You are on page 1of 42

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/304657714

Optimizing for Costs and Emissions in Vehicle Routing in Urban Areas

Technical Report · July 2016

CITATIONS READS
4 1,449

3 authors:

Jan Fabian Ehmke Ann Melissa Campbell


Otto-von-Guericke-Universität Magdeburg University of Iowa
89 PUBLICATIONS 1,764 CITATIONS 74 PUBLICATIONS 4,177 CITATIONS

SEE PROFILE SEE PROFILE

Barrett W. Thomas
University of Iowa
86 PUBLICATIONS 3,086 CITATIONS

SEE PROFILE

All content following this page was uploaded by Jan Fabian Ehmke on 01 July 2016.

The user has requested enhancement of the downloaded file.


Optimizing for Costs and Emissions in Vehicle Routing in Urban
Areas
Jan Fabian Ehmke (Corresp. Author)1 , Ann Melissa Campbell2 , and Barrett W. Thomas2
1
Business Information Systems, Freie Universität Berlin, Garystr. 21, D-14195 Berlin
1
Tel. +49-30-83858731, Fax +49-30-838458731, Email janfabian.ehmke@fu-berlin.de
2
Department of Management Sciences, The University of Iowa, 52240 Iowa City, USA

June 30, 2016

Abstract
The minimization of emissions has become an important topic in vehicle routing research. Still,
minimizing costs remains one of the most important objectives for logistics service providers, and it is
not clear how minimizing emissions impacts total costs. In urban areas, speeds can vary significantly
due to congestion over the course of the day, and the total cost and fuel consumption associated with
different paths between customers can vary based on travel speed and the load on the vehicle. To compute
cost-effective and environmental-friendly routes in urban areas appropriately, large amounts of detailed
speed data are needed in adapted vehicle routing algorithms. This paper compares the total cost (based
on combined driver and fuel costs), fuel consumption (emissions), distance, and travel time for routes
resulting from optimizing each of those measures. We also create routes that optimize fuel consumption
using a simplified fuel consumption function to understand the value of a detailed fuel model. We
explore the impact of multiple considerations on these measures as well as the structure of the routes
themselves. We consider driver hourly cost, fuel cost, customer geography (inner city, suburban, mixed),
customer load distribution (homogeneous versus heterogeneous), vehicle size (standard versus heavy),
fleet composition (homogeneous versus heterogeneous), traffic congestion (rush hour versus non-rush
hour), and whether or not the vehicle is delivering or picking up loads at customers. This extensive set
of experiments is conducted based on the LANTIME algorithm, which is a tabu search heuristic that we
have adapted to consider complex cost functions and load information. We use a large speed database
with millions of speed observations from the area of Stuttgart, Germany, as input for our computational
analysis. The results show that the total cost objective produces significantly different routes for mixed
fleets, heterogeneous customer demands, and suburban and mixed geography locations. We also observe
features about how the structures of the routes differ when total costs are minimized as compared with
fuel, such as how heavy vehicles are utilized in mixed fleet compositions and when customers with larger
loads are visited.
Key words: emissions, vehicle routing, green logistics, driver costs, urban logistics.

1 Introduction

With increasing attention given to the impact that delivery vehicles have on the environment, many compa-

nies are seeking to manage their environmental impact while maintaining cost competitiveness. As examples,

Walmart, General Mills, and Anheuser-Busch all have programs designed to reduce the fuel consumption

and thus emissions of their delivery fleets (Hardcastle, 2015). For many companies, programs to reduce fuel

1
consumption focus on reducing driving distance. In fact, Chuck Holland, Vice President of Industrial Engi-

neering at UPS, was recently quoted as saying, “We also have a saying here at UPS: we say the greenest mile is

the mile never driven” (Hardcastle, 2016). While reducing the miles driven can certainly reduce emissions,

optimizing for distance does not guarantee that fuel consumption and emissions are minimized. Perhaps

more importantly, minimizing distance does not guarantee that a company’s total costs are minimized.

In this paper, we seek to investigate the trade-offs that occur in total cost, time, distance, and fuel

consumption (emissions) in urban areas when optimizing for these different measures. We focus on urban

areas for two reasons. First, large cities, particularly European cities, have been the leaders in emissions

reduction efforts (Slavin, 2015). With trucks delivering the majority of goods in cities and trucks being

a major source of emissions (Smith et al., 2015), cities have been looking at changes in the operation of

the trucking industry to help meet emissions goals. For example, cities such as Amsterdam and London

have been utilizing congestion zones, where truck traffic is limited, to reduce emissions (see http://www.

milieuzones.nl and http://www.tfl.gov.uk/roadusers/lez/default.aspx). Second, because of time-

dependent traffic congestion and the variability in traffic speeds, urban areas are a challenging environment

to minimize total costs or emissions.

Specifically, this paper compares the total cost (based on combined driver and fuel costs), fuel consump-

tion (emissions), distance, and travel time for routes resulting from optimizing each of those measures. We

also create routes that optimize fuel consumption using a simplified fuel consumption function to understand

the value of a detailed fuel model. We explore the impact of multiple considerations on these measures as

well as the structure of the routes themselves. We consider driver hourly cost, fuel cost, customer geogra-

phy (inner city, suburban, mixed), customer load distribution (homogeneous versus heterogeneous), vehicle

size (standard versus heavy), fleet composition (homogeneous versus heterogeneous), traffic congestion (rush

hour versus non-rush hour), and whether or not the vehicle is delivering or picking up loads at customers.

Our results show that minimizing distance can lead to significantly higher costs than when minimizing for

total cost. Minimizing travel time can also lead to significantly higher fuel consumption than optimizing for

total cost, especially with heterogeneous fleets. In addition, our results show that minimizing for total cost

often increases fuel consumption only minimally over routes optimized for fuel consumption. We also observe

2
features about how the structures of the routes differ when total costs are minimized as compared with fuel

consumption, such as how heavy vehicles are utlilized in mixed fleet compositions, and when customers with

larger loads are visited.

To derive our conclusions, we base our experiments on the road network of Stuttgart, Germany and

a database of millions of speed observations. We solve instances on the network using an adaptation of

the LANTIME routing heuristic, a well known heuristic for time-dependent vehicle routing problems. Our

adaptation accounts for two features important to fuel consumption. First, the load of the vehicle changes

as it visits customers, and the load impacts the fuel consumption. Second, there is variability in the speeds

on arcs even at the same time of day. Because of the convexity of fuel-consumption curves, the failure to

account for the variability leads to an underestimation of the speed as shown by Jensen’s Inequality. These

two factors mean that the best path between two customers depends on the load and time of day of travel.

Thus, in determining routes optimized for total cost or fuel consumption, we must solve for the best paths

between customers on the routes. To improve computational efficiency, we extend an earlier result that

allows us to precompute and store a large number of these paths.

The remainder of this paper is outlined as follows. In the next section, we review the relevant literature.

Section 3 details our research questions and introduces our experimental design, and Section 4 explains our

methodology in detail. We present results of our experiments in Section 5 and conclusions in Section 6.

2 Literature Review

The vehicle routing problem is one of the most studied problems in operations research. However, until

recently, most of the work focused on linear objectives, usually related to minimizing distance or travel time

(Toth and Vigo, 2014). In recent years, researchers have begun to consider the impact of fuel consumption

or emissions. Both fuel consumption and emissions are generally modeled as nonlinear functions of vehicle

speed. In this paper, we model fuel consumption using the Comprehensive Emissions Model (CEM) described

in Barth and Boriboonsomsin (2008). This model was originally designed to model the emissions and

fuel consumption of heavy-goods vehicles, such as those of interest in this paper, and computes the fuel

consumption on an arc as a function of speed, vehicle weight, and numerous vehicle and arc-specific constants.

We specifically implement the time-dependent version of the CEM found in Franceschetti et al. (2013). For

3
comprehensive reviews of fuel consumption and emissions models, see Demir et al. (2011) and Demir et al.

(2014a).

The literature incorporating fuel consumption in routing and shortest path problems includes a wide

variety of models and methodological approaches. Reviews can be found in Bektas et al. (2016), Demir

et al. (2014a), and Lin et al. (2014). The work in this paper is most closely related to the work by Ehmke

et al. (2016a), Ehmke et al. (2016b), Hwang and Ouyang (2015), and Huang et al. (2016). All four papers

recognize that, when considering fuel consumption in the objective, the best paths between customers change

as a function of the congestion on the road network and the variability of speeds in the network. In the

manner of Huang et al. (2016), we refer to the ability to change paths in response to these factors as path

flexibility.

Ehmke et al. (2016a) introduce two methods for finding minimum fuel consumption paths. The first

method incorporates sampling methods into an A? -based algorithm. The second uses an averaging technique

that captures the impact of variability on the fuel consumption on an arc. The solution methodology for

the COST and FUEL objectives in this paper takes advantage of the latter method. We note that Wen

et al. (2014) consider the case of finding minimum cost paths in a time-dependent network with congestion

charges. However, unlike Ehmke et al. (2016a), Wen et al. (2014) do not consider variation in speeds nor

the nonlinearities introduced by minimizing fuel consumption.

Ehmke et al. (2016b) extend Ehmke et al. (2016a) to vehicle routing by exploring the impact of multiple

pickups and vehicle load on expected emissions-minimized routes. The load on the vehicle further impacts

the best paths between customers. To reduce the computational burden of computing these paths at runtime,

Ehmke et al. (2016b) introduce a result that allows a significant portion of the paths to be precomputed

and stored for access at runtime. We extend this result to the case of total cost. The results in Ehmke

et al. (2016b) show that the greatest increases in saving from explicitly modeling emissions occurs in longer

suburban routes with heavy vehicles. The results also show that expected emissions-minimized paths increase

route durations proportionately less than expected duration minimized routes increase emissions. This paper

extends the two prior papers by exploring the total cost of vehicle routes when considering both load and

the impact of variability on fuel consumption.

4
Like this paper, Hwang and Ouyang (2015) and Huang et al. (2016) seek to minimize total expected

costs. Hwang and Ouyang (2015) propose two approaches for the paths, the second similar to the second

approach presented in Ehmke et al. (2016a). Hwang and Ouyang (2015) present four small case studies and

conclude that modeling fuel consumption as a function of travel speed, particularly variable speeds, leads to

routes with total costs different than the routes chosen when the speed variability is ignored. In contrast to

Hwang and Ouyang (2015), this paper considers a significantly more comprehensive experimental design that

allows us to explore which network characteristics most likely lead to differences in various route measures.

Further, we base our work on a significantly larger network and base our time-dependent speed and travel

time distributions on a dataset of real speed observations.

Like Ehmke et al. (2016a), Ehmke et al. (2016b), and Hwang and Ouyang (2015), Huang et al. (2016)

observe that the fuel consumption from a path is a function of speed, speed variability, and load. In

contrast, the authors pre-compute a limited set of paths between each customer pair and select the best

path from this set as needed, though. As a result, there is no guarantee that the best route for a given

time of day and load is chosen. However, the model proposed by Huang et al. (2016) facilitates a scenario-

based stochastic programming approach. The majority of the discussion in Huang et al. (2016) ignores

the variability in travel speeds and instead assumes fixed, time-dependent speeds. The results in this paper

confirm and generalize the results presented in Huang et al. (2016). Our work considers the impact of pickups

in addition to deliveries, the distribution of load among customers, and different fleet compositions. Further

and importantly, our work focuses on a much larger network and generates speed and travel time data from

real speed observations. These individual observations allow us to compute paths, routes, and costs based

on time-dependent distributions of speed and travel times on every arc in the network rather than using a

limited set of scenarios.

Table 1 summarizes the previously described routing papers as well as the rest of the related litera-

ture. The table breaks the literature into three objective categories: comprehensive cost, fuel consump-

tion/emissions, and multi-objective. We note that there are variations in the specific objective calculations

among the papers in each category. The table next characterizes the model of emissions used in each paper.

Particularly, we look at whether or not a paper uses a validated fuel consumption model, such as those

5
discussed in Demir et al. (2011) and Demir et al. (2014a), and whether or not the fuel consumption model

includes the effect of the vehicle load. The table then categorizes the papers with regard to how speed is in-

corporated into the fuel consumption model. We first consider whether or not the speeds are time-dependent.

We also look at whether or not a paper treats speeds as deterministic by setting speeds on an arc to the mean

speed, treats the speeds as variable as we do in this paper, or seeks to optimize speed to reduce emissions.

A paper with no entry in these three categories is a paper that does not use an emissions model based on

speed. The next category considered is the solution method. We characterize each paper based on whether

or not the solution method is exact or heuristic. Finally, we consider whether or not the paper considers

time windows, heterogeneous fleets, path flexibility, and departure time optimization. The table shows the

limited number of papers that consider time dependence, variation in travel speeds, heterogeneous fleets,

and path flexibility, and that, other than this paper, no paper combines all these aspects with a total cost

objective.

3 Experimental Design

In this section, we first present the research questions we want to address in this paper. Second, we will

discuss the details of the experiments and test instances created to address these questions.

3.1 Research Questions

Our primary research question is how does a combined objective of driver and fuel costs impact routing

solutions, as compared with solutions found based on minimizing distance, time, and fuel? We will address

this question by solving the same problem instances with five different objectives, including cost, distance,

time, and fuel. We will then evaluate the solutions with four measures to understand how the solutions

differ. The details regarding the objectives and measures considered are in Section 3.2.1. In the following,

we describe the questions we study within the scope of the primary research question to help us understand

the impact of different factors on the problem and provide a high level summary of the types of instances

we created to solve these questions.

• How do the solutions vary as we consider only one vehicle (traveling salesman problem, TSP) vs.

multiple vehicles (vehicle routing problem, VRP)? How do the conclusions vary across different vehicle

6
Table 1: Literature Summary
Emissions Model Speed Solution Method Time Heterogeneous Path Departure Time
Objective Paper
Windows Fleets Flexibility Optimization
Validated Load Time Dependent Mean Variable Optimized Exact Heuristic

Bektas and Laporte (2011) X X X X X


Demir et al. (2011) X X X X X
Kramer and Subramanian (2015) X X X X X
Kramer et al. (2015) X X X X X X
Zhang et al. (2015) X X X
Dabia et al. (2014) X X X X X
Fukasawa et al. (2015b) X X X X X
Fukasawa et al. (2016) X X X X X
Comprehensive
Koç et al. (2014) X X X X X X X
Cost
Jabali et al. (2012) X X X X X
Franceschetti et al. (2013) X X X X X X X
Figliozzi (2010) X X X X X
Kwon et al. (2013) X X
Koc et al. (2016) X X X X X X
Hwang and Ouyang (2015) X X X X

7
Wen and Eglese (2015) X X X X
Huang et al. (2016) X X X X X X
This Paper X X X X X X X

Kara et al. (2007) X X


Fukasawa et al. (2015a) X X
Zachariadis et al. (2015) X X
Xiao et al. (2012) X X
Fuel Oberscheider et al. (2013) X X X X
Consumption/ Kuo (2010) X X X X
Emissions Kuo and Wang (2011) X X X X
Qian and Eglese (2016) X X X X X
Kopfer et al. (2014) X X X
Xiao and Konak (2016) X X X X X X X
Ehmke et al. (2016b) X X X X X X X

Sadegh-amalnick (2016) X X
Rao et al. (2016) X X X X
Multi-Objective
Demir et al. (2014b) X X X X X
Urquhart et al. (2010) X X X X
Carlos et al. (2014) X X X
types, as some vehicles have larger capacities but are less fuel efficient than others? How do the multiple

vehicle results change as we consider homogeneous vs. hetergeneous fleet compositions?

To this end, we consider experiments with one vehicle and a fleet of three vehicles. With one vehicle, we

consider two sizes (one standard or one heavy). With three vehicles, we experiment with the vehicles

being the same (three standard or three heavy) and with one heavy and two standard vehicles (mixed

fleet). Details are in Section 3.2.2.

• How do the conclusions vary across different customer geographies?

We will address this question by looking at customers found in the inner city of a metropolitan area,

the suburbs, and a mixture of both. Details of these instances are found in Section 3.2.3.

• How do the conclusions vary over the course of the day, as travel at rush hour periods is slower and

has more congestion than non-rush hour periods?

We will address this question by looking at departures at four different times of day, two representing

rush hour departure times and two representing non-rush hour departure times. Details of these

experiments are found in Section 3.2.4.

• How do the conclusions vary with different load sizes?

We will examine homogeneous load sizes for customer pickup quantities as well as three different

schemes for assigning heterogeneous pickup quantities to customers. Details are in Section 3.2.5.

• How sensitive are the results to different values of gas prices and driver wages?

We explore values representing low and high gas prices and low and high driver wages over the different

problem instances. Details on these cost parameters are in Section 3.2.6.

• How do the conclusions vary for pickups vs. deliveries?

For many of the instances discussed above, we solve a variant where the load at a customer represents

a pickup quantity as well as a variant where the load at a customer represents a delivery quantity.

8
3.2 Experimental Design

In the following, we explain how we design our experiments and test instances. We demonstrate and evaluate

the construction of routes with different objectives based on real speed data from the metropolitan area of

Stuttgart, a major city in southern Germany. The metropolitan area of Stuttgart is well-known for its

congestion, especially at peak times (Kröger, 2013). We use a database of approximately 230 million speed

observations from the years 2003-2005. These speed observations were collected by the German Aerospace

Center with FCD technology using a fleet of 700 taxis (Lorkowski et al., 2004). Due to the speed limitations

often imposed on freight vehicles, we set a speed limit of 100 km/h, i.e., whenever we observe a speed larger

than 100 km/h, we set it to 100 km/h. We focus on Tuesday data, as Tuesday is considered a “typical” day

by traffic engineers (Ehmke, 2012). We pair the database with a digital roadmap and derive a network that

consists of 5385 nodes and 8629 arcs.

3.2.1 Objectives

To analyze the impact of cost minimization versus more traditional objectives, we generate and compare

solutions for each of the following objectives:

Objective 1: COST minimize total costs, including driver and fuel costs.

Objective 2: TIME minimize total time-dependent travel time, focusing on the shortest working time,

which is a well-known objective in routing.

Objective 3: FUEL minimize time-dependent fuel consumption given detailed information about the load

at every arc in the course of a route, which is computationally challenging, but the most realistic model

with respect to fuel computation.

Objective 4: FUELS minimizing time-dependent fuel consumption given the gross weight of an empty

truck, which does not consider the vehicle’s changing load in the course of a route and also uses average

speeds on arcs to estimate fuel consumption. Average speeds are often much easier to obtain in practice

and ignoring both of these issues reduces the computational complexity of the objective computation.

Results from these experiments will indicate when the more complicated modeling of Objective 4 is

important to capture fuel usage. We have included this objective in our experiments because the

9
majority of the papers in Section 2 did not consider the variability of speed in their methodology and

many also ignored the impact of load.

Objective 5: DIST minimize total distance travelled, which is often taken as a proxy for minimizing

emissions particularly in early research and in practice.

To compare each of the objectives, we evaluate the solutions from Objectives 1, 2, 3, 4, and 5 with the

measures of total combined driver and fuel cost in euros, the total duration of the solution in minutes, total

fuel consumption in liters, the distance traveled by the solution in kilometers, and the runtime needed to

find the solution in seconds.

3.2.2 Fleet Composition

We generate routes for a fleet of standard vehicles (gross weight of 6350 kg) and heavy vehicles (gross weight

of 12700 kg). We assume that the capacity of the vehicle is equal to the gross weight of the vehicle, so that,

when fully loaded, the vehicle weighs twice as much as when empty. We also consider a mixed fleet scenario

with two standard vehicles and one heavy vehicle for the VRP experiments.

3.2.3 Geography

We develop several test instances that consider spatio-temporal dimensions relevant for freight transportation

in urban areas. In particular, we consider that inner city customers are closer to one another than suburban

customers are to one another, and that congestion patterns on inner city roads differ from congestion patterns

connecting the inner city and the suburbs and from patterns connecting the suburbs. To this end, based on

their geographical location, we divide the set of 5385 nodes into an inner city and a suburban node set with

1160 and 4225 nodes, respectively. Using this node division, three categories of test sets are generated:

• Inner city sets (I) each consisting of 10 (TSP) or 30 (VRP) nodes that are randomly drawn from the

nodes located in the inner city,

• Suburban sets (S) each consisting of 10/30 nodes that are randomly drawn from the nodes located

outside the inner city, and

• Mixed sets (M) each consisting of half of the nodes being contained in the inner city and suburban

sets.

10
For each geographical category, we produce five instances, resulting in instances denoted by I1. . . I5,

S1. . . S5, and M1. . . M5. The depot is the same for all instances; it is located in the suburbs in the North

Western part of the city, but close to the inner city area.

3.2.4 Departure Times

We consider different times for pickup and delivery services to analyze the impact of the different objectives

on on time of day. To this end, we consider the above node sets I, S and M with 10 and 30 customers, and

for each instance, we generate routes for four starting times: two rush hour times (06:30, 15:30) and two

non-rush hour times (12:30, 19:30).

3.2.5 Load Sizes

TSP Experiments

For the TSP experiments, we consider instances where all customers require a pickup or all customers require

a delivery. In both cases, the weight of the quantity at each customers is defined as

gross weight of vehicle


,
total number of customers

implying that a vehicle is “fully loaded” on the way from the last customer to the depot (for pickups) or

from the depot to the first customer (for deliveries).

We also consider instances with heterogeneous loads to test the impact of heavily varying load quantities

on route structures, as that had a big impact in Ehmke et al. (2016a). To this end, based on the I, S, and

M instances, we generate three load distributions for each of the gross truck weights (standard and heavy).

For two of the load distributions, we assign the loads for three selected customers such that their load to-

tals 90% of the vehicle’s capacity, and the remaining 10% is evenly divided among the other seven customers.

The three customers with heavy load quantities are chosen as follows:

Farthest three heavy: identification of the three customers farthest from the depot in terms of distance,

Closest three heavy: identification of the three customers closest to the depot in terms of distance.

We assign higher weights to only three loads to help build insights on where heavy loads can have the most

impact on solutions. For the third load distribution, we assign loads for all customers as follows:

11
Random decreasing load: all customers are given a load such that the second heaviest pickup (delivery)

is half the heaviest pickup (delivery), the third heaviest pickup (delivery) is half the second heaviest

pickup (delivery), etc., and we scale the sum of the pickup (delivery) weights such that it equals the

total capacity of the vehicle. These decreasing loads are randomly assigned to the customers. We use

this type of instance to represent a more typical case where customers’ pickup or delivery quantities

show a large diversity in sizes.

As with the homogeneous load instances, we run each of these heterogeneous load instances for two rush hour

times (06:30, 15:30) and two non-rush hour times (12:30, 19:30). In total, this represents 960 test instances

for the TSP that will be investigated for each of the five different objectives discussed above.

VRP Experiments

The VRP instances are based on the 30-customer instances. We consider both homogeneous fleets (three

standard or three heavy vehicles) and a mixed fleet of a heavy vehicle and two standard vehicles. We vary the

load at each customer so that three vehicles are required to process all pickups or all deliveries. Specifically,

we set the load for pickup or delivery for our homogeneous load experiments as follows:

Homogeneous loads: We define the mean load weight as


P
vehicle type gross weight of vehicle × number of trucks of each vehicle type
.
total number of customers

Thus, when considering only standard (heavy) vehicles, the numerator will be three times the weight

of a standard (heavy) vehicle. For the mixed fleet instances, the numerator will be the sum of two

standard weight vehicles and one heavy vehicle.

We also experiment with heterogeneous pickup and delivery quantities. These experiments should help us

gain insight into the characteristics of pickups and deliveries that are assigned to different vehicle types when

there is a choice. We consider three different load assignments for the mixed fleet experiments:

Farthest 50% heavy: The 15 customers closest to the depot have a load of 0.5 times the mean load weight,

and the 15 customers farthest from the depot have a load of 1.5 times the mean weight.

Closest 50% heavy: The 15 customers closest to the depot have a load of 1.5 times the mean load weight,

and the 15 customers farthest from the depot have a load of 0.5 times the mean weight.

12
Random 50% heavy: We randomly choose 15 customers to have a load of 1.5 times the mean load weight

and 15 customers to have a load of 0.5 times the mean load weight.

As with the TSP instances, for each I, S, and M VRP instance, we consider two rush hour times (06:30,

15:30) and two non-rush hour times (12:30, 19:30). Since we focus on the analysis of pickups only for the

VRP instances, we have a total of 720 test instances that will be investigated for each of the five objectives

discussed above.

3.2.6 Cost Parameters

For each of the test instances, we generate routes based on different cost parameters:

cheap gas/cheap labor (CC): this reflects the status quo, which is about e 1.19/l for fuel and e 11.46/h

for a truck driver in the area of Stuttgart, Germany;

expensive gas/cheap labor (EC): this reflects the highest fuel price of the last years, which was e 1.52/l;

cheap gas/expensive labor (CE): as truck drivers are becoming a scarce resource, we also consider a

labor cost of e 14.84/h, which is what professional drivers can earn.

Considering the number of different objectives above together with the different cost parameters, this creates

a total of 14400 TSP experiments and 10800 VRP experiments.

4 Methodology

In the following, we present our formal model of the routing problem and how we evaluate the cost of an arc

and of a particular path for the different objectives we consider. Then, we explain our approach to finding

solutions for each of the variants.

4.1 Model

We consider a set C of C customers such that each customer c = 1, . . . , C is associated with a pickup or

delivery of weight wc and service time stc . The customers are situated on a graph G = (N , A), where N is a

set of nodes, representing intersections and customers, with C ⊂ N , and A a set of directed arcs connecting

the nodes in N . We seek to serve the customers with a fleet of capacitated vehicles M consisting of M

vehicles.

13
To travel between two customers i and j in C, the vehicle uses a series of arcs, called a path. We represent

the cost of a path pij = (ap1 , . . . , apn ) starting at time t between customers i and j carrying load l as φpi,j (t, l).

We assume that there is no waiting at either customer or intersection nodes. Thus, for a path p, we can
Pn
compute φpi,j (τ, l) as φpi,j (t, l) = E [ k=1 ψkp (t(p, τ ), l)], where ψkp is the cost of the k th arc of path p and

t(p, τ ) makes explicit the dependence of the cost of k th arc on both the path p and the start time of path p,

which we denote τ .

A route r is a sequence of customers in C. We assume that ri is the ith customer on route r. The cost of

route r depends on the start time of the route and the paths chosen for travel from customer to customer

on the route. The total cost of a route r starting at time T is

C−1
X
Φr (T ) = φrk ,rk+1 (t(r, T ), l(r)), (1)
k=1

where t(r, T ) makes explicit the dependence of the cost of the path from rk to rk+1 on the start time of

route r and l(r) the load on the vehicle after serving customer rk on r. For convenience, we have dropped

the dependence on p from φ and assume that we choose the cost minimizing path between customers rk and

rk+1 . For routes performing pickups, the value of l(r) ranges from 0 for the initial path from the depot to

the first customer and increases for each subsequent path on the route based on the wrk values. Similarly,

for routes making deliveries, l(r) is initially set to the total amount to be delivered on a route and decreases

after each visit to a customer based on the values for wrk . The total cost of a set of routes R is the sum of

the costs of the routes r in R. We assume that there are M routes in R, one for each vehicle in the fleet M.

For a given start time T , the objective is find a set of routes R? such that
P P
r∈R? Φr (T ) ≤ r 0 ∈R Φr0 (T )

for any other set of routes R.

4.2 Computing the Cost of an Arc

Before we describe how we determine the paths between customers and the routes for particular vehicles, we

discuss the computation of the cost of an arc ψi,j (t, l) for each of the five objectives that we use. As noted

previously, this detail is necessary, because for a number of our objectives, the costs of using an arc are time

and load-dependent. The computation of the DIST objective is straightforward. For an arc a, the cost of

traversing the arc in the DIST objective is simply the distance da (in m) of arc a. The rest of the section

14
discusses the other four objectives.

4.2.1 FUEL and FUELS

We begin by discussing how to compute the cost of traversing an arc when using the FUEL objective. We

compute the fuel cost for the FUEL objective based on the fuel consumption provided by the CEM as

presented by Bektas and Laporte (2011) and Franceschetti et al. (2013). We use the same equations and

parameters as in Franceschetti et al. (2013) modified to account for time-dependent speeds. The CEM takes

as input the vehicle load l (in kg), the distance da associated with arc a, the speed on arc a (in m/s) at time

t, which for our purposes is a random variable, denoted va,t . The resulting time and load-dependent fuel

consumption on arc a is
  
da
ψaFUEL (t, l) = E λ kNe V 2
+ γβda va,t + γα(µ + l)da , (2)
va,t

where Ne , V , µ, λ, k, γ, β, and α are parameters related to the vehicle and its engine. Following Franceschetti

et al. (2013), we set the engine speed Ne = 33, the engine displacement V = 5, the engine friction factor

k = 0.2, and the curb-weight µ = 6350 for a standard vehicle and µ = 12700 for a heavy vehicle, respectively.

Note that α, β, γ and λ represent the product of multiple factors presented in Franceschetti et al. (2013).

The corresponding values are α = 0.0981, β = 1.6487, γ = 0.0028, and λ = 1/32428. For a given truck, the

parameter values are assumed to be identical for each arc. As in Bektas and Laporte (2011) and Franceschetti

et al. (2013), we assume that the gradient of the road is zero. The resulting fuel consumption can be easily

converted into a cost value by multiplying it with a given fuel price δ. To evaluate the expectation in

Equation 2, we use time-dependent speed samples from the Stuttgart speed database. A detailed discussion

of the procedure can be found in Ehmke et al. (2016a).

The FUELS objective is a simplification of Equation 2. In the FUELS objective, we replace the random

speeds va,t with their means v̂a,t and fix the load component to the gross weight of an empty truck (l = 0).

The result is
 
FUELS da 2
ψa (t, l) = λ kNe V + γβda v̂a,t + γαµda . (3)
v̂a,t

15
4.2.2 TIME

Like the FUEL objective, the arc cost associated with the TIME objective relies on the speed at which an

arc is traveled. We can express the expected travel on arc a at time t as


 
da
ψaTIME (t, l) = E . (4)
va,t

As with Equation 2, we use time-dependent speed samples from the Stuttgart speed database to evaluate

Equation 4.

4.2.3 COST

The computation of the cost of an arc for the COST objective is a straightforward extension of the arc costs

for the FUEL and TIME objectives. In the COST objective, arc costs are based on arc-related costs for fuel

and drivers. If we are leaving a customer node, we must also include the cost of the service time associated

with the stop, since this affects the driver costs. Fuel consumption is calculated exactly as is described for

the FUEL objective above. We assume a cost per unit of fuel of δ. The driver’s time on an arc is calculated

exactly as it is in the TIME objective. We let η be a cost of a unit of a driver’s time. We then compute the

cost of arc a for the COST objective as

ψaCOST (t, l) = δψaFUEL (t, l) + ηψaTIME (t, l) + ηsta , (5)

where ta denotes the service time at the tail node of a. If the tail node of arc a is not a customer, sta = 0.

4.3 Computing Minimum Cost Paths between Customers

As noted previously, some of the challenges faced in routing customers in this paper is the time dependence

of the travel times, the load dependence of fuel consumption, and the variability in the speeds. These issues

do not exist for the DIST objective. Thus, we can precompute the distance-minimum path between every

pair of customers based on the da values and store them in a lookup table.

The other four objectives are time dependent. To handle the time dependence, we follow the convention

of considering time in one-hour periods (Ehmke et al., 2012). We divide the samples for an arc among

one hour buckets (06:00–07:00, 07:00–08:00, 08:00–09:00, etc.). It is then assumed that the travel time and

emissions for traversing an arc starting at time t is given by the expected travel times and expected emissions

16
of the bucket that contains t. For example, if t = 06 : 30, then we would use values associated with the

bucket 06:00–07:00. If the neighboring bucket is entered while traversing an arc (e.g., when the start time is

close to the end of a bucket), the different speed levels are linearized in the transition area as described by

Fleischmann et al. (2004).

For the FUELS and TIME objectives, we can follow a standard approach in time-dependent routing (as

described in Fleischmann et al. (2004), for example) and compute the average travel speed v̂a,t for each

bucket and use these values to compute the cost of an arc as given in Equations 3 and 4, respectively. We

can then use a Dijkstra-like label-setting algorithm with these time-dependent arc values to efficiently find

the minimum FUELS and TIME paths between each pair of customers for each hour of the day. We store

the resulting path information in a time-indexed lookup table.

For the objectives of COST and FUEL, we cannot in general precompute the best path for every pair of

customer nodes at every hour of the day and store them in a lookup table. These paths depend not only

on the time at which the arc is being traversed, but also on the vehicle’s load at the time that that arc is

being traversed. In addition, as a result of Jensen’s Inequality, using the average speed on an arc as we do

in FUELS and TIME underestimate the true fuel usage. Further, fuel usage is impacted by the load on the

vehicle.

We take two steps to overcome this challenge. First, we take advantage of the results in Ehmke et al.

(2016a) to efficiently determine high quality, expected fuel-minimizing (or cost-minimizing) paths using a

Dijkstra-like approach. The path-finding method introduced in Ehmke et al. (2016a) relies on a transforma-

tion of the speed related terms of Equation 2. The transformation computes, for every value speed value k

in a bucket b, the value


da k 2
kNe V k
+ γβda (va,t ) . (6)
va,t
h  i
da 2
Taking the average over all of these values, we have an approximation of E λ kNe V va,t + γβda va,t . To

find the fuel-minimizing path, for each arc we must also add the load component of Equation 2, λγα(µ+l)da ,

to get the fuel usage on an arc. To find the total cost minimizing path, for each arc, we must convert the fuel

usage to fuel cost and add the driver cost. We note that this method is distinctly different than that used in

FUELS because it accounts for the variability in speed for a given time. However, the method does ignore

17
Algorithm 1 Time and Load-Dependent Routing Algorithm
1: Input: A solution s containing a set of routes and a start time t
2: for all routes r in s do
3: for all i in r do
4: if lookup(i, i + 1, b, l) = null then
5: Compute time and load-dependent cost-minimized path
6: Update table with expected costs of path for the given i, i + 1, b, l
7: end if
8: end for
9: end for

the potential arrival time variability to a node. Fortunately, the results in Ehmke et al. (2016a) demonstrate

that the effect of doing so is minimal relative to a computationally burdensome method that does account

for the arrival time variability.

Second, Ehmke et al. (2016b) presents a method to deal with the best path between customers being

dependent on the load on the vehicle. Ehmke et al. (2016b) addresses this issue in the case of FUEL

consumption. For this paper, we generalize the method to the case of the COST objective. The method

relies on a result that identifies the conditions under which a time-dependent path between two customers

is load invariant. Essentially, for a given time and two customers, if the path used by an empty vehicle

and the path used by a full vehicle are the same, then the path between customers is the same for all load

sizes. If the paths match, we can store the load-independent costs of the path in a lookup table that will

be used by the routing algorithm. Then, in runtime, we can compute the cost of the path by looking up

the load-independent portion of the path cost and adding the load portion of the cost. In the event that

the paths do not match, no values are placed in the precomputed lookup table, and we compute the cost of

each arc in a path (and the best path) at runtime. However, as the results in Ehmke et al. (2016b) show,

this method allows us to precompute a significant portion of the paths.

4.4 Routing Algorithm

Given the ability to evaluate arc costs and to determine the minimal cost paths between two customers,

this section presents our method of determining vehicle routes. To solve the routing problem for all five

objectives, we adapt the LANTIME tabu search algorithm, which was designed for time-dependent variants

of the VRP. Details of the algorithm are provided by Maden et al. (2010). An overview of the algorithm is

presented in Algorithm 1.

18
LANTIME requires a feasible initial solution as an input. We construct this initial feasible solution using

Solomon’s I1 heuristic (Solomon, 1987). In addition, as in the literature, our LANTIME implementation uses

the neighborhoods Adapted Cross Exchange, Insertion/Removal, One Exchange, and Swap. A neighboring

solution is considered superior if the number of vehicles is smaller (main objective for VRP optimization) or

the number of vehicles is the same and the total cost of all routes is smaller (secondary objective for VRP

optimization, only objective for TSP optimization).

5 Computational Results

In this section, we present our computational results and use these to find the answers to our research

questions presented in Section 3. Since we conducted an extensive number of experiments, our results will

primarily be in the form of graphical tables to promote easier interpretation. For the sake of completeness,

detailed results for all experiments can be found in the electronic appendix of this paper in conventional

table format.

Beginning with overall results, we will breakdown the results for relevant parameter settings, especially

where the measures and/or the resulting routes differ significantly from each other. For all experiments,

we assume that customers have a fixed service time of 10 minutes. Our experiments are performed on a

Windows 7 64-bit operating system with an Intel Core i5-3470 processor and 8 GB of RAM. Algorithms are

coded and executed in Java 64-bit with a memory allocation of 4 GB. We let the tabu search metaheuristic

run for 1000 iterations per experiment. All results are based on the best of five runs per experiment.

5.1 Summary of Results

We will begin with an analysis of overall TSP results for initial cost parameters in Section 5.1.1 and extend

the analysis to fleets of vehicles in Section 5.1.2. We then analyze varying geographies (Section 5.1.3),

different departure times (Section 5.1.4), varying load sizes (Section 5.1.5) and differences between pickup

and deliveries in Section 5.1.7. Finally, we consider results for varying cost parameters in Section 5.1.6.

5.1.1 One Vehicle (TSP)

We will start with looking at the average results across all single vehicle pickup experiments with initial cost

parameters, i.e., a price of e 1.19/l for fuel and e 11.46/h for a truck driver (”cheap gas / cheap labor”).

19
avg_cost avg_distance avg_duration avg_fuel

TIME

FUELS
method

FUEL

DIST

−3% 0% 3% 6% −3% 0% 3% 6% −3% 0% 3% 6% −3% 0% 3% 6%


value

Figure 1: Comparative objective results averaged over all TSP pickup experiments with initial cheap/cheap
cost parameters

The results for these experiments across the different objectives are portrayed graphically in Figure 1. The

first column of graphs represents the total costs of the solutions found by optimizing different objectives,

expressed relative to the best COST solution. The remaining columns represent the average distance, average

duration, and average fuel consumption found when optimizing for the different objectives, again expressed

relative to the COST solution. Negative values represent savings relative to the results for that particular

measure when optimizing for total costs, where positive values represent increases.

When looking at how the average total costs change for different objectives (in the first column), it is

easy to see that, of the traditional objectives, DIST creates the highest cost solutions (just under 5% more

than COST solutions). From looking at the detailed solutions, the DIST routes often involve congested

arcs that use both fuel and driver time. TIME, FUELS and FUEL have fairly similar total costs relative

to COST on average, which makes sense because COST represents a combination of these objectives. It is

important to note that FUEL requires only about 1% more cost than our COST solutions. Thus, minimizing

emissions does not create costly routing solutions, and alternatively, minimizing total cost is much better

than minimizing traditional objectives. We also notice that, on average, the COST routes have a slightly

20
avg_cost avg_distance avg_duration avg_fuel

TIME

FUELS
method

FUEL

DIST

−2.5% 0.0% 2.5% −2.5% 0.0% 2.5% −2.5% 0.0% 2.5% −2.5% 0.0% 2.5%
value

Figure 2: Comparative objective results averaged over all VRP pickup experiments with initial cheap/cheap
cost parameters

longer duration than TIME routes (duration of TIME routes is just under 1% less than COST routes) and use

more distance than DIST routes (DIST routes use approximately 5% less distance). One interesting pattern

to observe is how TIME solutions perform in terms of fuel usage (the fourth column). TIME solutions often

choose routes involving highways, requiring longer distances to travel at faster speeds. These characteristics

make these solutions use more fuel than FUEL and COST solutions (over 5% more than COST solutions).

This indicates that even though COST solutions are only slightly longer in duration than TIME solutions,

the arcs chosen use much less fuel. The final observation from Figure 1 is the approximation offered by the

FUELS objective. We can see, that at least for one vehicle, FUELS offers a solution of similar total cost to

FUEL, but the fuel usage differs by approximately 3%. This finding highlights the value of detailed speed

data values a detailed model of fuel consumption in route building.

5.1.2 Multiple Vehicles (VRP)

Next, we examine the average results across all multiple vehicle pickup experiments. To demonstrate this,

we examine Figure 2. We see the same basic patterns as in Figure 1. The big difference is in the magnitudes.

For example, the increase in average cost from using DIST solution is now about 3.5% instead of almost 5%.

21
avg_cost avg_distance avg_duration avg_fuel

TIME

FUELS

standard
FUEL

DIST

TIME

FUELS
method

heavy
FUEL

DIST

TIME

FUELS

mixed
FUEL

DIST

−5.0% −2.5% 0.0% 2.5% 5.0% 7.5%


−5.0% −2.5% 0.0% 2.5% 5.0% 7.5%
−5.0% −2.5% 0.0% 2.5% 5.0% 7.5%
−5.0% −2.5% 0.0% 2.5% 5.0% 7.5%
value

Figure 3: Comparative objective results averaged over VRP pickup experiments with standard, heavy, and
mixed fleets with initial cheap/cheap cost parameters

We breakdown the pickup results for the multiple vehicles by fleet composition in Figure 3. The first

two rows represent the homogeneous fleets with three standard and three heavy vehicles, respectively. The

third row represents the mixed fleet with two standard vehicles and one heavy vehicle. We see that the

biggest differences due to fleet composition occur with average fuel usage (fourth column). The TIME

solutions use much more fuel than COST and FUEL solutions with the mixed fleet. A heavy vehicle uses

more fuel to do the same work as a standard vehicle. Thus, the FUEL and COST optimized solutions have

to carefully consider how to use the heavy vehicle in the solutions, whereas the TIME optimized solution

does not consider this. Even for the FUEL and COST solutions, the heavy vehicle is used quite differently.

For instance, let us look at an example for the FUEL and COST routes for the heavy vehicle of the mixed

fleet (with homogeneous loads) in Figure 4. In this and the following figures, denotes the distance in

kilometers, denotes the total duration in minutes, denotes the fuel consumption in liters, and

denotes the total cost in EURO. In the FUEL solution, the heavy vehicle is used over a small area (black

route in the top right hand corner) so that the heavy vehicle is used as little as possible (only 54.68 km to

use 16.6 liters of fuel). With this solution, the two standard vehicles cover the rest of the service area and

22
Figure 4: Pickup FUEL and COST routes for S9 for the heavy vehicle of the mixed fleet

create large driver costs. In the COST solution, the heavy vehicle instead takes on more work, reflected by

the light gray route on the left hand side of the graph. In the COST solution, the heavy vehicle now travels

86.80 km (using 25.6 liters of fuel).

5.1.3 Geography

Next, we breakdown our combined results to understand what happens in different parts of the city. We

demonstrate the results from using a single heavy vehicle over the different instance geographies (inner city,

suburbs, and mixed) in Figure 5. We see the least relative differences between the different measures for

the inner city (I) instances. This is due to the fact that there is less distance between inner city customers

and thus fewer opportunities for different paths that create differences in the measures. When we consider

suburban (S) and mixed (M) instances, we see larger differences in all measures. Longer distances between

23
avg_cost avg_distance avg_duration avg_fuel

TIME

FUELS

I
FUEL

DIST

TIME

FUELS
method

S
FUEL

DIST

TIME

FUELS

M
FUEL

DIST

−5% 0% 5% −5% 0% 5% −5% 0% 5% −5% 0% 5%


value

Figure 5: Comparative objective results averaged over TSP pickup experiments with heavy vehicles broken
down (rows) by geography with initial cheap/cheap cost parameters

customers create the opportunities to choose between several paths, some of which are better in terms of

distance, time, fuel, or total cost.

Figure 6 offers an example of how the routes change for different objectives for a mixed customer TSP

instance. Since direction does not impact distance measures, the DIST route (in (a)) is the only route

that goes counter-clockwise. The TIME optimized clockwise route (in (b)) uses highways and increases the

distance by 9 km (52.33 km vs. 61.55 km) to reduce travel time by 16 minutes (228 vs. 212 minutes). The

FUEL optimized route (in (c)) also travels clockwise and uses less fewer highways than the TIME route,

bringing the distance down to 54.11 km and the fuel consumption down from 18.4 liters to 17.0 liters. Notably

the order of customers visited is quite similar between the TIME and FUEL routes, with the switching of

customers in positions 7 and 8 and positions 9 and 10. The FUEL optimized route takes 227 minutes, where

the COST route (in (d)) brings it down to 215 minutes to save on driver costs. The COST route follows

the same ordering of customers as the TIME route, but clearly takes some very different paths between the

customers (particularly in the top right hand side of the graph) to use less fuel.

24
(a) DIST route (b) TIME route

(c) FUEL route (d) COST route

Figure 6: Pickup routes for M5 at 12:30 with Homogeneous Loads and Heavy Vehicle for different objectives.
Aerial view provided by Google Earth.

25
avg_cost avg_distance avg_duration avg_fuel

TIME

FUELS

0630
FUEL

DIST

TIME

FUELS

1230
FUEL

DIST
method

TIME

FUELS

1530
FUEL

DIST

TIME

FUELS

1930
FUEL

DIST

−5% 0% 5% 10% 15% −5% 0% 5% 10% 15% −5% 0% 5% 10% 15% −5% 0% 5% 10% 15%
value

Figure 7: Comparative objective results averaged over VRP pickup experiments with mixed fleet with mixed
geography broken down (rows) by departure time with initial cheap/cheap cost parameters

5.1.4 Departure Times

Next, we examine the impact of different departure times. As you recall, we consider two departure times

that allow for travel at rush hour times (morning and afternoon rush hours) and two that allow for travel

primarily during non-rush hour times. We will look at the results for a fleet of three heavy vehicles with

customers from mixed geographies in Figure 7. The rush hour travel times are represented by the first and

third rows and the non-rush hour travel times are represented by the second and fourth rows. Looking at

average cost (the first column), we see that DIST performs slightly worse in terms of COST measures at

rush hour times. This makes sense since congestion becomes more of a factor in determining fuel and driver

costs during a rush hour, shifting the solution to slower paths. Similarly, the difference in duration from

the different routes (third column) is slightly more pronounced in rush hours where congestion can increase

route duration. Last, the fuel usage from using TIME optimized routes is highest at rush hour times, when

congestion can have a huge impact on fuel usage.

Next, we want to understand how the routes optimized for the COST objective change at rush hour

vs. non-rush hour times. It is not clear from Figure 7 if the solutions are the same but just have different

26
(a) rush hour route (b) non-rush hour route

Figure 8: Pickup COST routes for S5 at 15:30 (rush hour) and 19:30 (non-rush hour) with Homogeneous
Loads and Heavy Vehicle. Aerial view provided by Google Earth.

costs at different travel times or if the underlying solutions change. We look at an example of the pattern

we observed in Figure 8. We see that the ordering of the customers is exactly the same at rush hour (in

(a)) and non-rush hour times (in (b)), but the paths traveled between the customers change. In general, at

non-rush hour times, travel between customers is more likely to use highways to save time (which is why the

distance of the non-rush hour solution is longer at 102.72 km vs. 94.13 km for the rush hour solution), but

during rush hour times, the highways become congested. Alternate paths using more minor roads are used

at rush-hour times, but the total cost still goes up from 80.82 to 81.72 euros.

5.1.5 Load Sizes

We breakdown the mixed fleet results with mixed geography in Figure 9 to see the patterns in behavior

caused by different load schemes. For the first three columns, we do not see much difference in the relative

values of the different measures due to load scheme. The most noticeable difference is in the fuel usage (fourth

column). The least fuel usage for TIME routes occurs for the load scheme where the heavy customers are

those closest to the depot. This makes sense since a solution that picks up these heavy customers that cause

the most fuel usage last is more likely to be selected naturally when optimizing for TIME.

In looking at the detailed routes, particularly for the solutions with the heaviest customers furthest

from the depot, the heavy customers can be put in very different places on the route based on the selected

objective. To understand this, we provide the following single vehicle example in Figure 10 where heavy

27
avg_cost avg_distance avg_duration avg_fuel

TIME

homogeneous
FUELS

FUEL

DIST

TIME

farthest 50%
FUELS

FUEL

DIST
method

TIME

closest 50%
FUELS

FUEL

DIST

TIME

random 50%
FUELS

FUEL

DIST

−5% 0% 5% 10% 15% −5% 0% 5% 10% 15% −5% 0% 5% 10% 15% −5% 0% 5% 10% 15%
value

Figure 9: Comparative objective results averaged over VRP pickup experiments with mixed fleet with mixed
geography broken down (rows) by load sizes with initial cheap/cheap cost parameters

customers are indicated by boxes. To save on fuel, the FUEL solution (in (c)) puts the heavy customers near

the end of the route in positions 7, 8, and 9. The customers in positions 1 and 2 on this route are clearly

not visited in a way that is optimal in terms of distance (as compared with (a)) or travel time (as compared

with (b)). Due to the high driver cost implied by this ordering, the customers in position 1 and 2 on the

FUEL route are moved to position 9 and 10 in the COST solution (in (d)). Notably for this example, the

TIME and COST solutions are the same.

The differences in how the heavy customers are served become even more dramatic for mixed fleets. Of

particular interest is the assignment of the customers to the heavy vehicle given the different objectives. We

look at an example of how the heavy customers are served by the heavy vehicle in a mixed fleet in Figure 11.

Apparently, FUEL and COST routes only assign a small number of close and heavy customers to the heavy

vehicle and can hence keep fuel consumption of the heavy vehicle low (between 14.3 and 14.8 liters). TIME

and DIST routes ignore the particular load on the arcs between the customers and hence utilize the heavy

vehicle inefficiently from a fuel consumption as well as from a cost perspective. Hence, we can see that it is

very important to select the appropriate objective, especially when fleet composition and customer demand

28
(a) DIST route (b) TIME route

(c) FUEL route (d) COST route

Figure 10: Pickup routes for S4 at 15:30 with Farthest Three Heavy and Heavy Vehicle. Heavy customers
are denoted by squares. Aerial view provided by Google Earth.

29
(a) DIST route (b) TIME route

(c) FUEL route (d) COST route

Figure 11: Pickup routes for S5 at 15:30 with Closest 50% Heavy and the heavy vehicle from the mixed fleet.
Aerial view provided by Google Earth.

are both heterogeneous.

5.1.6 Cost Parameters

Last, we examine the impact of fuel and driver costs. We demonstrate the impact of different cost parameters

based on overall TSP pickup experiments in Figure 12 and VRP pickup experiments in Figure 13. For both

graphs, the rows indicate the cost parameters. Both graphs show that there is very little difference in the

relative measures from the different objectives with changes in the cost parameters. We would expect that

as fuel becomes more expensive (row 3), TIME routes would perform even worse in terms of fuel usage. This

is true but amounts to less than 1% increase from our initial parameter setting. These graphs again do a

30
avg_cost avg_distance avg_duration avg_fuel

TIME

cheap/cheap
FUELS

FUEL

DIST

TIME

cheap/expensive
FUELS
method

FUEL

DIST

TIME

expensive/cheap
FUELS

FUEL

DIST

−3% 0% 3% 6% −3% 0% 3% 6% −3% 0% 3% 6% −3% 0% 3% 6%


value

Figure 12: Comparative objective results averaged over all TSP pickup experiments broken down (rows) by
cost parameter combinations (cheap/cheap, cheap/expensive, etc.)

good job of demonstrating that the FUELS solution yields a similar total cost to the FUEL solution, even

though it does not do a great job of actually minimizing fuel usage.

Even though the relative performance measures remain mostly stable, the actual performance measures

of course change quite a bit with increases in fuel or driver costs. Interestingly, the routes themselves also

change quite a bit as well. To demonstrate one of the patterns to these changes, we look at Figure 14. Here,

we examine a single vehicle pickup COST route for different cost parameter combinations. In the upper left

(a), we have the baseline route (CC). In this instance, we have three heavy customers at the bottom of the

graph (indicated by boxes). The baseline route visits the heavy customers in positions 2, 3, and 4 because

it minimizes the combined costs. For this instance, the driver costs dominate the fuel costs, so the ordering

of customers is very similar when labor gets more expensive (in (b)). In the CE route, the main difference

is the longer distance path between (5) and (6) to decrease travel time but increase fuel usage. When gas

prices increase, fuel becomes more dominant in the objective so the solution prioritizes minimizing fuel. This

is reflected by changing the orientation of the route and putting the heavy customers at the end of the route

to use less fuel.

31
avg_cost avg_distance avg_duration avg_fuel

TIME

cheap/cheap
FUELS

FUEL

DIST

TIME

cheap/expensive
FUELS
method

FUEL

DIST

TIME

expensive/cheap
FUELS

FUEL

DIST

−5.0% −2.5% 0.0% 2.5% 5.0% 7.5%−5.0% −2.5% 0.0% 2.5% 5.0% 7.5%−5.0% −2.5% 0.0% 2.5% 5.0% 7.5%−5.0% −2.5% 0.0% 2.5% 5.0% 7.5%
value

Figure 13: Comparative objective results averaged over all VRP pickup experiments broken down (rows) by
cost parameter combinations (cheap/cheap, cheap/expensive, etc.)

5.1.7 Pickups vs. Deliveries

All of the results presented previously are for picking up loads at customers. Here, we examine how the

costs and routes differ for deliveries as compared with pickups. In Figure 15, we present the results for TSP

delivery experiments with large vehicles and mixed geography broken down by load sizes. In Figure 16, we

present the comparable results for TSP pickup experiments. Not surprisingly, the comparative results for

pickups and deliveries are quite similar. The noticeable differences are with the performance of the DIST

objective across the four measures with random loads. The DIST objective leads to solutions with higher

relative total costs, higher relative duration, and higher relative fuel costs. Looking into the detailed results,

it appears due to the locations of the heavy loads in many of the random instances. This indicates that

there is a lot of sensitivity in the various measures with random heavy loads.

When we look at the routes themselves, we find several differences in the appearance of the pickup and

delivery routes. Our prediction was that for FUEL and COST objectives, the pickup and delivery routes

would be reversals of each other to minimize the fuel consumption. In fact, this is rarely completely true

due to the time dependency of travel speeds and the impact of the time-dependent travel speeds on fuel

32
(a) CC route (b) CE route

(c) EC route

Figure 14: Pickup routes for M4 at 12:30 with Farthest Three Heavy and a heavy vehicle for different cost
parameters. Aerial view provided by Google Earth.

33
avg_cost avg_distance avg_duration avg_fuel

TIME

homogeneous
FUELS

FUEL

DIST

TIME

farthest three
FUELS

FUEL

DIST
method

TIME

closest three
FUELS

FUEL

DIST

TIME

FUELS

random
FUEL

DIST

0% 5% 10% 0% 5% 10% 0% 5% 10% 0% 5% 10%


value

Figure 15: Comparative objective results averaged over TSP delivery experiments with large vehicles with
mixed geography broken down (rows) by load sizes with initial cheap/cheap cost parameters

avg_cost avg_distance avg_duration avg_fuel

TIME

homogeneous
FUELS

FUEL

DIST

TIME

farthest three
FUELS

FUEL

DIST
method

TIME
closest three

FUELS

FUEL

DIST

TIME

FUELS
random

FUEL

DIST

−5% 0% 5% −5% 0% 5% −5% 0% 5% −5% 0% 5%


value

Figure 16: Comparative objective results averaged over TSP pickup experiments with large vehicles with
mixed geography broken down (rows) by load sizes with initial cheap/cheap cost parameters

34
(a) pickup (b) delivery

Figure 17: COST routes for M2 at 06:30 with Homogeneous Loads and Heavy Vehicle. Aerial view provided
by Google Earth.

usage. To demonstrate this, we look at Figure 17, which shows the best COST pickup route (in (a)) and the

best COST delivery route (in (b)). The pickup route travels counterclockwise across the top of the graph

(visiting 1-4) before crossing the graph to travel clockwise across the bottom of the graph (visiting 5-10).

The delivery route reverses the last six stops across the bottom of the graph but then travels through the

top of the graph in the same order as the pickup route. Interestingly, for this single example, the delivery

COST-based route is more efficient by means of all measures than the pickup route, which may be caused

by the slightly different orientation of the pickup route in combination with longer travel times caused by

congestion.

6 Conclusions

In this paper, we have examined the impact of minimizing combined fuel and driver costs to help understand

the impact on routes as compared with more traditional objectives as well as minimizing fuel exclusively.

Based on experiments that use a large database of real speed observations and the street network of the

urban area of Stuttgart, Germany, our results show that minimizing time and distance are poor substitutes

for minimizing route costs in urban areas. Minimizing distance performs particularly poorly with regard to

cost and relatively poorly with regard to fuel consumption and emissions. Likewise, minimizing total travel

time does a poor job of minimizing fuel usage and often chooses roads that often trade fuel for speed.

35
Interestingly, optimizing for total cost has the benefit of also lowering fuel consumption and thus emissions

relative to the traditional objectives of distance and duration. Further, we find that minimizing fuel does not

lead to significantly higher costs than considering a combined objective of fuel and driver costs. This finding

indicates that a focus on emissions should not substantially drive up operational costs. Simply, companies

wishing to minimize costs and particularly those facing pressure to reduce emissions for fleets in urban areas

need to seriously consider their objective choice when optimizing vehicle routes. The traditional objectives

of distance and time are not necessarily serving companies well in either regard.

Our results also indicate that the most substantial differences between the COST solutions and the

solutions of the other objectives occur in mixed and suburban geographies, much more so than for inner

city geographies. In addition, rush hour departures often lead to increased fuel consumption. We find that

explicitly modeling the impact of load on cost or fuel consumption can have a significant impact on where

heavy customers are placed on routes. Further, COST routes are quite sensitive to cost parameters. As fuel

costs increase, the routes start to resemble FUEL routes, and as driver costs increase, the routes come closer

to TIME routes. When we consider pickups versus deliveries, we found the costs for different objectives

to be quite similar, as expected, but the routes are often quite different. Importantly, while because of

time dependency the routes are not complete reversals of one another, substantial portions of the routes are

reversals.

One of the key insights from our work is that the advantage of the COST and FUEL objectives relative

to the DIST and TIME objectives results from the fact that the COST and FUEL objectives can account

for both the time dependence of speeds and the changing load of the vehicle as it traverses the a route. As a

result, the solutions to these two objectives follow different paths between customers depending on the order

of the customers and the time at which travel is taking place. These results suggest that researchers need to

be considering methods for integrating path finding into optimization algorithms for vehicle routing. While

the work in this paper, Ehmke et al. (2016b), and Huang et al. (2016) offer good starting points, there is an

opportunity for substantial future research in this area.

36
References

Barth, M. and Boriboonsomsin, K. (2008). Real-World Carbon Dioxide Impacts of Traffic Congestion.

Transportation Research Record: Journal of the Transportation Research Board, (2058):163–171.

Bektas, T., Demir, E., and Laporte, G. (2016). Green vehicle routing. In Psaraftis, H. N., editor, Green Trans-

portation Logistics: The Quest for Win-Win Solutions, volume 226 of International Series in Operations

Research & Management Science, pages 243–265. Springer International Publishing, Cham, Switzerland.

Bektas, T. and Laporte, G. (2011). The Pollution-Routing Problem. Transportation Research Part B:

Methodological, 45(8):1232–1250.

Carlos, J., Eguia, I., Racero, J., and Guerrero, F. (2014). Multi-objective vehicle routing problem with cost

and emission functions. Procedia - Social and Behavioral Sciences, 160:254–263.

Dabia, S., Demir, E., and Woensel, T. V. (2014). An Exact Approach for the Pollution-Routing Problem.

Beta Working Paper Series August, Eindhoven University of Technology.

Demir, E., Bektas, T., and Laporte, G. (2011). A comparative analysis of several vehicle emission models for

road freight transportation. Transportation Research Part D: Transport and Environment, 16(5):347–357.

Demir, E., Bektas, T., and Laporte, G. (2014a). A review of recent research on green road freight trans-

portation. European Journal of Operational Research, 237(3):775–793.

Demir, E., Bektas, T., and Laporte, G. (2014b). The bi-objective pollution-routing problem. European

Journal of Operational Research, 232(3):464–478.

Ehmke, J. F. (2012). Integration of Information and Optimization Models for Routing in City Logistics,

volume 177 of International Series in Operations Research & Management Science. Springer, New York.

Ehmke, J. F., Campbell, A. M., and Thomas, B. W. (2016a). Data-driven approaches for emissions-minimized

paths in urban areas. Computers and Operations Research, 67:34 – 47.

Ehmke, J. F., Campbell, A. M., and Thomas, B. W. (2016b). Vehicle routing to minimize time-dependent

emissions in urban areas. European Journal of Operational Research, 251(2):478–494.

37
Ehmke, J. F., Meisel, S., and Mattfeld, D. C. (2012). Floating car based travel times for city logistics.

Transportation Research Part C: Emerging Technologies, 21(1):338–352.

Figliozzi, M. (2010). Vehicle Routing Problem for Emissions Minimization. Transportation Research Record:

Journal of the Transportation Research Board, (2197).

Fleischmann, B., Gietz, M., and Gnutzmann, S. (2004). Time-varying travel times in vehicle routing.

Transportation Science, 38(2):160–172.

Franceschetti, A., Honhon, D., Van Woensel, T., Bektaş, T., and Laporte, G. (2013). The time-dependent

pollution-routing problem. Transportation Research Part B: Methodological, 56:265–293.

Fukasawa, R., He, Q., Santos, F., and Song, Y. (2016). A joint routing and speed optimization problem.

Available from http://arxiv.org/abs/1602.08508.

Fukasawa, R., He, Q., and Song, Y. (2015a). A branch-cut-and-price algorithm for the energy minimization

vehicle routing problem. To appear in Transportation Science.

Fukasawa, R., He, Q., and Song, Y. (2015b). A disjunctive convex programming approach to the pollution

routing problem. Available from http://www.optimization-online.org/DB{\_}FILE/2015/06/4978.

pdf, accessed on April 12, 2016.

Hardcastle, J. L. (2015). Walmart, General Mills, Anheuser-Busch improve freight efficiency, cut

emissions. Environmental Leader, available from http://www.environmentalleader.com/2015/

05/13/walmart-general-mills-anheuser-busch-improve-freight-efficiency-cut-emissions/

#ixzz473YFXy9e, accessed on June 28, 2016.

Hardcastle, J. L. (2016). How UPS, DHL drive emissions cuts, efficiency improvements in transportation

and logistics. Environmental Leader, available from http://www.environmentalleader.com/2015/

05/13/walmart-general-mills-anheuser-busch-improve-freight-efficiency-cut-emissions/

#ixzz473YFXy9e, accessed on June 28, 2016.

Huang, Y., Zhao, L., Van Woensel, T., and Gross, J.-P. (2016). The time-dependent vehicle routing problem

with path flexibility. Submitted for publication.

38
Hwang, T. and Ouyang, Y. (2015). Urban freight truck routing under stochastic congestion and emission

considerations. Sustainability, 7(6):6610–6625.

Jabali, O., Van Woensel, T., and de Kok, A. (2012). Analysis of Travel Times and CO2 Emissions in

Time-Dependent Vehicle Routing. Production and Operations Management, 21(6):1060–1074.

Kara, İ., Kara, B. Y., and Yetis, M. K. (2007). Energy minimizing vehicle routing problem. In Dress, A.,

Xu, Y., and Zhu, B., editors, Combinatorial Optimization and Applications, volume 4616 of Lecture Notes

in Computer Science, pages 62–71. Springer Berlin Heidelberg.

Koç, Ç., Bektaş, T., Jabali, O., and Laporte, G. (2014). The fleet size and mix pollution-routing problem.

Transportation Research Part B: Methodological, 70:239–254.

Koc, C., Bektas, T., Jabali, O., and Laporte, G. (2016). The Impact of Location , Fleet Composition and

Routing on Emissions in Urban Freight Distribution. Transportation Research Part B: Methodological,

84:81–102.

Kopfer, H. W., Schönberger, J., and Kopfer, H. (2014). Reducing greenhouse gas emissions of a heterogeneous

vehicle fleet. Flexible Services and Manufacturing Journal, 26(1-2):221–248.

Kramer, R., Maculan, N., Subramanian, A., and Vidal, T. (2015). A speed and departure time optimization

algorithm for the pollution-routing problem. European Journal of Operational Research, 247(3):782–787.

Kramer, R. and Subramanian, A. (2015). A matheuristic approach for the Pollution-Routing Problem.

European Journal of Operational Research, 243(2):523–539.

Kröger, M. (2013). Stau-Analyse: Stuttgart quält seine Pendler am meisten (Analysis of traffic jams:

Stuttgart tortures its commuters the most). Spiegel Online.

Kuo, Y. (2010). Using simulated annealing to minimize fuel consumption for the time-dependent vehicle

routing problem. Computers and Industrial Engineering, 59(1):157–165.

Kuo, Y. and Wang, C.-C. (2011). Optimizing the VRP by minimizing fuel consumption. Management of

Environmental Quality: An International Journal, 22(4):440–450.

39
Kwon, Y. J., Choi, Y. J., and Lee, D. H. (2013). Heterogeneous fixed fleet vehicle routing considering carbon

emission. Transportation Research Part D: Transport and Environment, 23:81–89.

Lin, C., Choy, K., Ho, G., Chung, S., and Lam, H. (2014). Survey of Green Vehicle Routing Problem: Past

and future trends. Expert Systems with Applications, 41(4):1118–1138.

Lorkowski, S., Mieth, P., Thiessenhusen, K.-U., Chauhan, D., Passfeld, B., and Schäfer, R.-P. (2004). To-

wards Area-Wide Monitoring-Applications derived from Probe Vehicle Data. In AATT 2004, pages 389–

394.

Maden, W., Eglese, R., and Black, D. (2010). Vehicle routing and scheduling with time-varying data: A

case study. Journal of the Operational Research Society, 61:515–522.

Oberscheider, M., Zazgornik, J., Henriksen, C. B., Gronalt, M., and Hirsch, P. (2013). Minimizing driving

times and greenhouse gas emissions in timber transport with a near-exact solution approach. Scandinavian

Journal of Forest Research, 28(5):493–506.

Qian, J. and Eglese, R. (2016). Fuel emission optimization in vehicle routing problems with time-varying

speeds. European Journal of Operational Research, 248(3):840 – 848.

Rao, W., Liu, F., and Wang, S. (2016). An Efficient Two-Objective Hybrid Local Search Algorithm for

Solving the Fuel Consumption Vehicle Routing Problem. Applied Computational Intelligence and Soft

Computing, 2016:1–16.

Sadegh-amalnick, N. N. M. (2016). Modified particle swarm optimization in a time-dependent vehicle routing

problem : minimizing fuel consumption. Optimization Letters.

Slavin, T. (2015). The european cities moving faster on clean energy than their governments.

The Guardian, available from https://www.theguardian.com/public-leaders-network/2015/jul/06/

european-cities-clean-energy-governments-eu-climate-targets, accessed on June 29, 2016.

Smith, T. W. P., Jalkanen, J. P., Anderson, B. A., Corbett, J. J., Faber, J., Hanayama, S., O’Keeffe, E.,

Parker, S., Johansson, L., Aldous, L., Raucci, C., Traut, M., Ettinger, S., Nelissen, D., Lee, D. S., Ng, S.,

40
Agrawal, A., Winebrake, J. J., Hoen, M., Chesworth, S., and Pandey, A. (2015). Third IMO Greenhouse

Gas Study 2014. International Maritime Organization, London.

Solomon, M. M. (1987). Algorithms for the vehicle routing and scheduling problems with time window

constraints. Operations Research, 35:254–265.

Toth, P. and Vigo, D., editors (2014). Vehicle Routing: Problems, Methods, and Applications. MOS-

SIAM Series on Optimization. Society for Industrial and Applied Mathematics and the Mathematical

Optimization Society, Philadelphia, second edition.

Urquhart, N., Hart, E., and Scott, C. (2010). Building low co2 solutions to the vehicle routing problem

with time windows using an evolutionary algorithm. In Evolutionary Computation (CEC), 2010 IEEE

Congress on, pages 1–6.

Wen, L., Çatay, B., and Eglese, R. (2014). Finding a minimum cost path between a pair of nodes in a time-

varying road network with a congestion charge. European Journal of Operational Research, 236(3):915–923.

Wen, L. and Eglese, R. (2015). Minimum cost VRP with time-dependent speed data and congestion charge.

Computers & Operations Research, 56:41–50.

Xiao, Y. and Konak, A. (2016). The heterogeneous green vehicle routing and scheduling problem with

time-varying traffic congestion. Transportation Research Part E: Logistics and Transportation Review,

88:146–166.

Xiao, Y., Zhao, Q., Kaku, I., and Xu, Y. (2012). Development of a fuel consumption optimization model for

the capacitated vehicle routing problem. Computers and Operations Research, 39(7):1419–1431.

Zachariadis, E. E., Tarantilis, C. D., and Kiranoudis, C. T. (2015). The load-dependent vehicle routing

problem and its pick-up and delivery extension. Transportation Research Part B: Methodological, 71:158–

181.

Zhang, J., Zhao, Y., Xue, W., and Li, J. (2015). Vehicle Routing Problem with Fuel Consumption and

Carbon Emission. International Journal of Production Economics, 170(Part A):234–242.

41

View publication stats

You might also like