Accepted Manuscript

A genetic algorithm with exact dynamic programming for the green vehicle routing &
scheduling problem

Yiyong Xiao, Abdullah Konak

PII: S0959-6526(16)31967-9
DOI: 10.1016/j.jclepro.2016.11.115
Reference: JCLP 8504

To appear in: Journal of Cleaner Production

Received Date: 8 June 2016

Revised Date: 29 September 2016
Accepted Date: 18 November 2016

Please cite this article as: Xiao Y, Konak A, A genetic algorithm with exact dynamic programming for
the green vehicle routing & scheduling problem, Journal of Cleaner Production (2016), doi: 10.1016/
j.jclepro.2016.11.115.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT
Word count result:
Words: 12,577
Characters (no spaces): 64,541
Characters (with spaces): 76,169

A genetic algorithm with exact dynamic programming for the green vehicle
routing & scheduling problem

PT
Yiyong Xiao1, Abdullah Konak2
1
School of Reliability and System Engineering, Beihang University, Beijing 100191, China

RI
2
Information Sciences and Technology, Penn State Berks, Tulpehocken Road, P.O. Box 7009, Reading, PA 19610-6009, United States

Abstract: Traffic congestion significantly increases CO2 (a well-known greenhouse gas) emissions of vehicles in

SC
road transportation and causes other environmental costs as well. A road-based delivery company can reduce its
CO2 emissions through operational decisions such as efficient vehicle routes and delivery schedules by considering
time-varying traffic congestion in its service area. In this paper, we study the time-dependent vehicle routing &

U
scheduling problem with CO2 emissions optimization (TD-VRSP-CO2) and develop an exact dynamic
programming algorithm to determine the optimal vehicle schedules for given vehicle routes. A hybrid solution
AN
approach that combines a genetic algorithm with the exact dynamic programming procedure (GA-DP) is proposed
as an efficient solution approach for the TD-VRSP-CO2. Computational experiments on 30 small-sized instances
and 14 large-sized instances are used to study the efficiency and effectiveness of the proposed hybrid optimization
M

approach with promising results. Contributions of this study can help road-based delivery companies be ready for a
low-carbon economy and also help individual vehicle drivers make better vehicle scheduling plans with lower CO2
emissions and fuel consumption.
D

Keywords: CO2 emissions; green logistics; sustainability; dynamic programming; hybrid optimization
TE

1. Introduction
The 2014 National Climate Assessment report (Melillo et al, 2014) unambiguously reaffirms that the global
warming of the past 50 year has been caused by the human activity, particularly burning of coal, oil, and gas that
EP

leads the concentration of heat-trapping gases, mainly the carbon dioxide (CO2), in the atmosphere. The keeling
curve in (http://keelingcurve.ucsd.edu/), which has been perpetually measuring the concentration of CO2 in Earth's
atmosphere since 1958, has reached the highest, a record setting level and indicates an accelerated increase in the
C

past decade. If the World continues along a business-as-usual energy path, the Intergovernmental Panel on Climate
Change predicts that there is a fifty-fifty chance that the temperature will exceed 5 degrees by the end of this
AC

century, and many species, including Humans, will have a hard time in adapting to the changing climate (Chu,
2009). The world economic development should be steered towards a more sustainable model.
According to a report published by the International Energy Agency (IEA, 2015), the transportation sector was
the second-largest contributor of CO2 emissions, counting for 23% of the global CO2 emissions in 2015, road
transportation sector was responsible for almost three-quarters of the total emissions due to transportation activities.
In urban areas, traffic congestion is one of the factors that substantially increase CO2 emissions by vehicles. Based
on the results of an empirical study to investigate the impact of traffic congestion on CO2 emissions in Southern
California, Barth and Boriboonsomsin (2008) point out that CO2 emissions can be reduced by up to 20% through
congestion mitigation strategies. Supply chain researchers suggest that commercial vehicle emissions can be
significantly reduced by implementing operation research techniques to avoid traffic congestion and cooperation
 ACCEPTED MANUSCRIPT
strategies in supply chain (Benjaafar et al., 2013). In context of the vehicle routing problem (VRP) which deals
with operational level fleet management decisions, researchers have proposed new optimization models that
consider both traditional monetary costs and CO2 emissions (Figliozzi, 2010; Bektas and Laporte, 2011; Jabali et al.,
2012; Kwon et al., 2013; Franceschetti et al., 2013; Demir et al. 2014b; Zhu et al., 2014; Xiao and Konak, 2015b,
2016). In practice, traffic conditions in urban areas are often time-varying and non-uniformly distributed over
different regions. While determining of the routes and schedules of fleet vehicles, considering variations in traffic
congestion in addition to distance-based performance metrics can lead to substantial reductions in CO2 emissions.
This study is motivated to develop efficient solution approaches for the routing and scheduling optimization for

PT
a heterogeneous vehicle fleet operating in a time-varying traffic environment, with objective of minimizing
the total CO2 emissions. The specified problem is also called the time-dependent vehicle routing and scheduling
problem with CO2 emissions optimization (TD-VRSP-CO2 for short). The formulated problem is important for the

RI
economic and environmental operations of logistic companies located in urban areas. This paper is an extension
study based on Xiao and Konak (2015b) with three main contributions as follows: (i) an enhanced mixed integer
linear programming (MILP) model for the TD-VRSP-CO2 is presented to consider the impact of varying payload

SC
weight on vehicle’s emissions, (ii) an exact algorithm based on dynamic programming (DP) is proposed for the
scheduling part the MILP model, and (iii) a hybrid genetic algorithm (GA) combined with the exact DP algorithm,
named GA-DP, is proposed as an efficient solution approach for the TD-VRSP-CO2.

U
The rest of the paper is organized as follows. In Section 2, related papers in literature are reviewed. In Section 3,
AN
The TD-VRSP-CO2 problem is reformulated as a MILP model in addition to considering the weight impact and
tardiness penalty. In Section 4, we propose an exact dynamic programming method to schedule vehicles with
optimal discrete departure times, arriving times, and travel distances in different periods. In addition, a genetic
algorithm with dynamic programming (GA-DP) is presented to solve the TD-VRSP-CO2. In Section 5,
M

computational experiments are carried out to examine the MILP model and the performance of the proposed hybrid
algorithm. Finally, we conclude the paper in Section 6.
D

2. Related literature review

TE

In recent years, green logistics, which aims at improving environmental sustainability in the transport sector by
reducing fossil fuel consumption, has been a subject undergoing intense studies and attracted the attention of
Operations Research professionals (Dekker et al., 2012; Erdogan and Miller-Hooks, 2012; Govindan et al., 2014;
EP

Yang et al., 2016; etc.). Use of clean energy vehicles in supply chain systems is also encouraged and practiced in
real-life applications (Günther et al., 2015; Kushwah and Sharma, 2016). Although developments of new
technologies on alternative-energy vehicles, such as electronic, hydrogen, solar, etc., have been strongly supported
C

in many countries for many years, the transportation sectors of the today’s world are still heavily dependent on
fossil-based fuels. Responding to the emission regulatory policies imposed by governments, firms in the same
AC

supply chain devise collaboration initiatives to reduce the carbon footprint of the entire supply chain through
coordinated operational decisions (Benjaafar et al., 2013). Recent theoretical models and practices on reducing CO2
emissions in various industries can be found in (Validi et al., 2012, 2014a, 2014b, and 2015) and (Soysal et al.
2015).
The vehicle routing problem (VRP) is a classic optimization problem in road transportation and has been widely
studied and practiced since it was first introduced by Dantzig and Ramser (1959). In general, the traditional VRP
involves determining the optimal routes for a fleet of vehicles started from a depot with the objective of satisfying
customers’ requirements at minimum cost that is usually expressed as a function of the total distance traveled.
Various versions of VRPs have been developed for different applications, such as pickup and delivery VRP,
capacitated VRP (CVRP), multi-depot VRP, VRP with time windows (VRPTW), split delivery VRP,
 ACCEPTED MANUSCRIPT
time-dependent VRP (TD-VRP), etc. Recent surveys on various VRP formulations and related heuristic
techniques/algorithms can be found in Laporte et al. (2000), Toth and Vigo (2002), Golden et al. (2008), and
Eksioglu et al. (2009).
Another important variant of the VRP in recent years focuses on fuel consumption optimization. As stated in
Xiao et al. (2012), the amount of fuel consumed during operation of a fleet is in fact a greater concern to
transportation companies than the total distance traveled. Fuel cost is a significant part of the transportation cost
and may account for more than 60% of the total transportation cost (Sahin et al., 2009). Ericsson et al. (2006)
identified the impact of traffic disturbance events on fuel consumption and proposed a model for estimating the

PT
potential reduction in fuel consumption through route optimization. Kara et al. (2007) first proposed the Energy
Minimizing VRP that aims to minimize the total energy consumed along vehicle routes instead of the traditional
total distance. Kuo (2010) proposed a VRP model with a fuel consumption objective in a time-dependent traffic

RI
environment and developed a simulating annealing (SA) algorithm as the solution approach. Xiao et al. (2012)
incorporated the fuel consumption rate resulted from payload, which can be decreasing or increasing along the tour,
into the CVRP and developed an optimization model considering fuel consumption rate. Gaur et al. (2013) studied

SC
the cumulative VRPs with the objective of minimizing fuel consumption and proposed an approximation algorithm
for a version of the problem in which vehicles have finite capacity and an arbitrary number of depot offloads are
allowed. Zhu et al. (2014) proposed a mathematical model for a routing problem that requires the identification of a

U
safest route in terms of the use of roads and bridges with geometric and fuel consumption limitations.
AN
The earliest work addressing CO2 emissions optimization in the context of the VRP is the paper by Figliozzi
(2010) who proposed a partial emissions model to optimize vehicle departure times calculated based on the
traditional time-dependent VRP. Jabali et al. (2012) developed a model for an emissions-based time-dependent
VRP where a uniform speed limit is applied to all arcs. Kwon et al. (2013) developed a heterogeneous, fixed fleet
M

VRP model that takes carbon-trading costs into account. Ren et al. (2016) carried out a case study for evaluating
the vehicle emissions due to inter-city passenger transport in Shenyang Metropolitan Area. Xiao and Konak (2015b)
D

presented the green vehicle routing and scheduling problem (GVRSP) considering the scenario of time varying
traffic congestion and formulated it as a mixed integer linear programming (MILP) model. The GVRSP aims at
TE

minimizing CO2 emissions through better planning of routes and schedules for delivery/pickup assignments of a
fleet of heterogeneous vehicles in a time-varying traffic environment with multiple periods. In Xiao and Konak
(2016), a more comprehensive MILP model was formulated for the GVRSP considering dynamic vehicle payloads.
EP

Ehmke et al. (2016) also assumed that vehicles move at the speed of traffic under varying traffic congestions and
demonstrated that significant reduction in CO2 can be achieved if emission-based objectives are used instead of the
traditional time-based or distance-based objectives particularly in urban settings.
C

Time-varying traffic conditions cause travel times, vehicle speeds, and fuel consumption to be time-dependent,
and the VRP under this case is called time-dependent VRP (TD-VRP) (Malandraki and Daskin, 1992; Chen et al.,
AC

2006; Maden et al., 2010; Figliozzi, 2012). Compared with the traditional distance-based VRP, the TD-VRP also
considers the departure time of each vehicle from each node visited by the vehicle as a decision variable, and it is
usually studied with an alternative objective function such as minimizing the total travel time, the total fuel
consumption, or CO2 emissions (Figliozzi, 2010; Kuo 2010). The TD-VRP is computationally much more
challenging than the traditional VRP because the solution space is exponentially expanded by the introduction of
the departure time (or service starting time) decisions that are often modeled as integer decision variables.
Malandraki and Dial (1996) first proposed a restricted dynamic programming heuristic for constructing the route
for time-dependent traveling salesperson problem (TDTSP). Kok et al. (2010) proposed a restricted dynamic
programming heuristic for constructing the vehicle routes to minimize the total duty time and developed a heuristic
for optimizing the vehicle’s departure times considering driving hours regulations. Kok et al. (2011) presented the
 ACCEPTED MANUSCRIPT
vehicle departure time optimization (VDO) problem and proposed a MILP formulation to minimize the total duty
time for the TD-VRP.
Recently, the pollution routing problem (PRP), another stream of environmental-sustainability-oriented variant of
the VRP raised by Bektas and Laporte (2011), has attracted the attention of researchers. The PRP extends the
classical VRP with time windows (VRPTW) where the pick-up from and/or delivery to each customer node must
take place within a given time-window of the customer. The main departure of PRP from the VRPTW is that the
PRP treats the travel speed of a vehicle on each arc of its route as a decision variable and considers the dynamic
payload of the vehicle through its route in order to determine the CO2 emissions rate precisely. Demir et al. (2012)

PT
developed a two-stage heuristic algorithm to solve the PRP of large-size instances with up to 200 nodes. In this
two-stage approach, the PRP is solved as the traditional VRPTW in the first stage by a large neighborhood search
(LNS) heuristic, and then the optimal speed for each arc of the route is determined in the second stage.

RI
Franceschetti et al. (2013) extended the PRP model of Bektas and Laporte (2011) to the Time-Dependent PRP
model by considering two periods of traffic congestion. Demir et al. (2014b) proposed an adaptive large
neighborhood search algorithm (ALNS) to minimize the fuel consumption and the driving time with Pareto

SC
optimality. Koc et al. (2014) introduced the fleet size and mix pollution-routing problem that extends the PRP by
considering a heterogeneous vehicle fleet. Kramer et al. (2015) proposed a new speed and departure time
optimization algorithm for the PRP, which runs in quadratic time and returns an optimal schedule.

U
In this study, we propose an exact dynamic programming (DP) algorithm to optimize travel schedules (i.e., the
AN
departure/arrival time and waiting/idle time in different periods) of given routes of heterogeneous vehicles
operating in a time-varying traffic environment. The DP algorithm has a polynomial complexity of O((nT)2) for the
worst case and is also applicable to other TD-VRPs in the optimization of the departure times. To this sense, with
the newly developed DP algorithm, traditional TD-VRPs would be no longer more complex than traditional VRPs
M

in theory. We develop a hybrid genetic algorithm incorporated with the DP algorithm for the GVRSP in Xiao and
Konak (2015b, 2016), and examine its solution efficiency and effectiveness by using synthesized problems and
D

well-known existing benchmark problems as well. The experimental results show the proposed algorithm is
effective and efficient.
TE

3. The green vehicle routing & scheduling problem (GVRSP)

The GVRSP was originally presented in Xiao and Konak (2015a), and a more comprehensive version of the
EP

GVRSP with heterogeneous fleet can be found in Xiao and Konak (2016). In this paper, we consider the GVRSP
with tardiness objective that occurs frequently when it is impossible to satisfy all customers’ orders on time due to
limited time/capacity. The tardiness objective smoothes the solution space of the problem and is often used in the
C

scheduling literature to model the timeliness of customer service. The defined problem is called TD-VRSP-CO2
because it is also an extension of traditional TD-VRP with an alternative objective of CO2 emissions.
AC

The TD-VRSP-CO2 can be described as follows. A fleet of heterogeneous vehicles is going to visit n customers
randomly located in a region. All vehicles start their tours from a single depot and must return to the depot after
completing their tours. The TD-VRSP-CO2 is defined on a directed network G(N,A) where N={0,1,…,n} denotes
set of the nodes, and A={(i, j): i ≠j and i, j∈N} denotes the set of arcs. The depot is represented by Node 0, and
N′=N\{0} is the node set of customers. Each arc (i, j) represents the shortest path from node i to j, and the
distance of arc (i, j) is denoted by Dij. Each customer i is associated with a demand Ri (measured in weight), a
service time gi, and a due-time Ei before which the service must start. A late visit of customer i will result in a
tardiness penalty weighted by a coefficient ωi in the objective function. The planning horizon is divided into m
periods, each of which can have different lengths. The travel speeds of arcs are dynamic, i.e., the travel speed on
an arc changes from a period to another one. During a period, however, the travel speed of the arc is assumed to
 ACCEPTED MANUSCRIPT
be known and constant. Fleet vehicles can be of different types, and each vehicle has a maximum payload
capacity, a maximum travel range, and a known CO2 emissions rate (e.g., Lb/mile) function with respect to
different travel speeds.
According to the comprehensive module emissions model (CMEM) by Barth et al. (2005) and Barth and
Boriboonsomsin (2008), the payload of a vehicle affects its CO2 emissions by increasing the vehicle’s rolling
friction on the road, which is only proportional to the traveled distance. Therefore, we can calculate the total
amount of CO2 emitted by vehicle h while it travels arc (i, j) as follows:

∑c ⋅ d ijkh + ϕ h ⋅ fijh ⋅ Dij

PT
ijkh
k ∈K

where cijkh the amount of CO2 emitted by vehicle h per unit of distance traveled on arc (i, j) in time period k, dijkh is
the total distance that vehicle h travels on arc (i, j) in period k, and fijh is the payload weight of vehicle h on arc (i, j),

RI
and φh is the marginal CO2 emissions of vehicle h per unit payload weight per unit distance. The first term is the
amount of CO2 emitted by an empty vehicle while traveling on the arc in a given period. Because the average
travel speed on an arc (i, j) depends on the time of the day and can be estimated based on historical data, it is

SC
possible to pre-calculate cijkh and use it in the model. The second term is related with CO2 emissions due to the
payload of vehicle h. Note that the payload weight is dynamic and depends on the order that the vehicle visits its
assigned customers during its tour.

U
Based on the scenario and assumptions described above, the TD-VRSP-CO2 is formulated as a MILP model.
The parameters used in the MILP model are listed in advance as follows.
AN
F objective function, indicating the total amount of CO2 emitted by all vehicles.
i, j index of nodes, i=0,1,2,…,n (the depot is represented by 0)
n total number of customer nodes
M

N set of nodes including the depot, i ∈ N

N' set of customers excluding the depot, N ′ = N \{0}
set of arcs formed by all pairs of nodes, (i, j) ∈ A, ∀i ∈ N, j ∈ N, i ≠ j
D

A
Dij distance of arc (i, j)
TE

Ri demand of customer i
Ei due-time of servicing customer i
ωi tardiness penalty coefficient of customer i
EP

gi service time needed for customer i

h index of vehicle, h=1,2,…,q
q number of vehicles
set of vehicles, h ∈ H
C

H
Ch maximum payload capacity of vehicle h
AC

Lh maximum travel length of vehicle h

k index of time period, k=1,2,…,m
K set of time periods, k ∈ K
m number of time periods, m=card(K)
[bk, ek] beginning and ending time of time period k
vijk vehicle’s travel speed on arc (i, j) within time period k.
cijkh CO2 emission of vehicle h for per unit of traveled distance on arc (i, j) in time period k
φh additional amount of CO2 emitted by vehicle h when traveling with one more unit of payload
The binary decision and continuous decision variables used in the MILP model are as follows:
Xij {0,1} variable indicating whether arc (i, j) is traveled (1) or not (0)
yijh {0,1} variable indicating whether arc (i, j) is traveled by vehicle h (in any period) (1) or not (0)
 ACCEPTED MANUSCRIPT
xijkh {0,1} variable indicating whether arc (i, j) is traveled in period k by vehicle h (1) or not (0)
dijkh non-negative continuous variable indicating the length of arc (i, j) scheduled to be travelled by
vehicle h in period k
τijkh non-negative continuous variable indicating the travel time corresponding to dijkh
fijh non-negative continuous variable indicating the payload of vehicle h when traveling arc (i, j)
li non-negative continuous variable indicating the departure time of a vehicle from node i (the earliest
leave for i=0)
ai non-negative continuous variable indicating the arrival time of a vehicle at node i (the latest arrival

PT
for i=0)
Oi non-negative continuous variable indicating tardiness for serving customer i

The MILP formulation for the TD-VRSP-CO2 is given as follows:

RI
Problem TD-VRSP-CO2:

Min F= ∑ ∑ ∑c ijkh ⋅ dijkh + ∑ ∑ ϕh ⋅ fijh ⋅ Dij + ∑ ωi ⋅ Oi (1)

SC
h∈H ( i , j )∈ A k ∈K h∈H ( i , j )∈ A i∈N ′

Subject to:

(1) ∑X ij =1 ∀j ∈ N ′

U
i∈N

∑X =1 ∀i ∈ N ′
AN
(2) ij
j∈N

(3) X ij = ∑ yijh ∀(i, j ) ∈ A

h∈H
M

(4) yijh ≥ xijkh ∀(i, j ) ∈ A, k ∈ K , h ∈ H

(5) yijh ≤ ∑ xijkh ∀(i, j ) ∈ A, h ∈ H

D

k ∈K
TE

(6) ∑
( i , j )∈ A
yijh = ∑
( j ,i )∈ A
y jih ∀h ∈ H , j ∈ N

(7) ∑ y0 jh ≤ 1 ∀h ∈ H
EP

(0, j )∈ A

(8) dijkh ≤ Dij ⋅ xijkh ∀(i, j ) ∈ A, k ∈ K , h ∈ H

∑ ∑d = X ij Dij ∀(i, j ) ∈ A
C

(9) ijkh
k∈K h∈H
AC

(10) τ ijkh = 60 ⋅ dijkh / vijk ∀(i, j ) ∈ A, k ∈ K , h ∈ H

(11) ∑τ
(i , j )∈ A
ijkh ≤ ek − bk ∀k ∈ K , h ∈ H

(12) li ≤ ek − τ ijkh +em ⋅ (1 − xijkh ) ∀(i, j ) ∈ A, k ∈ K , h ∈ H

(13) a j ≥ bk + τ ijkh − em ⋅ (1 − xijkh ) ∀(i, j ) ∈ A, k ∈ K , h ∈ H

(14) a j ≥ li + ∑ ∑ τ ijkh − em ⋅ (1 − X ij ) ∀(i, j ) ∈ A

k∈K h∈H

(15) ai + gi ≤ li ∀i ∈ N ′
(16) a0 ≤ em
 ACCEPTED MANUSCRIPT

(17) ∑
h∈H ,i∈N ,i ≠ j
f ijh − ∑
h∈H ,i∈ N ,i ≠ j
f jih = R j ∀j ∈ N ′

(18) fijh ≤ Ch ⋅ yijh ∀h ∈ H ,(i, j ) ∈ A

(19) ∑
( i , j )∈ A
Dij ⋅ yijh ≤ Lh ∀h ∈ H

(20) Oi ≥ ai − Ei ∀i ∈ N ′
In the model above, the objective function includes three parts: (1) the total amount of CO2 emissions of vehicles

PT
without any payload, (2) the total amount of CO2 emissions incurred by vehicles’ payload, and (3) the total of
weighted tardiness penalties. Constraint (1) and Constraint (2) are common VRP constraints that guarantee each
customer node being visited once. Constraint (3) bounds the relationship between variables Xij and yijh.

RI
Constraints (4) and (5) is a linearization of the relation between yijh and xijkh , which is yijh = max{xijkh | k ∈ K } .
Constraint (6) requires that a vehicle visiting a customer node must also depart from the node. Constraint (7)
indicates that each vehicle departs from the depot at most once and prevents the vehicle’s being used again after it

SC
n
returns to the depot. Note that ∑y
j =1
0 jh =0 indicates vehicle h is not utilized. Constraints (8) forces dijkh to be zero

U
if xijkh=0 and enforce dijkh to be less than or equal to Dij for xijkh=1. Constraint (9) ensures that the total distance of
each selected arc (i, j) (i.e., Xij=1) is traveled. This constraint also enforces dijkh to be zero for non-selected arcs (i.e.,
AN
Xij=0). Constraint (10) calculates the time τ ijkh that vehicle h takes to travel a segment (dijkh) of arc (i, j) in period
k. In this constraint, constant 60 converts the time unit from hours to minutes. Note that Constraint (10) is linear
because vijk is a given parameter in the model. Constraint (11) ensures that, for each vehicle and each time period,
M

the total of travel time of vehicle h in period k does not exceed the length of that time period. This constraint
prevents periods from being overused. Constraint (12) bounds the maximum departure time for node i based on
the periods used by the vehicle that visits node i. Similarly, Constraint (13) bounds the minimum arrival time of
D

node i. Note that Constraints (12) and (13) are active only for xijkh=1. When xijkh=0, these two constraints will
always hold because em is a large enough constant to guarantee the validity of the inequalities. Constraint (14)
TE

represents the relationship between the departure time of node i (li) and the arrival time of node j (aj) if arc (i, j) is
selected in a tour (i.e., Xij=1). Constraint (15) ensures that the departure time of each customer node must be after
the arrival time plus the service time. Constraint (16) requires that vehicles return the depot before the end time of
EP

the last period. Note that different vehicles may have different return times but Constraint (16) enforces only the
latest return time. Constraint (17) computes the vehicle’s payload along the tour and eliminates any sub-tours.
Constraint (18) limits the vehicle’s maximum payload by its capacity Ch and forces fijh to be zero if arc (i, j) is not
C

traveled by vehicle h (i.e., yijh=0). Constraint (19) restricts the total travel distance of a tour within the range limit
of the vehicle. Constraint (20) is a linear constraint for calculating the tardiness value. Note that Oi is a
AC

non-negative continuous variable so it will have a zero value if ai is smaller than Ei.
Note that although the objective function includes only CO2 emissions, it can be easily transformed into a
comprehensive cost function including any CO2 emissions related cost, e.g., fuel, tax, carbon emission permit, etc.
In addition, traditional VRP performance metrics such as any travel time/distance related costs, e.g., driver wage,
vehicle rent/depreciation, etc., can be included without any change in the current constraints. One can also define
multiple objective functions, such as monetary cost, environmental concerns, customer satisfaction, etc., to utilize
the same set of decision variables and constraints. The solution approaches introduced in the later sections are also
applicable to such cases.
Note that the TD-VRSP-CO2 is clearly a NP-hard problem because the traditional VRP, its simplified version, is
a NP-hard problem. There does not exist an exact solution approach with polynomial computational time for the
 ACCEPTED MANUSCRIPT
TD-VRSP-CO2. Therefore, a heuristic algorithm is justified to find good solutions for large problem instances.

4 A Genetic Algorithm with Dynamic programming to solve the TD-VRP-CO2

The linearly formulated Problem TD-VRSP-CO2 in Eq.(1) and Constraints (1)-(20) can be optimally solved by
using commercial optimizers such as CPLEX, Lingo etc., but only for small-sized instances because the problem
itself is NP-hard. For large-sized problems, near-optimal solutions can be found by heuristic approaches such as

PT
Genetic Algorithm (GA), Simulated Annealing (SA), Variable Neighborhood Search (VNS), etc. However, one of
the challenges of applying heuristic approaches to solve the TD-VRSP-CO2 is devising an effective encoding
schema to represent multiple types of binary and continuous decision variables at the same time. To deal with this

RI
challenge, we solve the problem in two phases as a routing problem and a scheduling problem. The routing phase
involves selecting arcs to construct a tour for each heterogeneous vehicle, optimizing only the traditional VRP
decision variables Xij and yijh, while the scheduling phase determines when to travel the selected arcs, involving

SC
decision variables xijkh, dijkh, τijkh, li, ai, and Oi. In this section, we first introduce the string representation of routes,
and then propose an exact dynamic programming (DP) method to determine vehicle schedules. Since the proposed
DP has polynomial computational complexity, the complexity of the TD-VRSP-CO2 can be reduced. Therefore,

U
heuristic algorithms like GA are applicable as a solution approach. Finally, we present a GA integrated with
AN
Dynamic Programming (GA-DP).
4.1 String representation of routes
We use a string vector used in Xiao et al. (2012) to encode the route of a vehicle in the GA-DP algorithm. The
M

string vector of a single vehicle h is represented by S = {0, i1 , i2 ,..., inh −1 ,0} , where i1 , i2 ,..., inh −1 are customer IDs
h

to be visited by vehicle h in the order from i1 to inh − 1 by starting from and ending at the depot (i.e., 0), n h − 1 is
D

the number of customers, and n h is the number of arcs in the route. Routes of multiple vehicles can be encoded
TE

by concatenating the string vectors of all vehicles as S = {S + S + ... + S } after removing the last zero repeating
1 2 h

in each sub-string.
EP

Bellow we define additional notations that are used to describe the string vector and the schedule decisions for a
given route. Since the schedule decisions of different vehicles are independent from one another, we just focus on
C

one vehicle for the sake of presentation clarity. These additional notations are in consistent with those defined in
MILP model.
AC

Sh route string of vehicle h, S h = {iph | p = 0,1,2,..., nh } .

i ph node ID locating at pth position of the string Sh, we have i0h = 0 , inhh = 0 , and i ph > 0 for

1 ≤ p ≤ nh − 1 .
p index of the element in string Sh , 0 ≤ p ≤ nh .
nh number of arcs in string Sh, nh=card( S h )-1.
index of the arc in string Sh that 1 ≤ r ≤ n . r=1 represents the first arc (i0h , i1h ) that departs from
h
r
depot to the first customer and r = n h represents the last arc (inhh −1 , inhh ) return to depot from the last
customer.
 ACCEPTED MANUSCRIPT
fr payload of vehicle h when travelling the rth arc of Sh.

Dr distance of the rth arc

vrk constant travel speed of the rth arc in time period k (mile/hour)
crkh CO2 emission rate of empty vehicle h for traveling arc (i, j) in time period k
j(r) node ID at the head of the rth arc
lrh integer of departure time of vehicle h to travel the rth arc of Sh.
arh integer of arrival time of vehicle h when finished traveling the rth arc of Sh.

PT
d rkh continuous distance of the rth arc traveled by vehicle h in period k.

Note that the departure time lrh and arrival time arh are defined as integers, which is the discretization of the

RI
continuous variables li and ai defined in TD-VRSP-CO2. In our experiments, we use the time unit as small as one
minute in the discretization and show that the difference between the continuous and discrete schedules can be

SC
ignored in real-life applications.

4.2 Dynamic programming formulation for determining vehicle schedules

U
Based on the definition of the string representation for the vehicle routing and scheduling in Section 3.1, in this
section, we introduce a dynamic programming (DP) procedure to determine the optimal schedules of a vehicle h
AN
with a given route Sh. In other words, the DP procedure calculates the optimal departure time lrh , the optimal

arrival time arh , and the optimal distance to be traveled in each period (drkh) for each arc r in Sh. The basic
M

framework of applying the DP procedure to determine the optimal schedules of a given vehicle route is based on a
multiple-stage optimization strategy where each arc of the route is considered as a stage of optimization. The DP
D

procedure starts from the last arc of the route and moves backward recursively towards the first one. In each stage,

say arc r ( r = n , n − 1,...,2,1 ), the optimal departure time lrh and the corresponding objective value will be
TE

h h

determined based on previous optimization on arc r+1, and this process repeats until the first stage (i.e., r=1).
A key and time-consuming part of the DP procedure is to calculate the minimum objective value for traversing
EP

h
an arc r within the specified departure time t and arrival time t' by vehicle h. Let z rtt ′ denote the optimal

objective function value when vehicle h travels arc r between times t and t'. With a given pair of (t, t'), the
C

h
optimal value of z rtt ′ can be determined by a linear programming model using d rkh as the decision variable as
AC

follows:

Min zrtth ′ = ∑ crkh ⋅ d rkh + ϕh ⋅ f r ⋅ Dr (2)

k ∈K

Subject to:
(21) 0 ≤ d rkh ≤ vrk ⋅ µktt ′ / 60 ∀k ∈ K

(22) Dr = ∑ d rkh ∀k ∈ K
k ∈K

In the formulation above, the objective function consists of two terms: (1) the CO2 emissions by vehicle without
payload and (2) additional CO2 emissions caused by vehicle’s payload. Constraint (1) bounds the value range of
 ACCEPTED MANUSCRIPT

variable d rkh . Let µktt ′ indicate the length of the time interval overlapped between period k and interval [t, t'], i.e.,

µktt ′ = max{0, min{ek , t ′} − max{bk , t}} . The right side of Constraint (1) calculates the maximum length (in minutes)

of distance the vehicle can be scheduled to travel in time period k. Constraint (2) requires the arc r is exactly
traversed.
Solving this linear programming model is straightforward. The first term of the objective function can be
optimally determined by the dispatch rule that assigns the maximum possible value to variables drkh in the

PT
descending order of crkh until constraint Dr = ∑ d rkh is met. The rule means that vehicle h first travels the arc as
k∈K

much as possible in the period that will cause the lowest emissions. If the constraint is not satisfied, then vehicle h

RI
travels the arc in the period that will cause the second lowest emissions, and so forth until the constraint is satisfied.
During this process, if the time interval [t, t'] is not long enough for the vehicle to traverse the full distance Dr, then

SC
we set zrtt′ = max{zr ,t1 ,t2 | t1 < t2 } , representing the worst objective cost for travelling arc r by vehicle h.
h h

h
After determining all zrtt′ values using the simple heuristic defined above, we can use them in the DP procedure

U
to deduce the optimal objective function for each arc starting in the backward order from r= nh to 1 in Sh. To
AN
achieve this, the planning horizon (i.e., [0, em], which contains m time periods) needs to be divided into T integer
time units, t=0, 1, 2, …, T, which represent the discretized time points when a vehicle can depart from a node and
arrive at the next one in Sh. Each pair of (t, t') that t, t′ ∈[0,T ] and t < t′ can be considered as the departure time
M

(i.e., t) and arrival time (i.e., t') of each arc r in a given route Sh. Before providing the DP equations, we define
some important notations as follows:
D

f rth optimal objective function value of traveling the rth arc and all the arcs behind it, i.e., the rth,
(r+1)th,…, (nh)th arcs, by vehicle h when starting at time point t, i.e., lrh = t .
TE

ηrth arrival time of traveling the rth arc when starting at time point t, i.e., lrh = t and arh = ηrth , for
achieving the f rth .
Fh optimal hobjective function value of travelling the whole route Sh by vehicle h, calculating by
EP

n
Fh = ∑
r =1
f rth .

Based on the above definition and the heuristic for calculating zrtth ′ , we can use the following backward
C

recursive equations to deduce f rth and ηrth .

AC

min{zrtth ′ | t < t ′ ≤ T } if r = nh
f rth =  ∀t ∈ [0, t ], ∀r = nh ,..., 2,1 (3)
min{zrtt ′ + f r +1,t ′ | t < t ′ ≤ T }
h h
else

By using the above DP equation, all values of f rth and ηrth can be deduced backwardly starting from r=nh to 1
with respect to t=0, 1, 2,…,T. Thus, the optimal objective value for vehicle h, noted as F h , can be determined by
F h = min{ f1th | t = 0,1, 2,..., T } , and the corresponding arriving time ηrth for all arcs under this objective value can
also be deduced. After that, we can determine the optimal departure time l1h and arrival time a1h for the first arc
(r=1), and then the optimal arh and lrh for r=2 to nh by utilizing the matrix ηrth . Note that the DP procedure can
be applied independently to all other vehicles, and the final objective function value is a sum of objective values of
 ACCEPTED MANUSCRIPT
q
all vehicles calculated as F = ∑ F .
h
In Fig. 1, we give an example of 9 customers served by 3 vehicles in 3
h =1

periods to illustrate how the DP procedure works. The three periods are simulating a typical rush-hour pattern. The
DP procedure determines the departure time, arrival time, and to-be-traveled distance in each period for each arc
and even schedules idle time on road (e.g., vehicle 1) and waiting time + service time at customer place (e.g.,
vehicle 2). In Fig. 2, we provide the pseudo-code of the DP procedure, where the input is the given routes of
multiple vehicles: S={Sh} and the output is the optimal F, {Fh},{arh } and {lrh}.

PT
[Insert Figure 1 about here]
Fig. 1 Optimizing vehicle’s time schedule by dynamic programming

RI
[Insert Figure 2 about here]
Fig.2 Pseudo-code of the DP procedure

SC
4.3 Computer Implementation of the dynamic programming model

U
The dynamic programming procedure in Fig.2 has a polynomial complexity of O((nT)2) for the worst case. The
most time-consuming part of the DP procedure is between steps 6 and 9 because all values of zrtth ′ have to be
AN
repeatedly calculated through the linear programming model given in Section 3.2. To avoid this, we introduce a
h
pre-calculated matrix denoted as Zijtt ′ that stores the minimum feasible objective function values of vehicle h to
M

h
travel each arc (i, j) in every pair of interval [t, t'], i.e., departing at t and arriving at t'. The matrix Zijtt ′ needs to be

calculated in advance and only once before starting to solve a TD-VRSP-CO2 problem, and then the DP procedure
D

h h
can obtain the value of zrtt ′ by referring directly to Zijtt ′ . We give the detailed procedure of pre-calculating

h h
TE

Z ijtt ′ in Fig. 3, which determines the optimum value for Z ijtt ′ using a heuristic based on the linear programming

model given in Section 3.2. Note that an arc indexed by r on a route can always find a pair of nodes indexed as (i,
j).
EP

[Insert Figure 3 about here]

h
Fig.3 Function of calculating the optimal values for Zijtt ′
C
AC

The procedure in Fig.3 divides the total distance Dij of arc (i, j) into multiple sections that are traveled in different
periods that overlap with time interval [t, t']. Step 1 calculates the length of the overlapped time interval between [t,
t'] and [bk, ek], where k∈K. Step 4 finds out the available period, noted as k*, that has the lowest emission rate.
Steps 6 to 12 determine the length that should be traveled in period k* and calculates the amount of total CO2
emissions on arc (i, j). In Step 14, the tardiness penalty is added to the objective function value if the arrival time t'
h
is later than the due time of the customer. It is notable that Z ijtt ′ can be very large for large-sized problems.
h
Nonetheless, Z ijtt ′ has a very high ratio of identical values in the matrix, such that storage techniques for sparse
matrix can be used to store it in the computer memory.
h
Next, we provide two properties on Z ijtt ′ and f rth that improve the DP procedure with a higher efficiency on
 ACCEPTED MANUSCRIPT
searching the matrix.
h
Property 1: For an arc (i, j) and a given arrival time t', the Z ijtt ′ , if existed, is a monotonically increasing

function of departure time t (t<t') such that:

(23) Zi , j ,t1 ,t ′ ≤ Zi , j ,t2 ,t ′

h h
∀h, i, j, t ′ ≤ t1 < t2 .

Property 2: For a tour Sh and an arc r in Sh, the f rth , if existed, is a monotonically increasing function of the

PT
departure time t such that:

(24) frt1 ≤ f rt2 ∀h, r, t1 < t2 .

h h

RI
Property 1 indicates that an earlier starting time t1 always results in lower CO2 emissions than a later starting time
t while traversing the arc (i, j) with arrival time of t' at node j. This result is obvious because an earlier starting

SC
time allows the vehicle to have longer length of time interval to arrange its travel schedules. Property 2 indicates
the same rule for traversing the rth arc and all arcs behind the rth arc on tour Sh. Therefore, using Property 1, we can
determine the earliest arrival time (noted as t1*) that leads to optimal zrtt
h
* (in step 6 of Fig. 2) for each discrete

U
time point in [0, T] in advance. In addition, using Property 2, we can determine the latest departure time (noted as
AN
* h
t2 ) that leads to optimal f rt in previous calculation for arc r+1. Therefore, each arc r in the tour is associated
with a pair of time points (t1 (r), t2*(r)), indicating the earliest arrival time and the latest departure time. The
*

optimal time point t' starting to travel the rth arc and all arcs behind it and resulting optimal f rth must be in the
M

time interval [t1*, t2*]. Thus, we can use the following recursive backward equation to determine f rth for all arcs in
the tour.
D

 zrh,t ,t* ( r ) if t1* ( r ) ≤ t2* ( r ) or r = n h

f rth =  1 ∀r ∈ [0, T ], ∀r = n h ,..., 2,1 , (4)
 min{z rtt ′ + f r +1,t ′ | t2 ( r ) ≤ t ′ ≤ t1 ( r )}
TE

h h * *
else

where r = n indicates the last arc connected to the depot.

h
Thereby, the calculation of f rth can be either
EP

accessed directly from pre-calculated matrix (for the cases t1* (r ) ≤ t2* (r ) or r = nh ) or by the minimum zrtth ′ + f rh+1,t ′

such that t' is within [t2* (r ), t1* (r )] . Note that if t1* ( r ) ≤ t2* (r ) , then the vehicle can have an idle time of
t2* (r ) − t1* (r ) before starting to travel the rth arc without any impact on the objective function value.
C

Note that if f rth cannot be computed even for t=0 (i.e., the rth arc and the arcs behind the rth arc cannot be
AC

traveled even if the route starts at the beginning of plan horizon [0, T]), then the route is infeasible. However,
infeasible solutions are valuable in GAs because they can include good solution structures which may be improved
through mutation or contribute to other solutions through crossover. In most problems, optimal solutions reside on
the boundary between infeasible and feasible regions. Therefore, infeasible solutions are allowed in the population
of GA-DP, but their objective values are penalized. For each t ∈ [0, T ] and r in Sh, if f rth cannot be computed by
Eq.(5), its value is set as frth = β max{zrtth ′ | ∀t , t ′} + frh+1,t* ( r ) . The first term of penalized f rth is the worst objective value
2

of traveling arc r multiplied by a dynamic penalty coefficient β, which will be introduced in next sub-section.

4.4 Genetic algorithm with dynamic programming (GA-DP)

 ACCEPTED MANUSCRIPT
In this section, we integrate the dynamic programming algorithm into a genetic algorithm to solve large-sized
problems. The GA with the proposed DP procedure is called GA-DP. Genetic algorithms (GAs) mimic the
process of natural evolution in a population of chromosomes with the objective of reserving those individuals who
are most adaptive (i.e., the fittest one) to the environment. In the literature, GAs have been successfully applied to
solve various VRPs. In this paper, a GA is selected as the meta-heuristic because of its ability to handle feasible
and infeasible solutions concurrently in its population. Due to the dependency between the routing and scheduling
decisions, the solution space of the TD-VRSP-CO2 includes many infeasible solutions that are difficult to repair.
Through the crossover operator, a GA can combine infeasible and feasible solutions to create new ones, and a

PT
fitness function with a dynamic penalty term can guide the search toward the boundaries of feasible and infeasible
regions in the solution space where best solutions tend to reside. In the computational experiments, the performance
of the adaptive penalty function is demonstrated.

RI
i) Solution Encoding
In the GA-DP, a solution is represented by tuple <S, ∏> where S is the string representing the routes of the
solution as discussed before, and ∏ holds the schedule of the vehicles for each arc in S, including the departure

SC
times, arrival times, and the travel distances on each arc in each time period. The general strategy of the GA-DP
is: (1) use traditional GA operators to evolve the route parts of solutions and (2) optimize the schedule parts of
solutions and determine their fitness values using the DP procedure.

U
ii) Initialization
AN
The population is initialized by a greedy insertion heuristic. The insertion heuristic starts from an empty string S
with only q+1 zeros, i.e., S={0, 0, 0, …, 0}, and then inserts a customer into a position that will result in the least
increase in the objective function value. This process repeats one customer at a time until all customers are inserted
into the string. Once a solution is constructed, it will serves as a seed to generate λ individuals randomly through
M

application of Insert and Swap operators on the seed solution.

iii) Genetic operators
D

In order to generate new solutions for the next population, Insert, Swap, and Crossover operators are applied on
the route part of the existing solutions. The insert operator simply relocates a randomly selected node to another
TE

randomly selected position in S by shifting the positions of the other nodes accordingly. The swap operator
exchanges the positions of two randomly selected nodes in S. In order words, the insert and swap operators
performs the function of mutation in the GA-DP.
EP

The crossover operator exchanges two random segments of two random parents to generate two offspring. For
the crossover operator, the GA-DP uses the one-point Partially Matched Crossover (PMX) (Goldberg, 1989), which
recombines two randomly selected parents to produce two offspring while making sure the integrity of the routes of
C

the offspring without requiring a repair operation. The details of PMX can be found in (Goldberg, 1989).
iv) Fitness function and selection
AC

The fitness of a solution i is calculated as follows:

æ F - F ö÷
Pi = exp ççç- i min ÷
÷, (5)
çè Favg - Fmin ø÷
÷

where Fi is objective function value for individual i, and Fmin and Favg are the minimum and average objective
function values of the population, respectively. Note that Fi also includes a penalty term calculated within the DP
procedure for infeasible solutions. Fitness Pi defines the probability that solution i will remain in the population
for the next generation. In addition, the clustering method of Dellaert and Jeunet (2000) is used in the GA-DP to
maintain the diversity of the population. In this approach, solutions with the identical objective value are grouped
as a cluster, and at most ten solutions in the best cluster and only one solution in any other clusters are selected for
 ACCEPTED MANUSCRIPT
the next generation.
v) Dynamic penalty function
Because of the many constraints involved, finding even a feasible solution to Problem TD-VRSP-CO2 is
challenging. Although the crossover operator ensures to generate valid routes in all times, it is likely that the
population will include many infeasible solutions due to two main reasons: (i) the assigned route to a vehicle
cannot be completed within the planning horizon and (ii) the capacity and/or travel distance limits of a vehicles are

exceeded during the tour. For the first case, the objective function is penalized while calculating f rth as discussed

PT
in Section 3.2. For the second case, a penalty term proportional to the extent to which the routes violate the load
capacities and travel distance limits is added to the objective function. In the GA-DP, penalty coefficient β is
dynamically determined based on the feasibility ratio of the population, which is defined as λ * / λ where λ * is

RI
the number of feasible solutions in the population of size λ . In the GA-DP, penalty coefficient β is updated in
each generation as follows:

SC
 λ*
 1.1β ( g ) if < 0.85
β ( g + 1) =  λ ∀g = 1, 2,... , (6)
 β ( g ) if λ > 0.95
*

 1.1

U
λ
where g indicates the index of generation. The dynamic penalty function in Eq (6) is calculated after the selection
AN
process and before the fitness calculation of the new generation in the GA-DP algorithm. The feasibility ratio of the
population is encouraged to stay within an empirical range of [0.85, 0.95]. If λ * / λ of the current generation is
less than 0.85, then β is increased by 10% for the next generation, and if it is greater than 0.95, then β is reduced by
M

10%. The maximum and minimum values of β are also restricted to an empirical range of [0.1, 100]. This dynamic
penalty update procedure attempts to maintain the population close to boundaries of feasible and infeasible regions,
where optimal solutions tend to be located in the solution space.
D

Thus, the objective function for given routes of the heterogeneous fleet can be generally expressed in Eq.(7),
TE

where the dynamic penalty coefficient β is enforced to work on the violations of the load capacity limit and travel
distance limit, as follows.

Min F= ∑F h
= ∑ f rh0
EP

h∈ H h∈ H

nh nh nh
= ∑∑z h
r , lrh , arh
+ β ( ∑ ψ h max{0, ∑ R j ( r ) − Ch } + ∑ π h max{0, ∑ D j ( r ) − Lh }) , (7)
h∈H r =1 h∈H r =1 h∈H r =1
C

nh
where ψ h and π h are the penalty coefficients for the capacity limit violations (i.e., max{0, ∑ R j ( r ) − Ch } ) and
r =1
AC

nh
the travel distance limit violations ( max{0, ∑ R j ( r ) − Ch } ) by vehicle h, respectively. Note that calculation of
r =1

zrh,l h ,a h term also includes penalty term as discussed in the DP procedure.

r r

vi) Overall Procedure of the GA-DP

The pseudo code of the GA-DP is given in Fig. 4. Respectively, parameters λ and g are for the population size
and the stopping condition (the maximum number of consecutive generations that the feasible solution has not been
improved). The GA-AP is first initialized with a population of λ solutions, and all genetic operators are applied for
λ times. As a result, the population size increases to 5λ from λ. After evaluating the fitness value of each solution, λ
 ACCEPTED MANUSCRIPT
solutions are selected for the next generation.

[Insert Figure 4 about here]

Fig.4 Framework of the GA-DP

5. Computational experiments

In this section, the performance of the GA-DP is gauged against a construction heuristic as well as an exact
method (AMPL/CPLEX) based on the MILP model for Problem TD-VRSP-CO2. In the first set of computational

PT
experiments, the performance of the GA-DP algorithm is compared with respect to the optimal solutions found by
Problem TD-VRSP-CO2. We generated 30 small-sized test problems with 5, 10, and 15 customers served by
vehicles of different types so that we were able to use both the GA-DP algorithm and the AMPL/CPLEX to solve

RI
these small-sized problems. These test problems were generated to simulate a general case of time-dependent
road conditions. In this problem set, customers are randomly located in an area of 20x20 square miles, and the

SC
average travel speeds of the arcs in different time periods are randomly generated between 15 and 75 mph. The
planning horizon includes five 30-minute periods (150 minutes in total). There are four vehicles types as
specified in Table 1, and the number of available vehicles in each type is set to n/5. In this first set of

U
computational experiments, the demand of each customer is assumed to be zero, and no distance limit is imposed
on the travel range of the vehicles. The CO2 emissions rates (Lb/mile) of the vehicle types under different speeds
AN
are calculated by the emissions model by Hickman (1999) as follows:

D E F
e = K + Av + Bv 2 + Cv3 + + + , (8)
v v 2 v3
M

where e is the rate of CO2 emissions in g/km for the average speed of v (km/h) on a road with a slope of 0%.
Table 1 lists the parameter of the model provided by Hickman (1999) for various vehicle gross weights. We ignore
D

CO2 emissions caused by payload by setting ϕh =0. Fig.5 illustrates the relationships between CO2 emissions
TE

(Lb/mile) and vehicle’s travel speed (mile/hour) for the four vehicle types used in the computational experiments.

[Insert Table 1 about here]

EP

[Insert Figure 5 about here]

Fig. 5 CO2 emissions rates as a function of travel speeds for heavy-duty vehicles
C

The optimal solutions to the 30 small-sized instances were found by solving Problem TD-VRSP-CO2 using
AC

AMPL/CPLEX (version 12.4.0.1) in an MAC computer with 2.50GHz CPU, and these problems were also solved
by the GA-DP algorithm coded in Visual C++ 6.0 and run in a PC with 3.4GHz CPU. In GA-DP, the initial
population size was λ=50, and the stopping condition was g=50 for all problems.
In order to study the effect of discretizing the continuous decision variables in the DP procedure on the final
solution quality, we first solved the test problems using only CPLEX with continuous departure/arrival time
decision variables and then using both CPLEX and the GA-DP algorithm with three different levels of discretized
time units: (i) one-minute, (ii) two-minute, and (iii) five-minute. In order words, the sensitivity of optimal
solutions with respect to discrete time units used in the DP procedure was investigated. In order to solve the
problems with different discrete time units using AMPL/CPLEX, the following Constraints (25) and (26) were
added to the MILP model for Problem TD-VRSP-CO2 described in Eq.(1) and Constraints (1)-(20).
(25) ai′ = ai / κ ∀i ∈ N ′
 ACCEPTED MANUSCRIPT
(26) li′ = li / κ ∀i ∈ N ′ ,
where li and ai are integer decision variables representing the departure and arrival time of customer i, and li′
and ai′ are also integer variables that enforce li and ai to have integer values divisible by integer κ , which
indicates the time unite for discretization. For example κ =5, constraints (25) and (25) enforce decision variable li
and ai to take values 0, 5, 10, 15,…, and as results a vehicle can depart from and arrive at a node only at time
points divisible by five.
The computational results are listed and compared in Table 2, where the solutions in columns CPLEX were
obtained by solving Problem TD-VRSP-CO2 by CPLEX within a one-hour CPU time limit, and the solutions in

PT
columns 10 runs of GA-DP were obtained by the GA-DP algorithm. For the solutions obtained by CPLEX, the
optimality gap is also provided in columns Gap (note that 0% gap means an optimal solution). In the Table, it can
be observed that the solutions found by CPLEX and the best solutions found by 10 runs of GA-DP for all the

RI
discrete time units are identical for problem groups 5x5x150 and 10x5x50, indicating that the GA-DP algorithm
was able to find the optimal solutions for these problems. The average and best of 10 runs of the GA-DP are also

SC
the identical for problem group 5x5x150 and very close to one-another for problem group 10x5x150, indicating that
the GA-DP algorithm is robust for these instances. We can also note that the discretized solutions are slightly worse
than the continuous solutions. For problem group 5x5x150, the deviations from continuous solutions are 0.37%,

U
1.38%, and 3.12% for one-minute, two-minute, and five-minute discretized time units, respectively. For problem
group 10x5x150, the deviations are 2.62%, 4.24%, and 9.91%, respectively, indicating that a larger unit of time in
AN
the discretization results a higher deviation of the results from continuous optimal solutions.
Another observation is that the GA-DP requires significantly shorter CPU times than that used by CPLEX. For
all instances of problem groups 5x5x150 and 10x5x150, the GA-DP used averagely only about 1 second while
M

CPLEX used averagely more than 1 minute for problem group 5x5x150 and 10 minutes for problem group
10x5x150, respectively. And, unfortunately CPLEX could not solve optimally the larger problem group 15x5x150
in hours while the GA-DP could also delivery quite good solution (21% better on the average) within a few seconds
D

for this problem set. The solutions and optimality gaps listed for problem group 15x5x150 in Table 2 were obtained
with one hour of time limit. Interestingly, in this case some solutions with a larger unit of time discretization are
TE

even better than those with a smaller one. Such a phenomenon occurs mainly due to fact that a coarse unit of time
discretization reduces the solution space significantly, and the branch-and-bound procedure in CPLEX can
investigate more solutions or find solutions with lower optimality gaps to the optimal ones within given time limit.
EP

In summary, based on the results in Table 2, it can be claimed that the GA-DP has a very high computational
efficiency in finding optimal solutions for small-sized problems and near-optimal solutions for larger problem using
relatively short time. However, selecting a proper unit of time discretization is important for balancing solution
C

quality and computational effort to solve the problems.

AC
 ACCEPTED MANUSCRIPT

[Insert Table 2 about here]

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT
Next, the computational efficiency and effectiveness of the GA-DP algorithm was tested on large-sized problems
from the well-known CVRP benchmark instances by Christofides and Eilon (1969). There are 14 large-sized CVRP
problem instances with the number of nodes ranging from 51 to 200. The problem instances include constraints on
vehicle’s capacity and maximum route distance, but they are not designed for a time-dependent case. Therefore, we
modified them to include time-dependency as follows: (1) all vehicles must complete their service within a
planning horizon of [0, 300] with five 60-minute periods, (2) each arc in each period is assigned to a random traffic
speed between 15mph and 75mph, and (3) each customer is assigned to a random due time Ei within [0, 300] and a
tardiness penalty coefficient between 0.5 and 1.5. The heterogeneous fleet is composed of four heavy-duty vehicle

PT
types with different emissions rates as given in Fig. 5. In all instances, we set the type of vehicle h as ((h mod 4)
+1) in Table 1. In the DP procedure, five-minute time intervals were used to discretize vehicle arrival and departure
times. The GA-DP was run for 10 random replications with parameters λ=100 (population size) and g=100

RI
(stopping condition). In order to demonstrate the effect of the dynamic penalty function on the performance of the
algorithm, the GA-DP was first run by allowing only feasible solutions in the population (denoted as 10 runs of
GA-DP* in Table 3), and then it was run with the dynamic penalty function (denoted as 10 runs of GA-DP in Table

SC
3).
The experimental results are listed and compared in Table 3, where columns Ava. Veh. and Used Veh. show the
number available vehicles and used vehicles in the best solution of 10 runs, respectively, column Avg. G. is the

U
average number of total generations, and columns Cap. and Len. present the capacity and length limits of vehicles,
AN
respectively. It is clear that the GA-DP outperformed the GA-DP* in terms of both average and best objective
function values. The GA-DP found 13 best solutions while GA-DP* found only one best solution for the 14 test
instances. The GA-DP* did not find feasible solutions in the first generation of instance No. 9. In Table 3, the
best solutions found by a greedy construction heuristic are also provided for benchmarking the performance of the
M

GA-DP. The column Dev.% shows that the deviations of these solutions from the best solutions found by the
GA-DP (or GA-DP*) are quite large, ranging from 39.5% to 301.1% and 112.4% on the average. Furthermore,
D

the best solutions found by the GA-DP are compared to the optimal solutions found by a restricted version of
Problem TD-VRSP-CO2. In this approach, only the arcs among the nearest five-neighbors of each node were
TE

considered in the model. In other words, the model was solved in CPLEX by including only a sub-set of all
possible arcs. In Table 3, the solutions found by this restricted-CPLEX method for instances No.1, No. 6, and
No.14 are provided, and feasible solutions of the other instances solutions could not be found within a 24-hour
EP

CPU time. The GA-DP clearly outperformed the restricted-CPLEX method in these three instances by 26.1%,
38.8%, and 51.4%, respectively. The experimental results also show that the GA-DP is computationally efficient
since all large-sized instances were solved in reasonable CPU times. The largest problem with 200 nodes and 60
C

discrete time points was solved in 18 minutes.

In Fig.6, the average computing time (in seconds) is plotted against the problem size (measured by number of
AC

customers). It can be observed that with the same parameters (population size λ=100 and stop condition g=100 ),
the computational time used by the GA-DP is nearly proportional to the problem size (with R2=0.9736), which
suggests that the GA-DP algorithm is scalable and the empirical computational time is linear although the DP
procedure has polynomial time complexity. It should be noted that this linear performance is mainly due to the
implementation of the DP procedure. Since all CO2 emissions of traveling arcs by different vehicle types on
different time intervals are pre-calculated, so the DP procedure can maintain high computational efficiency for
large-sized problems.
 ACCEPTED MANUSCRIPT

[Insert Table 3 about here]

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT

[Insert Figure 6 about here]

Fig. 6 The trend of average computing time against the problem size

6. Conclusions

In this paper, we studied the GVRSP with heterogeneous fleet and tardiness objective, named as the
time-dependent vehicle routing and scheduling problem with CO2 emissions optimization (TD-VRSP-CO2) in a

PT
time-varying traffic environment. The MILP model can be directly solved to optimality using a MIP solver for
small-sized problems. For large-sized problems, we divided the problem into two parts: (1) the routing part and (2)
the scheduling part. We presented an exact dynamic programming (DP) method with a complexity of O((nT)2) for

RI
the scheduling part, which can find optimal schedules for all traveled arcs with discrete departure/arriving times
and continuous to-be-traveled distance in different periods. We developed a hybrid heuristic solution approach,
named as the GA-DP, by combining a genetic algorithm with the DP procedure. The GA-DP algorithm

SC
demonstrates how this difficult problem with multiple binary and continuous variables can be reduced to a
workable structure for heuristics using a hybrid approach. Computational experiments on 30 small-sized problem
instances and 14 well-known CVRP benchmark instances (up to 199 customers) show that the proposed approach is

U
effective and efficient.
We also pointed out that the dynamic programming method has a slight drawback–it may lead to sub-optimal
AN
solutions due to the time domain discretization (averagely 1.5% in comparison to the MILP model). However, this
drawback will be insignificant for real-life applications. Future research on this topic may focus on three directions
as follows.
M

(1) By incorporating the exact DP procedure into other meta-heuristics to find better solution approaches for
the TD-VRSP-CO2 and for traditional time-dependent VRPs as well.
D

(2) Modeling the TD-VRSP-CO2 with more real-world information such as routing customer arcs on actual
road network, using real-time traffic patterns, considering maximum continuous driving time by law, and
TE

etc.
(3) Solving the problem under a multi-objective framework, including CO2 emissions related cost/benefit,
such as fuel cost, carbon tax, gains in carbon trade, etc., travel time related cost/benefit, such as driver
EP

wage, vehicle rent cost, customer satisfaction, etc., and other cost/benefit.

Acknowledgements
This work is partly supported by the National Natural Science Foundation of China under Grants No.
C

71271009.
AC

Reference
1. Barth, M., J., Boriboonsomsin, K., 2008. Real-world CO2 impacts of traffic congestion. Transportation Research Record: Journal
of the Transportation Research Board 2058, 163–171.
2. Barth, M., J., Younglove, T., Scora, G., 2005. Development of a heavy-duty diesel modal emissions and fuel consumption model.
Technical Report. UCB-ITSPRR-2005-1, California PATH Program, Institute of Transportation Studies, University of California
at Berkeley.
3. Bektas, T., Laporte, G., 2011. The Pollution-Routing Problem. Transportation Research Part B 45, 1232–1250.
4. Benjaafar, S., Li, Y., Daskin, M., 2013. Carbon Footprint and the Management of Supply Chains: Insights from Simple Models.
IEEE Transactions on Automation Science and Engineering, 10 (1), 99–116.
5. Chen, H., K., Hsueh, C., F., Chang, M., S., 2006. The real-time time-dependent vehicle routing problem. Transportation Research
Part E 42, 383–408.
6. Chu S. 2009. U.S. Energy Secretary Steven Chu’s address at Harvard’s Afternoon Exercises
 ACCEPTED MANUSCRIPT
<http://news.harvard.edu/gazette/story/2009/06/u-s-energy-secretary-steven-chus-address-at-harvards-afternoon-exercises/>
[accessed 16.02.14]
7. Dantzig, G., B., Ramser, J., H., 1959. The truck dispatching problem. Management Science 6(1), 80–91.
8. Dellaert, N., & Jeunet, J., 2000. Solving large unconstrained multilevel lot-sizing problems using a hybrid genetic algorithm.
International Journal of Production Research 38, 1083-1099.
9. Dekker, R., Bloemhof, J., Mallidis, I., 2012. Operations Research for green logistics – An overview of aspects, issues,
contributions and challenges. European Journal of Operational Research 219, 671–679.
10. Demir, E., Bektas, T., Laporte, G., 2014a. A review of recent research on green road freight transportation. European Journal of
Operational Research 237(3), 775–793.

PT
11. Demir, E., Bektas, T., Laporte, G., 2014b. The bi-objective Pollution-Routing Problem, European Journal of Operational
Research 232 (3), 464–478.
12. Eksioglu, B., Vural, A., V., Reisman, A., 2009. The vehicle routing problem: A taxonomic review. Computers & Industrial
Engineering 57(4), 1472–1483.

RI
13. Ehmke, J.F., Campbell, A.M., and Thomas, B.W., 2016. Vehicle routing to minimize time-dependent emissions in urban areas.
European Journal of Operational Research 251(2), 478–494.
14. Erdogan, s., Miller-Hooks, E., 2012. A Green Vehicle Routing Problem. Transportation Research Part E 48, 100–114.

SC
15. Ericsson, E., Larsson, H., Brundell-Freij, K., 2006. Optimizing route choice for lowest fuel consumption—potential effects of a
new driver support tool. Transportation Research Part C—Emerging Technologies 14, 369–83.
16. Figliozzi, M.A., 2010. Vehicle routing problem for emissions minimization. Transportation Research Record: Journal of the
Transportation Research Board 2197, 1–7.

U
17. Figliozzi, M., A., 2012. The time dependent vehicle routing problem with time windows: Benchmark problems, an efficient
solution algorithm, and solution characteristics. Transportation Research Part E 48, 616–636
AN
18. Franceschetti, A., Honhon, D., Woensel, T., V., Bektas, T., Laporte, G., 2013. The time-dependent pollution-routing problem.
Transportation Research Part B 56, 265–293.
19. Gaur, D., R., Mudgal, A., Singh, R., R., 2013. Routing vehicles to minimize fuel consumption. Operations Research Letters 41,
576-580.
M

20. Goldberg, D.E., 1989. Genetic Algorithms in search, optimization, and machine learning. Reading.Mass:Addison-Wesley;1989.
21. Golden, B., L., Raghavan, S., Wasil, E., 2008. The vehicle routing problem: latest advances and new challenges. Berlin: Springer,
2008.
D

22. Govindan, K., Azevedo, S., G., Carvalho, H., Cruz-Machado, V., 2014. Impact of supply chain management practices on
sustainability. Journal of Cleaner Production 85, 212-225.
TE

23. Günther, H.-O, Kannegiesser, M., Autenrieb, N., 2015. The role of electric vehicles for supply chain sustainability in the
automotive industry. Journal of Cleaner Production 90, 220-233.
24. Hickman, A., J., 1999. Methodology for calculating transport emissions and energy consumption. TRANSPORT RESEARCH
LABORATORY, pp.69-73.
EP

25. International Energy Agency, 2015. CO2 emissions from fuel combustion—highlights (2015 Edition).
<https://www.iea.org/publications/freepublications/publication/CO2EmissionsFromFuelCombustionHighlights2015.pdf>
[Accessed January 5, 2016].
26. Jabali, O., Van Woensel, T., Kok, A.G. de, 2012. Analysis of travel times and CO2 emissions in time-dependent vehicle routing.
C

Production and Operations Management 21(6), 1060–1074.

27. Kara, I., Kara, B., Y., Yetis, M., K., 2007. Energy minimizing vehicle routing problem. In: Dress A, Xu Y, Zhu B, editors.
AC

Combinatorial Optimization and Applications: FIRST International Conference, LNCS 4616. Berlin/Heidelberg: Springer; 2007.
p. 62–71.
28. Kok, A., L., Hans, E., E., Schutten, J., M., J., Schutten, J., M., J.,2010. A dynamic programming heuristic for vehicle routing with
time-dependent travel times and required breaks. Flexible Services and Manufacturing Journal 22, 83-108.
29. Koc, C., Bektas, T., Jabali, O., and Laporte, G., 2014. The fleet size and mix pollution-routing problem. Transportation Research
Part B 70, 239–254.
30. Kok, A., L., Hans, E., E., Schutten, J., M., J., 2011. Optimizing departure times in vehicle routes. European Journal of
Operational Research 210, 579-587.
31. Kramer, R., Maculan, N., Subramanian, A., Vidal, T., 2015. A speed and departure time optimization algorithm for the
pollution-routing problem. European Journal of Operational Research 247, 782–787.
32. Kuo, Y., 2010. Using simulated annealing to minimize fuel consumption for the time-dependent vehicle routing problem.
Computers & Industrial Engineering 59, 157–165.
 ACCEPTED MANUSCRIPT
33. Kushwah, G., S., Sharma, N., K., 2015. Green initiatives: a step towards sustainable development and firm's performance in the
automobile industry. Journal of Cleaner Production 121, 116-129.
34. Kwon, Y., J., Choi, Y., J., Lee, D., H., 2013. Heterogeneous fixed fleet vehicle routing considering carbon emission.
Transportation Research Part D 23, 81–89.
35. Laporte, G., Gendreau, M., Potvin, J., Y., Semet, F., 2000. Classical and modern heuristics for the vehicle routing problem.
International Transactions in Operational Research 7(4–5), 285–300.
36. Lin, C., Choy, K., L., Ho, G., T., S., Chung, S., H., Lam, H., Y., 2014. Survey of green vehicle routing problem: past and future
trends. Expert Systems with Applications 41, 1118–1138.
37. Maden, W., Eglese, R., Black, D., 2010. Vehicle routing and scheduling with time-varying data: a case study. Journal of the
Operational Research Society 61(3), 515–522.

PT
38. Melillo, Jerry M., Terese (T.C.) Richmond, and Gary W. Yohe, Eds., 2014. Climate Change Impacts in the United States: The
Third National Climate Assessment. U.S. Global Change Research Program, 841 pp. doi:10.7930/J0Z31WJ2.
39. Malandraki, C., Daskin, M., S., 1992. Time-dependent vehicle-routing problems: formulations, properties and heuristic

RI
algorithms. Transportation Science 26(3), 185–200.
40. Malandraki, C., Dial, R., B., 1996. A restricted dynamic programming heuristic algorithm for the time dependent traveling
salesman problem. European Journal of Operational Research 90, 45-55.

SC
41. Ren, W., Xue, B., Geng Y., Lu, C., Zhang Y., Zhang, L., Fujita T., Hao, H., 2016. Inter-city passenger transport in larger urban
agglomeration area: emissions and health impacts. Journal of Cleaner Production 114, 412-419.
42. Sahin, B., Yilmaz, H., Ust, Y., Guneri, A., F., Gulsun, B., 2009. An approach for analysing transportation costs and a case study.
European Journal of Operational Research 193, 1–11.

U
43. Soysal, M., Bloemhof-Ruwaard, J.M., van der Vorst, J.G.A.J., 2014. Modelling food logistics networks with emission
considerations: The case of an international beef supply chain. International Journal of Production Economics, 152, 57–70.
AN
44. Toth, P., Vigo, D., 2002. An overview of vehicle routing problems. In: Toth P, Vigo D, editors. The vehicle routing problem.
SIAM monographs on discrete mathematics and applications, vol. 9. Siam; 2002. p. 1–26.
45. Validi, S., Bhattacharya, A., Byrne, P., J., 2012. Greening the Irish Food Market Supply Chain through Minimal Carbon
Emission: An Integrated Multi-Objective Location-Routing Approach. In: International Conference on Manufacturing Research
M

(ICMR 2012), 11th - 13th September 2012, Aston University, UK.

46. Validi, S., Bhattacharya, A., Byrne, P., J., 2015. A solution method for a two-layer sustainable supply chain distribution model.
Computers & Operations Research 54, 204-217.
D

47. Validi, S., Bhattacharya, A., Byrne, P., J., 2014a. Integrated low-carbon distribution system for the demand side of a product
distribution supply chain: a DoE-guided MOPSO optimiser-based solution approach. International Journal of Production
TE

Research, 52(10), 3074-3096.

48. Validi, S., Bhattacharya, A., Byrne, P., J., 2014b. A case analysis of a sustainable food supply chain distribution system—A
multi-objective approach. International Journal of Production Economics, 152, 71–87.
49. Xiao,Y., Zhao, Q., Kaku, I. Xu, Y., 2012. Development of a fuel consumption optimization model for the capacitated vehicle
EP

routing problem. Computers & Operations Research 39, 1419-1431.

50. Xiao Y., Konak A. 2015a. A simulating annealing algorithm to solve the green vehicle routing & scheduling problem with
hierarchical objectives and weighted tardiness. Applied Soft Computing 34, 372-388.
51. Xiao Y., Konak A. 2015b. Green Vehicle Routing Problem with Time-Varying Traffic Congestion. 14th INFORMS Computing
C

Society Conference, Richmond, Virginia, 2015, p. 134-148

52. Xiao Y., Konak, A. 2016. The heterogeneous green vehicle routing and scheduling problem with time-varying traffic congestion.
AC

Transportation Research Part E: Logistics and Transportation Review 88, 146-166.

53. Yang, J., Guo, J., Ma, S., 2016. Low-carbon city logistics distribution network design with resource deployment. Journal of
Cleaner Production 119, 223-228.
54. Zhu, X., Garcia-Diaz, A., Jin, M., Zhang, Y., 2014. Vehicle fuel consumption minimization in routing over-dimensioned and
overweight trucks in capacitated transportation networks. Journal of Cleaner Production 85, 331-336.
 ACCEPTED MANUSCRIPT

Table 1. Constant and coefficients with respect to vehicle gross weight

Gross weight K A B C D E F
3.5 to 7.5 tones 110 0 0 0.000375 8,702 0 0
7.5 to 16 tones 871 -16.0 0.143 0 0 32,031 0
16 to 32 tones 765 -7.04 0 0.000632 8,334 0 0
32 to 40 tones 1,576 -17.6 0 0.00117 0 36,067 0

PT
RI
U SC
AN
M
D
TE
C EP
AC
 ACCEPTED MANUSCRIPT

Table 2. Experiments with 30 small-sized instances

(i) Continuous (ii) Discrete -One Minute Unit of Time (iii) Discrete -Two-Minute Unit of Time (iv) Discrete -Five Minute Unit of Time
Prob.size Prob. Vehicle CPLEX 10 runs of GA-DP CPLEX 10 runs of GA-DP CPLEX 10 runs of GA-DP
(n×m×T) ID Ava. Used CPLEX Gap% Obj. Gap% Dev.% Avg. Min. Dev.% CPLEX Gap% Dev.% Avg. Min. Dev.% CPLEX Gap% Dev.% Avg. Min. Dev.%
5×5×150 0 1 1 58.64 0.0 58.95 0.0 0.53 58.95 58.95 0.53 61.74 0.0 5.29 61.74 61.74 5.29 62.06 0.0 5.83 62.06 62.06 5.83
1 1 1 66.26 0.0 66.26 0.0 0.00 66.26 66.26 0.00 66.26 0.0 0.00 66.26 66.26 0.00 67.21 0.0 1.43 67.21 67.21 1.43

PT
2 1 1 85.31 0.0 85.60 0.0 0.34 85.60 85.60 0.34 85.78 0.0 0.55 85.78 85.78 0.55 86.19 0.0 1.03 86.19 86.19 1.03
3 1 1 55.08 0.0 55.10 0.0 0.04 55.10 55.10 0.04 56.53 0.0 2.63 56.53 56.53 2.63 56.79 0.0 3.10 56.79 56.79 3.10
4 1 1 62.11 0.0 62.11 0.0 0.00 62.11 62.11 0.00 62.11 0.0 0.00 62.11 62.11 0.00 62.15 0.0 0.06 62.15 62.15 0.06

RI
5 1 1 55.24 0.0 55.25 0.0 0.02 55.25 55.25 0.02 55.29 0.0 0.09 55.29 55.29 0.09 55.29 0.0 0.09 55.29 55.29 0.09
6 1 1 82.19 0.0 82.50 0.0 0.38 82.50 82.50 0.38 82.55 0.0 0.44 82.55 82.55 0.44 86.96 0.0 5.80 86.96 86.96 5.80
7 1 1 68.96 0.0 70.60 0.0 2.38 70.60 70.60 2.38 70.69 0.0 2.51 70.69 70.69 2.51 70.65 0.0 2.45 70.65 70.65 2.45

SC
8 1 1 51.39 0.0 51.39 0.0 0.00 51.39 51.39 0.00 52.49 0.0 2.14 52.49 52.49 2.14 53.71 0.0 4.51 53.71 53.71 4.51
9 1 1 56.72 0.0 56.75 0.0 0.05 56.75 56.75 0.05 56.80 0.0 0.14 56.80 56.80 0.14 60.64 0.0 6.91 60.64 60.64 6.91
CPU Time 64.19 t<1s 64.45 t<1s 0.37 64.45 t=1.0s 0.37 65.02 t<1s 1.38 65.02 t<1s 1.38 66.17 t<1s 3.12 66.17 t<1s 3.12

U
10×5×150 0 2 2 102.46 0.0 103.96 0.0 1.46 104.7 103.96 1.46 103.97 0.0 1.47 104.59 103.97 1.47 104.11 0.0 1.61 105.53 104.11 1.61
1 2 1 56.79 0.0 56.79 0.0 0.00 57.37 56.79 0.00 56.83 0.0 0.07 57.29 56.83 0.07 62.18 0.0 9.49 62.18 62.18 9.49

AN
2 2 2 137.30 0.0 140.26 0.0 2.16 140.7 140.26 2.16 140.28 0.0 2.17 140.53 140.28 2.17 142.49 0.0 3.78 142.63 142.49 3.78
3 2 1 74.80 0.0 74.89 0.0 0.12 74.89 74.89 0.12 75.02 0.0 0.29 75.02 75.02 0.29 86.73 0.0 15.95 86.78 86.73 15.95
4 2 1 70.84 0.0 71.77 0.0 1.31 71.77 71.77 1.31 73.30 0.0 3.47 73.30 73.30 3.47 77.48 0.0 9.37 77.48 77.48 9.37

M
5 2 1 102.67 0.0 106.43 0.0 3.66 107.14 106.43 3.66 108.72 0.0 5.89 108.99 108.72 5.89 109.34 14 6.50 110.89 109.34 6.50
6 2 2 89.19 0.0 93.74 0.0 5.10 93.74 93.74 5.10 93.74 0.0 5.10 93.74 93.74 5.10 95.20 0.0 6.74 95.20 95.20 6.74
7 2 1 88.82 0.0 89.05 0.0 0.26 89.05 89.05 0.26 89.49 0.0 0.75 89.49 89.49 0.75 89.65 0.0 0.93 89.65 89.65 0.93

D
8 2 1 77.49 0.0 82.76 0.0 6.80 82.76 82.76 6.80 89.71 0.0 15.77 89.71 89.71 15.77 100.39 0.0 29.55 100.39 100.39 29.55
9 2 2 102.88 0.0 108.32 0.0 5.29 108.66 108.32 5.29 110.47 0.0 7.38 111.33 110.47 7.38 118.48 0.0 15.16 118.92 118.48 15.16

TE
CPU Time 90.32 336s 92.80 t=410s 2.62 93.08 t=2.5s 2.62 94.15 t=1126s 4.24 94.40 t=1.9s 4.24 98.60 t=930s 9.91 98.97 t<1s 9.91
15×5×150 0 3 2 123.41 37 156.89 51 -- 129.26 125.01 -- 196.48 62 -- 135.97 129.36 -- 175.96 56 -- 153.81 147.41 --
1 3 2 120.48 43 126.66 47 -- 100.98 94.91 -- 172.28 61 -- 101.08 95.19 -- 118.62 44 -- 106.25 102.94 --
EP
2 3 2 86.55 11 82.87 5 -- 85.55 82.87 -- 93.39 22 -- 83.99 83.99 -- 148.43 52 -- 97.96 92.33 --
3 3 2 171.70 37 181.52 40 -- 167.06 163.65 -- 223.63 53 -- 167.17 163.92 -- 192.77 47 -- 177.18 175.67 --
4 3 2 120.52 10 169.19 42 -- 125.85 124.37 -- 134.28 25 -- 126.71 124.91 -- 148.92 34 -- 140.65 134.2 --
C

5 3 1 108.13 16 153.66 49 -- 102.41 102.03 -- 154.51 47 -- 104.71 104.71 -- 161.08 50 -- 110.97 107.79 --
AC

6 3 2 149.89 43 237.59 66 -- 139.67 137.91 -- 154.52 45 -- 140.85 137.95 -- 276.17 71 -- 142.24 140.29 --
7 3 2 125.02 22 134.89 29 -- 120.24 117.91 -- 118.55 17 -- 120.19 118.55 -- 242.16 64 -- 135.95 132.58 --
8 3 2 131.89 34 171.68 52 -- 135.1 133.14 -- 164.93 48 -- 138.96 135.98 -- 260.59 68 -- 153.60 149.49 --
9 3 1 92.20 9 114.20 29 -- 96.54 94.74 -- 150.59 49 -- 106.98 103.39 -- 132.91 39 -- 117.72 115.97 --
CPU Time t=1h t=1h 120.27 t=6.7s t=2.5s t<1s

Note: bold face indicates the optimal solution

 ACCEPTED MANUSCRIPT

Table 3. Experiments on 14 instances in Christofides and Eilon (1969)

Prob. Prob. size Ava. Used Greedy Construction Heuristic Restricted-CPLEX 10 runs of GA-DP* 10 runs of GA-DP
ID n×m×T Veh. Veh. Cap. Len. Obj Dev.% Time Obj Dev.% Time Avg. Min. Avg. G. Avg. t. Avg. Min. Avg. G. Avg. t.
1 50×5×60 7 7 160 -- 1,559.6 41.9 6s 1386.6 26.1 5h 1,187.1 1,148.9 262 31s 1,189.5 1,099.1 518 51s
2 75×5×60 13 11 140 -- 2,483.8 53.7 16s 1,824.7 1,696.9 494 1.2m 1,728.0 1,615.6 1091 2.7m
3 100×5×60 12 10 200 -- 2,878.0 49.4 38s 2,038.0 1,925.9 1090 3.0m 2,108.7 1,953.8 1031 3.0m
4 150×5×60 16 14 200 -- 3,755.2 47.0 1.8m 2,724.2 2,601.2 1482 7.4m 2,772.2 2,555.0 2190 9.5m

PT
5 199×5×60 20 18 200 -- 4,344.1 39.5 3.8m 3,429.5 3,353.0 1133 7.5m 3,444.1 3,115.0 3133 17.6m
6 50×5×60 7 6 160 200 1,953.3 65.7 6s 1636.2 38.8 5h 1,296.7 1,245.6 541 54s 1,251.7 1,178.5 458 44s
7 75×5×60 13 11 140 160 2,602.9 59.6 15s 1,888.3 1,749.9 834 1.8m 1,705.0 1,631.4 1179 2.7m

RI
8 100×5×60 11 9 200 230 3,094.0 70.6 43s 2,088.0 1,965.5 805 2.5m 1,939.0 1,814.0 1067 3.3m
9 150×5×60 15 13 200 200 -- -- -- -- -- -- 2,794.2 2,536.8 1979 9.1m
10 199×5×60 21 18 200 200 4,059.7 34.0 4.0m 3,178.4 3,099.5 1466 9.2m 3,210.3 3,028.7 1629 10.5m

SC
11 120×5×60 11 9 200 -- 14,840.4 301.1 1.2s 3,905.1 3,594.6 1132 4.7m 3,943.8 3,700.2 1393 5.4m
12 100×5×60 12 10 200 -- 2,813.0 58.6 40s 2,033.9 1,795.6 772 2.6m 2,088.5 1,773.8 664 2.2m
13 120×5×60 12 10 200 720 14,799.5 281.5 1.2m 10,304.2 4,007.5 873 4.2m 6,019.9 3,879.4 1210 5.2m

U
14 100×5×60 12 10 200 1040 3,294.7 83.0 42s 2726.0 51.4 5h 2,047.8 1,861.4 1025 3.8m 1,958.0 1,800.8 994 5.8m
Avg 4,806 112.4 1.2m 2,919 2,311 916 3.8m 2582 2,263 1324 5.6m

AN
Note: bold face indicates best solutions found.

M
D
TE
C EP
AC
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC