this type of cut uses variables to bound the recourse
these bounds are compared later in Section 7.
for sets of customers 𝑆. The LBFs of Côté et al. (2020) use a different
set of variables than those used for optimality cuts, which makes the
5.3.1. Single-route lower bound
LBFs independent from the optimality cuts. Instead, in this work, we
The lower bound 𝐿1 (𝑆) introduced in this section applies to sets of
use the same 𝜃𝑖 variables for both types of cuts. This has the advantage
customers that respect the monotonicity condition and can be assigned
of bounding the recourse more effectively and for a much broader range
to a single vehicle. Let  1 (𝑆) denote the set of all possible paths
of solutions. The new LBFs bound the recourse using all the 𝜃𝑖 variables
that visit the customers of 𝑆 consecutively. Proposition 12 shows that
associated with a set of customers 𝑆. A cut is active whenever a path
sorting these customers in non-increasing order of their distance to the
visits all the customers of this set consecutively. The new LBFs are
depot produces the path of  1 (𝑆) that minimizes the recourse cost.
referred to as set cuts (S-Cuts) and are as follows:
( ⌈∑ ⌉ ) The proof is given in Appendix B. By the monotonicity condition, the
∑ 𝑖∈𝑆 𝜇𝑖 probability that a restocking trip occurs at a given customer increases
𝜃𝑖 ≥ 𝐿(𝑆) 𝑥(𝑆) − |𝑆| + +1 , (36)
𝑖∈𝑆
𝑄 with its position in the path. The idea of this bound is thus to place
∑ customers that are far from the depot at the beginning of the path to
where 𝑥(𝑆) = (𝑖,𝑗)∈𝐸(𝑆) 𝑥𝑖𝑗 . The S-Cut (36) of a set 𝑆 is based on the
avoid costly restocking trips.
least possible recourse cost of visiting the customers of 𝑆. Such a lower
bound 𝐿(𝑆) is valid if, for any set of feasible paths {𝑝1 , … , 𝑝𝑚 } that
Proposition 12. Let 𝑆 ⊆ 𝑁 be a set of customers respecting the
forms a partition of 𝑆 and uses the smallest number of vehicles allowed
monotonicity condition. Let 𝑝∗ = (𝑖∗1 , 𝑖∗2 , … , 𝑖∗|𝑆| ) ∈  1 (𝑆) be a path such
by the capacity constraints to visit the customers of 𝑆, the following
that 𝑐0𝑖∗ ≥ 𝑐0𝑖∗ ≥ ⋯ ≥ 𝑐0𝑖∗ . The expected recourse (𝑝∗ ) of path 𝑝∗ is a
inequality is respected. 1 2 |𝑆|
valid lower bound on the expected recourse (𝑝) of any path 𝑝 ∈  1 (𝑆).
∑
𝑚
𝐿(𝑆) ≤ (𝑝𝑗 ) (37)
𝑗=1 Proof. See Appendix B. □

The definition of the single-route lower bound follows.

Proposition 11. The S-Cut (36) is valid if the problem has the monotonic-
ity property. 𝐿1 (𝑆) = (𝑝∗ ) (38)

526
L. Parada et al. European Journal of Operational Research 318 (2024) 520–533

By construction, 𝐿1 (𝑆) returns the lowest possible recourse cost Then, function 𝐻𝑘 (𝑞𝑘 ) finds the optimal assignment of 𝑞𝑘 units to the
for visiting the customers of 𝑆. It is thus the highest lower bound first 𝑘 vehicles. It is defined as follows.
that can achieved without additional knowledge of the first stage. ⎧
In practice, this bound is significantly better than previous bounds ⎪∞, if 𝑞𝑘 > 𝑘𝑄,
⎪
from the literature because it considers that a failure can happen at 𝐻𝑘 (𝑞𝑘 ) = ⎨𝐺𝑘 (𝑞𝑘 ), for 𝑘 = 1 and 𝑞𝑘 ≤ 𝑄,
any customer of path 𝑝∗ , whereas those from the literature typically ⎪
⎪min0≤𝑥𝑘 ≤min{𝑞𝑘 ,𝑄} {𝐺𝑘 (𝑥𝑘 ) + 𝐻𝑘−1 (𝑞𝑘 − 𝑥𝑘 )}, if 2 ≤ 𝑘 ≤ 𝑚.
consider that failures can only happen at the last customer of a path. ⎩
Finally, the expected recourse cost of the optimal assignment of the
5.3.2. Dynamic programming algorithm demands is given by:
The lower bound proposed in this section applies to sets of cus-
𝐿2 (𝑆) = 𝐻𝑚 (𝐷) , (39)
tomers 𝑆 that can be assigned to 𝑚 ≥ 2 vehicles. It is inspired by the
∑
lower bound of Laporte et al. (2002), where the demands are seen as where 𝐷 = 𝑖∈𝑆 𝜇𝑖 . This value can be computed in 𝑂(𝑚𝐷𝑄).
continuous quantities that must be distributed among the vehicles. We
assume that 𝑄, 𝜇𝑖 and 𝜎𝑖 are integer. If a 𝜇𝑖 or a 𝜎𝑖 is fractional, we 5.3.3. A set-covering formulation
round it down; if 𝑄 is fractional, we round it up. This new bound, The last lower bound is based on a set-covering formulation, where
which requires the monotonicity of the recourse function, improves the linear relaxation is solved by column generation. Each column
upon (Laporte et al., 2002) in two ways. First, the value of the lower represents a path that visits some customers of 𝑆 and respects the
bound can be greatly increased by considering that multiple failures expected capacity constraint. The problem is to find a set of feasible
can happen for a vehicle. This is done by computing for each vehicle paths of cardinality 𝑚 that visits all customers of 𝑆 and minimizes the
𝑘 and for each quantity of expected demand 𝑞 ∈ [0, 𝑄] the least- total expected recourse cost. Let (𝑆) be the set of all feasible paths vis-
cost sequence of customers by taking advantage of Proposition 12. iting only customers of 𝑆. We define 𝑧𝑝 as a binary variable indicating
The second improvement is that our search for an assignment of the whether path 𝑝 ∈ (𝑆) is taken or not. Also, the binary coefficient 𝑏𝑗𝑝
demands can be restricted to integer solutions through the use of a new indicates whether customer 𝑗 is in path 𝑝. The formulation is as follows:
dynamic programming algorithm. The new bound is computed in two ∑
𝐿3 (𝑆) = min (𝑝)𝑧𝑝 (40)
steps.
𝑝∈(𝑆)
In the first step, the least-cost sequence for assigning 𝑞ℎ ∈ [0, 𝑄] ∑
units of expected demand to a vehicle using the |𝑆| − ℎ + 1 farthest s.t. 𝑏𝑖𝑝 𝑧𝑝 ≥ 1, 𝑖 ∈ 𝑆, (41)
𝑝∈(𝑆)
customers of 𝑆 is computed. To do so, the customers of 𝑆 are sorted ∑
by non-decreasing distance to the depot. The least-cost sequence is 𝑧𝑝 = 𝑚, (42)
computed by dynamic programming over a set of |𝑆| stages indexed 𝑝∈(𝑆)

by ℎ = 1 to |𝑆|. The state 𝑞ℎ represents the number of units of expected 𝑧𝑝 ∈ {0, 1}, 𝑝 ∈ (𝑆). (43)
demand assigned in stage ℎ. In state 𝑞ℎ , we denote the total demand
The objective function (40) minimizes the sum of the expected
of the assigned customers by the RV 𝜉𝑞ℎ . Its distribution is given by
recourse costs of the selected paths. Constraints (41) state that each
𝜉𝑞ℎ ∼ Poisson(𝑞ℎ ) if the demands are modeled as Poisson variables and
customer of 𝑆 must be visited. Constraint (42) ensures that exactly
by 𝜉𝑞ℎ ∼  (𝑞ℎ , 𝑣𝑎𝑟(𝑞ℎ )) if the demands are normally distributed. In the
∑|𝑆| ∑|𝑆| 𝑚 paths are selected, and constraints (43) define the domain of the
case of normal demands, we set 𝑣𝑎𝑟(𝑞ℎ ) = min{ 𝑖=ℎ 𝜎𝑖2 𝑧𝑖 | 𝑖=ℎ 𝜇𝑖 𝑧𝑖 =
variables.
𝑞ℎ , 𝑧𝑖 ∈ {0, 1}} to underestimate the probability of failure, which is
As the set of paths (𝑆) is too large to be generated upfront, we rely
the smallest possible variance for the total demand of the assigned
on column generation to solve the problem. An initial set of columns
customers. Function 𝐺ℎ (𝑞ℎ ) finds the value of the least-cost sequence.
is first generated to obtain a feasible solution for the linear relaxation.
Then, a pricing problem is solved to find a column having a negative
{
⎧min 𝐺ℎ+1 (𝑞ℎ ), 𝐺ℎ+1 (𝑞ℎ − 𝜇ℎ )+ reduced cost. If such a column is found, it is added to the model, and
⎪
⎪ 2𝑐 ∑∞ P(𝜉 } the linear relaxation is solved again. This process continues until no
𝑞ℎ −𝜇ℎ ≤ 𝑙𝑄 < 𝜉𝑞ℎ ) , for ℎ < |𝑆| and 𝑞ℎ ≥ 𝜇ℎ ,
⎪ 0ℎ 𝑙=1 more negative reduced cost columns can be found.
⎪
⎪𝐺ℎ+1 (𝑞ℎ ), for ℎ < |𝑆| and 𝑞ℎ < 𝜇ℎ , The pricing problem of finding the least-cost path can be formulated
𝐺ℎ (𝑞ℎ ) = ⎨
using the dynamic program of the previous section. The first equation,
⎪0, for ℎ = |𝑆| and 𝑞ℎ = 0,
⎪ when adding customer ℎ, is modified by subtracting the value of the
⎪2𝑐 P(𝜉 > 𝑄), for ℎ = |𝑆| and 𝑞ℎ = 𝜇ℎ , dual variable associated with the constraint (41) of customer ℎ. A
⎪ 0ℎ 𝑞ℎ
⎪ negative cost column is found if the value of the least-cost sequence
⎩∞, for ℎ = |𝑆| and 𝑞ℎ ∉ {0, 𝜇ℎ }.
is less than the value of the dual variable associated with constraint
The first case selects the lowest-cost assignment between the previ- (42).
ous stage in state 𝑞ℎ and the previous stage in state 𝑞ℎ − 𝜇ℎ plus the Lastly, 𝐿3 (𝑆) can be implemented for instances that do not re-
expected recourse cost paid at customer ℎ. The second case indicates spect the monotonicity property because the inequality constraints (41)
that customer ℎ cannot be used to fulfill 𝑞ℎ < 𝜇ℎ units of expected make it possible for customers to be present in multiple routes, hence
demand, only those after ℎ can thus be used. The last three cases cover allowing longer and cheaper routes to be selected if needed.
the initial stage. The farthest customer ℎ = |𝑆| can be either visited
or not, which corresponds to assigning either 0 or 𝜇ℎ units of expected 6. Implementation of the DL-shaped method
demand to the vehicle.
In the second step, the bound is obtained by computing the least- The DL-shaped method is implemented like a branch-and-cut al-
∑
cost assignment of the 𝑖∈𝑆 𝜇𝑖 units of expected demand to the vehicles. gorithm, where a mixed-integer solver is in charge of solving the
This is done by defining another dynamic program that uses the pre- linear relaxations and the branching to find integer solutions. User-
viously computed functions 𝐺ℎ (𝑞) and only allows the ℎth customer to defined methods, also known as callbacks, are automatically called
be assigned to vehicle 𝑘 if 𝑘 ≤ ℎ. This dynamic program is defined over by the solver at each node of the branch-and-bound tree to separate
𝑚 stages indexed by 𝑘 = 1 to 𝑚. The state 𝑞𝑘 represents the number violated inequalities. In our case, these include the subtour and capacity
of units of expected demand that are assigned to the first 𝑘 vehicles. constraints, the S-cuts, and the optimality cuts. At the beginning, we
At stage 𝑘, variable 𝑥𝑘 decides which quantity is assigned to vehicle 𝑘. perform the following steps to initiate the solver:

527
L. Parada et al. European Journal of Operational Research 318 (2024) 520–533

1. Compute 𝐿𝑚 = 𝐿3 (𝑁) for each number of vehicles 𝑚 ∈ 𝑀. Table 1

Relative gap of lower bounds on the instances of Jabali et al. (2014).
2. Build model (2)–(8) with the objective function (32) and add
𝑛 # LLVH02 LSG18 𝐿1 𝐿2 𝐿3 𝐿3 (s)
constraint (35).
40 30 99.9 99.9 3.5 89.7 60.0 0.9
We use two types of user-defined methods: one for fractional solu- 50 60 96.0 99.6 6.9 85.1 55.1 2.9
tions and one for integer solutions. We perform the following steps each 60 90 88.6 97.2 7.4 78.5 50.8 6.4
70 60 86.3 96.7 9.0 73.8 44.2 13.7
time our methods are called for a solution 𝑥𝜈 :
80 30 76.6 94.5 10.2 61.4 30.4 21.7
1. Verify if there are any set 𝑆 ⊆ 𝑁 that violates the subtour or the Average 89.5 97.6 7.4 77.7 48.1 8.8
capacity constraint (4) using the CVRPSEP package of Lysgaard
(2003). Add the violated inequalities (4) to the model. Addition-
Table 2
ally, for each of these 𝑆, check if its associated S-Cut is violated, Relative gap of lower bounds on the CVRP instances.
and if so, add the inequality to the model. Set # LLVH02 LSG18 𝐿1 𝐿2 𝐿3 𝐿3 (s)
2. If 𝑥𝜈 is fractional, enumerate the connected components of the
A 27 67.6 100.0 1.9 67.0 64.9 0.5
subgraph 𝐺(𝑥𝜈 ) of 𝐺 where only the non-zero-valued edges of 𝑥𝜈 B 23 58.4 99.9 0.5 57.9 55.9 0.5
are present. For each component, verify if an S-Cut is violated, E 13 60.4 99.5 4.0 59.4 56.9 0.7
and if one is found, add the S-Cut to the model. Furthermore, F 3 97.6 100.0 3.2 97.6 93.8 6.3
if the number of nodes in the component equals the number of M 5 50.8 99.8 2.2 49.3 43.3 7.0
P 24 38.5 95.5 2.0 37.2 35.2 0.4
non-zero-valued edges plus one, then the component is a path.
Average 57.1 98.8 1.9 56.3 53.9 1.0
Then, check if the P-Cut associated with the path is violated, and
if so, add it to the model.
3. If no violated inequalities were found in the previous steps, add
the violated P&S-Cuts to the model for each route in the solution. 𝜉𝑏 ∼  (𝜇𝑏 , 𝜎𝑏2 ), and 𝜉̃ ∼  (𝜇,
̃ 𝜎̃ 2 ). We ignored the case 𝑙 ≥ 4 since the
The procedure stops when all the nodes of the brand-and-bound probability of more than three restocking trips occurring is negligible.
tree have been explored. The incumbent solution is then optimal. In The second set is taken from classical CVRP instances and contains
our implementation, the S-Cut associated with a set of customers 𝑆 is the 95 instances of the sets A, B, E, F, M, and P of the CVRPLIB
generated using the lower bound 𝐿1 (𝑆) if the sum of their expected repository (Uchoa et al., 2017). In these instances, the fleets are un-
demands does not exceed the vehicle capacity. Otherwise, the lower limited, and the customer demands are modeled as Poisson RVs. The
∑ expectation 𝜇𝑖 of each customer 𝑖 ∈ 𝑁 is set to the deterministic
bound 𝐿2 (𝑆) is used, with 𝑚 = ⌈( 𝑖∈𝑆 𝜇𝑖 )∕𝑄⌉. To compute the P&S-Cuts
of Step 7, we iterate on the routes of the current solution and generate demand of the CVRP instance. Also, as done by Christiansen and
all the subpaths in which at most five customers are removed from the Lysgaard (2007), Gauvin, Desaulniers, and Gendreau (2014), and Florio
original route. The recourse cost and the single-route lower bound 𝐿1 et al. (2020), demands and the capacity of each instance are divided by
are then computed for each subpath. The five most violated P-Cuts and their greatest common divisor.
the five most violated S-Cuts are added to the model. For the new set, we generated 1980 instances using the method
of Jabali et al. (2014). To obtain a challenging set of instances, we
produced ten instances for each combination of the following prob-
7. Computational experiments
lem parameters: 𝑛 ∈ {20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120}, 𝑚̄ ∈
{2, 3, 4, 5, 6, 7}, and 𝑓̄ ∈ {0.85, 0.90, 0.95}. The coefficient of dispersion
This section analyzes the performance of the DL-shaped method of the demand of each customer 𝑖 ∈ 𝑁 is given by 𝐷 = 𝜎𝑖2 ∕𝜇𝑖 = 1.
through computational experiments. All the results were produced with By Proposition 4, these instances therefore respect the monotonicity
a C++ implementation with Cplex 12.10 on a machine with an AMD property.
Rome 7532 @ 2.40 GHz CPU. For all the computational experiments,
we defined a maximum time of one hour per instance. A quick upper
bound is also provided to the solver by applying the adaptive large 7.2. Lower bounds on the recourse
neighborhood search of Ropke and Pisinger (2006). The detailed results
can be found at https://sites.google.com/view/jfcote/. This section compares our new lower bounds on the recourse to
Section 7.1 describes the existing instances that are used in our those previously used in the literature based on the relative gap they
experiments and introduces a new set of instances. In Section 7.2, the provide when applied to an optimal first-stage solution. This measure
new lower bounds on the recourse are compared with classical bounds is defined as follows.
from the literature. Sections Section 7.3 compare our algorithm to Recourse in the best or optimal solution − lower bound
Gap% = × 100
previous methods on two different instance sets. Finally, Section 7.4 Recourse in the best or optimal Solution
presents the performance of our method on the new set of instances
For the first set of instances of Jabali et al. (2014), the results are
and summarizes the key takeaways of our experiments.
grouped by number of customers, whereas the results are presented
separately for each set A, B, E, M, and P for the CVRP instances. In
7.1. Characterization of the instances Tables 1 and 2, the columns LLVH02 and LSG18 present the gaps ob-
tained from the implementation of the lower bound from Laporte et al.
Computational experiments are based on three sets of instances: two (2002) and Louveaux and Salazar-González (2018), respectively. The
from the literature and a new one. The first set is taken from Jabali et al. performance of our three lower bounds, 𝐿1 , 𝐿2 , and 𝐿3 is also reported.
∑
(2014) and contains 270 instances. In each instance, 𝑛 ranges from 40 to Column 𝐿1 presents the gaps obtained from computing ∗𝑟∈ 𝐿1 (𝑟),
80, and the demands are normally distributed, with parameters (𝜇𝑖 , 𝜎𝑖 ) ∗
where  is the set of routes obtained in either the optimal or best
such that 𝜇𝑖 = 3𝜎𝑖 and 𝜇𝑖 ∈ {1, … , 10}. The number
∑ of vehicles is fixed solution found at the end of the DL-shaped method. Columns 𝐿2 and 𝐿3
𝜇𝑖
to 𝑚̄ ∈ {2, 3, 4} and the filling coefficient 𝑓̄ = 𝑖∈𝑁
𝑚𝑄
̄
ranges from 0.80 are obtained by computing 𝐿2 (𝑁) and 𝐿3 (𝑁). The average computing
to 0.95. The DL-shaped method can be used on this set since it respects time of 𝐿3 (𝑁) is also reported. The computing time of all the other
the monotonicity property. We verified that inequality (12) holds for bounds is negligible.
each capacity 𝑄 in this set of instances, for each number of restocking The lower bound of Louveaux and Salazar-González (2018) is sys-
trips 𝑙 ∈ {1, 2, 3}, and for each possible distribution of 𝜉𝑎 ∼  (𝜇𝑎 , 𝜎𝑎2 ), tematically the weakest, followed by that of Laporte et al. (2002). Their

528
L. Parada et al. European Journal of Operational Research 318 (2024) 520–533

Table 3
Performance comparison with Hoogendoorn and Spliet (2023), Jabali et al. (2014) on the instances of Jabali et al. (2014).
n 𝑚̄ JRGL14 HS23 DL-Shaped
Opt Sec Gap% Opt Sec Gap% Opt Sec
40 4 9 1240.7 1.5 28 91.9 1.6 30 2.0
50 3 16 6918.0 0.7 29 101.5 2.9 30 9.8
50 4 5 1360.8 1.9 25 224.8 1.7 30 30.9
60 2 24 1393.0 0.4 30 72.2 – 30 1.2
60 3 6 2766.0 0.7 27 191.4 1.2 30 16.7
60 4 3 4922.0 2.0 25 262.1 1.9 30 25.3
70 2 17 2577.5 0.5 30 99.2 – 30 2.0
70 3 9 1753.3 1.5 24 563.7 1.5 30 46.2
80 2 13 1809.2 0.5 28 193.7 1.0 30 19.8
Total or avg. 102 2711.4 1.2 246 185.7 1.6 270 17.1

average gaps increase with both the fleet size and the number of cus- instances, and some use the OR policy instead of the DTD policy. Never-
tomers. The bound of Louveaux and Salazar-González (2018) assumes theless, the fact that the DL-shaped method can solve more than a third
a single large vehicle and therefore becomes a poor approximation of of the 95 instances of this set is a new milestone for a B&C approach. In-
the recourse as the fleet size increases. Regarding the bound of Laporte deed, the previous integer L-shaped algorithms from the literature have
et al. (2002), it suffers from an underestimation of the variance of the not been successful in solving these instances. Jabali et al. (2014), Lou-
total demand on a route, which worsens as the number of customers veaux and Salazar-González (2018), Salavati-Khoshghalb et al. (2019)
grows. did not report any results, and HS23 can only solve 14 of them
These results demonstrate that our new bounds significantly im- under the OR policy. Our method achieves comparable performance as
prove over the previous bounds from the literature. The set covering GDG14 for the B and E sets and generally produces solutions of good
bound 𝐿3 achieves an average gap of approximately 50% on average on quality, with the average gaps being at most 10% for each set. Also, the
both sets of instances. The DP bound 𝐿2 is generally slightly less tight DL-shaped method can solve two of the three instances of set F. In one
than 𝐿3 but is significantly cheaper to compute. The good computa- of those, the vehicle capacity is 30,000. For comparison, the highest
tional performance of 𝐿2 can be attributed to the fact that the least-cost vehicle capacity is 8,000 among the instances solved to optimality by
sequence can be computed by sorting the customers according to their HS23, and 3,000 for FHM20. These results illustrate the advantage of
distance to the depot. The single-route lower bound 𝐿1 is by far the the DL-shaped method over previous algorithms from the literature
tightest, with an average gap of 7.4% for the first set and 1.4% for for instances with very high vehicle capacities and, more generally,
the second. For a given set of customers, the quality of bounds 𝐿1 that B&C methods are less sensitive than B&P methods to large vehicle
and 𝐿2 directly depends on the similarity of the least-cost sequence capacities and expected customer demands. The last aspect in which
for the recourse cost provided by Proposition 12 and the total least- our method stands out is the length of the routes that can be handled.
cost sequence that takes the first-stage routing cost into account. In The last row of Table 4 reports the maximal ratio of customers to
particular, the performance of bound 𝐿1 reported in Table 2 indicates vehicles among the instances that can be solved to optimality by each
the substantial similarity of these sequences for the CVRP instances. method. This value is 25 for the DL-shaped method, more than twice
that of FHM20 and HS23. It was achieved on an instance of set P
7.3. Results for the instances of the literature that contains 100 customers and whose optimal solution comprises four
routes, the longest of which visits 29 customers.
This section reports results on the two sets of instances of the
literature. It compares the efficiency of the DL-shaped method against
other B&C methods and against B&P methods. 7.4. Results for the new instances
Table 3 presents the results against the integer L-shaped methods
of Jabali et al. (2014) and Hoogendoorn and Spliet (2023). The perfor- This section presents the results for our new set of instances. For
mance of the three methods is presented under the columns JRGL14, each number of customers and vehicles, Table 5 reports the number
HS23, and DL-Shaped of Table 3. Each row of Table 3 contains 30 of instances, out of 30, that could be solved within the time limit
instances with the same number of customers 𝑛 and the same fleet size and their average solving time. In total, 1175 instances were solved,
̄ The number of optimal solutions found by each method and their
𝑚. and the average optimality gap of the 805 unsolved instances was less
average computing time are reported, as well as the average optimality than 5% in each group. Even for as many as 120 customers, instances
gap of the unsolved instances. with two vehicles were generally solved easily, with an average time
Table 3 indicates that the DL-shaped method achieves state-of-the- of one minute per solved instance. Although the difficulty of the new
art results by solving to optimality all the 270 instances of the first instances appears to increase with the number of vehicles, some of the
set, while Jabali et al. (2014) and Hoogendoorn and Spliet (2023) instances that were solved by the DL-shaped method are significantly
respectively solve 102 and 246 instances. Furthermore, our average larger than those previously found in the literature. These include two
solving time is 17.1 s per instance, compared to 2711.4 and 185.72 s instances with 50 customers and seven vehicles and one instance with
for the instances that were solved to optimality by Jabali et al. (2014) 120 customers and five vehicles.
and Hoogendoorn and Spliet (2023), respectively. These results high- Fig. 1 summarizes the parameters of the largest instances that
light the advantage of disaggregating the recourse in the context of could be solved and the smallest instances that remained unsolved
the integer L-shaped method, as well as the capacity of our LBFs to by different B&C methods from the literature. For each number of
approximate the recourse function efficiently. customers and each number of vehicles, the highest filling coefficient 𝑓̄
In Table 4, the DL-shaped method is compared to the algorithms for which instances have been solved and the smallest filling coefficient
of Christiansen and Lysgaard (2007) (CL07), Gauvin et al. (2014) leading to instances that could not be solved are reported. To increase
(GDG14), Florio et al. (2020) (FHM20), and Hoogendoorn and Spliet readability, we only present what we consider to be the most impor-
(2023) (HS23) on the second set of instances. tant configurations per paper. We also indicate whether the customer
A rigorous comparison of these algorithms is difficult to achieve, demands were modeled as deterministic (D), Poisson RVs (P), or normal
as the previous works do not all report results for the complete set of RVs (N) in each set of instances. Results from Jabali et al. (2014) are

529
L. Parada et al. European Journal of Operational Research 318 (2024) 520–533

Table 4
Performance comparison with Christiansen and Lysgaard (2007), Florio et al. (2020), Gauvin et al. (2014) on the CVRP instances.
Set # CL07 GDG14 FHM20a HS23a DL-Shaped Gap% Time (s)
A 27 6/19 22/26 15/27 1 9 6.8 202.1
B 23 – 7/23 – – 6 4.3 1067.1
E 13 2/3 5/11 3/8 6 7 7.3 127.2
F 3 – – – – 2 1.8 18.7
M 5 – 1/1 – – 0 9.9 –
P 24 11/18 20/22 17/23 7 11 6.0 143.9
Total or avg. 95 19/40 55/83 35/58 14 35 6.1 306.7
Max. 𝑛∕𝑚 6.3 18.5 10.5 10.5 25
a
Indicate that the authors use the OR policy.

Table 5
Solved instances and computation times of the DL-shaped method for the new set.
𝑛∖𝑚̄ Number of optimal solutions Average computing times (s)
2 3 4 5 6 7 All 2 3 4 5 6 7 All
20 30 30 30 30 30 30 180 0.0 0.4 1.3 7.9 32.4 10.7 8.8
30 30 30 30 30 24 20 164 0.1 0.6 5.6 276.1 807.3 863.9 275.2
40 30 30 30 28 13 3 134 0.3 1.4 42.2 498.0 867.0 1718.3 236.5
50 30 30 28 23 4 2 117 0.6 45.1 232.2 516.1 913.6 1297.1 222.1
60 30 30 24 17 3 0 104 3.9 12.3 275.3 1114.8 1926.9 – 306.0
70 30 30 22 7 1 0 90 1.6 286.1 272.3 1653.7 400.0 – 295.5
80 30 30 19 10 2 0 91 17.7 198.1 1024.2 1455.7 1667.5 – 481.6
90 30 28 19 6 1 0 84 15.8 233.3 710.4 611.6 2184.9 – 313.8
100 30 25 18 6 0 0 79 16.3 396.9 956.0 1382.5 – – 454.6
110 30 26 9 5 0 0 70 101.9 520.4 234.7 1484.3 – – 373.2
120 29 21 11 1 0 0 62 62.2 306.3 317.4 1468.2 – – 212.8
All 329 310 240 163 78 55 1175 19.9 170.1 318.1 615.2 602.2 460.9 262.3

Fig. 1. Largest solved and smallest unsolved instances in the literature with their filling coefficient. .

omitted because they are slightly inferior to those of Hoogendoorn and 8. Conclusion
Spliet (2023).
Although the figure should be taken with a grain of salt, these This paper presents a new approach for solving a class of stochastic
results indicate that the DL-shaped method constitutes a significant integer programs in which the first-stage solutions can be decomposed
advancement in the ability to solve instances with a high number of into disjoint components with monotonic recourse. The the DL-shaped
customers and vehicles, and a high filling coefficient. Previous methods method is applied to the vehicle routing problem with stochastic de-
from the literature can solve some instances with a relatively high mands under the detour-to-depot recourse policy. Our computational
number of customers, yet they also fail to solve instances with as few experiments show that it achieves state-of-the-art results on instances
as 30 or 40 customers. from the literature and is particularly efficient on instances with long

530
L. Parada et al. European Journal of Operational Research 318 (2024) 520–533

routes and few vehicles. Our approach is based on new optimality cuts such that 𝜁𝑘 ∼ Exp(𝜆). From there, the sufficient condition (13) can be
and exploits new lower bounds on the recourse that are used by new rewritten as follows.
lower bounding functionals. Computational experiments show that our
lower bounds improve over the existing ones from the literature. 𝑃𝑙𝑄 (𝑛̃ + 𝑛𝑎 + 𝑗) ≥ 𝑃𝑙𝑄 (𝑛̃ + 𝑗) ∀𝑙 ∈ N, ∀𝑗 ∈ {1, … , 𝑛𝑏 } (A.2)
Regarding future lines of research, the DL-shaped method could be
It follows from Kreimer and Dror (1990) that, for any 𝑙 ∈ N, the
applied to other variants of the stochastic vehicle routing problem and ⌈𝜆𝑙𝑄⌉ ∑
other two-stage stochastic integer programs whose recourse present a ̃ 𝑎 +𝑛𝑏 ≤ 𝑖∈𝑆 𝑛𝑖 ≤
sequence {𝑃𝑙𝑄 (𝑠)}𝑠=1 is non-decreasing. Since 𝑛+𝑛
monotonic structure. Since the efficiency of this method depends on the 𝜆𝑄, this allows to conclude that (A.2) holds. □
availability of tight bounds on disjoint components of the recourse, this
work may motivate the derivation of such bounds for other problems of Proposition 7. Any set 𝑆 of customers whose demands are given by
independent negative binomial RVs 𝜉𝑖 ∼ NB(𝑟𝑖 , 𝑝) sharing a common
interest. Finally, it would be interesting to conduct a systematic study ∑
of the monotonicity property for other two-stage stochastic programs. success probability 𝑝 with total expected value 𝑖∈𝑆 𝑟𝑖 (1−𝑝)∕𝑝 ≤ 𝑄
respects the monotonicity condition.
CRediT authorship contribution statement
Proof. Consider a pair (𝑎, 𝑏) ∈ 𝑆 × 𝑆 such that 𝑎 ≠ 𝑏, a subset 𝑆̃ ⊆
∑
Lucas Parada: Data curation, Formal analysis, Investigation, 𝑆 ⧵ {𝑎, 𝑏} with total expected demand 𝑟̃(1−𝑝)∕𝑝, where 𝑟̃ = 𝑖∈𝑆̃ 𝑟𝑖 and
Methodology, Writing – original draft. Robin Legault: Formal a positive integer 𝑙 ∈ N. We use the fact that the sum of 𝑠 independent
analysis, Methodology, Writing – original draft, Writing – review negative binomial RVs with parameters 𝑟1 , … , 𝑟𝑠 and the same success
probability 𝑝 is also a negative binomial RV with parameters 𝑟 =
& editing. Jean-François Côté: Conceptualization, Formal analysis, ∑𝑠
Funding acquisition, Investigation, Methodology, Project administra- 𝑘=1 𝑟𝑘 and 𝑝 to rewrite inequality (12) as follows.
tion, Supervision, Validation, Writing – original draft, Writing – ( ) ( )
P 𝜉̃ + 𝜉𝑎 ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑎 + 𝜉𝑏 ≥ P 𝜉̃ ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑏 (A.3)
review & editing. Michel Gendreau: Formal analysis, Methodology, (𝑙𝑄+̃𝑟+𝑟 ) (𝑙𝑄+̃𝑟 )
Project administration, Supervision, Writing – review & editing. ∑𝑎 ∑
𝑙𝑄+̃𝑟+𝑟𝑎 +𝑟𝑏
∑ ∑
𝑙𝑄+̃𝑟+𝑟𝑏
⟺P 𝜁𝑘 ≤ 𝑙𝑄 < 𝜁𝑘 ≥ P 𝜁𝑘 ≤ 𝑙𝑄 < 𝜁𝑘
𝑘=1 𝑘=1 𝑘=1 𝑘=1
Acknowledgments (A.4)

We thank the Digital Research Alliance of Canada for providing To obtain inequality (A.4), we use the fact that the CDF of a negative
high-performance computing facilities. Financial support for this work binomial with parameters 𝑟 and 𝑝 evaluated at 𝑙𝑄 ∈ N is equal to
was provided by the Canadian Natural Sciences and Engineering Re- the CDF of a binomial distribution with 𝑙𝑄 + 𝑟 trials and probability
search Council (NSERC) under Grant 2021-04037. This support is of success 1 − 𝑝 evaluated at 𝑙𝑄. In inequality (A.4), these binomial
gratefully acknowledged. RVs are decomposed into sums of i.i.d. RVs 𝜁𝑘 ∼ Bern(1−𝑝). Using this
Appendix A. Proofs of Proposition 5, 6 and 7 reformulation of the monotonicity condition based on the RVs 𝜁𝑘 , it
suffices to show that 𝑃𝑙𝑄 (𝑙𝑄 + 𝑟̃ + 𝑟𝑎 + 𝑗) ≥ 𝑃𝑙𝑄 (𝑙𝑄 + 𝑟̃ + 𝑗) ∀𝑗 ∈ {1, … , 𝑟𝑏 }
This appendix presents the proofs that the monotonicity property is to conclude the proof, by Proposition 1. For a given 𝑗 ∈ {1, … , 𝑟𝑏 }, this
respected, under some conditions, when customer demands follow in- inequality can be developed as follows:
dependent binomial, Erlang and negative binomial probability distribu-
tions. For the sake of completeness, we also provide the corresponding 𝑃𝑙𝑄 (𝑙𝑄 + 𝑟̃ + 𝑟𝑎 + 𝑗) ≥ 𝑃𝑙𝑄 (𝑙𝑄 + 𝑟̃ + 𝑗) (A.5)
(𝑙𝑄+̃𝑟+𝑟 +𝑗−1 )
propositions. ∑𝑎 𝑙𝑄+̃𝑟+𝑟𝑎 +𝑗
∑
⟺P 𝜁𝑘 ≤ 𝑙𝑄 < 𝜁𝑘
Proposition 5. Any set 𝑆 of customers whose demands are given by 𝑘=1 𝑘=1
(𝑙𝑄+̃𝑟+𝑗−1 )
independent binomial RVs 𝜉𝑖 ∼ Bin(𝑛𝑖 , 𝑝) sharing a common success ∑ ∑𝑟+𝑗
𝑙𝑄+̃
∑ ≥P 𝜁𝑘 ≤ 𝑙𝑄 < 𝜁𝑘 (A.6)
probability 𝑝 with total expected value 𝑖∈𝑆 𝑛𝑖 𝑝 ≤ 𝑄 respects the
𝑘=1 𝑘=1
monotonicity condition. ( ) ( )
⟺ P 𝜒 ≤ 𝑙𝑄 < 𝜒 + 𝜁1 ≥ P 𝜒̄ ≤ 𝑙𝑄 < 𝜒̄ + 𝜁1 , (A.7)
Proof. Consider a pair (𝑎, 𝑏) ∈ 𝑆 × 𝑆 such that 𝑎 ≠ 𝑏 and a subset
∑ where 𝜒 ∼ Bin(𝑙𝑄 + 𝑟̃+𝑟𝑎 +𝑗−1, 1−𝑝) and 𝜒̄ ∼ Bin(𝑙𝑄+̃𝑟+𝑗−1, 1−𝑝). Note
𝑆̃ ⊆ 𝑆 ⧵ {𝑎, 𝑏} with total expected demand 𝑛𝑝,̃ where 𝑛̃ = 𝑖∈𝑆̃ 𝑛𝑖 . Using
̃ that the event in the left-hand side probability of inequality (A.7) occurs
the notation of Proposition 1, the demands 𝜉, 𝜉𝑎 and 𝜉𝑏 can respectively
if and only if 𝜒=𝑙𝑄 and 𝜁1 =1. Since 𝜒 and 𝜁1 are independent, the joint
be decomposed as the sum of 𝜌= ̃ 𝜌𝑎 =𝑛𝑎 and 𝜌𝑏 =𝑛𝑏 i.i.d. RVs 𝜉𝑘 , 𝑘 ∈ N,
̃ 𝑛,
probability of these two events is the product of their probabilities.
such that 𝜁𝑘 ∼ Bern(𝑝). From there, the sufficient condition (13) can be
Applying the same reasoning to the left-hand side, inequality (A.7) can
rewritten as follows.
be written as follows.
𝑃𝑙𝑄 (𝑛̃ + 𝑛𝑎 + 𝑗) ≥ 𝑃𝑙𝑄 (𝑛̃ + 𝑗) ∀𝑙 ∈ N, ∀𝑗 ∈ {1, … , 𝑛𝑏 } (A.1) ( ) ( )
P (𝜒 = 𝑙𝑄) P 𝜁1 = 1 ≥ P (𝜒̄ = 𝑙𝑄) P 𝜁1 = 1 (A.8)
It follows from⌈ Kreimer
⌉ and Dror (1990) that, for any 𝑙 ∈ N, the
𝑙𝑄 ⟺ P (𝜒 = 𝑙𝑄) ≥ P (𝜒̄ = 𝑙𝑄) (A.9)
𝑝 ∑
̃ 𝑎 +𝑛𝑏 ≤ 𝑖∈𝑆 𝑛𝑖 ≤
sequence {𝑃𝑙𝑄 (𝑠)}𝑠=1 is non-decreasing. Since 𝑛+𝑛 (𝑙𝑄 + 𝑟̃ + 𝑟𝑎 + 𝑗 − 1)!
𝑄∕𝑝, this allows to conclude that (A.1) holds. □ ⟺ (1 − 𝑝)𝑙𝑄 𝑝𝑟̃+𝑟𝑎 +𝑗−1
(𝑙𝑄)!(̃𝑟 + 𝑟𝑎 + 𝑗 − 1)!
(𝑙𝑄 + 𝑟̃ + 𝑗 − 1)!
Proposition 6. Any set 𝑆 of customers whose demands are given ≥ (1 − 𝑝)𝑙𝑄 𝑝𝑟̃+𝑗−1 (A.10)
(𝑙𝑄)!(̃𝑟 + 𝑗 − 1)!
by independent Erlang RVs 𝜉𝑖 ∼ Erlang(𝑛𝑖 , 𝜆) sharing a common rate ∏𝑟𝑎
∑ (𝑙𝑄 + 𝑟̃ + 𝑖 + 𝑗 − 1)
parameter 𝜆 with total expected value 𝑖∈𝑆 𝑛𝑖 ∕𝜆 ≤ 𝑄 respects the ⟺ 𝑖=1∏𝑟𝑎 𝑝𝑟𝑎 ≥ 1 (A.11)
monotonicity condition. 𝑖=1
(̃
𝑟 + 𝑖 + 𝑗 − 1)
𝑟𝑎 ( )
∏ 𝑝(𝑙𝑄 + 𝑟̃ + 𝑖 + 𝑗 − 1)
Proof. Consider a pair (𝑎, 𝑏) ∈ 𝑆 × 𝑆 such that 𝑎 ≠ 𝑏 and a subset ⟺ ≥1 (A.12)
∑ 𝑖=1
𝑟̃ + 𝑖 + 𝑗 − 1
𝑆̃ ⊆ 𝑆 ⧵{𝑎, 𝑏} with total expected demand 𝑛∕𝜆,
̃ where 𝑛̃ = 𝑖∈𝑆̃ 𝑛𝑖 . Using
the notation of Proposition 1, the demands 𝜉, ̃ 𝜉𝑎 and 𝜉𝑏 can respectively To complete the proof, we demonstrate that inequality (A.12) holds
be decomposed as the sum of 𝜌= ̃ 𝜌𝑎 =𝑛𝑎 and 𝜌𝑏 =𝑛𝑏 i.i.d. RVs 𝜉𝑘 , 𝑘 ∈ N,
̃ 𝑛, by showing that each term in the product is greater than or equal to 1.

531
L. Parada et al. European Journal of Operational Research 318 (2024) 520–533

∑
∞ ( )
For 𝑖 ∈ {1, … , 𝑟𝑎 }, we have: +2 P 𝜉̃ ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑖ℎ+1 𝑐0𝑖ℎ+1
𝑝(𝑙𝑄 + 𝑟̃ + 𝑖 + 𝑗 − 1) 𝑝𝑙𝑄 𝑙=1
= +𝑝 (A.13) ( )
𝑟̃ + 𝑖 + 𝑗 − 1 𝑟̃ + 𝑖 + 𝑗 − 1 ∑
∞
+2 P 𝜉̃ + 𝜉𝑖ℎ+1 ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑖ℎ+1 + 𝜉𝑖ℎ 𝑐0𝑖ℎ
𝑝𝑄
≥ +𝑝 (A.14) 𝑙=1
𝑟̃ + 𝑟𝑎 + 𝑟𝑏 |𝑆|
( 𝑗−1 )
∑ ∑
∞ ∑ ∑
𝑗
𝑝𝑄 +2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖𝑗
≥ ∑ +𝑝 (A.15)
𝑖∈𝑆 𝑟𝑖 𝑗=ℎ+2 𝑙=1 𝑘=1 𝑘=1
(1 − 𝑝)𝑝𝑄 Since the expected recourse cost at customers 𝑖1 to 𝑖ℎ−1 and at customers
≥ +𝑝 (A.16)
𝑝𝑄 𝑖ℎ+2 to 𝑖|𝑆| is identical in both paths, the inequality (𝑝) ≥ (𝑝′ )
=1 (A.17) simplifies to:
To obtain inequality (A.14), we notice that 𝑙 ≥ 1 and 𝑖+𝑗−1 ≤ 𝑟𝑎 +𝑟𝑏 −1 ≤ (𝑝) ≥ (𝑝′ )
𝑟𝑎 +𝑟𝑏 . Inequality (A.16) follows from the assumption that the total ∑
∞
( )
expected demand of the customers of set 𝑆 does not exceed the vehicle ⟺2 P 𝜉̃ ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑖ℎ 𝑐0𝑖ℎ
𝑙=1
capacity. □
∑
∞ ( )
+2 P 𝜉̃ + 𝜉𝑖ℎ ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑖ℎ + 𝜉𝑖ℎ+1 𝑐0𝑖ℎ+1
Appendix B. Proof of Proposition 12 𝑙=1
∑
∞ ( )
≥2 P 𝜉̃ ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑖ℎ+1 𝑐0𝑖ℎ+1
This appendix presents the proof that arranging a set of customers 𝑙=1
adhering to the monotonicity property, in non-increasing order based ∑
∞ ( )
on their individual distances to the depot, provides a lower bound for +2 P 𝜉̃ + 𝜉𝑖ℎ+1 ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑖ℎ+1 + 𝜉𝑖ℎ 𝑐0𝑖ℎ
𝑙=1
the expected recourse cost.
∞ ( ( ))
∑ ( )
⟺ P 𝜉̃ ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑖ℎ − P 𝜉̃ + 𝜉𝑖ℎ+1 ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑖ℎ+1 + 𝜉𝑖ℎ 𝑐0𝑖ℎ
Proposition 12. Let 𝑆 ⊆ 𝑁 be a set of customers respecting the 𝑙=1
∞ ( ( ) ( ))
monotonicity condition. Let 𝑝∗ = (𝑖∗1 , 𝑖∗2 , … , 𝑖∗|𝑆| ) ∈  1 (𝑆) be a path such ∑
that 𝑐0𝑖∗ ≥ 𝑐0𝑖∗ ≥ ⋯ ≥ 𝑐0𝑖∗ . The expected recourse (𝑝∗ ) of path 𝑝∗ is a ≥ P 𝜉̃ ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑖ℎ+1 − P 𝜉̃ + 𝜉𝑖ℎ ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑖ℎ + 𝜉𝑖ℎ+1 𝑐0𝑖ℎ+1
1 2 |𝑆| 𝑙=1
valid lower bound on the expected recourse (𝑝) of any path 𝑝 ∈  1 (𝑆). ∞ (
∑ ( ) ( ) ( )
⟺ P 𝜉̃ + 𝜉𝑖ℎ > 𝑙𝑄 − P 𝜉̃ > 𝑙𝑄 − P 𝜉̃ + 𝜉𝑖ℎ+1 + 𝜉𝑖ℎ > 𝑙𝑄
𝑙=1
Proof. Let us consider a path 𝑝 = (𝑖1 , … , 𝑖|𝑆| ) ∈  1 (𝑆) in which the ( ))
customers are not sorted in non-increasing order of distances to the + P 𝜉̃ + 𝜉𝑖ℎ+1 > 𝑙𝑄 𝑐0𝑖ℎ
∞ ( ( ) ( )
depot. In particular, this implies that there exists at least one index ∑ ( )
ℎ ∈ {1, … , |𝑆|−1} such that 𝑐0𝑖ℎ < 𝑐0𝑖ℎ+1 . We show that the expected ≥ P 𝜉̃ + 𝜉𝑖ℎ+1 > 𝑙𝑄 − P 𝜉̃ > 𝑙𝑄 − P 𝜉̃ + 𝜉𝑖ℎ + 𝜉𝑖ℎ+1 > 𝑙𝑄
𝑙=1
recourse of path 𝑝′ = (𝑖1 , … , 𝑖ℎ+1 , 𝑖ℎ , … , 𝑖|𝑆| ) ∈  1 (𝑆), in which the order ( ))
+ P 𝜉̃ + 𝜉𝑖ℎ > 𝑙𝑄 𝑐0𝑖ℎ+1
of visit of customers 𝑖ℎ and 𝑖ℎ+1 is inverted, is less than or equal to that
∞ ( ( )
of path 𝑝. ∑ ( ) ( )
⟺ −P 𝜉̃ + 𝜉𝑖ℎ ≤ 𝑙𝑄 + P 𝜉̃ ≤ 𝑙𝑄 + P 𝜉̃ + 𝜉𝑖ℎ+1 + 𝜉𝑖ℎ ≤ 𝑙𝑄
We first develop the expression of (𝑝): 𝑙=1
( 𝑗−1 ) ( ))
|𝑆| ∑
∑ ∞ ∑ ∑ 𝑗 − P 𝜉̃ + 𝜉𝑖ℎ+1 ≤ 𝑙𝑄 𝑐0𝑖ℎ
(𝑝) = 2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖𝑗 ∞ ( ( ) ( )
∑ ( )
𝑗=1 𝑙=1 𝑘=1 𝑘=1
( 𝑗−1 ) ≥ −P 𝜉̃ + 𝜉𝑖ℎ+1 ≤ 𝑙𝑄 + P 𝜉̃ ≤ 𝑙𝑄 + P 𝜉̃ + 𝜉𝑖ℎ + 𝜉𝑖ℎ+1 ≤ 𝑙𝑄
∑∑
ℎ−1 ∞ ∑ ∑
𝑗
(
𝑙=1
))
=2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖𝑗 − P 𝜉̃ + 𝜉𝑖ℎ ≤ 𝑙𝑄 𝑐0𝑖ℎ+1
𝑗=1 𝑙=1 𝑘=1 𝑘=1
(ℎ−1 ) ∑
∞
( ( ) ( )
∑
∞ ∑ ∑
ℎ ⟺ (𝑐0𝑖ℎ+1 − 𝑐0𝑖ℎ ) P 𝜉̃ + 𝜉𝑖ℎ ≤ 𝑙𝑄 − P 𝜉̃ ≤ 𝑙𝑄
+2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖ℎ 𝑙=1
( ) ( ))
𝑙=1 𝑘=1 𝑘=1
( ) − P 𝜉̃ + 𝜉𝑖ℎ+1 + 𝜉𝑖ℎ ≤ 𝑙𝑄 + P 𝜉̃ + 𝜉𝑖ℎ+1 ≤ 𝑙𝑄
∑
∞ ∑
ℎ ∑
ℎ+1
+2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖ℎ+1 ≥0
𝑙=1 𝑘=1 𝑘=1 ( ) ( )
|𝑆| ∑
( 𝑗−1 ) ⟸ P 𝜉̃ + 𝜉𝑖ℎ ≤ 𝑙𝑄 − P 𝜉̃ + 𝜉𝑖ℎ+1 + 𝜉𝑖ℎ ≤ 𝑙𝑄
∑ ∞ ∑ ∑
𝑗
( )
+2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖𝑗 ( )
≥ P 𝜉̃ ≤ 𝑙𝑄 − P 𝜉̃ + 𝜉𝑖ℎ+1 ≤ 𝑙𝑄 , ∀𝑙 ≥ 1
𝑗=ℎ+2 𝑙=1 𝑘=1 𝑘=1
( 𝑗−1 )
∑∑
ℎ−1 ∞ ∑ ∑
𝑗 Since 𝑆 respects the monotonicity condition, the last inequality,
=2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖𝑗 which corresponds to inequality (12) with 𝑆̃ = {𝑖1 , … , 𝑖ℎ−1 } and (𝑎, 𝑏) =
𝑗=1 𝑙=1 𝑘=1 𝑘=1
(𝑖ℎ , 𝑖ℎ+1 ), is verified for any 𝑙 ∈ N.
∑∞ ( )
This implies that, starting from any path 𝑝 ∈  1 (𝑆), one can
+2 P 𝜉̃ ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑖ℎ 𝑐0𝑖ℎ
𝑙=1
iteratively perform inversions of consecutive customers (𝑎, 𝑏) ∈ 𝑆 × 𝑆
∑
∞ ( ) such that 𝑐𝑎 < 𝑐𝑏 until all the customers are sorted in non-increasing
+2 P 𝜉̃ + 𝜉𝑖ℎ ≤ 𝑙𝑄 < 𝜉̃ + 𝜉𝑖ℎ + 𝜉𝑖ℎ+1 𝑐0𝑖ℎ+1 order of their distance to the depot without increasing the expected
𝑙=1
( 𝑗−1 ) recourse cost. Therefore, the set arg min𝑝∈ 1 (𝑆) (𝑝) contains at least
|𝑆|
∑ ∑
∞ ∑ ∑
𝑗
one path 𝑝̃ = (𝑖̃1 , … , 𝑖̃|𝑆| ) such that 𝑐0𝑖̃1 ≥ 𝑐0𝑖̃2 ≥ ⋯ ≥ 𝑐0𝑖̃|𝑆| .
+2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖𝑗 , To conclude that 𝑝∗ ∈ arg min𝑝∈ 1 (𝑆) (𝑝), and thus that (𝑝∗ ) is a
𝑗=ℎ+2 𝑙=1 𝑘=1 𝑘=1
valid lower bound on the expected recourse of any path 𝑝 ∈  1 (𝑆),
∑
where 𝜉̃ = ℎ−1 ′
𝑘=1 𝜉𝑖𝑘 . Doing the same for (𝑝 ), we obtain:
it remains to show that (𝑝) ̃ = (𝑝∗ ). To do so, we sort the distances
( 𝑗−1 ) separating the customers of 𝑆 from the depot and remove the duplicates
∑∑
ℎ−1 ∞ ∑ ∑
𝑗
to obtain an ordered list 𝑐0(1) > ⋯ > 𝑐0(𝑡) of length 𝑡 ≤ |𝑆|. For each
(𝑝′ ) = 2 P 𝜉𝑖𝑘 ≤ 𝑙𝑄 < 𝜉𝑖𝑘 𝑐0𝑖𝑗
𝑗=1 𝑙=1 𝑘=1 𝑘=1
𝑠 ∈ {1, … , 𝑡}, let us denote by 𝐶𝑠 = {𝑖 ∈ 𝑆 ∶ 𝑐0𝑖 = 𝑐0(𝑠) } the set

532
L. Parada et al. European Journal of Operational Research 318 (2024) 520–533

of customers that share the same distance 𝑐0(𝑠) from the depot and Dror, M., Laporte, G., & Trudeau, P. (1989). Vehicle routing with stochastic demands:
∑
by 𝜉 𝑠 = 𝑖∈𝐶𝑠 𝜉𝑖 the sum of their demands. By construction, in both Properties and solution frameworks. Transportation Science, 23(3), 166–176.
Elçi, Ö., & Hooker, J. (2022). Stochastic planning and scheduling with logic-based
paths 𝑝̃ and 𝑝∗ , a permutation of 𝐶1 will first be visited, followed by a
benders decomposition. INFORMS Journal on Computing, 34(5), 2428–2442.
permutation of 𝐶2 and so on until 𝐶𝑡 . Denoting the number of customers Florio, A. M., Feillet, D., Poggi, M., & Vidal, T. (2022). Vehicle routing with stochastic
in group 𝐶𝑠 as 𝑣𝑠 = |𝐶𝑠 | and the total number of customers visited demands and partial reoptimization. Transportation Science, 56(5), 1393–1408.
∑
before moving to group 𝐶𝑠+1 as 𝑉𝑠 = 𝑠𝑞=1 𝑣𝑟 , we can demonstrate the Florio, A., Hartl, R., & Minner, S. (2020). New exact algorithm for the vehicle routing
expected equality. problem with stochastic demands. Transportation Science, 54(4), 1073–1090.
Gauvin, C., Desaulniers, G., & Gendreau, M. (2014). A branch-cut-and-price algorithm
|𝑆| ∑
( 𝑗−1 )
∑ ∞ ∑ ∑
𝑗 for the vehicle routing problem with stochastic demands. Computers & Operations
(𝑝)
̃ =2 P 𝜉𝑖̃𝑘 ≤ 𝑙𝑄 < 𝜉𝑖̃𝑘 𝑐0𝑖̃𝑗 (B.1) Research, 50, 141–153.
𝑗=1 𝑙=1 𝑘=1 𝑘=1 Gendreau, M., Jabali, O., & Rei, W. (2016). 50Th anniversary invited article—future
𝑉𝑠
( 𝑗−1 ) research directions in stochastic vehicle routing. Transportation Science, 50(4),
∑
𝑡 ∑ ∑
∞ ∑ ∑
𝑗
1163–1173.
=2 P 𝜉𝑖̃𝑘 ≤ 𝑙𝑄 < 𝜉𝑖̃𝑘 𝑐0(𝑠) (B.2)
Gendreau, M., Laporte, G., & Séguin, R. (1995). An exact algorithm for the vehicle
𝑠=1 𝑗=𝑉𝑠−1 +1 𝑙=1 𝑘=1 𝑘=1
( 𝑉𝑠
( 𝑗−1 )) routing problem with stochastic demands and customers. Transportation Science,
∑
𝑡 ∑
∞ ∑ ∑ ∑
𝑗 29(2), 143–155.
=2 𝑐0(𝑠) P 𝜉𝑖̃𝑘 ≤ 𝑙𝑄 < 𝜉𝑖̃𝑘 (B.3) Hjorring, C., & Holt, J. (1999). New optimality cuts for a single-vehicle stochastic
𝑠=1 𝑙=1 𝑗=𝑉𝑠−1 +1 𝑘=1 𝑘=1 routing problem. Annals of Operations Research, 86, 569–584.
( 𝑠−1 ) Hoogendoorn, Y., & Spliet, R. (2023). An improved integer L-shaped method for the
∑
𝑡 ∑
∞ ∑ ∑
𝑠
=2 𝑐0(𝑠) P 𝑞
𝜉 ≤ 𝑙𝑄 < 𝜉 𝑞
(B.4) vehicle routing problem with stochastic demands. INFORMS Journal on Computing,
35(2), 423–439.
𝑠=1 𝑙=1 𝑞=1 𝑞=1
( 𝑉𝑠
( 𝑗−1 )) Hooker, J., & Ottosson, G. (2003). Logic-based benders decomposition. Mathematical
∑
𝑡 ∑
∞ ∑ ∑ ∑
𝑗
Programming, 96(1), 33–60.
=2 𝑐0(𝑠) P 𝜉𝑖∗ ≤ 𝑙𝑄 < 𝜉𝑖∗ (B.5) Jabali, O., Rei, W., Gendreau, M., & Laporte, G. (2014). Partial-route inequalities
𝑘 𝑘
𝑠=1 𝑙=1 𝑗=𝑉𝑠−1 +1 𝑘=1 𝑘=1 for the multi-vehicle routing problem with stochastic demands. Discrete Applied
𝑉𝑠
( 𝑗−1 ) Mathematics, 177, 121–136.
∑
𝑡 ∑ ∑
∞ ∑ ∑
𝑗
=2 P 𝜉𝑖∗ ≤ 𝑙𝑄 < 𝜉𝑖∗ 𝑐0(𝑠) (B.6) Kreimer, J., & Dror, M. (1990). The monotonicity of the threshold detection probability
𝑘 𝑘 in a stochastic accumulation process. Computers & Operations Research, 17(1),
𝑠=1 𝑗=𝑉𝑠−1 +1 𝑙=1 𝑘=1 𝑘=1
|𝑆| ∑
( 𝑗−1 ) 63–71.
∑ ∞ ∑ ∑
𝑗
Laporte, G., & Louveaux, F. (1993). The integer L-shaped method for stochastic integer
=2 P 𝜉𝑖∗ ≤ 𝑙𝑄 < 𝜉𝑖∗ 𝑐0𝑖∗ (B.7) programs with complete recourse. Operations Research Letters, 13(3), 133–142.
𝑘 𝑘 𝑗
𝑗=1 𝑙=1 𝑘=1 𝑘=1 Laporte, G., Louveaux, F., & Van Hamme, L. (2002). An integer L-shaped algorithm
for the capacitated vehicle routing problem with stochastic demands. Operations
= (𝑝∗ ) (B.8)
Research, 50(3), 415–423.
The equivalence of equalities (B.3) and (B.4) and, analogously, of Li, Y., Côté, J.-F., Callegari-Coelho, L., & Wu, P. (2022). Novel formulations and logic-
based benders decomposition for the integrated parallel machine scheduling and
equalities (B.4) and (B.5), comes from the fact that the 𝑙th restocking location problem. INFORMS Journal on Computing, 34(2), 1048–1069.
trip will occur at a customer of set 𝐶𝑠 if and only if the total demand Louveaux, F., & Salazar-González, J.-J. (2018). Exact approach for the vehicle routing
of the customers of sets 𝐶1 to 𝐶𝑠−1 is less than or equal to 𝑙𝑄 and the problem with stochastic demands and preventive returns. Transportation Science,
total demand of the customers of sets 𝐶1 to 𝐶𝑠 exceeds 𝑙𝑄. The sum of 52(6), 1463–1478.
Lysgaard, J. (2003). CVRPSEP: A package of separation routines for the capacitated vehicle
the probabilities of the mutually exclusive events that are considered
routing problem: Technical report, Department of Management Science and Logistics,
for each 𝑠 ∈ {1, … , 𝑡} and for each 𝑙 ∈ N in both Eqs. (B.3) and (B.5) Aarhus School of Business.
is thus equal to the probability of their disjunction, which is given in Pillac, V., Gendreau, M., Guéret, C., & Medaglia, A. (2013). A review of dynamic vehicle
(B.4). □ routing problems. European Journal of Operational Research, 225(1), 1–11.
Rahmaniani, R., Crainic, T. G., Gendreau, M., & Rei, W. (2017). The benders decom-
position algorithm: A literature review. European Journal of Operational Research,
References
259(3), 801–817.
Ropke, S., & Pisinger, D. (2006). An adaptive large neighborhood search heuristic for
Ak, A., & Erera, A. L. (2007). A paired-vehicle recourse strategy for the vehicle-routing the pickup and delivery problem with time windows. Transportation Science, 40(4),
problem with stochastic demands. Transportation Science, 41(2), 222–237. 455–472.
Benders, J. (1962). Partitioning procedures for solving mixed-variables programming Salavati-Khoshghalb, M., Gendreau, M., Jabali, O., & Rei, W. (2019). An exact algorithm
problems. Numerische Mathematik, 4(1), 238–252. to solve the vehicle routing problem with stochastic demands under an optimal
Bertsimas, D., Jaillet, P., & Odoni, A. (1990). A priori optimization. Operations Research, restocking policy. European Journal of Operational Research, 273(1), 175–189.
38(6), 1019–1033. Secomandi, N., & Margot, F. (2009). Reoptimization approaches for the vehicle-routing
Birge, J. R., & Wets, R. J.-B. (1986). Designing approximation schemes for stochastic problem with stochastic demands. Operations Research, 57(1), 214–230.
optimization problems, in particular for stochastic programs with recourse. In Séguin, R. (1996). Problèmes stochastiques de tournées de véhicules (Ph.D. thesis),
Stochastic programming 84 Part I (pp. 54–102). Berlin, Heidelberg: Springer Berlin Université de Montréal, Montréal.
Heidelberg. Tillman, F. (1969). The multiple terminal delivery problem with probabilistic demands.
Christiansen, C., & Lysgaard, J. (2007). A branch-and-price algorithm for the capaci- Transportation Science, 3(3), 192–204.
tated vehicle routing problem with stochastic demands. Operations Research Letters, Uchoa, E., Pecin, D., Pessoa, A., Poggi, M., Vidal, T., & Subramanian, A. (2017). New
35(6), 773–781. benchmark instances for the capacitated vehicle routing problem. European Journal
Côté, J.-F., Gendreau, M., & Potvin, J.-Y. (2020). The vehicle routing problem with of Operational Research, 257(3), 845–858.
stochastic two-dimensional items. Transportation Science, 54(2), 453–469. Van Slyke, R. M., & Wets, R. (1969). L-shaped linear programs with applications to
Denton, B., Miller, A., Balasubramanian, H., & Huschka, T. (2010). Optimal allocation optimal control and stochastic programming. SIAM Journal on Applied Mathematics,
of surgery blocks to operating rooms under uncertainty. Operations Research, 17(4), 638–663.
58(4-part-1), 802–816. Wets, R. (1966). Programming under uncertainty: The equivalent convex program. SIAM
Dror, M., & Ball, M. (1987). Inventory/routing: Reduction from an annual to a Journal on Applied Mathematics, 14(1), 89–105.
short-period problem. Naval Research Logistics, 34(6), 891–905. Yang, W.-H., Mathur, K., & Ballou, R. (2000). Stochastic vehicle routing problem with
restocking. Transportation Science, 34(1), 99–112.

533