Professional Documents
Culture Documents
Research Paper
Research Paper
Research Paper
With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research
(O.R.) and analytics professionals and students. INFORMS provides unique networking and learning opportunities for
individual professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and
methods to transform strategic visions and achieve better outcomes.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
INFORMS Journal on Computing
Vol. 26, No. 1, Winter 2014, pp. 103–120
ISSN 1091-9856 (print) | ISSN 1526-5528 (online)
http://dx.doi.org/10.1287/ijoc.2013.0550
© 2014 INFORMS
Raf Jans
HEC Montré al and GERAD, Montré al, H3T 2A7, Canada, raf.jans@hec.ca
T he inventory routing problem (IRP) and the production routing problem (PRP) are two difficult problems
arising in the planning of integrated supply chains. These problems are solved in an attempt to jointly
optimize production, inventory, distribution, and routing decisions. Although several studies have proposed
exact algorithms to solve the single-vehicle problems, the multivehicle aspect is often neglected because of its
complexity. We introduce multivehicle PRP and IRP formulations, with and without a vehicle index, to solve
the problems under both the maximum level (ML) and order-up-to level (OU) inventory replenishment
policies. The vehicle index formulations are further improved using symmetry breaking constraints; the
nonvehicle index formulations are strengthened by several cuts. A heuristic based on an adaptive large
neighborhood search technique is also developed to determine initial solutions, and branch-and-cut
algorithms are proposed to solve the different formulations. The results show that the vehicle index
formulations are superior in finding optimal solutions, whereas the nonvehicle index formulations are
generally better at providing good lower bounds on larger instances. IRP and PRP instances with up to 35
customers, three periods, and three vehicles can be solved to optimality within two hours for the ML policy.
By using parallel computing, the algorithms could solve the instances for the same policy with up to 45 and
50 customers, three periods, and three vehicles for the IRP and PRP, respectively. For the more difficult IRP
(PRP) under the OU policy, the algorithms could handle instances with up to 30 customers, three (six)
periods, and three vehicles on a single core machine, and up to 45 (35)
customers, three (six) periods, and three vehicles on a multicore machine.
Key words: integrated supply chain planning; inventory routing; production routing; multivehicle; symmetry
breaking; branch-and-cut
History : Accepted by Karen Aardal, Area Editor for Design and Analysis of Algorithms; received April 2012;
revised October 2012; accepted January 2013. Published online in Articles in Advance June 14, 2013.
predefined target stock level (TSL). The second pol- up to 10c/2p/5v were solved to optimality within 30
icy, maximum level (ML), allows delivery quantities minutes. The emphasis of the study was instead on a
to be any positive value, but the inventory at each heuristic procedure using the branch-and-price frame-
customer cannot exceed its maximum stock level. work. Archetti et al. (2011) adapted the branch-and-
The third policy is similar to the ML policy, but cut approach of Archetti et al. (2007) for the PRP-ML
there is no maximum stock level imposed at the with uncapacitated production and a single vehicle.
customers. Archetti et al. (2007) used different Several valid inequalities were also used to strengthen
inequalities to strengthen the formulation for each the formulation. However, computational testing was
Downloaded from informs.org by [14.139.108.35] on 11 July 2023, at 22:27 . For personal use only, all rights reserved.
policy and could solve instances up to 45c/3p/1v only performed on 14c/6p/1v instances, and not all
and 30c/6p/1v to optimality within two hours for instances were solved to optimality within two hours.
the IRP with the OU and ML policy, respectively. Table 1 presents a summary of the exact algorithms
Solyalı and Sü ral (2011) proposed a stronger for the PRP and IRP in the literature. We classify the
formulation for the single-vehicle IRP-OU using a problems along three dimensions: IRP versus PRP, the
shortest path network representa- tion of the OU replenishment policy (ML versus OU), and the num-
policy at each customer, and used a similar ber of vehicles (single versus multiple). The size of
branch-and-cut approach as Archetti et al. (2007). the problems that can be solved to optimality and the
They could solve instances up to 60c/3p/1v and computing time limit (in hours) are shown in brack-
15c/12p/1v to optimality within four hours. ets. Table 1 clearly shows an important gap in the
A closely related problem, the PRP, has also existing literature. The only exact algorithm for the
received increased attention in the last few years. The multiple-vehicle PRP is that of Bard and Nananukul
benefits of coordination in the PRP were first dis- (2010), which only solved relatively small instances to
cussed by Chandra and Fisher (1994). As for the IRP, optimality, compared with the results on the single-
most of the previous studies employed heuristic pro- vehicle case.
cedures to solve the problem (see Adulyasak et al. In this paper, we consider a single product and
2012 for a recent review of heuristic procedures for the a production-distribution network that consists of a
PRP). Few studies have introduced exact algorithms production plant and multiple customers that have
or even methods to compute strong lower bounds their own storage area. At the beginning of the plan-
for the PRP. Fumero and Vercellis (1999) developed ning horizon, the production plant and the customers
a Lagrangian relaxation approach to obtain lower may have initial inventory. In each period, each cus-
bounds and heuristic solutions for a variant of the tomer must have sufficient inventory to satisfy its
PRP, where unit transportation costs are assumed and demand. In the case of the PRP, the plant must decide
the routing decisions can be determined by solv- whether to produce the product and the quantity to
ing a minimum cost flow (MCF) problem. Instances be produced. If production takes place, fixed setup
with up to 12 customers, 8 periods, and 10 products and unit production costs are incurred. The
were tested, and the algorithm could obtain solutions produced quantities can be transported by a limited
with an average optimality gap of 5.5%. A similar number of capacitated vehicles to the customers’
Lagrangian relaxation approach was used by Solyalı area, and rout- ing costs are paid. The product can
and Sü ral (2009) to solve the PRP-OU. However, the also be stored at the plant or at the customers’ area,
lower bounds obtained by this approach were weak. and unit inven- tory holding costs are incurred. We
In the instances with 8c/5p/1v, the lower bound pro- consider the cases where the customer
duced by the Lagrangian relaxation has an average replenishment is based on the ML and OU policies.
deviation of 33.16% from the optimal value. Solyalı The hypotheses we adopt are generally in line with
and Sü ral also tested the performance of the formula- Chandra and Fisher (1994) and Fumero and Vercellis
tion based on the multicommodity fixed charge net- (1999) for the ML policy and Solyalı and Sü ral
work flow problem using the same instances, and (2009) for the OU policy. There is a slight
the longest computing time was approximately 20 difference with respect to the imposed max- imum
hours to obtain the optimal solution. Ruokokoski et al. inventory in this paper compared to Archetti et al.
(2010) explored the performance of different lot-sizing (2011). To be well aligned with the concept of the OU
reformulation schemes for the PRP-ML with unca- policy, we set the delivery quantity to each customer
pacitated production and a single uncapacitated vehi- equal to the difference between its current stock
cle and further employed a branch-and-cut approach level and its TSL before demand consumption. In
similar to that of Archetti et al. (2007) to solve the contrast, the TSL is imposed after demand con-
problem. Bard and Nananukul (2010) introduced a sumption (which typically is not known in advance
branch-and-price procedure for the PRP-ML with in practice) in Archetti et al. (2011). This also applies
multiple vehicles. Because their subtour elimination to the ML policy, where the maximum inventory level
constraints are in the form of the Miller-Tucker- at the customers’ area is imposed before demand
Zemlin inequalities (Miller et al. 1960), they obtained con- sumption. It should also be noted that the
rather weak lower bounds, and only the instances replenish- ment practice in our PRP and the noted
literature is
Adulyasak, Cordeau, and Jans: Formulations and Branch-and-Cut for the
MVPRP 1
Table 1 Summary of Exact Algorithms for the Deterministic PRP and IRP with Single Product, Single Plant,
and Multiple Retailers
slightly different from the IRP presented in Archetti discussion of computational experiments in §6 and
et al. (2007) and Solyalı and Sü ral (2011). In the lat- conclusions in §7.
ter studies, the delivery to the customers must take
place before the distribution facility is replenished in
each period, whereas in our PRP, the quantity pro- 2. MVPRP Formulations
duced in period t can be delivered to customers to This section presents the main notation and mathe-
satisfy their demand in the same period. These two matical formulations of the MVPRP with the ML and
practices, however, can be converted into each other, OU policies.
as we show in the online supplement (available as
supplemental material at http://dx.doi.org/10.1287/ 2.1. Notation
ijoc.2013.0550). Because the PRP is a generalization of The PRP can be defined on a complete undirected
the IRP, we prefer to use the name PRP in the graph G = (N, E) with the following notation.
remain- der of this paper to represent both the IRP Sets:
and PRP unless stated otherwise. Note that MVPRP
is used to represent the PRP with the multivehicle T Set of time periods, indexed by t∈ {1,. . . , l}
(MV) aspect. The main contributions of our study are N Set of plant and customers, indexed by i ∈
fourfold. First, we present strong formulations and {0,. . . , n}, where the plant is represented by
exact algo- rithms for both the IRP and PRP with node 0 and Nc = N \ {0} is the subset of n
multiple vehi- cles, thereby filling several important customers
gaps in the literature, as shown in Table 1. Several E Set of edges, E= {(i, j): i, j∈ N, i < j}
formulations are presented and branch-and-cut K Set of identical vehicles, indexed by k ∈
algorithms are pro- posed to solve the problems {1,. . . , m}
under both the OU and ML policy. Second, we E(S) Set of edges (i, j) ∈E such that i, j ∈S, where
propose several valid inequal- ities and symmetry S ⊆ N is a given set of nodes
breaking constraints to strengthen the formulations (S) Set of edges incident to a node set S, (S) =
and test the effect of these inequal- ities. Third, we {(i, j) ∈E: i ∈S, j ØS or i ØS, j ∈S} (for simplic-
use an adaptive large neighborhood search (ALNS) ity, we write (i) to represent the set of edges
procedure that was previously devel- oped for the incident to node i).
MVPRP-ML (Adulyasak et al. 2012) and extend it
to the MVPRP-OU, MVIRP-ML, and MVIRP-OU. Decision variables:
Fourth, we provide extensive computa- tional
pt Production quantity in period t
results of the new formulations and further explore
Iit Inventory at node i at the end of period t
the performance of the algorithm on a multi-
core machine. yt Equal to 1 if there is production at the plant in
period t, 0 otherwise
The rest of the paper is organized as follows. Sec-
tion 2 presents different formulations of the MVPRP. zikt Equal to 1 if node i is visited by vehicle k in
period t, 0 otherwise
Section 3 describes the valid inequalities that are
applied to the formulations. The details of the xijkt If vehicle k travels directly between node i and
node j in period t, 0 otherwise
branch- and-cut approaches are discussed in §4, and
the details of the heuristic algorithm to calculate qikt Quantity delivered to customer i with vehicle k
upper bounds are presented in §5. This is followed in period t.
by the
Adulyasak, Cordeau, and Jans: Formulations and Branch-and-Cut for the
1 MVPRP
Li Maximum or target inventory level at node i tion capacity constraints at the plant: they force the
Ii0 Initial inventory available at node i. setup variable to be 1 if production takes place and
limit the production quantity to the minimum of
2.2. Multivehicle Formulations for the ML Policy the production capacity and the total demand in the
In this section, we introduce two formulations for remaining periods. The inventory quantity at the
the MVPRP-ML: one with and one without a vehicle pro- duction facility at the end of each period is
index. limited by constraints (5) and the inventory quantities
2.2.1. Formulation With a Vehicle Index for the at the cus- tomers’ area after delivery cannot exceed
ML Policy. To formulate the MVPRP-ML with a vehi- their inven- tory capacities (6). The total quantity
cle index, we extend the single-vehicle PRP formula- loaded in each vehicle can be at most the vehicle
tion used by Archetti et al. (2007, 2011), as follows: capacity as specified by (7). Constraints (8) allow
each customer to be vis-
Σ Σ Σ Σ
min upt + fyt + hiIit + cijxijkt (1) ited at most once in each period. Constraints (9) allow
t∈T i∈N (i, j)∈E k∈K a positive delivery quantity from vehicle k to node i
in period t only if this node is visited by the vehi-
s.t.
Σ cle in period t. In the ML policy, there always exists
I0, t−1 + pt = Σ qikt + I0t ∀ t ∈ T , (2) an optimal solution with 0 ending inventory, because
i∈Nc k∈K
Σ it is never beneficial to carry inventory at the end of
Ii, t−1 + qikt = dit + Iit ∀ i ∈ Nc, ∀ t ∈ T , (3) the planning horizon if the holding costs are nonneg-
k∈K ative.
limitedTherefore, the delivery
by the minimum quantity
value among tothe
a customer is
inventory
Σ l }
pt ≤ min C, Σ di capacity at the customer, the vehicle capacity, and the
i∈Nc j=t
yt ∀t ∈T, (4) total demand of the customer in the remaining peri-
j
I0t ≤ L0 ∀t ∈T, (5) ods. Constraints (10) are the degree constraints; they
Σ require the number of edges incident to node i to be
Ii, t−1 + qikt ≤ Li ∀ i ∈ Nc, ∀ t ∈ T , (6) 2 if it is visited. Constraints (11) eliminate subtours
k∈K
Σ for each vehicle.
qikt ≤ Qz0kt ∀ k ∈ K, ∀ t ∈ T , (7) Archetti et al. (2007, 2011) also strengthen the for-
i∈Nc
mulation using several valid inequalities. We present
Σ
zikt ≤ 1 ∀ i ∈ Nc, ∀ t ∈ T , (8) here the inequalities that are valid for the PRP
k∈K with capacitated production and extend them to the
}
qikt ≤ min Σ multivehicle case. Denote by tr and trr the earliest
Li, l di zikt
period when the plant must produce and the ear-
Q, j liest period when at least one customer must be
j=t
∀ i ∈ Nc, ∀ k ∈ K, ∀ t ∈ T , (9) replenished to prevent a stockout, respectively; i.e.,
Σ Σ ij
Σ tr = arg min 1≤t≤l { i∈N max{0, t d — } − I00 > 0}, rr
xjjr kt = 2zikt and rr c
j=1
ij
Ii0 Σ{t
r j )∈ (i) t = min i∈N t , iwhere t =i arg min 1≤t≤l
rr rr d −
(j, c j=1
Σ
∀ i ∈ N, ∀ k ∈ K, ∀ t ∈ T , (10) i.e., w = i∈Nc max{0, d — Ii0 }. First, two inequal-
Σtrrthe minimum
Ii0 > 0}. Let w be
ij
shipping quantity in t ;
Σ Σ j=1
Second, the following inequalities are imposed to Constraints (22)–(26) are equivalent to (2)–(3), (6),
strengthen the customer replenishments: and (9)–(10), respectively. Constraints (27) limit the
s Σ s number of vehicles leaving the production facility to
Ii, t−s−1 ≥ di, t−j zik,t−j the number of available vehicles in each period. Con-
Σj=0 1− k ∈K straints (28) are the subtour elimination and vehicle
Σ
j=0
capacity constraints. If one divides the inequalities
∀ i ∈ Nc, ∀ t ∈ T , s = 0, 1,. . . , t − 1. (18) by Q, these constraints have a form similar to the
Finally, the following inequalities are imposed for the generalized fractional subtour elimination constraints
routing part: (GFSECs) for the vehicle routing problem (VRP)
(Toth and Vigo 2002). Unlike GFSECs in the VRP,
zikt ≤ z0kt ∀ i ∈ Nc, ∀ k ∈ K, ∀ t ∈ T , (19)
how- ever, we cannot round up the value of the term
xijkt ≤ zikt and xijkt ≤ zjkt qit/Q because it contains the qit variable, and
∀ (i, j) ∈ E(Nc), ∀ k ∈ K, ∀ t ∈ T . (20) therefore they do not provide strong linear
programming (LP) relax- ation bounds. We prefer to
The formulation (1)–(20) will be referred to as
use the form (28), because preliminary tests have
F (ML) | k. indicated that the original form of GFSECs is
2.2.2. Formulation Without a Vehicle Index for numerically unstable because of the frac- tional
the ML Policy. The previous formulation has the right-hand side.
drawback that the number of variables grows in pro- We can also rewrite inequalities (17)–(20) for the
portion to the number of vehicles. Alternatively, one nonvehicle index formulation as follows:
can express the routing constraints with variables that , ,
do not comprise a vehicle index. The formulation is Σtrr
w ,
z0j ≥ Q (34)
j=1
s
written using the variables q, z, and x with the same s
notation as in the previous section, but the vehicle Ii, t−s−1 ≥ Σ
di, t−j zi, t−j
index k is dropped. The only exception is the variable Σj=0
1−
j=0
z0t , which is changed to be an integer variable rep- ∀ i ∈ Nc, ∀ t ∈ T , s = 0, 1,. . . , t − 1, (35)
resenting the number of vehicles leaving the plant in
z ≤ z0t ∀ i ∈ Nc, ∀ t ∈ T ,
it
(36)
period t. The formulation without the vehicle index
can be stated as follows: xijt ≤ zit and xijt ≤ zjt ∀ (i, j) ∈ E(Nc), ∀ t ∈ T . (37)
Σ Σ Σ
min upt + fyt + hiIit + cijxijt (21) The nonvehicle index formulation, together with
t∈T i∈N (i, j)∈E the inequalities (34)–(37) in this section and (16), will
s.t. (4)–(5) and be referred to as F (ML) nk.
|
Σ We also remark here on reformulation schemes for
I0, t−1 + pt = qit + I0t ∀t ∈T, (22) the PRP-ML. We have tested the facility location refor-
i∈Nc mulation, called four-index facility location (FIFL),
Ii, t−1 + qit = dit + Iit ∀ i ∈ Nc, ∀ t ∈ T , (23) proposed by Ruokokoski et al. (2010). The prelimi-
nary results show that, in our case, where produc-
Ii, t−1 + qit ≤ Li ∀ i ∈ Nc, ∀ t ∈ T , (24) tion, inventory, and vehicle capacities are imposed,
Σ
l }
qit ≤ min the facility location reformulation is slightly inferior
Li, di zit ∀ i ∈ Nc, ∀ t ∈ T , (25) in terms of computing times to the basic formula-
Q, j tion with the inequalities used in Archetti et al. (2007,
j=t
Σ
xjjr t = 2zit ∀ i ∈ N, ∀ t ∈ T , (26) 2011). The main reasons are, first, that the inequalities
(j, j r )∈ (i) (18) already substantially strengthen the formulation,
and second, that the FIFL formulation has a much
z 0t ≤ m ∀ t ∈ T , (27) larger number of variables than the basic formulation.
Σ Σ
Q xijt ≤ (Qzit − qit)
(i, j)∈E(S) 2.3. Multivehicle Formulations for the OU Policy
i∈S
This section presents two formulations for the
∀ S ⊆ Nc, |S| ≥ 2, ∀ t ∈ T , (28) MVPRP-OU: one with and one without a vehicle
ptt, Iitit, qit ≥ 0 ∀ i ∈ N, c∀ t ∈ T , (29) y , z ∈ {0, 1} ∀ i ∈ N , ∀ t ∈ T , (30)
Adulyasak, Cordeau, and Jans: Formulations and Branch-and-Cut for the
1 MVPRP
index. 2.3.1. Formulation With a Vehicle Index for the
OU Policy. In the OU policy, when a customer is vis-
z0t ∈ A+ ∀t ∈T, (31) ited, the inventory before demand consumption must
be replenished to reach its TSL. Archetti et al. (2007,
xijt ∈ {0, 1} ∀ (i, j) ∈ E: i /= 0, ∀ t ∈ T , (32) 2011) added constraints to the formulation F (ML) | k
x0jt ∈ {0, 1, 2} ∀ j ∈ N c, ∀ t ∈ T . (33) to solve the single-vehicle IRP with the OU policy.
Downloaded from informs.org by [14.139.108.35] on 11 July 2023, at 22:27 . For personal use only, all rights reserved.
Adulyasak, Cordeau, and Jans: Formulations and Branch-and-Cut for the
MVPRP 1
However, it has been shown by Solyalı and Sü ral and givt > Li. The strong formulation, referred to as
(2011) that a stronger version of the IRP-OU can be
F (OU) | k, is as follows:
obtained by using a shortest path network represen-
Σ Σ Σ
tation for the customer inventory replenishment part. min upt + fyt + h0I0t + cijxijkt
This reformulation scheme exploits the characteristic t∈T
}k∈K
(i, j)∈E
of the OU policy that the delivery quantity for a cus- Σ
t−1
Σ Σ eivthikvt
tomer i visited in period t is equal to the total +
Σ (38)
demand
−
consumption in the interval between t 1 and the i∈Nc k∈K t∈T r v=v(i, t)
pre-
vious visit in period v < t. In our preliminary test, we s.t. (5), (10)–(15) and
have also observed that this reformulation is far supe-
rior to the formulations of Archetti et al. (2007, 2011) Σ Σ Σ
t−1
for multiple vehicles. As a consequence, we adopt the I0, t−1 + pt = givthikvt + I0t ∀t ∈T, (39)
reformulation presented in Solyalı and Sü ral (2011) i∈Nc k∈K v=v(i, t)
and extend it using a vehicle index. We define di0 =
di, l+1 = 0, T r = T ∪ {l + 1} and use the following addi- pt ≤ Cyt ∀t ∈T, (40)
t−1
Σ
tional notation:
givthikvt ≤ Qz0kt ∀ k ∈ K, ∀ t ∈ T , (41)
Σc v=v(i, t)
i∈N
hikvt Binary variable, equal to 1 if node i is visited t−1
Σ
by vehicle k in period t and the previous visit hikvt = zikt ∀ i ∈ Nc, ∀ k ∈ K, ∀ t ∈ T , (42)
is in period v; 0 otherwise v= v(
Σ
i, t)
Σ Σ
∈K
eivt Total inventory holding cost when customer t−1 Σ µ(i, t)
hikvt − Σ
i is visited in period t and the previous visit hiktv = 0 ∀ i ∈ Nc, ∀ t ∈ T , (44)
is in period v k∈K v=v(i, t) k∈K v=t+1
µ(i, t) The latest period after period t when l
Σ
customer i can be replenished next with-
out having a stockout; i.e., µ(i, t) = Σ hikt, l+1 = 1 ∀ i ∈ Nc, (45)
arg maxt<v≤l+1{givt Li} k∈K t=v(i, l+1)
v(i, t) The earliest period ≤before period t when cus- hikvt ∈ {0, 1} ∀ i ∈ Nc, ∀ k ∈ K, ∀ v, t ∈ T r. (46)
tomer i can be replenished without having a
stockout; i.e., v(i, t) = arg min0≤v<t{givt ≤ Li}. The objective function (38) and constraints (39)–(41)
are equivalent to (1), (2), (4), and (7), respectively.
The parameters givt and eivt can be calculated as Constraints (42) provide the link between the hikvt
follows:
and zikt variables. Constraints (43)–(45) represent the
short- est path network of the OU policy at each
Σ
t−1
customer.
As in the formulation presented by Solyalı and
Sü ral (2011), the inequalities (19) and (20) are also
dij + (Li − Ii0) if v = 0, added to strengthen the routing part of the formu-
j=1
lation. We further add (16) and (17) to reinforce the
t−1 production part.
givt = Σ
dij if 0 < v < t ≤ l, 2.3.2. Formulation Without a Vehicle Index for
j=v
the OU Policy. The nonvehicle index formulation for
0 if t = l + 1, the OU policy can be written using the same notation
j |
Σ
t−1
eivt = j=1
Ii −
0
i
Adulyasak, Cordeau, and Jans: Formulations and Branch-and-Cut for the
1 MVPRP
Σ j dir i 0 a e formulation F (ML) nk, using the variable h
s as in the previous section but without the vehicle
f = t index k. The formulation, referred to as F (OU) | nk, is
r =1 h as follows: Σ
t−1
hi Σ Li − r dir if 0 < v < t ≤ l + min Σ upt + fyt + h0I0t + cijxijt
j=v Σ (i, j
1. t∈T
Σ Σ
=v
+ Σ )∈E
}
Downloaded from informs.org by [14.139.108.35] on 11 July 2023, at 22:27 . For personal use only, all rights reserved.
eivthivt (47)
Adulyasak, Cordeau, and Jans: Formulations and Branch-and-Cut for the
MVPRP 1
s.t. (5), (40), (26)–(27), (29)–(33) and each dispatched vehicle. These two types of symme-
try are present in each period, and hence there can be
m m m
Σ Σ
t−1
[ m¯ m¯ 1 !][m¯ m¯ 2 !] ·m· · [l m¯ !] equivalent solutions. For
I0, t−1 + pt =
¯l
givthivt + I0t ∀t ∈T, (48) 1 2
example, for an instance with three periods and three
i∈Nc v=v(i, t)
t−1 vehicles, if two vehicles are used in each period,
Σ hivt = zit ∀ i ∈ Nc, ∀ t ∈ T , (49) there are3 [ 3 2!] = 216 equivalent
2
v=v(i, t) solutions that can
be obtained by re-indexing the vehicles. Such symme-
µ(i, 0)
Σ try issues typically slow down the branch-and-bound
process because of the duplications in the search pro-
hi0t = 1 ∀ i ∈ Nc, (50) cess (Sherali and Smith 2001).
t=1
To break the first type of symmetry, we can use
+
Adulyasak, Cordeau, and Jans: Formulations and Branch-and-Cut for the
1 MVPRP
µ(i, t)
Σ
t−1
hivt − Σ h = 0 ∀i∈N, ∀t ∈T, (51) the following symmetry breaking constraints (SBCs)
itv c
v=t+1
to allow vehicle k 1 to be dispatched only if vehicle k
v=v(i, t)
is also dispatched:
l
Σ (SBC0) z0kt ≥ z0, k+1, t 1 ≤ k ≤ m − 1, ∀ t ∈ T . (55)
hit, l+1 = 1 ∀ i ∈ Nc, (52)
To address the second symmetry issue, we can use
t=v (i,
l+1)
Σ
Q xijt ≤ Σ Qzit − Σ
t−1 different sets of SBCs. These sets cannot be imposed
givthivt
Downloaded from informs.org by [14.139.108.35] on 11 July 2023, at 22:27 . For personal use only, all rights reserved.
∀ S ⊆ Nc, |S| ≥ 2, ∀ t ∈ T , (53) by ordering them according to their total route costs:
Σ cx Σ
hivt ∈ {0, 1} ∀ i ∈ Nc, ∀ v, t ∈ T r. (54) (SBC1) ij ijkt ≥ cijxij, k+1, t
(i, j)∈E (i, j)∈E
Constraints (48)–(52) are equivalent to (39) and 1 ≤ k ≤ m − 1, ∀ t ∈ T .
(42)–(45), respectively. Constraints (53) are equivalent
to (28). Note that inequalities (16), (34), and (36)–(37) Alternatively, one can impose that the vehicles be
are also added a priori in our implementation to make ordered according to their total delivery quantity:
a fair comparison with the other formulations. Σ Σ
(SBC2) qikt ≥ i, +1, t
k
q
2.4. Formulations for the MVIRP i∈Nc i∈Nc
3.2. Valid Inequalities for the Nonvehicle root node lower bound, as can be seen in the compu-
Index Formulations
tational experiments.
To strengthen the nonvehicle index formulations, we Denote by p(S, r) the minimum number of vehi-
add the following inequalities a priori. cles that must be dispatched to carry the demands in
customer set S during periods 1 to r, calculated as
Σ Σr
Qz0t ≥ qit ∀ t ∈ T for F (ML) | nk or (56) p(S, r) = i∈S max{0,
dit − Ii0 }/Q . The following
i∈N t=1
Σ
「
Downloaded from informs.org by [14.139.108.35] on 11 July 2023, at 22:27 . For personal use only, all rights reserved.
t−1
c
inequalities are also imposed:
Σ Σ Σr Σ
Qz0t ≥ givt h ∀ t ∈ T for F (OU) | nk. (57) xijt ≥ 2p(S,r) ∀ S ⊆ Nc, |S|≥ 2, r ∈ T . (61)
ivt
i∈Nc v=v(i, t) t=1 (i, j)∈ (S)
Constraints (56) and (57) require the number of These constraints ensure that the total number of
vehi- cles leaving the production facility to be vehicles entering and leaving the set of customers S
sufficient to carry delivery quantities to all from period 1 to period r must be sufficient to carry
customers in each period. the demands during these periods.
Because GFSECs (28) and (53) are generally weak,
we further strengthen the nonvehicle index formu-
lations by adding the following subtour elimination
4. Branch-and-Cut Approaches
Because all the formulations contain an exponentially
constraints
Σ (SECs): large number of subtour elimination constraints, a
x ≤ Σz − zijt it et natural way to solve the problems is to use a branch-
(i, j )∈E(S)
i∈S and-cut technique. In this process, the subtour elimina-
tion constraints, i.e., constraints (11) for the F (ML) | k
∀ S ⊆ Nc, |S| ≥ 2, e ∈ S, ∀ t ∈ T . (58) and F (OU) | k and GFSECs for the F (ML) | nk and
These cuts are used to prevent subtours in each F (OU) nk, are dropped from the formulations and
period, but they do not take into account the vehi- | iteratively when they are violated at each
are added
cle capacity. Therefore, they have to be used together node of the branch-and-bound tree. In this section
with GFSEC (28) or (53) to generate feasible multi- we provide the details of our branch-and-cut
vehicle routes. approaches for both types of formulations. For the
To take into account the periodic routing decisions variable selec- tion, we first branch on the y
of the MVPRP, we also add another set of constraints variables, next on the z variables, and finally on the x
to the formulation, called multiperiod generalized variables. Among the variables with the same
fractional subtour elimination constraints (MGFSECs), priority, we use the default settings in CPLEX 12.3 to
as follows: select a specific variable to branch on. The remaining
parameters are set to their
default values.
QΣ ΣΣ
t∈R (i, j)∈E(S)
xijt ≤ (Qzit − qit)
Σ t∈R i∈S 4.1. Branch-and-Cut for the Vehicle
Index Formulations
∀ S ⊆ Nc, |S| ≥ 2, ∀ R ⊆ T or (59) To solve the vehicle index formulations F (ML) | k and
Σ Σ ΣΣ
t−1
Σ F (OU) | k, we use an exact separation algorithm that
Q xijt ≤ givthivt solves a minimum s − t cut problem to detect vio-
Qzit −
t∈R (i, j)∈E(S) t∈R i∈S v=v(i, t) lated subtour elimination constraints for each vehicle
∀ S ⊆ Nc, |S| ≥ 2, ∀ R ⊆ T . (60) improvements in the
Ruokokoski et al. (2010) and use the minimum s t stored in an ordered list. Then an empty set of cus-
−
cut algorithm of the Concorde callable library (Apple- tomers S and an empty set of violated sets ((S) are
gate et al. 2011). Although this separation algorithm created. The separation algorithm starts by adding
is different from that used by Archetti et al. (2007, the first customer in the ordered list to S and checks
2011) and Solyalı and Sü ral (2011), it has proven to be
whether the MGFSEC of the set S is violated. The
efficient in our branch-and-cut algorithm.
next customer is then added to S and the algorithm
We have also tested different options for adding
checks for the MGFSEC again, and so on. The violated
the subtour elimination constraints to the formula-
Downloaded from informs.org by [14.139.108.35] on 11 July 2023, at 22:27 . For personal use only, all rights reserved.
two main phases: initialization and improvement. an attempt to approximately take transportation costs
The initialization phase is used to generate a number into account, without explicitly modeling the routing
of initial solutions with different production setup decisions. Then the routes for the vehicles are deter-
con- figurations. These solutions are created by an mined by solving the traveling salesman problem
Downloaded from informs.org by [14.139.108.35] on 11 July 2023, at 22:27 . For personal use only, all rights reserved.
embed- ded enumeration scheme using inequalities (TSP) for each vehicle individually. We construct the
inspired from local branching (Fischetti and Lodi TSP tours using the GENIUS procedure (Gendreau
2003). In the improvement phase, the generated et al. 1992) and improve them using the three-opt
initial solu- tions are further improved by the pro- cedure (Lin 1965). Additionally, to solve the
ALNS operators. In this process, the customer visit MCF for the OU policy, we set the inventory level of a
and routing vari- ables are handled by two types of visited customer i in period =t as− Iit Li dit . The
operators, selection remainder of the Op-ALNS algorithm for the OU
and transformation; the remaining continuous vari- policy is the same as for the ML policy.
ables associated with production, inventory, and
ship- ment quantities are set by solving a network 5.2. Adaptation of the Op-ALNS for the MVIRP
flow model. The computational experiments Because production setups are irrelevant for the IRP,
reported by Adulyasak et al. (2012) show that the we generate initial solutions with different customer
visit decisions. Denote by s the
¯it value of the cus-
Op-ALNS outperformed existing heuristics on small
tomer visit variable zit in solution
z s. The original local
to very large instances. We provide more details on
branching inequality used in the initialization phase
the Op- ALNS in the online supplement, and explain
is replaced with the following inequality to generate
how this algorithm is adapted for the MVPRP-OU and an initial solution s¯:
the MVIRP (both ML and OU) in §§5.1 and 5.2, , ,
respec- tively. Note that only the initialization Σ ΣΣ
¯≥
phase of the
original Op-ALNS is modified in these adaptations. Σ (1 − zit ) + s
zit 0.25 z it
it zit | z¯its i∈Nc t∈T
5.1. Adaptation of the Op-ALNS for the zit |
=0 s = 1,. . . , s¯ − 1. (64)
MVPRP-OU z¯s =1
In the original Op-ALNS, infeasible routes are
The inequality (64) forces at least 25% of the total cus-
allowed when generating initial solutions because
tomer visits over the horizon to change. The maxi-
the selection and transformation operators can han-
mum number of initial solutions in the pool is set to
dle such routes effectively by repeatedly reallocating
10 for the MVIRP, and the other parts of the
delivery quantities. However, it is much more difficult algorithm remain the same.
to remove and reinsert node candidates from infeasi-
ble routes in the OU policy because the delivery quan-
tity is defined by the difference between the inventory 6. Computational Experiments
level and the TSL. Therefore, it is easier to start from The branch-and-cut algorithms were coded in C# on
initial solutions with feasible routes. To ensure that MonoDevelop 2.2 using CPLEX 12.3 under Scientific
feasible solutions are obtained at the end of the ini- Linux 6.1. The experiments were performed on a
tialization process, we take the capacity restriction of workstation with an Intel Xeon 2.67 GHz processor
each vehicle into account in the first subproblem. We and 24 GB of RAM. A multiple core processor, with
denote this by σi = min{2c0i , minj, k∈N , j/=k (cij +cik )}, i.e., each core having the same specifications, was also
the minimum value between the cost of making a used for the parallel computing experiments in §6.4.
round trip from the production facility and the cost The Op-ALNS heuristic, adapted from Adulyasak
to the two nearest neighbors of customer i. The first et al. (2012), was coded in C# using Microsoft Visual
sub- problem in the initialization phase for the Studio 2008 and executed on a workstation with a 2.10
MVPRP-OU is as follows: GHz CPU and 2 GB of RAM under Windows XP.
Σ Σ Σ In Tables 2 and 5–7, we report the average CPU
min upt + fyt + h0I0t + σizikt times in seconds and the average number of nodes
t∈T
i∈Nc k∈K in the columns CPU and Nodes, respectively. Column
}
t−1 %LB shows the final lower bound as a percentage of
+ Σ Σ Σ e h the best upper bound found by all approaches. Bold-
Σ
ivt ikvt (63)
face letters are used to indicate the best results. In
i∈Nc k∈K t∈T r v=v(i, t)
all tables, if all instances of a given problem size are
subject to (4), (8), (12)–(13), (39)–(46). solved to optimality, we put boldface letters on the
Adulyasak, Cordeau, and Jans: Formulations and Branch-and-Cut for the
1 MVPRP
1,367
8,564
2,567
1,208
7,893
1,765
98.9 4,954
3 94.9 3,438.3 45,567
98.4 1,033.7 52,474 104 99.3 617.310,737 103 99.1 655.5 11,723
42 99.7 588.1
43 100.0 193.5
43 99.8 477.4
40 100.0 147.2
2 98.1 2,792.7
0 97.8 3,600.0
SBC0 + 4
4 95.6 3,261.541,581
6 95.4 2,981.223,282
Node
1,662
7,991
2,189
744.3 12,195 94
3,600.0 34,414 0
Node
650.0
134.5
423.3
2,818.9
3,600.0
44 100.0
0 97.7
44 99.8
1,802
8,978
2,400
48 43 99.8 524.0
4 0 97.6 3,600.0
48 36
100 86
120 86
100 96
120 96
8 2
20 0
20 0
Not optimal
Not optimal
Not optimal
MVPRP-OU
MVPRP-ML
MVIRP-ML
Optimal
Optimal
Optimal
Optimal
Total
Total
Total
Table 2
Total
limit was set to one hour. The detailed results are pro-
vided in the online supplement.
Adulyasak, Cordeau, and Jans: Formulations and Branch-and-Cut for the
MVPRP 1
Table 4 Effects of the Valid Inequalities for the Nonvehicle Index Formulations on
Average Lower Bounds at the Root Node for the MVPRP and MVIRP Instances
Problem None (56) (58) (59) (61) All None (57) (58) (60) (61) All
MVPRP 93.3 93.3 94.7 93.5 93.7 95.2 94.0 94.6 94.5 94.4 94.1 95.2
MVIRP 83.5 83.6 85.4 84.8 87.4 89.0 85.7 86.1 88.3 87.1 87.4 89.7
Downloaded from informs.org by [14.139.108.35] on 11 July 2023, at 22:27 . For personal use only, all rights reserved.
6.2.1. Effect of Vehicle Symmetry Breaking Con- for the MVPRP-ML, MVPRP-OU, MVIRP-ML, and
straints on the Vehicle Index Formulations. We ana- MVIRP-OU instances, respectively. By adding SBC0 +
lyze the effect of symmetry breaking constraints SBC0 SBC3 to the F (ML) k| and F (OU) k | formulations,
alone and SBC0 together with one of the other con- the algorithm could also solve 36 instances that
straints SBC1–SBC4 for the formulations F (ML) k could not be solved to optimality within one hour
and F (OU) k. The average results on the MVPRP and |
| using the formulations without these inequalities.
MVIRP instances are shown in Table 2. In our tests, We further evaluated the performance of the
we used the default settings of CPLEX that allow the different SBCs on instances with a larger number of
solver to detect and generate its own symmetry break- customers, i.e., the first instance in each group with
ing constraints. 25–50 customers and three periods. The +
Table 3 provides a summary of the time reduction formulations with SBC0 SBC3 provided the
factors for each approach, calculated as the average best results with 36 (of 84) instances solved to
computing time spent to solve an instance size with- optimality, compared with only 24 instances without
out using any of our SBCs, divided by the average these SBCs. For the instances not solved to
computing time of using each cut strategy. A time fac- optimality in one hour, the + average lower bound for
tor equal to 2 means the algorithm spent, on average, the formulations with SBC0 SBC3 is 92.4%; the aver-
only half the time by using the SBC strategy com- age lower bound is 90.2% when these SBCs are not |
pared to using no additional SBCs. added. We thus consider the
| + formulations F (ML) k
The results clearly show the benefits of using the and F (OU) k with SBC0 SBC3 in the remaining
SBCs, compared to relying only on those generated computational experiments.
by CPLEX. The original formulation without any addi- 6.2.2. Effect of Valid Inequalities for the Non-
tional SBC provides the worst results; adding SBC0 vehicle Index Formulations. In this section we
could generally improve the computing times and analyze the effect of the valid inequalities that
reduce the number of nodes in the branch-and- we implemented for the formulations F (ML)| nk
bound tree. The combination of SBC0 together and F (OU) |nk. First, we evaluate the effects of
with one of the other SBCs could further speed up the valid inequalities on the lower bounds at the
the solu- tion process, except for SBC2, where some root node of the branch-and-bound tree. To avoid
results are worse than using the CPLEX cuts alone. misinterpretation caused by the impact of the CPLEX
The cut strategies + SBC0 SBC3 and SBC0 + SBC4 cuts, we conducted the experiments without these
pro- vide good results, but+SBC0 SBC3 is slightly cuts. The average lower bounds are shown in Table
better overall. The average time factor reductions 4. The numbers presented are equal to the average
obtained using+ SBC0 SBC3 within the maximum lower bounds at the root node compared to the
computing time limit of one hour are 4.53, 2.03, optimal solutions or the best upper bounds if the
9.70, and 3.13 instances were not solved to optimality. Each
column shows
Table 5 Effects of the Valid Inequalities for the Nonvehicle Index Formulations on the Branch-and-Cut Algorithm for the MVPRP and MVIRP Instances
Maximum level (ML) Order-up-to level (OU)
F (ML) | nk F (ML) | nk+ F (OU) | nk F (OU) | nk+
Problem # #O %LB CPU Nodes #O %LB CPU Nodes # #O %LB CPU Nodes #O %LB CPU Nodes
MVPRP 32 29 100.0 397.3 28,639 32 100.0 161.3 10,337 40 40 100.0 109.7 7,325 40 100.0 69.6 5,105
Optimal
Not optimal 16 0 98.9 3,600.0 83,703 3 99.3 2,954.0 47,366 8 1 98.8 3,530.5 65,816 2 98.9 3,255.2 43,935
Total 48 29 99.6 1,464.9 46,994 35 99.8 1,092.2 22,680 48 41 99.8 679.8 17,073 42 99.8 600.6 11,576
MVIRP
Optimal 80 78 99.9 215.8 36,379 80 100.0 130.9 23,356 80 80 100.0 66.1 5,573 80 100.0 14.0 1,215
Not optimal 40 8 95.4 3,039.4 154,296 8 96.5 3,077.9 124,501 40 9 95.0 3,105.8 155,052 10 96.5 3,014.3 113,077
Total 120 86 98.4 1,157.0 75,685 88 98.8 1,113.2 57,071 120 89 98.3 1,079.4 55,399 90 98.8 1,014.1 38,502
Downloaded from informs.org by [14.139.108.35] on 11 July 2023, at 22:27 . For personal use only, all rights reserved.
10 3 2 100.0 0.2 98.4 0.1 100.0 0.1 98.4 0.1 100.0 0.4 98.1 0.2 100.0 0.1 98.6 0.1
10 3 3 100.0 0.6 98.4 0.2 100.0 0.3 98.9 0.1 100.0 0.6 98.2 0.2 100.0 0.2 98.8 0.1
15 3 2 100.0 2.5 98.0 0.4 100.0 8.6 98.4 0.4 100.0 17.6 93.4 0.9 100.0 54.2 93.4 0.8
15 3 3 100.0 26.6 96.6 1.6 100.0 35.4 97.9 0.7 100.0 17.6 93.1 1.2 100.0 234.4 93.8 0.6
20 3 2 100.0 2.6 99.0 0.7 100.0 4.0 99.1 0.4 100.0 59.6 98.3 1.2 100.0 91.5 98.6 0.9
20 3 3 100.0 51.1 98.3 3.9 100.0 70.9 99.0 1.2 100.0 692.5 97.1 1.9 100.0 1,193.7 97.7 1.2
25 3 2 100.0 3.9 99.5 1.0 100.0 9.5 99.2 0.3 100.0 2,854.8 95.4 2.9 99.5(4) 7,200.0 95.5 1.9
25 3 3 100.0 85.8 98.2 8.1 100.0 589.5 98.7 2.3 98.4(4) 7,200.0 94.6 4.7 98.8(4) 7,200.0 94.8 3.3 ©
30 3 3 100.0 194.2 98.8 32.4 99.9(1) 1,954.5 99.2 7.2 97.8(4) 7,200.0 93.6 19.6 98.0(4) 7,200.0 93.8 8.8
2014 INFORMS
30 3 4 100.0 2,027.8 97.9 62.5 99.8(2) 4,463.3 98.6 6.7 96.3(4) 7,200.0 92.3 28.0 97.1(4) 7,200.0 92.7 9.8
35 3 3 100.0 1,221.2 98.5 52.7 99.5(2) 4,279.9 98.7 13.2 95.6(4) 7,200.0 94.6 25.4 96.0(4) 7,200.0 94.8 11.6
35 3 4 99.0(3) 6,515.4 97.5 70.0 99.0(4) 7,200.0 98.2 12.6 94.6(4) 7,200.0 93.6 36.8 95.4(4) 7,200.0 93.9 11.5
40 3 3 99.7(1) 4,097.6 98.3 100.1 99.4(4) 7,200.0 98.7 35.8 97.6(4) 7,200.0 95.0 88.6 97.8(4) 7,200.0 95.1 16.6
40 3 4 98.3(3) 5,970.7 97.4 168.9 98.9(4) 7,200.0 97.9 35.7 95.3(4) 7,200.0 93.1 100.0 95.8(4) 7,200.0 93.3 27.1
45 3 3 99.1(3) 5,647.0 98.2 324.6 99.4(4) 7,200.0 98.7 50.3 96.7(4) 7,200.0 95.2 84.3 97.0(4) 7,200.0 95.4 36.2
45 3 4 97.2(4) 7,200.0 96.4 482.6 97.8(4) 7,200.0 97.2 71.3 95.0(4) 7,200.0 93.8 129.5 95.7(4) 7,200.0 94.1 48.4
50 3 3 99.3(2) 5,243.6 98.6 215.1 99.3(4) 7,200.0 98.8 51.8 96.2(4) 7,200.0 94.8 315.6 96.5(4) 7,200.0 94.7 57.5
50 3 4 98.3(3) 7,058.4 97.7 407.0 98.5(4) 7,200.0 97.9 64.7 94.5(4) 7,200.0 93.1 597.5 95.1(4) 7,200.0 93.2 93.6
10 6 2 100.0 1.9 97.4 0.5 100.0 0.6 97.6 0.3 100.0 0.5 96.3 0.3 100.0 0.3 96.9 0.2
10 6 3 100.0 12.5 96.6 1.2 100.0 15.1 96.5 0.5 100.0 1.3 96.4 0.5 100.0 0.2 98.4 0.2
15 6 2 100.0 97.5 96.8 3.6 99.8(1) 1,940.9 97.3 2.3 100.0 8.6 96.4 0.8 100.0 35.3 97.2 0.9
15 6 3 100.0 1,105.8 95.6 8.6 99.5(4) 7,200.0 96.5 2.7 100.0 73.8 95.4 3.0 100.0 146.8 96.7 1.1
20 6 2 100.0 84.9 97.4 11.0 99.8(1) 2,501.2 98.1 4.2 100.0 23.7 96.8 1.9 100.0 281.6 97.6 2.3
20 6 3 100.0 806.5 96.9 29.8 99.5(2) 3,608.5 97.6 5.3 100.0 269.7 95.8 4.5 100.0 297.2 97.0 2.5
25 6 2 100.0 170.6 97.8 45.3 99.9(2) 3,663.4 98.5 15.1 100.0 328.8 96.4 5.5 100.0 1,552.4 96.4 3.7
25 6 3 99.9(1) 2,811.0 97.2 54.2 99.4(2) 4,150.6 98.0 14.2 99.3(4) 7,200.0 95.4 12.4 99.4(4) 7,200.0 95.6 3.7
30 6 3 99.4(2) 4,347.1 97.1 301.3 99.3(4) 7,200.0 98.1 41.7 100.0 1,634.9 97.5 24.3 99.7(2) 4,002.2 97.7 8.7
30 6 4 97.9(4) 7,200.0 95.9 527.5 98.8(4) 7,200.0 97.4 39.2 98.9(4) 7,200.0 96.8 83.1 99.2(4) 7,200.0 96.9 9.1
10 9 2 100.0 20.5 98.4 1.5 100.0 30.2 98.7 1.1 100.0 15.3 98.0 1.0 100.0 4.5 98.2 0.8
10 9 3 100.0 405.2 97.6 2.5 100.0 1,095.9 98.4 1.2 100.0 75.3 97.0 1.7 100.0 35.5 97.4 0.7
15 9 2 99.6(2) 3,797.7 96.8 8.2 99.5(4) 7,200.0 97.5 5.2 100.0 887.1 96.9 3.5 99.4(2) 5,143.6 97.1 2.4
15 9 3 97.4(4) 7,200.0 95.0 18.5 98.7(4) 7,200.0 96.3 7.0 98.6(2) 5,612.0 95.5 9.4 98.9(4) 7,200.0 96.2 3.7
20 9 2 100.0 1,480.0 98.1 43.2 99.6(2) 5,859.9 98.6 10.4 100.0 590.9 97.1 6.4 99.7(2) 4,081.3 97.5 4.8
20 9 3 98.9(2) 4,546.3 97.3 50.2 99.2(4) 7,200.0 97.7 13.5 99.2(3) 6,430.5 96.1 12.9 99.3(3) 5,737.2 96.8 6.9
Optimal 100.0 371.5 97.8 14.8 99.9 1,777.7 98.3 3.6 100.0 397.5 96.5 3.3 99.9 1,281.8 97.1 1.8
Not optimal 98.8 5,510.4 97.2 209.9 99.0 6,965.4 97.9 34.1 96.9 7,042.8 94.5 103.2 97.3 7,102.5 94.8 23.2
Total 99.5 2,336.4 97.6 89.4 99.6 3,761.2 98.1 15.2 98.6 3,329.3 95.6 47.3 98.8 3,849.8 96.0 11.2
(−)
Indicates the number of instances (out of 4) that were not solved to optimality.
115
Downloaded from informs.org by [14.139.108.35] on 11 July 2023, at 22:27 . For personal use only, all rights reserved.
116
Table 7 Average Results on MVIRP Instances
MVIRP-ML MVIRP-OU
F (ML) | k F (ML) | nk F (OU) | k F (OU) | nk
n l m hi %LB CPU %RLB RCPU %LB CPU %RLB RCPU %LB CPU %RLB RCPU %LB CPU %RLB RCPU
5 3 2 L 100.0 0.1 90.7 0.1 100.0 0.1 93.2 0.0 100.0 0.0 99.7 0.0 100.0 0.1 99.4 0.1
5 3 3 L 100.0 0.3 88.4 0.1 100.0 0.2 94.3 0.0 100.0 0.0 99.5 0.0 100.0 0.1 97.3 0.0
10 3 2 L 100.0 1.3 87.4 0.3 100.0 3.3 92.4 0.2 100.0 2.0 85.5 0.5 100.0 3.0 88.9 0.4
10 3 3 L 100.0 8.3 79.5 0.5 100.0 19.2 89.7 0.2 100.0 4.8 86.0 1.0 100.0 4.4 90.8 0.3
15 3 2 L 100.0 4.4 90.1 0.8 100.0 26.7 92.6 0.5 100.0 5.4 87.7 1.2 100.0 26.7 87.4 0.9
15 3 3 L 100.0 38.5 85.0 1.7 100.0 1,701.0 89.3 0.6 100.0 19.1 87.2 2.1 100.0 83.3 87.6 1.0
20 3 2 L 100.0 31.2 89.5 2.4 100.0 1,054.5 92.5 1.2 100.0 67.1 85.3 3.7 99.8(1) 1,791.9 87.2 2.8
20 3 3 L 100.0 386.8 83.0 4.0 96.9(3) 4,540.2 89.5 2.0 100.0 509.0 81.9 5.7 98.4(2) 4,400.1 87.2 3.3
25 3 2 L 100.0 71.0 92.5 3.6 99.1(1) 1,837.9 94.6 2.8 100.0 127.8 90.2 5.8 100.0 503.6 90.7 5.4
25 3 3 L 100.0 978.3 86.8 6.4 96.2(5) 7,200.0 91.1 4.0 100.0 1,035.3 85.6 11.0 98.7(3) 5,764.7 86.0 4.9
30 3 3 L 100.0 1,962.3 87.5 16.7 94.5(4) 6,091.3 91.3 8.0 100.0 3,054.1 85.6 34.0 97.1(4) 6,593.6 86.8 10.2
30 3 4 L 93.4(4) 6,203.8 82.1 26.1 92.8(5) 7,200.0 89.6 11.2 91.8(4) 6,586.0 82.0 39.7 94.4(5) 7,200.0 85.0 10.8
Adulyasak, Cordeau, and Jans: Formulations and Branch-and-Cut for the MVPRP
35 3 3 L 100.0 4,124.9 87.9 26.8 94.1(5) 7,200.0 91.6 12.1 96.4(3) 5,262.4 86.4 56.1 96.1(5) 7,200.0 88.3 25.2
35 3 4 L 90.4(5) 7,200.0 83.0 51.8 91.5(5) 7,200.0 89.3 13.4 86.9(5) 7,200.0 81.3 81.5 91.5(5) 7,200.0 84.1 27.6
40 3 3 L 97.3(3) 5,699.9 89.0 37.5 94.1(5) 7,200.0 92.2 17.8 97.1(4) 7,078.1 87.7 89.4 95.9(4) 7,200.0 87.8 32.1
40 3 4 L 86.7(5) 7,200.0 82.7 66.2 90.3(5) 7,200.0 88.2 27.7 83.9(5) 7,200.0 80.7 184.8 89.5(5) 7,200.0 81.8 30.8
5 6 2 L 100.0 4.0 86.4 0.2 100.0 15.6 92.8 0.1 100.0 1.5 95.2 0.8 100.0 4.6 90.0 0.2
5 6 3 L 100.0 140.9 80.8 0.2 100.0 273.3 92.9 0.1 100.0 0.2 99.6 0.1 100.0 3.6 94.1 0.1
2014 INFORMS
Adulyasak, Cordeau, and Jans: Formulations and Branch-and-Cut for the
MVPRP 1
the results of using each cut presented in §3.2, where
RCPU
0.2
0.1
0.8
1.0
2.3
2.8
2.0
6.9
29.3
30.2
17.4
the columns None and All present the results without
using any additional cuts (i.e., only GFSECs (28) and
%RLB
(53) are applied) and with all cuts together, respec-
95.8
92.8
93.7
96.8
91.8
89.5
93.2
90.1
92.3
88.9
91.2
tively. The results show that adding all the cuts
together generally provides the best lower bounds,
4.4
4.0
6,567.8
7,200.0
2,428.4
5,298.3
7,200.0
7,200.0
1,935.4
6,893.9
3,513.1
and it has more effect on larger instances. We also
CPU
lower bound at the root node for the MVIRP than for
F (OU) | nk
the MVPRP.
98.7(4)
99.8(1)
97.9(3)
98.0(5)
95.6(5)
95.8(5)
%LB
95.5
98.3
100.0
100.0
99.5
In Table 5 we report the average results of using all
MVIRP-OU
0.7
0.1
1.3
2.5
3.8
7.2
4.4
64.6
23.6
102.9
167.4
|
inequalities of §3.2 could provide significant
95.7
92.4
96.9
99.9
89.4
87.1
93.0
88.9
91.8
87.3
90.4
improve-
ments in the branch-and-cut procedure for both for-
6,700.1
7,200.0
1.4
0.2
272.4
3,561.8
763.1
7,200.0
392.6
6,269.5
2,262.6
99.3(1)
96.4(5)
%LB
98.3
100.0
100.0
100.0
100.0
100.0
94.8
0.1
0.1
0.5
0.7
1.9
2.9
2.2
5.4
27.5
35.6
13.8
97.4
95.7
95.2
95.3
94.6
93.2
95.0
92.1
93.9
92.2
93.4
3,013.5
5,668.3
7,200.0
7,200.0
2,507.1
6,979.3
3,726.8
|
further computational experiments. |
CPU
F (ML) | nk
99.1(2)
97.7(3)
98.3(4)
96.2(5)
96.6(5)
100.0
100.0
99.0
95.1
97.9
MVIRP-ML
34.2
12.4
RCPU
96.4
93.8
91.3
87.1
91.4
87.3
94.4
90.6
90.1
88.0
89.6
97.9(2)
98.1(5)
%LB
95.5
98.8
100.0
100.0
100.0
100.0
100.0
H
H
H
H
H
H
H
H
Not Optimal
10
10
15
15
5
5
could generally provide better lower bounds at the known production quantity is made available in each
root node and also better overall gaps (%LB) on the period, the algorithm has to be modified to put more
instances that were not solved to optimality within emphasis on the inventory and distribution part.
two hours, except for the MVIRP-ML instances. The We also found that by expanding the neighborhood
average %LB of the nonvehicle index formulations (i.e., the number of customer-period combinations) by
after two hours is slightly better than for the vehicle 15%–30%, the average solution quality can be further
index formulations on the MVPRP instances, whereas improved by approximately 1%–2% but with double
the lower bounds produced by the vehicle index the average computing time. However, the current
Downloaded from informs.org by [14.139.108.35] on 11 July 2023, at 22:27 . For personal use only, all rights reserved.
formulation are better on the MVIRP-ML instances setting is preferable because it has been shown in
and as good as the nonvehicle index formulation Adulyasak et al. (2012) that the Op-ALNS procedure
on the MVIRP-OU. The computing times on the could handle small to very large PRP instances effec-
MVIRP instances increase in a greater ratio when tively. It is worth noting that the Op-ALNS is much
the number of periods increases, compared to the faster than the MVIRP heuristic presented by Coelho
MVPRP instances. The largest instance sizes that can et al. (2012), which had an average computing time
be solved to optimality are 35c/3p/3v for MVPRP- of 2,000 seconds for the instances with three periods
ML and MVIRP-ML, 30c/3p/3v for MVIRP-OU, and and 8,000 seconds for the instances with six periods.
25c/3p/2v for the MVPRP-OU. In Table 8, we also We also remark that, on the instances with n≤ 15 that
report some statistics for the best found solutions. were solved to optimality, setting the initial upper
Note that the solutions of the MVPRP-OU instances bounds in CPLEX leads to an average improvement
with l 3 have on average more setups than l 6 of 11.8% and 6.2% in computing time on the MVPRP
because= they have lower initial inventory levels,=as and MVIRP instances, respectively.
described in the online supplement.
We further remark the performance of the heuris- 6.4. Performance of the Branch-and-Cut
tic used to set the initial upper bounds for the Algorithm on Multicore Processors
branch-and-cut algorithms. The Op-ALNS could gen- Nowadays, most computers have multiple core pro-
erally provide high quality solutions on the MVPRP cessors, and modern solvers like CPLEX have the
instances, where the deviations from the optimal capability to perform parallel optimization. In this
solutions or best upper bounds for the MVPRP-ML section, we report the results of experiments per-
and MVPRP-OU are 1.2% and 0.8%, respectively, and formed on an eight-core processor. We first show
the corresponding CPU times are 30.4 and 31.0 sec- the comparison of the average results using single
onds, respectively. The results on the MVIRP are not and multiple core processors on the instances that
as good as for the MVPRP, where gaps of the best were solved to optimality by the single core machine
upper bound for the MVIRP-ML and MVIRP-OU are in the previous section (see Table 9). In the results
3.5% and 4.8%, respectively, and the corresponding for the eight-core processor, columns CPU and WC
CPU times are 17.4 and 19.8 seconds, respectively. show the average total aggregate CPU time of all
Note that the details on the quality of the initial cores and average wall clock time, respectively.
upper bounds obtained by the Op-ALNS procedure Because we did not run other tasks when running the
can be found in the online supplement. algorithm, the wall clock time is a good
Because the Op-ALNS was originally developed for approximation of the max- imum CPU time spent
the PRP—where it has to take the production and set- on all cores. We also calcu- late the ratio of the
up costs into account—the operators mainly attempt aggregate CPU time and wall clock time (CPU/WC)
to find a solution where production and distribution in column C/W to show the benefit of parallel
costs are minimized. In the IRP, however, because a computing on multicore processors. Because the
main purpose of the experiments is to
MVPRP MVIRP
ML OU ML OU
Values l=3 l=6 l=9 l=3 l=6 l=9 l=3 l=6 l=3 l=6
Average number of setups 1.10 2.00 3.00 2.71 2.00 3.08 n/a n/a n/a n/a
Average number of visits per customer 1.23 1.71 3.08 1.00 1.31 2.07 1.08 2.68 1.12 2.73
Average number of dispatched vehicles
Instances with 2 vehicles 1.38 1.41 1.63 2.00 1.51 1.62 1.59 1.69 1.40 1.53
Instances with 3 vehicles 1.75 1.81 1.84 2.82 2.08 2.18 2.14 2.29 1.95 2.32
Instances with 4 vehicles 2.27 2.33 n/a 3.65 2.83 n/a 2.57 n/a 2.43 n/a
Average vehicle usage (%) 76.50 83.40 84.20 89.20 76.80 73.70 86.80 90.00 82.40 82.40
Adulyasak, Cordeau, and Jans: Formulations and Branch-and-Cut for the
MVPRP 1
Table 9 Average Results Using Single and Multicore Processors on MVPRP and MVIRP Instances Solved to Optimality
Table 10 Summary of the Largest Instances Solved to Optimality Using Single and Multiple Core Processors