Softree 3

Horizontal Alignment Optimization in
Road Design
by
Sukanto Mondal
B.Sc., Rajshahi University of Engineering & Technology, 2010
A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF

THE REQUIREMENTS FOR THE DEGREE OF
MASTER OF SCIENCE
in
THE COLLEGE OF GRADUATE STUDIES
(Interdisciplinary Studies - Optimization)
THE UNIVERSITY OF BRITISH COLUMBIA

(Okanagan)
July 2014

c Sukanto Mondal, 2014
Abstract
The horizontal alignment optimization problem in road design is a com-
plex problem. Usually, classic optimization techniques cannot be used to
address the problem. A few studies investigated the problem mainly us-
ing heuristics. Unfortunately, all of the previously studied heuristic based
methods do not guarantee optimality. In this study, we develop a novel opti-
mization model to solve the horizontal alignment optimization problem in a
specified corridor. The cost of a horizontal alignment is significantly affected
by the associated vertical alignment cost. So in order to formulate the cost
function of the model, we consider both the vertical alignment and earth-
work allocation associated with a horizontal alignment. The representation
of a horizontal alignment in our model satisfies all of the geometrical spec-
ifications used by engineers. Our model is suitable for both backtracking
and non-backtracking horizontal alignments. Derivative-free optimization
algorithms are used to solve the problem and guarantee the local optimal-
ity of our solution. The numerical experiment results of a set of practical
problems are reported.
ii
Table of Contents
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . viii
Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
Chapter 1: Introduction . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Road design optimization . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Earthwork optimization . . . . . . . . . . . . . . . . . 4
1.2.2 Vertical alignment optimization . . . . . . . . . . . . . 5
1.2.3 Horizontal alignment optimization . . . . . . . . . . . 5
1.3 Background and literature review . . . . . . . . . . . . . . . 7
1.3.1 Earthwork and vertical alignment optimization model 7
1.3.2 Horizontal alignment optimization model . . . . . . . 8
1.3.3 Three dimensional alignment optimization model . . . 10
1.4 Our research approach . . . . . . . . . . . . . . . . . . . . . . 11
1.5 Organization of the Thesis . . . . . . . . . . . . . . . . . . . . 11
Chapter 2: A basic approach to solve the horizontal align-

ment optimization problem . . . . . . . . . . . . . . 13
2.1 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Solution Approach . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . 18
iii
TABLE OF CONTENTS
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Chapter 3: Horizontal alignment optimization model . . . . . 20

3.1 Geometric representation of horizontal alignment . . . . . . . 20
3.2 Model description . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2.2 The optimization model . . . . . . . . . . . . . . . . . 27
3.3 Model Summary . . . . . . . . . . . . . . . . . . . . . . . . . 40
Chapter 4: Numerical results . . . . . . . . . . . . . . . . . . . 41

4.1 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2 The NOMAD and HOPSPACK solvers . . . . . . . . . . . . . 42
4.3 Results for the test problems . . . . . . . . . . . . . . . . . . 42
4.4 Summary of the result . . . . . . . . . . . . . . . . . . . . . . 49
Chapter 5: Conclusion . . . . . . . . . . . . . . . . . . . . . . . . 50
5.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.2 Recommendations for future research . . . . . . . . . . . . . . 51
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
Appendix A: Tables . . . . . . . . . . . . . . . . . . . . . . . . . . 60
A.1 Results for basic model . . . . . . . . . . . . . . . . . . . . . 60
A.2 Optimized alignments of the test problems . . . . . . . . . . . 62
Appendix B: Figures . . . . . . . . . . . . . . . . . . . . . . . . . . 69
iv
List of Tables
Table 1.1 Classification of the construction costs. . . . . . . . . . 4
Table 4.1 Specifications of the test problems . . . . . . . . . . . 41

Table 4.2 Cost improvement, no. of black-box evaluations and
wall-clock time required to solve the test problems us-
ing the NOMAD solver. . . . . . . . . . . . . . . . . . 43
Table 4.3 Comparison of optimum objective function values ob-
tained by the HOPSPACK and NOMAD solvers to
solve the test problems. . . . . . . . . . . . . . . . . . 46
Table 4.4 Overall comparison of the HOPSPACK solver and the
NOMAD solver with the optimum objective function
values. . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Table 4.5 Comparison of the no. of black-box evaluations re-
quired for the HOPSPACK and NOMAD solvers to
solve the test problems. . . . . . . . . . . . . . . . . . 48
Table A.1 Computational Experience (Hart Rd Small.csv) . . . . 60

Table A.2 Computational Experience (Diamond Road align-1.csv) 61
Table A.3 Computational Experience (Diamond Road align-2.csv) 61
Table A.4 Computational Experience (bluff road.csv) . . . . . . . 62
Table A.5 Computational Experience (spur 3 demo.csv) . . . . . 62
Table A.6 The optimal alignments of the Road A obtained by
the NOMAD and HOPSPACK solvers. . . . . . . . . . 63
Table A.7 The optimal alignments of the Road B obtained by the
NOMAD and HOPSPACK solvers. . . . . . . . . . . . 64
Table A.8 The optimal alignments of the Road C obtained by the
Table A.9 The optimal alignments of the Road D obtained by
the NOMAD and HOPSPACK solvers. . . . . . . . . . 66
Table A.10 The optimal alignments of the Road E obtained by the
v
List of Figures
Figure 1.1 A three dimensional alignment (blue curve) showing

its projection onto the XY -plane. The projected red
curve is the horizontal alignment. . . . . . . . . . . . 3
Figure 1.2 Example of cut and fill in road construction. . . . . . 5
Figure 1.3 Two potential vertical alignments of a fixed horizontal
alignment. . . . . . . . . . . . . . . . . . . . . . . . . 6
Figure 1.4 Some potential horizontal alignments in a specified
corridor. . . . . . . . . . . . . . . . . . . . . . . . . . 6
Figure 2.1 Corridor of a horizontal alignment. Base data points

are in red, offset data points are in black. Points cir-
cled in green constitute an example of a station. The
orange and purple curves are two potential horizontal
alignments in the corridor. . . . . . . . . . . . . . . . 14
Figure 2.2 Three dimensional alignment. . . . . . . . . . . . . . 15
Figure 2.3 Blackbox optimization in horizontal alignment opti-
mization problem. . . . . . . . . . . . . . . . . . . . . 16
Figure 2.4 Flowchart of the solution approach. . . . . . . . . . . 17
Figure 2.5 Required time with respect to number of stations for
solved problem only. . . . . . . . . . . . . . . . . . . . 18
Figure 2.6 Growth of no. of function calls with respect to num-
ber of stations. . . . . . . . . . . . . . . . . . . . . . . 19
Figure 3.1 Geometric representation of horizontal alignment. The

alignment displayed is represented as ((S, 0), (P1 , r1 ),
(P2 , r2 ), (P2 , r2 ), (E, 0)). . . . . . . . . . . . . . . . . 21
Figure 3.2 Geometric specifications of a circular curve. . . . . . . 21
Figure 3.3 Road segment representation in a specified corridor. . 26
vi
LIST OF FIGURES
Figure 3.4 Horizontal alignment in a specified corridor. The red

curve is an example horizontal alignment. The blue
stars are the intersection points (IP) and the dotted
blue rectangular boxes are feasible region (for mov-
ing) associated with the intersection points. . . . . . . 26
Figure 3.5 Piece-wise linear representation of horizontal align-
ment segment considering all of the cross-section lines. 28
Figure 3.6 An example road of four segments showing the asso-
ciated variables of the optimization model. The green
cross-section lines separate the road segments. The
starting point and the end point of the alignment are
(px1,1 , py1,1 ) and (px4,3 , py4,3 ), respectively, which are
fixed. . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Figure 3.7 A segment of a road showing an associated horizontal
alignment. . . . . . . . . . . . . . . . . . . . . . . . . 31
Figure 3.8 An example of the quadrants issue in generation of a
circular arc. . . . . . . . . . . . . . . . . . . . . . . . 34
Figure 3.9 Discontinuity in a horizontal alignment. . . . . . . . . 38
Figure 4.1 A non-backtracking alignment (the test problem as-

sociated with Road A) showing the initial alignment
and the optimized alignments obtained by the NO-
MAD and HOPSPACK solvers. . . . . . . . . . . . . 44
Figure 4.2 A backtracking alignment (the test problem associ-
ated with Road D) showing the initial alignment and
the optimized alignments obtained by the NOMAD
and HOPSPACK solvers. . . . . . . . . . . . . . . . . 45
Figure B.1 Optimum alignments of the Road B obtained by the

NOMAD solver and the HOPSPACK solver. . . . . . 69
Figure B.2 Optimum alignments of the Road C obtained by the
Figure B.3 Optimum alignments of the Road E obtained by the
vii
Acknowledgements
First and foremost, I would like to thank my supervisors, Dr. Yves
Lucet and Dr. Solomon Tesfamariam, for their continuous help and en-
couragement. This thesis would not have become a reality without their
guidance and advice. More importantly, thank you to my supervisors for
allowing me to knock on their door at any time.
I would like to thank Dr. Warren Hare, who actively advised me during
my research. His insightful comments on my research work helped me to
develop better ideas.
A big thank you goes out to our industrial partner, Softree Technical Sys-
tems Inc., for providing engineering details and practical datasets. It must
be mentioned, in the weekly meetings, we had many productive discussions
with David Mills, Craig Speirs, and Alexis Guigue on different engineering
aspects of the problem. More specifically, during the development of the
optimization model, their valuable comments helped me a lot to build a
better model.
I would also like to thank my colleagues from different parts of the world,
affiliated with the Center for Optimization and Convex Analysis and Non-
smooth Analysis, for our off topics discussions to know their diverse native
cultures and lifestyles. My dear colleagues, I really enjoyed your compan-
ionship.
This research is partially funded by a Collaborative Research and De-
velopment (CRD) Grants from the Natural Sciences and Engineering Re-
search Council (NSERC) sponsored by Softree Technical Systems Inc. The
research was performed in the Computer-Aided Convex Analysis (CA2 ) lab-
oratory funded by a Leaders Opportunity Fund (LOF) from the Canadian
Foundation for Innovation (CFI) and by a British Columbia Knowledge De-
velopment Fund (BCKDF).
viii
Dedication
Dedicated to underprivileged talents of the world.
ix
Chapter 1
Introduction
In this chapter, we briefly describe the background of the road design
optimization research and our motivation to solve the horizontal alignment
optimization problem.
1.1 Motivation
Since the early days of human civilization, the transportation system is
considered an integral part of sustainable socioeconomic development. The
gradual development of human civilization has led us to invent different
modes of transportation, such as land transportation, sea transportation,
and air transportation. The invention of the wheel revolutionized the land
transportation system and accelerated the economic development manyfold.
As of today, the transportation system is continuously contributing to our
economy significantly. For example, the transportation sector of Canada
contributed about 4.2% of Canada’s gross domestic product (GDP) in 2005
[DAAMP06]. In particular, more than one third (about 35%) of the GDP
generated by the transportation sector in 2005 came from the truck trans-
portation industry [DAAMP06]. The truck transportation system uses a
total of 1,042,300 km of roads in Canada.
The alignment of a road is the route connecting two given end-points.
An alignment consists of the vertical and horizontal alignments. Intuitively,
a good alignment is one which minimizes the construction costs satisfying
the design constraints. In the traditional road design process, engineers use
their professional judgment to determine several selected candidate align-
ments and then manually try to find the best one. In fact, a large number of
alternative alignments exist that should be considered in the design process.
In the conventional design process, finding the best alignment requires repet-
itive manual iterations. So it is almost impossible for engineers to consider
all of the possible alternative alignments. Hence, engineers cannot ensure
that the chosen alignment is (even locally) optimal.
In order to overcome the difficulties in the road design process, it is im-
perative to develop a computer-aided process to find the optimal alignment.
1
1.2. Road design optimization
In the literature, many studies proposed computerized models to find the

optimal horizontal alignment. Unfortunately, all of the models have some
drawbacks that prevent them to be used effectively in practice (for further
discussion see Section 1.3.2). In our research, we solve the horizontal align-
ment optimization problem and provide a practical model that guarantees
(local) optimality. Our approach is validated by computing several real-
world alignments.
1.2 Road design optimization

Road design optimization is the problem of finding a curve connect-
ing two given end-points that minimizes the cost while satisfying all of the
desired design specifications. The problem is usually divided into three in-
terrelated sub-problems, namely, horizontal alignment optimization, vertical
alignment optimization, and earthwork optimization.
In the three dimensional space, a curve can be projected on the horizontal
plane (i.e., XY -plane). In road design literature, the curve is called the
alignment, and the horizontal projection is called the horizontal alignment.
In Figure 1.1, for a three dimensional alignment, the associated horizontal
alignment is depicted.
Along the horizontal alignment shown in Figure 1.1, we can measure the
distance from the starting point to the end point. Corresponding to any
measured distance (using the x-values and the y-values of a three dimen-
sional alignment), we have a z-value (ground elevation). If we draw the
z-values with respect to the distance from the starting point in two dimen-
sional space, then the resulting curve is called the vertical alignment.
Finding the optimal alignment connecting the two end-points is hard
because of the complex cost structure associated with an alignment and the
requirement for satisfying design constraints [JSJ06]. Moreover, the con-
tinuous search space of the problem gives an infinite number of alternative
alignments. Since the terrain might have an irregular surface, a small change
in an alignment may result in a significant change in the total cost.
In order to find a good alignment, an engineer considers five major costs
[JSJ06] as follows
– planning and administrative cost,
– construction cost,
– maintenance cost,
2
Figure 1.1: A three dimensional alignment (blue curve) showing its projec-
tion onto the XY -plane. The projected red curve is the horizontal alignment.
– user cost, and
– social and environment cost.
Planning and administrative costs are not considered in the alignment

optimization because these costs are insensitive to alignment alternatives
[JSJ06]. In the study of Chew et al. [CGF89], construction costs are clas-
sified into six categories. Table 1.1 lists the cost components of the con-
struction costs and the associate approximate contributions toward the total
construction cost.
The percentage of each cost component in the total construction cost is
not fixed [JSJ06]. Depending on the road location, it could be significantly
different. For example, the land acquisition cost might be higher in urban
areas, whereas in mountainous areas, the earthwork cost is substantially
higher than other costs.
Maintenance costs have many classifications (at least eight) such as road-
way surface, bridges, tunnels, roadside features, drainage, shoulders and ap-
proaches maintenance, snow and ice control, and traffic control devices. The
net maintenance cost over 30 years is about 5% of the total construction cost
[OEC73].
User costs (vehicle operating costs) consist of the cost of vehicle main-
tenance, the value of travel time and the cost of traffic accidents. The net
user cost over 30 years varies approximately from 300% to 1000% of the
total construction cost [OEC73].
3
Table 1.1: Classification of the construction costs.
Cost components Contributions ( %)

Land 5%
Miscellaneous items 10 %
Drainage 10 %
Bridges 20 %
Earthwork 25%
Pavement 30%
Social and environment costs are the negative impacts of the road on the
environmental and social features of a particular region. In some extreme
cases, environmental and social issues might be very critical and even dom-
inates other cost. In practice, social and environment costs are very hard to
quantify. Typically, social and environment costs are carefully considered in
the planning stage when the preliminary corridor is selected.
However, among the five major cost components of the total cost, in
a selected corridor, constructions costs form an important component of
the total cost function. In particular, construction costs (excluding land
acquisition costs), mainly consist of excavation costs, embankment costs, and
hauling costs for the construction materials. In our research, we minimize
the total excavation costs, embankment costs, and hauling costs satisfying all
design constraints for the road. Note that the design constraints contribute
to the user costs and the maintenance costs [BDE10]. For instance, a long
and gentle vertical alignment provides a great sight distance which minimizes
road accidents.
In a traditional engineering approach, finding a good alignment is a
repetitive and complex process. It involves a series of phases, starting from
feasibility studies followed by planning, then narrowing down to the selection
of several possible corridors, and finally focusing on the details of an align-
ment including earthwork minimization, and horizontal and vertical design
constraints.
1.2.1 Earthwork optimization

Earthwork optimization is the problem of minimizing the total excava-
tion cost, embankment cost, and hauling cost for a fixed alignment. In road
construction, earthwork is the major task which involves excavation, em-
bankment and hauling of large quantities of earth materials. The cut and
4
Cut area Fill area
Ground profile Road profile
Figure 1.2: Example of cut and fill in road construction.
fill areas are determined by the intersection of the ground profile and the
road profile (see Figure 1.2). If the ground profile is below the road profile,
then more material is needed to fill. On the other hand, if the ground pro-
file is above the road profile, then additional material has to be cut. The
excavation and embankment costs are incurred in the cut and fill areas re-
spectively. The hauling cost is the cost associated with moving the material
from the cut areas to fill areas. The earthwork allocation cost is defined as
the combination of the excavation, embankment, and hauling costs.
1.2.2 Vertical alignment optimization

A large number of vertical alignments can be built for a fixed horizontal
alignment (see Figure 1.3). For each vertical alignment, the earthwork al-
location cost (which is the minimum cost corresponding to that fixed align-
ment) can be calculated by solving the earthwork optimization problem.
Thus the vertical alignment optimization problem is to find the vertical
alignment which has the minimum earthwork allocation cost satisfying the
vertical alignment design constraints for a fixed horizontal alignment. The
major design criteria for a vertical alignment are the allowable grades and the
rate of curvature [AAS04]. The maximum and minimum allowable grades
depend on the design speed (i.e., maximum speed limit) and the traffic
composition. The rate of curvature is the length of the vertical curve per
percent algebraic difference between the grade at the two end-points of the
curve [AAS04].
1.2.3 Horizontal alignment optimization

In a specified corridor, a large number of horizontal alignments can be
built (see Figure 1.4). As we described earlier, in the vertical alignment
5
Ground elevation
Vertical alignment
Elevation
XY plane
Figure 1.3: Two potential vertical alignments of a fixed horizontal alignment.
optimization process the horizontal alignment is fixed. Thus, for each fea-
sible horizontal alignment, the vertical alignment optimization problem can
be solved to get the optimal vertical alignment. The vertical alignment
optimization yields the minimum earthwork allocation cost satisfying the
vertical alignment design constraints. From the previous two assertions,
immediately it follows that each horizontal alignment has a cost (in partic-
ular, the minimum earthwork allocation cost) which is computed by solving
the corresponding vertical alignment optimization problem. Therefore, the
horizontal alignment optimization problem can be defined as the problem
of finding the horizontal alignment which has the minimum earthwork al-
location cost satisfying the horizontal alignment design constraints and the
associated vertical alignment design constraints.
Corridor
Horizontal alignments
Y
Figure 1.4: Some potential horizontal alignments in a specified corridor.
A horizontal alignment consists of some tangential line segments followed

by some circular arcs. The most important design constraint of the hori-
zontal alignment is the minimum radius of curvature of the circular arcs for
6
1.3. Background and literature review
the safety requirement.
1.3 Background and literature review

Computer-aided road design optimization started during the 1960’s and
1970’s. Due to the limited computational power, it was hard to solve op-
timization problems precisely. The advent of the modern computer and
GIS technology has helped us to solve the problem accurately. The road
construction cost can be minimized substantially by solving the mathemat-
ical models using modern computer programs. Many optimization models
have been developed to address the problem from different perspectives.
Although existing models work well in certain aspects, they still have some
drawbacks that prevent them to be used in practice.
1.3.1 Earthwork and vertical alignment optimization model

Many different approaches have been considered to model the vertical
alignment optimization problem. Numerical search [Hay70, Pea73, Rob73,
GCF88], and dynamic programming [Hua73, Mur73, Fwa89, GLA05] are
notably used to solve the problem.
Numerical search models are the earliest techniques to deal with the ver-
tical alignment optimization problem. Usually, a resulting numerical search
model of the vertical optimization problem becomes a non-linear non-convex
optimization problem which is hard to solve [JSJ06, page 28, section 2.5.4].
The dynamic programming approach yields a piece-wise linear alignment
rather than a smooth alignment. Rahman [Rah12, page 9, section 1.2.3]
noted that the dynamic programming approach is not suitable to solve the
vertical alignment optimization problem.
In traditional engineering, a mass haul diagram is used to minimize the
earthwork allocation. The mass haul diagram cannot be used in practical
situations [MS81]. The mass diagram has mainly the following limitations, it
cannot handle nonlinear hauling costs and the costs associated with borrow
and waste pits. To overcome the previous limitations, a linear programming
model was developed by Mayer and Stark [MS81]. Later, the earthwork
allocation and the vertical road profile are modelled together by the Trans-
portation and Road Research Laboratory in the United Kingdom and the
model was further modified by the Ontario Ministry of Transportation and
Communications. This model is the first attempt to integrate the earthwork
and vertical alignment optimization together. Unfortunately, the model does
not guarantee optimality.
7
Easa [Eas88] developed a linear programming model considering the

earthwork allocation and the vertical alignment. His approach works in
three steps: parameterizing all technically feasible vertical alignments, cal-
culating the cut and fill requirements for each vertical alignment, and then
using linear programming to optimize the earthwork allocation. The model
gives the global optimum through an exhaustive enumeration of possible
vertical profile.
Moreb [Mor96] developed the earthwork allocation and vertical road pro-
file in a single linear programming model. Both Easa’s model and Moreb’s
model output a piecewise linear vertical curve. But a smooth alignment is
desired. In order to remove the sharp connectivity of the piecewise linear
vertical alignment, the optimum result might be tempered. Engineers use
quadratic spline for the vertical alignment. Moreb and Aljohani [MA04]
modified the model developed in [Mor96] considering the vertical alignment
as a quadratic spline. Moreb [Mor09] further improved the previous model
by adding some additional constraints to ensure the smoothness with any
degree of polynomial spline. However, Koch and Lucet [KL10] proved that
the linearity of the model can only be maintained up to quadratic splines.
Recently, Hare, Koch, and Lucet [HKL11] developed a mixed integer lin-
ear programming model for the earthwork optimization considering blocks.
Rahman [Rah12] extended several models: [Mor09], [KL10], [HKL11] (un-
published manuscript [HHLM11] is also mentioned in [Rah12] but I could
not access it) to formulate the vertical alignment optimization problem as a
mixed integer linear program model. Hare et al. [HHLR14] further improved
the model to reduce the solution time. In our research, we use the model
developed in [HHLR14] as the objective function to solve the horizontal
alignment optimization problem.
1.3.2 Horizontal alignment optimization model

The horizontal alignment optimization problem is more complicated than
the vertical alignment optimization problem [JSJ06]. The main reasons are
that the horizontal alignment requires more data, and the cost of the hori-
zontal alignment is dependent on the vertical alignment cost, political, so-
cioeconomic, and environmental issues. In the literature, mainly three basic
approaches have been studied: calculus of variation, network optimization,
and dynamic programming.
Calculus of variation tries to find a curve connecting two end points in
space which minimizes the integral of a function [Wan95]. The nature of the
alignment optimization problem allows us to use the concept of calculus of
8
variation to find the optimal alignment. Howard et al. [HBS68] used the idea
of calculus of variation to develop the Optimum Curvature Principle (OCP),
which specifies the optimal vertical and horizontal curvatures at any point.
In order to apply the OCP, Shaw and Howard [SH81] proposed two numerical
integration methods, namely, the arc of circle algorithm and the intrinsic
equation procedure. The OCP was applied by Shaw and Howard to find the
optimal alignment of an expressway in South Florida [SH82]. Two major
requirements to use the OCP are the followings: first, the cost function has
to be continuous and second, the cost function has to be twice continuously
differentiable. In practice, the cost function might not be continuous [JSJ06,
page 8, section 2.4.1]. Although the OCP guarantees global optimality, it
requires some assumptions that make it impractical.
Another well-known approach to model the horizontal alignment opti-
mization problem is network optimization. In the network optimization
approach, a network is designed to represent a region through which a road
could pass. The region is divided into small cells to make a grid. Each cell in
the grid represents a node of the network. The nodes are connected through
the arcs. Each arc in the network is assigned a weight considering the cost
associated with the two connecting cells. An alignment is defined as a set
of connecting arcs from the starting node to the ending node.
In the early 1970’s, the idea of network optimization was used by Turner
and Miles [TM71] to model the route selection problem. This model [TM71]
considered the square grid to define the network. Considering all of the cost
factors, for each cell in the grid, a smooth surface was constructed. They
[TM71] developed the Generalized Computer Aid Route Selection (GCARS)
system to generate a set of ranked alignments. Turner [Tur78] further im-
proved the GCARS system by incorporating the environmental impacts as
a cost factor.
Athanassoulis and Calogero [AC73] formulated the route selection prob-
lem as a modified transportation problem. Note that both Turner’s and
Athanassoulis’s models did not consider the vertical profile. In practice, it
is highly expected to incorporate the vertical alignment cost in the horizontal
alignment optimization process.
Parker [Par77] and Trietsch [Tri87b, Tri87a] developed the two stages ap-
proach considering the vertical profile and used network optimization to find
the optimal alignment. While Parker studied only the square search grid,
Trietsch studied four different types of search grids: rectangular, square,
ellipse, and honeycomb. However, the resulting horizontal alignments of
the models ([Par77, Tri87b, Tri87a]) are piece-wise linear curves which are
unrealistic.
9
A basic shortcoming of the network optimization models is that the

optimal alignment is a piecewise linear trajectory. In practice, a nonsmooth
alignment is to be avoided for safety reasons. Moreover, in order to get a
more precise alignment, more nodes are needed which increases the problem
size and eventually the problem becomes hard to solve in a reasonable time.
A few studies [OEC73, Hog73, NEW76] applied the dynamic program-
ming approach to optimize the alignment. Similar to the network opti-
mization approach, the dynamic programming approach yields a nonsmooth
alignment. Moreover, in the dynamic programming approach, the backward
bending feature of the roads (backtracking road) introduces difficulties to
handle the alignment [Nic73, page 123, Chapter 5], [JSJ06, page 21, Section
2.4.3], [Par77].
Jong et al. [Jon98, JJS00] developed a horizontal alignment optimization
model which was solved by a genetic algorithm. The horizontal alignments
represented in [Jon98, JJS00] are very different from a practical alignment
[LTL09].
Lee et al. [LTL09] presented a heuristic based method to optimize the
horizontal alignment that works in two stages. In the first stage, the heuris-
tic tries to approximate a piecewise linear alignment and then in the second
stage, it refines the solution to make the previously generated piecewise
linear alignment compatible to a real road alignment. The solution align-
ment of the model ([LTL09]) yields a practical alignment. Since a heuristic
algorithm was used to solve the model, optimality is not guaranteed.
1.3.3 Three dimensional alignment optimization model

In three dimensional alignment optimization, the vertical and horizon-
tal alignments are optimized simultaneously. Modeling three dimensional
problem is a complex problem [JSJ06] and most of the studies use heuristic
based algorithms.
Chew et al. [CGF89] developed a model to solve the three dimensional
alignment using the concept of optimal control theory. Chew’s model is the
first model that yields a smooth three dimensional alignment. The objective
function of Chew’s model involves integrals which are hard to compute.
Tat and Tao [TT03] proposed a three dimensional alignment optimiza-
tion model and used genetic algorithm to solve it. This model [TT03] consid-
ers all of the major constraints of the road design. Akay [Aka06] developed
a model for three dimensional alignment optimization for forest roads. A
simulated annealing algorithm was used to solve the model. A tabu search
method was presented by [Aru05] for optimizing the three dimensional align-
10
1.4. Our research approach
ment of forest roads.

A criteria-based decision support system for three dimensional alignment
optimization was developed by Jha [Jha03] considering the environmental
costs. Jong et al. [JS03] presented an evolutionary model for optimizing the
vertical and horizontal alignment simultaneously. The previous two models
[Jha03, JS03] were further improved in [JK06] by considering accessibility,
proximity, and land-use changes in the road alignment planning process.
Recently, Cheng and Lee [CL06] also proposed a heuristic based model
for three dimensional alignment optimization. The heuristic solves the mod-
els in three steps: first, it generates a good general horizontal alignment by
adding, deleting, or moving the intersection points one by one, then it deter-
mines an improved horizontal alignment by adjusting the intersection points
based on the previously generated horizontal alignment, and finally, it finds
a better three dimensional alignment by tuning the vertical alignment cor-
responding to the previously obtained horizontal alignment.
All of the above mentioned three dimensional alignment optimization
models excluding the model in [CGF89] use a heuristic based algorithm
which does not guarantee optimality. Unfortunately, the heuristic based
algorithms do not ensure any mathematical proof of convergence.
1.4 Our research approach

In this research, we formulate the horizontal alignment optimization
problem as a bi-level optimization problem. Since the horizontal alignment
optimization problem is interrelated to the vertical alignment optimization
problem, in the inner level, the optimization model solves a vertical align-
ment optimization problem corresponding to a given horizontal alignment
(which comes from the outer level). The outer level of the problem opti-
mizes the horizontal alignment. An alignment in our model considers all
of the geometric specifications used by engineers. We used two derivative
free optimization algorithms to solve the problem. Our approach requires
an initial alignment to start a derivative free optimization algorithm. The
resulting solution alignment of our model is locally optimum.
1.5 Organization of the Thesis

The rest of the thesis is organized as follows. In Chapter 2, we describe
a basic formulation of the horizontal alignment optimization problem. We
also discuss the solution approach to solve the problem. Some numerical
11
1.5. Organization of the Thesis
experiment results are reported for the basic model. It is shown that, for a
road of reasonable length, the basic model cannot be solved in a reasonable
time.
In Chapter 3, we develop a model using the concept of the basic model
developed in Chapter 2. This model considers all of the geometric spec-
ifications used by engineers in practice. The resulting model is a bi-level
optimization problem, in which the vertical alignment optimization prob-
lem is considered as an inner problem.
In Chapter 4, we report the numerical results obtained by solving the
model developed in Chapter 3. The model was solved by two derivative-free
optimization solvers. The performance of the two solvers are reported. Fi-
nally, in Chapter 5 we summarize the contribution of the thesis and highlight
some future works.
12
Chapter 2
A basic approach to solve

the horizontal alignment
optimization problem
In this chapter, we describe a basic optimization problem formulation
for the horizontal alignment optimization. We report numerical experiments
that lead to the improved model described in Chapter 3.
2.1 Terminology
Horizontal alignment optimization consists of finding an optimal curve
within a designated corridor. Each corridor has a baseline, which defines the
horizontal and vertical alignment. There are two given end-points within a
specified corridor. The curve connecting the two end-points is the baseline
of the corridor. The ground profile data is given for some discrete points
within the corridor, which are called data points.
There are two types of data points, namely, base data points and offset
data points. The base data points are the points along the baseline, i.e., the
engineer’s original horizontal alignment. The offset data points represent the
horizontal displacement from the base data points. The base data points are
selected a few units apart between the two end-points along the baseline.
Each of the base data points has some associated offset data points in both
the left and the right directions. The red points and the black points in
Figure 2.1 are the base data points and offset data points, respectively.
The baseline of a corridor is a piece-wise linear curve connecting the
base data points. A base data point together with the associated offset data
points is defined as a station.
13
2.2. Problem Formulation
-2 -1 0 1 2
Base - offset Base Base + offset
Figure 2.1: Corridor of a horizontal alignment. Base data points are in red,
offset data points are in black. Points circled in green constitute an example
of a station. The orange and purple curves are two potential 1 horizontal
alignments in the corridor.
2.2 Problem Formulation

Each data point within the corridor, either a base data point or an off-
set data point, has some associated ground profile data. Therefore, for the
vertical road profile, we can move vertically up and down for each hori-
zontal offset data point. Altogether a three dimensional alignment data
is visualized as shown in Figure 2.2. Thus the horizontal and the vertical
displacements from the baseline make a discrete grid for each station (see
Figure 2.2). The horizontal displacement allows to move along the x and y
axes of the grid points and the vertical displacement allows to move along
the z axis. Our goal is to find, for each station, a horizontal offset that
generates a horizontal alignment and a vertical alignment which is (locally)
optimal.
Since every point in the designated corridor has the ground profile data,
we can make an alignment by taking a point from each station and optimiz-
ing it as a vertical alignment optimization problem by fixing a horizontal
alignment. For instance, the orange curve and the purple curve in Figure 2.1
could make a horizontal alignment and by fixing that particular alignment
we get a vertical alignment optimization problem. Each station has some
horizontal offset values associated with the offset data points. Note that
at the base data point of each station the offset value is zero and at the
left and right sides of base data points the offset is positive and negative,
respectively. To formulate the horizontal alignment optimization problem,
we can now consider the vertical alignment and the horizontal alignment
14
2.2. Problem Formulation
Station i+2
-2 Station i+1
-1
Vertical offset
0 Data point
1
Station i
Figure 2.2: Three dimensional alignment.
together in a way that optimizes the vertical alignment cost by varying the
horizontal offset value.
Let S = {1, 2, 3, . . . , n} be the index set for the stations. The decision
variable xi is the horizontal offset value at station i ∈ S. The lower bound
and the upper bound of the horizontal offset value of station i ∈ S are
defined as li and ui , respectively. In vector form, the horizontal offset X,
the lower bound L, and the upper bound U can be written as
X = (x1 , x2 , x3 , · · · , xn )| ,
L = (l1 , l2 , l3 , · · · , ln )| , and (2.1)
|
U = (u1 , u2 , u3 , · · · , un ) .
The dimension of the vector X is the number of stations. Now the problem
can be written mathematically as follows:
min f (X) = CVA (X)

s.t
(2.2)
L ≤ X ≤ U,
X ∈ Rn .
where CVA (X) is a function that returns the cost of an optimal vertical
15
2.3. Solution Approach
alignment for a horizontal offset X. So the objective function of the pro-

posed formulation is a vertical alignment optimization problem that can be
calculated by the method proposed in [HHLR14].
2.3 Solution Approach

The objective function of the problem is an optimization problem itself
that is a large scale mixed integer program. So it is very hard to access the
derivative information of the objective function (if it exists). We applied a
derivative-free optimization (DFO) approach to deal with the problem. In
DFO approach, we can put the objective function and the bound constraints
in a blackbox (see Figure 2.3) and optimize the problem without knowing
much information on the objective function. Only a vector X is given as an
input in the blackbox and an output value is obtained without knowing how
the output is computed. To solve our problem we used the NOMAD [LD11]
(see http://www.gerad.ca/nomad) solver that implements a mesh adaptive
direct search (MADS) algorithm developed in [AD06]. The NOMAD solver
is an open source solver that is proven to be competitive in solving DFO
Solution Approach
problems [RS13]. DFO We can integrate it as a static library. The flowchart of
Technique
the solving technique is illustrated in Figure 2.4.
NOMAD Solver
HAOptimization
Problem
HA Obj =
X VAOptimization value
HA Constraints
Black Box
1
Figure 2.3: Blackbox optimization in horizontal alignment optimization
problem.
The algorithm takes a three dimensional alignment data as the input

data. The VAOptimization block makes the blackbox for the MADS algo-
rithm. The MADS algorithm is an iterative algorithm that continues until
the optimal solution (i.e. the mesh is small enough) is found. Every iter-
ation of the MADS algorithm needs to evaluate the blackbox. So we are
solving a vertical alignment optimization problem at each iteration. There-
fore the optimal solution of the defined problem in Equation (2.2) gives us
the optimal vertical and the horizontal alignment.
16
2.3. Solution Approach
Start
Input Data
Extract VA data
VAOptimization
YES HA NO
Optimal ?
NO
Exceed terminating
Optimal VA & HA parameters limit
YES
Optimal VA for a HA
End
Figure 2.4: Flowchart of the solution approach.
The NOMAD solver has some terminating parameters to stop the MADS
algorithm. For instance, we can set the maximum number of iterations for
the MADS algorithm. So if the limit of the terminating parameters is ex-
ceeded, the algorithm terminates and eventually ends up with a non-optimal
horizontal alignment but the optimal vertical alignment is guaranteed to be
(globally) optimal for the given horizontal alignment.
The horizontal offsets given by the input data are discrete values but
our formulation has continuous offset variables. The lowest and highest
horizontal offsets are the lower and higher bounds, respectively. To get the
ground profile data for any offset within the bound, linear interpolation is
used. Let ya and yb be the ground profile data for the horizontal offset xa
and xb , respectively. For any offset x within xa and xb the ground profile
data is interpolated using the following equation:
x − xa
y = ya + (yb − ya ) . (2.3)
xb − xa
17
2.4. Numerical Results
2.4 Numerical Results

The experiments were setup with 50 test problems with a different num-
ber of stations. We used the derivative free optimization solver NOMAD
(version 3.5, available in http://www.gerad.ca/nomad) to solve the opti-
mization problems. All of the numerical experiments were performed on a
Dell workstation with an Intel(R) Xenon(R) 2.40 GHz (2 cores) processor,
24 GB of RAM and a 64-bit Windows 7 Enterprise operating system.
We performed numerical experiments on five different roads. For each
road, we created ten different test problems by varying the number of sta-
tions. To analyze the performance, we consider the number of function calls
(blackbox evaluations) and the wall-clock time for each test problem. The
full set of numerical data is included in Appendix A.1.
200
180
160
140
Time (in minute)
120
100
80
60
40
20
0
0 10 20 30 40 50 60
No. of Stations
Figure 2.5: Required time with respect to number of stations for solved
problem only.
Figures 2.5 and 2.6 shows the wall-clock time and the number of function
calls required to solve the test problems. Since the numerical experiments
were performed on five different roads, for every number of stations, we have
five different test problems. As a terminating condition of the algorithm,
we set timeout to 3 hours. When the number of stations is up to 30, all
of the five test problems can be solved in 3 hours. From Figures 2.5, we
can see that when the number of stations increased to 45 and 50 then only
one problem can be solve within the time-limit. From Figure 2.6, we can
18
2.5. Summary
25000
20000
No. of Function Calls
15000
10000
5000
0
0 10 20 30 40 50 60
No. of Stations
Figure 2.6: Growth of no. of function calls with respect to number of

stations.
also observe that for a road of 45 or 50 stations, the solver required around
17250 function calls, which is a large number of function calls to solve a
problem corresponding to a small road of 45 or 50 stations. It would be
worth mentioning that a typical road is closer to 200 stations.
2.5 Summary
We have discussed a very straight-forward formulation of the horizontal
alignment optimization problem. We introduced a derivative-free optimiza-
tion approach to solve the problem. In this basic optimization model, the
problem size increases as the number of stations increases. Usually, the
NOMAD solver can only effectively handle a problem of a small number
of variables; i.e., the problem size is less than 50 [LD11]. In Addition, the
model yields a piece-wise linear curve, which is not used by engineers in
practice. Considering all of the engineering specifications of a horizontal
alignment, a more precise formulation could be developed for a road of a
large number of stations (i.e., more than 100 stations). In the next chapter,
we explain how to build a practical model.
19
Chapter 3
Horizontal alignment
optimization model
In this chapter, we describe a horizontal alignment optimization model
that produces a piece-wise linear-circular curve instead of the piecewise lin-
ear curve outputed by the basic model introduced in Chapter 2. We also
describe in detail the geometric specifications of the model.
3.1 Geometric representation of horizontal

alignment
In this model, a horizontal alignment consists of a sequence of circular
curves and tangential lines. The circular curves and tangential lines are
defined by some intersection points and the radius of curvature associated
with each intersection point. In Figure 3.1, S and E are the start and end
points of the alignment, respectively. The intersection points of the align-
ment are P1 , P2 , and P3 . Each intersection point has a radius of curvature
that defines the circular curve. The radius of curvature associated with the
intersection points P1 , P2 , and P3 are r1 , r2 , and r3 . The purple and red
portions of the alignment in Figure 3.1 are the circular curve and tangential
line segments, respectively.
Let i be the index of the intersection points and np be the number of
intersection points. Since the intersection point Pi has an associated radius
of curvature ri , we define an intersection point with radius of curvature as
(Pi , ri ), where Pi ∈ R2 and ri ∈ R. Without loss of generality, we can say
that the start and end points are a point in R2 with zero radius of curvature.
The starting and end points are denoted as (P0 , 0) and (Pnp +1 , 0). So we
represent a horizontal alignment HA as the sequence
HA = ((P0 , 0), (P1 , r1 ), (P2 , r2 ), . . . (Pnp , rnp ), (Pnp +1 , 0)). (3.1)
To determine the actual horizontal alignment, we need to calculate the

circular curves and tangential line segments from the given intersection
20
3.1. Geometric representation of horizontal alignment
Figure 3.1: Geometric representation of horizontal alignment. The align-

ment displayed is represented as ((S, 0), (P1 , r1 ), (P2 , r2 ), (P2 , r2 ), (E, 0)).
𝑃𝑖
𝐸𝑖 𝐹𝑖
𝐶𝑖
𝑃𝑖−1 𝑃𝑖+1
𝑄𝑖
Figure 3.2: Geometric specifications of a circular curve.
21
points and the associated radius of curvatures. So for each intersection

point Pi , we need to find out the two tangential points and the center of
curvature. Let Ei and Fi be the left and right tangential point, respectively.
In Figure 3.2 the green points are the two tangential points. Let Ci (red
point in Figure 3.2) be the center of curvature that corresponds to the inter-
section point Pi . We can calculate Ei , Fi , and Ci for each intersection point
Pi using the radius of curvature ri and the three consecutive intersecting
points Pi−1 , Pi , and Pi+1 .
Tangential point calculation

We define the following variables:
– θi is the angle at Pi using three consecutive intersection point Pi−1 ,

Pi , and Pi+1 .
– Ui is the vector from the point Pi to Pi−1 .
– Vi is the vector from the point Pi to Pi+1 .
– Wi is the vector from the point Pi−1 to Pi+1 .
– Qi is the intersection point of the angle bisector of θi and the line

joining Pi−1 and Pi+1 .
By definition, we have
Ui = Pi−1 − Pi ,
Vi = Pi+1 − Pi , (3.2)
Wi = Pi+1 − Pi−1 .
We can calculate the angle θi by using the dot product of the vector Ui and
Vi .
Ui · Vi = kUi kkVi k cos θi . (3.3)

Using Equation (3.3) we have

Ui · Vi
θi = arccos . (3.4)
kUi kkVi k
Since the angle bisector Pi Qi bisects the angle θi , we have
θi
∠Pi−1 Pi Qi = = ∠Pi+1 Pi Qi . (3.5)
2
22
Pi Pi−1 and Pi Pi+1 are the tangent to the circle at the tangential point Ei and
Fi . Thus we have Ci Ei ⊥Pi Pi−1 and Ci Fi ⊥Pi Pi+1 . The triangle 4Pi Ei Ci
and 4Pi Fi Ci are right angle triangles. The segment Pi Ci is the common
side of 4Pi Ei Ci and 4Pi Fi Ci . Since Ei Ci =Fi Ci , we have Pi Ei = Pi Fi . Let
lt be the length of Pi Ei . In the triangle 4Pi Ei Ci , we have
θi Ei Ci ri
tan = = . (3.6)
2 Pi Ei lt
So
ri
lt = . (3.7)
tan θ2i
Let êUi and êVi be the two unit vector of Ui and Vi . The tangential point
Ei and Fi can be calculated as follows:
Ei = Pi + lt êUi , (3.8)
Fi = Pi + lt êVi . (3.9)
Center point calculation

Fact 3.1 ([Byr47, Book VI Proposition III]). The angle bisector of an angle
in a triangle divides the opposite side in the same ratio as the sides adjacent
to the angle.
Let lb be the length of the Qi Pi−1 . The length of Pi Pi−1 , Pi Pi+1 and
Pi−1 Pi+1 are kUi k, kVi k, and kWi k, respectively. Pi Qi is the angle bisector
of θi in 4Pi−1 Pi Pi+1 . So using Fact 3.1 we have
kUi k lb
= . (3.10)
kVi k kWi k − lb
So
kUi kkWi k
lb = . (3.11)
kUi k + kVi k
Let êWi be the unit vector of Wi . The point Qi can be calculated as follows:
Qi = Pi−1 + lb êWi . (3.12)

Define Xi as
X i = Qi − P i . (3.13)
23
3.2. Model description
Let lx be the length of Pi Ci . In the triangle 4Pi Ei Ci , we have

θi Pi Ei lt
cos = = . (3.14)
2 P i Ci lx
So
lt
lx = . (3.15)
cos θ2i
Let êXi be the unit vector of Xi . So the center point Ci can be calculated
as follows:
Ci = Pi + lx êXi . (3.16)
3.2 Model description

3.2.1 Definitions
In order to model the horizontal alignment, we group a set of consecutive
stations to make a segment. We divide the entire corridor into m segments,
which are indexed by a set IG = {1, 2, 3 . . . , m}. Every segment consists
of a set of stations. The g th segment has ng stations. For all g ∈ IG ,
the stations associated with the g th segment are indexedPby the set ISG =
{1, 2, 3, . . . , ng }. So the total number of stations is n = g∈IG ng . The g th
segment (g ∈ IG ) j th station (j ∈ ISG ) is denoted by Sg,j . The stations of
the corridor are indexed by the set IS = {1, 2, 3, . . . n}. The ith (i ∈ IS )
station of the corridor is denoted by si . The set of all stations is S =
{s1 , s2 , s3 , . . . sn }. We define a function to map a station index of a segment
to the actual station index (i.e. station index of the corridor) as follows:
δ : (IG , ISG ) 7→ IS . (3.17)
Clearly, for all g ∈ IG , j ∈ ISG , i = δ (g, j) = g−1

P
p=1 np + j ∈ IS . Therefore,
at station si ∈ S, we have Sg,j = si for all g ∈ IG , j ∈ IGS . Each station
has nd data points, which are indexed by the set ID = {1, 2, 3 . . . , nd }. The
ith station k th (k ∈ ID ) data point is denoted by Di,k . The set of nd data
points at station si is Di = {Di,1 , Di,2 , Di,3 . . . Di,nd }. Corresponding to
each data point, we have vertical road profile data which is defined by a
vector VAi,k . The vector VAi,k consists of the cut-fill areas of a material at
different heights. For example, the data point (x, y) has cut-fill areas of a
material at different heights (i.e., at some different z-values). The cut-fill
areas constitute the vectorVAi,k corresponding to the data point Di,k . At
24
each station si ∈ S, we have the leftmost and rightmost offset data points
that define the boundary of the corridor. The leftmost and the rightmost
data points of the ith station are denoted by ui and vi and defined as follows:
ui = Di,1 ∀i ∈ IS ,
(3.18)
vi = Di,nd ∀i ∈ IS .
The set of the leftmost offset data points is U = {u1 , u2 , u3 , . . . , un }. The
set of the rightmost offset data points is V = {v1 , v2 , v3 , . . . , vn }. At the g th
segment j th station, the leftmost and the rightmost data points are denoted
by Ug,j and Vg,j , respectively. So when δ (g, j) = i, we have
Ug,j = ui ,
(3.19)
Vg,j = vi .
Every two consecutive segments share a common station. The last station
of a segment is the first station of the next segment (see Figure 3.3). The
following equations are satisfied for g = 1, 2, 3, . . . , m − 1.
Sg,ng = Sg+1,1 ,
Ug,ng = Ug+1,1 ,
(3.20)
Vg,ng = Vg+1,1 ,
Dδ(g,ng ),k = Dδ(g+1,1),k .
At each station, the line passing through the leftmost offset data point
and the rightmost offset data point is defined as a cross-section line of the
station, see Figure 2.1. The parametric equation of the cross-section line
Li (t) of the ith station si ∈ S with the leftmost offset data point ui and the
rightmost offset data point vi is
Li (t) = (1 − t)ui + tvi for t ∈ R. (3.21)

Using the mapping in Equation (3.17), for the g th segment j th station, the
cross-section line Lg,j (tlg,j ) is
Lg,j (t) = (1 − t)Ug,j + tVg,j for t ∈ R. (3.22)

In Equation (3.22), if the parameter t is restricted to [0, 1], we obtain a
segment. The equation of the cross-section segment with end-points Ug,j
and Vg,j is
L̄g,j (t) = (1 − t)Ug,j + tVg,j for t ∈ [0, 1] . (3.23)
25
Figure 3.3: Road segment representation in a specified corridor.
Figure 3.4: Horizontal alignment in a specified corridor. The red curve is

an example horizontal alignment. The blue stars are the intersection points
(IP) and the dotted blue rectangular boxes are feasible region (for moving)
associated with the intersection points.
We assume that each segment has exactly three intersection points (ex-
cept for the last segment which may have two) named Pg,1 , Pg,2 , and Pg,3
26
for the g th segment. Each intersection point of a segment has a radius of

curvature. For the g th segment rg,1 , rg,2 , and rg,3 are defined as the radius
of curvatures corresponding to the intersection points Pg,1 ,Pg,1 , and Pg,3 , re-
spectively. In Figure 3.4, the blue points are the intersection points and the
red curve is the associated horizontal alignment. By moving the intersection
points, we can build a wide variety of horizontal alignments.
Each intersection point has a feasible region. We define the feasible
region of each intersection point by a rectangular box. The dotted blue
rectangles in Figure 3.4, are the feasible region of the intersection points.
A rectangular box is defined by the leftmost bottom corner point and the
rightmost top corner point. Let Bg,1 and Bg,1 be the leftmost bottom corner
point and the rightmost top corner point of the rectangular box associated
with the intersection point Pg,1 . So the rectangular box associated with Pg,1
is defined as (Bg,1 , Bg,1 ). Similarly, the rectangular boxes associated with
Pg,2 and Pg,2 are defined as (Bg,2 , Bg,2 ) and (Bg,3 , Bg,3 ), respectively.
3.2.2 The optimization model

The road design problem is formulated as a bi-level optimization prob-
lem. We can solve a vertical alignment optimization problem for a fixed
horizontal alignment. So the main idea is to move the horizontal alignment
and then minimize the vertical alignment.
Basic approach: one variable per station

At each station si ∈ S we are given the data points Di,k and correspond-
ing vertical road profile data VAi,k (∀k ∈ ID ) as input. At a station si , for
an arbitrary data point Di,a the vertical road profile data VAi,a between two
consecutive data points Di,k and Di,k+1 along the cross-section line segment
L̄i (t) is interpolated using the following equation
kDi,a − Di,k k
VAi,a = VAi,k + (VAi,k+1 − VAi,k ) . (3.24)
kDi,k+1 − Di,k k
Therefore, we have the vertical road profile data along every cross-section
line segment (L̄g,j (t) ∀g ∈ IG , ∀j ∈ ISG ). We can build a horizontal align-
ment by taking a point from each cross-section line segment. For instance,
the red and purple piecewise linear curve in Figure 3.5 shows two different
horizontal alignments for a road segment. For a fixed horizontal alignment,
the optimal vertical alignment cost can be calculated by solving the vertical
alignment optimization problem formulated in [HHLR14]. So the cost of a
27
horizontal alignment is the optimal vertical alignment cost corresponding to

that horizontal alignment. Since we are given Di,k and VAi,k for all stations,
we could build the cross-section line Lg,j (t). Thus, by changing t between
0 to 1 of the Lg,j (t), we can obtain any point Di,a on Lg,j (t) and the as-
sociated VAi,a using Equation (3.24). Let tlg,j be the parameter of the g th
segment j th station. We can obtain all possible horizontal alignments by
moving the parameter tlg,j along the line Lg,j (tlg,j ). The parameter values
of the Lg,j (tlg,j ) defines a horizontal alignment. So the parameters of the
cross-section line can be used as variables. Let Tl be the variable vector
defining the horizontal alignment as a piecewise linear function. We define
Tl = htl1,1 , tl1,2 , . . . , tl1,n1 , tl2,1 , tl2,2 , . . . , tl2,n2 , . . . . . . , tlm,1 , tlm,2 , . . . , tlm,nm i.

| {z } | {z } | {z }
Segment 1 Segment 2 Segment m
Figure 3.5: Piece-wise linear representation of horizontal alignment segment

considering all of the cross-section lines.
The cost function of a horizontal alignment optimization problem asso-

ciated with variable Tl is defined as CVA (Tl ). The function CVA (Tl ) gives
the optimal vertical alignment cost for Tl . So the horizontal alignment op-
timization problem becomes
min CVA (Tl ). (3.25)

0≤Tl ≤1
In practice, if we consider all cross-section lines (i.e., all of the stations),

we
Pm have ng variables for each segment. So the total number of variables is
g=1 ng = n, which can be large. We can reduce the number of variables
by defining the horizontal alignment geometrically.
28
Advanced approach: three variables per intersection point

Objective function and variables
We define a horizontal alignment using the intersection points and asso-
ciated radius of curvatures. We have three intersecting points Pg,1 ,Pg,2 , and
Pg,3 for the g th segment. By moving the intersection points and with differ-
ent associated radius of curvatures, a wide variety of horizontal alignments
can be built. So we can have the intersection points and the associated
radius of curvature as the variables.
All of the intersection points are on the XY -plane. We define the inter-
section points Pg,1 , Pg,2 , and Pg,3 in Cartesian coordinate as follows:
Pg,1 = (pxg,1 , pyg,1 ),

Pg,2 = (pxg,2 , pyg,2 ), (3.26)
Pg,3 = (pxg,3 , pyg,3 ).
Thus pxg,1 , pyg,1 , pxg,2 , pyg,2 , pxg,3 , pyg,3 , rg,1 ,rg,2 , and rg,3 are the variable
for the g th segment. We assume that the starting and end points of the
alignment are fixed.
Two adjacent segments share an intersection point. Let P̂g,g+1 =
(p̂xg,g+1 , p̂yg,g+1 ) be the common intersection point between the g th and
(g + 1)th segments. So for two consecutive segments we have
Pg,3 = Pg+1,1 ,
P̂g,g+1 = Pg,3 , (3.27)
P̂g,g+1 = Pg+1,1 .
Let r̂g,g+1 be the radius of curvature corresponding to the point P̂g,g+1 . So
for two consecutive segments we have
rg,3 = rg+1,1 ,
r̂g,g+1 = rg,3 , (3.28)
r̂g,g+1 = rg+1,1 .
Let X be the variable vector of the optimization problem. We have
X = hpx1,2 , py1,2 , p̂x1,2 , p̂y1,2 , r1,2 , r̂1,2 , px2,2 , py2,2 , p̂x2,3 , p̂y2,3 , r2,2 , r̂2,3 . . . . . .
| {z } | {z }
Segment 1 Segment 2
pxm−1,2 , pym−1,2 , p̂xm,m−1 , p̂ym−1,m , rm−1,2 , r̂m−1,m , pxm,2 , pym,2 , rm,2 i.

| {z } | {z }
Segment m-1 Segment m
29
(𝑝𝑥3,4 , 𝑝𝑦3,4 )
(𝑝𝑥3,2 , 𝑝𝑦3,2 )
(𝑝𝑥2,3 , 𝑝𝑦2,3 )
Figure 3.6: An example road of four segments showing the associated vari-
ables of the optimization model. The green cross-section lines separate the
road segments. The starting point and the end point of the alignment are
(px1,1 , py1,1 ) and (px4,3 , py4,3 ), respectively, which are fixed.
In Figure 3.6, for a road of four segments, the associated variables are
depicted. Note that the last segment might have two intersection points
(when the number of intersection points (IP) is not divisible by 3). In
that case, the last segment has only variables corresponding to the common
intersection point. For instance, if we assume Figure 3.6 does not have the
intersection point (py4,2 , py4,2 ); i.e., the number of intersection points is 8,
then the last segment has only the variables (p̂x3,4 , p̂y3,4 ) and r̂3,4 , which are
also considered in the previous (third) segment.
A horizontal alignment for a road segment consists of some circular
curves and tangential lines. For each segment the horizontal alignment
curve has three circular curves and two tangential lines (see Figure 3.7).
The purple portions and the red portions of the horizontal alignment curve
in Figure 3.7 are circular curves and tangential lines, respectively. Let Eg,1 ,
Eg,2 , and Eg,3 be the left tangential points; and Fg,1 , Fg,2 , and Fg,3 be the
right tangential points correspond to the intersection points Pg,1 , Pg,2 , and
30
Figure 3.7: A segment of a road showing an associated horizontal alignment.
Pg,3 . The left and right tangential points can be calculated using Equa-
tion (3.8) and Equation (3.9) . Let Cg,1 , Cg,2 , and Cg,3 be the center of
curvature corresponds to the intersection point Pg,1 ,Pg,2 , and Pg,2 . The cen-
ter of curvature can be calculated using Equation (3.16). Since all points
are in XY -plane, we define
Cg,1 = (cxg,1 , cyg,1 ), Cg,2 = (cxg,2 , cyg,2 ), Cg,3 = (cxg,3 , cyg,3 ),

Eg,1 = (exg,1 , eyg,1 ), Eg,2 = (exg,2 , eyg,2 ), Eg,3 = (exg,3 , eyg,3 ), (3.29)
Fg,1 = (fxg,1 , fyg,1 ), Fg,2 = (fxg,2 , fyg,2 ), Fg,3 = (fxg,3 , fyg,3 ).
Let Hg,1 Hg,2 , Hg,3 , Hg,4 , and Hg,5 be five parametric pieces of the horizontal
alignment curve for the g th segment. Hg,1 , Hg,2 , and Hg,3 are the circular
arc corresponding to the intersection points Pg,1 , Pg,2 , and Pg,3 . Hg,4 is the
tangential lines connecting the two arcs Hg,1 and Hg,2 . Hg,5 is the tangential
lines connecting the two arcs Hg,2 and Hg,3 . Let tcg,1 , tcg,2 , tcg,3 , tcg,4 , and tcg,5
be the parameters of Hg,1 Hg,2 , Hg,3 , Hg,4 , and Hg,5 , respectively.
The parametric equation of the circle corresponding to the intersection
point Pg,1 can be written as follows:
c
rg,1 cos(tcg,1 ) + cxg,1

x(tg,1 )
= for tcg,1 ∈ [0, 2π] . (3.30)
y(tcg,1 ) rg,1 sin(tcg,1 ) + cyg,1
Equation (3.30) gives the full circle but we need a circular arc with two
endpoints. So we need to calculate the bounds of tcg,1 in Equation (3.30) to
get the endpoints of a circular arc. The tangential point Eg,1 and Fg,1 are
the two endpoints of the circular arc associated with the intersection point
31
Pg,1 . Since Eg,1 and Fg,1 are on the circle, we have

exg,1
= , (3.31)
rg,1 sin(tcg,1 ) + cyg,1 eyg,1
and

fxg,1
= . (3.32)
rg,1 sin(tcg,1 ) + cyg,1 fyg,1
Equation (3.31) and Equation (3.32) give the two values of t that make
the bounds for the circular arc with two end points Eg,1 and Fg,1 . From
Equation (3.31) we deduce
exg,1 − cxg,1 ey − cyg,1
tcg,1 = arccos or arcsin g,1 . (3.33)
rg,1 rg,1
Similarly, from Equation (3.32) we have
fxg,1 − cxg,1 fy − cyg,1

rg,1 rg,1
The ranges of arccos and arcsin are [0, π] and − π2 , π2 , respectively. But the

parameter tcg,1 of the full circle in Equation (3.30) varies from 0 to 2π. In
order to handle this issue, first we need to identify the quadrant of the circle
in which the two endpoints Eg,1 and Fg,1 lie on, and then adjust the value
of tcg,1 to calculate the actual value with respect to the full circle. Let teg,1
and tfg,1 be the two parameter values corresponding to the two endpoints
Eg,1 and Fg,1 , respectively. The value of the parameter associated with the
endpoint Eg,1 can be written as follows:
ex −cx ex −cx ey −cy



 arcsin g,1rg,1 g,1 if g,1rg,1 g,1 ≥ 0 and g,1rg,1 g,1 ≥ 0,
π − arcsin exg,1 −cxg,1 ex −cx ey −cy

if g,1rg,1 g,1 < 0 and g,1rg,1 g,1 ≥ 0,

e rg,1
tg,1 = ex −cx ex −cx ey −cy


 π − arcsin g,1rg,1 g,1 if g,1rg,1 g,1 < 0 and g,1rg,1 g,1 < 0,
 ex −cx ex −cx ey −cy
2π + arcsin g,1rg,1 g,1 if g,1rg,1 g,1 ≥ 0 and g,1rg,1 g,1 < 0.


(3.35)
32
Similarly, associated with the endpoint Fg,1 , the value of the parameter is
 fxg,1 −cxg,1 fx −cx fy −cy
arcsin

 rg,1 if g,1rg,1 g,1 ≥ 0 and g,1rg,1 g,1 ≥ 0,
π − arcsin fxg,1 −cxg,1 fx −cx fy −cy

if g,1rg,1 g,1 < 0 and g,1rg,1 g,1 ≥ 0,

f rg,1
tg,1 = fx −cx fx −cx fy −cy


 π − arcsin g,1rg,1 g,1 if g,1rg,1 g,1 < 0 and g,1rg,1 g,1 < 0,
fx −cx fx −cx fy −cy


2π + arcsin g,1rg,1 g,1 if g,1rg,1 g,1 ≥ 0 and g,1rg,1 g,1 < 0.

(3.36)
Note that the four cases in Equation (3.35) and Equation (3.36) represent
the first, second, third and fourth quadrants of the circle, respectively.
In Equation (3.35) and Equation (3.36), we do not know which value
is the upper bound or lower bound of the circular arc connecting the two
endpoints Eg,1 and Fg,1 . So the lowest and highest of the two values are
the lower bound and the upper bound, respectively. Let tg,1 and tg,1 be the
upper bound and the lower bound for the circular arc Hg,1 . We have
n o
tg,1 = min teg,1 , tfg,1 ,
n o (3.37)
tg,1 = max teg,1 , tfg,1 .
Note that if any of the two endpoints lies on the fourth quadrant and
the other endpoint lies on the first quadrant then the value of the parameter
corresponding to the first quadrant has to be added by 2π, otherwise a wrong
arc will be generated. In Figure 3.8 an example of this issue is illustrated.
For the example shown in Figure 3.8, using the formulas in Equations (3.35),
(3.36), and (3.37), we can calculate the lower bound and upper bound of the
arc as π4 and 7π 4 which generates the red arc. In Figure 3.8, we can observe
that the correct arc is the green arc rather than the red arc. So in order
to generate the green arc, we have to add 2π to the associated parameter
value of the endpoint Eg,1 (which lies on the first quadrant). After adding
2π to π4 (the value associated with Eg,1 ), we get the new lower bound and
upper bound as 7π 9π
4 and 4 , respectively, which generates the green arc in
Figure 3.8.
Let tg,2 , tg,3 be the lower bounds and tg,2 , tg,3 be the upper bounds for
the circular arcs Hg,2 and Hg,3 , respectively. The lower and upper bound
for Hg,2 and Hg,3 can be calculated similarly as in Equation (3.37). So the
parametric equation of the circular arcs Hg,1 , Hg,2 , and Hg,3 are

h i
c c
Hg,1 (tg,1 ) = for tg,1 ∈ tg,1 g,1 ,
, t (3.38)
rg,1 sin(tcg,1 ) + cyg,1
33
𝑃𝑔,1
𝑐
𝑡𝑔,1 = 𝜋/4
𝐸𝑔,1
𝑃𝑔,2
𝐹𝑔,1
𝑐
𝑡𝑔,1 = 7𝜋/4
𝑃𝑔,3
Figure 3.8: An example of the quadrants issue in generation of a circular

arc.

h i
Hg,2 (tcg,2 ) = for tcg,2 ∈ tg,2 , tg,2 , (3.39)
and
h i
Hg,3 (tcg,2 ) = for tcg,3 ∈ tg,3 , tg,3 . (3.40)
The tangential line segments Hg,4 connects the endpoints Fg,1 and Eg,2 .
The parametric equation of Hg,4 is
Hg,4 (tcg,4 ) = (1 − tcg,4 )Fg,1 + tcg,4 Eg,2 for tcg,4 ∈ [0, 1] . (3.41)
34
Similarly, the tangential line segments Hg,5 connects the endpoints Fg,2 and
Eg,3 . The parametric equation of Hg,5 is
Hg,5 (tcg,5 ) = (1 − tcg,5 )Fg,2 + tcg,5 Eg,3 for tcg,5 ∈ [0, 1] . (3.42)
In order to calculate the cost of the horizontal alignment, we need to
compute parameter tlg,j of the cross-section lines. We have two different
cases: finding the intersection parameters for a circular arc and a line seg-
ment and finding the intersection parameters for two line segments (i.e., a
cross section line segment and a tangential line segment).
For each segment, we have a set of cross-section lines Lg,1 , . . . , Lg,ng .
A cross-section line Lg,j (tlg,j ), j ∈ ISG of the g th segment is
Lg,j (tlg,j ) = (1 − tlg,j )Ug,j + tlg,j Vg,j for tlg,j ∈ R. (3.43)

Since Ug,j and Vg,j are in the XY -plane, we define Ug,j = (uxg,j , uyg,j ) and
Vg,j = (vxg,j , vyg,j ). Equation (3.43) can be written as follows:
x = uxg,j + (vxg,j − uxg,j )tlg,j ,

(3.44)
y = uyg,j + (vyg,j − uyg,j )tlg,j .
The equation of the circle that corresponds to the intersection point Pg,1 in
implicit form is
(x − cxg,1 )2 + (y − cyg,1 )2 − rg,1 2 = 0. (3.45)
Substituting for x and y from Equation (3.44) into the Equation (3.45) gives
a quadratic equation of tlg,j :
(uxg,j +(vxg,j −uxg,j )tlg,j −cxg,1 )2 +(uyg,j +(vyg,j −uyg,j )tlg,j −cyg,1 )2 −rg,1 2 = 0.
(3.46)
The two roots of Equation (3.46) give the points on the line that cuts the
circle. We deduce
√
l a± ∆
tg,j = , (3.47)
b
where
a =(vxg,j − uxg,j )(cxg,1 − uxg,j ) + (vyg,j − uyg,j )(cyg,1 − uyg,j ),
∆ =rg,1 2 ((vxg,j − uxg,j )2 + (vyg,j − uyg,j )2 )
− ((vxg,j − uxg,j )(cyg,1 − uyg,j ) − (vyg,j − uyg,j )(cxg,1 − uxg,j ))2 , and
b =(vxg,j − uxg,j )2 + (vyg,j − uyg,j )2 .
35
The two different values of tlg,j give the two intersection points of Equa-
tion (3.44) and Equation (3.45) on the line Lg,j (tlg,j ). The roots maybe
similar, in that case, the line intersect only in a single point. If the roots
are imaginary then there is no intersection. Let Sg,j = (sxg,j , syg,j ) be an
intersection point obtained by tlg,j . To calculate the parameter of the circle
in Equation (3.30) corresponding to the point Sg,j , we deduce

sxg,1
= . (3.48)
rg,1 sin(tcg,1 ) + cyg,1 syg,1
So
sxg,1 − cxg,1 sy − cyg,1
rg,1 rg,1
As we described earlier, due to the quadrant issue of the circle, the value
of the parameter tcg,1 has to be adjusted. Let tsg,1 be the parameter value
corresponding to the point Sg,j . We have
 sx −cx sx −cx sy −cy

arcsin g,1rg,1 g,1 if g,1rg,1 g,1 ≥ 0 and g,1rg,1 g,1 ≥ 0,

π − arcsin fxg,1 −cxg,1 sx −cx sy −cy

if g,1rg,1 g,1 < 0 and g,1rg,1 g,1 ≥ 0,

s rg,1
tg,1 = fx −cx sx −cx sy −cy


π − arcsin g,1rg,1 g,1 if g,1rg,1 g,1 < 0 and g,1rg,1 g,1 < 0,
2π + arcsin fxg,1 −cxg,1 if sxg,1 −cxg,1 ≥ 0 and syg,1 −cyg,1 < 0.


rg,1 rg,1 rg,1
(3.50)
h i
If tlg,j ∈ [0, 1] and tsg,1 ∈ tg,1 , tg,1 then the intersection point is in the
corridor. In this case, we accept the value of the parameter tlg,j , otherwise
we reject the value. Similarly, for all other circular arcs we can calculate the
value of the parameter tlg,j .
Now we need to calculate the intersection point between the cross-
sectional line and the tangential line segment. The equation of the tangential
line segment in Equation (3.41) can be written as follows:
x = fxg,1 + (exg,2 − fxg,1 )tcg,4 ,

(3.51)
y = fyg,1 + (eyg,2 − fyg,1 )tcg,4 .
To calculate the intersection parameter of a cross-section line and a tan-
gential line, from Equation (3.51) and Equation (3.44) we have the following
linear system of equations:
uxg,j + (vxg,j − uxg,j )tlg,j = fxg,1 + (exg,2 − fxg,1 )tcg,4 ,

(3.52)
uyg,j + (vyg,j − uyg,j )tlg,j = fyg,1 + (eyg,2 − fyg,1 )tcg,4 .
36
Equations (3.52) in matrix form are
vxg,j − uxg,j exg,2 − fxg,1 tlg,j

fxg,1 − uxg,j
= . (3.53)
vyg,j − uyg,j eyg,2 − fyg,1 tcg,4 fyg,1 − uyg,j
The solution for tlg,j and tcg,4 is
−1
tlg,j

v − uxg,j exg,2 − fxg,1 fxg,1 − uxg,j
= xg,j . (3.54)
tcg,4 vyg,j − uyg,j eyg,2 − fyg,1 fyg,1 − uyg,j
In order to calculate the solution defined in Equation (3.54), the coefficient

matrix has to be invertible. Thus in Equation (3.52), if the coefficient matrix
is not full-rank, the system has no solution (i.e. the two lines are parallel). If
the value of both tlg,j and tcg,4 are in [0, 1], we accept the solution. Otherwise,
the solution point is outside of the corridor.
If all tlg,j ∈ [0, 1] we can compute the optimal vertical alignment cost for
a horizontal alignment defined by X. On the other hand, if any tlg,j ∈ / [0, 1],
the alignment is outside of the corridor. In this case, we set the optimal
vertical alignment cost to infinity. Finally, the objective function of the
horizontal alignment optimization can be written as
if tlg,j ∈ [0, 1] ∀g ∈ IG , j ∈ ISG ,

CVA (X)
f (X) =
∞ otherwise. (3.55)
Constraints
A horizontal curve consists of the tangential line segments followed by
the circular arcs. Two consecutive circular arcs are connected by a tangential
line. The horizontal alignment will be discontinuous when two circular arcs
overlap along the tangential line (see Figure 3.9).
For the g th segment, we have two tangential lines. The line passing
through the intersection points Pg,1 and Pg,2 and the line passing through
the intersection points Pg,2 and Pg,3 are the two tangential lines. In order
to maintain continuity on the line passing through the intersection points
Pg,1 and Pg,2 , the length of Pg,1 Pg,2 must be greater than or equal to the
summation of the length of Pg,1 Fg,1 and the length of Pg,2 Eg,2 . So we can
write the continuity constraints as follows:
kPg,2 − Pg,1 k ≥ kPg,1 − Fg,1 k + kPg,2 − Eg,2 k. (3.56)
37
𝑃𝑔,2
𝑃𝑔,3
Figure 3.9: Discontinuity in a horizontal alignment.
Similarly, on the line passing through the intersection points Pg,3 and Pg,2 ,
the continuity constraint becomes
kPg,2 − Pg,3 k ≥ kPg,3 − Eg,3 k + kPg,2 − Fg,2 k. (3.57)
Each intersection point has a feasible region. The feasible region is de-
fined by a rectangular box. We define the box corner points in Cartesian
coordinate as follows:
Bg,1 = (bxg,1 , byg,1 ), Bg,1 = (bxg,1 , byg,1 );
Bg,2 = (bxg,2 , byg,2 ), Bg,2 = (bxg,2 , byg,2 ); (3.58)
Bg,3 = (bxg,3 , byg,3 ), Bg,3 = (bxg,3 , byg,3 ).
So, in order to bound the intersection points inside of the rectangular boxes
we have the following constraints for the g th segment:
bxg,1 ≤ pxg,1 ≤ bxg,1 ,
byg,1 ≤ pyg,1 ≤ byg,1 ,
bxg,2 ≤ pxg,2 ≤ bxg,2 ,
(3.59)
byg,2 ≤ pyg,2 ≤ byg,2 ,
bxg,3 ≤ pxg,3 ≤ bxg,3 ,
byg,3 ≤ pyg,3 ≤ byg,3 .
38
Two adjacent segments have a common intersection point P̂g,g+1 . Let

(B̂g,g+1 , B̂g,g+1 ) be the rectangular box corresponding to the common inter-
section point P̂g,g+1 . So for two consecutive segments we have
(B̂g,g+1 , B̂g,g+1 ) = (Bg,1 , Bg,1 ),

(3.60)
(B̂g,g+1 , B̂g,g+1 ) = (Bg,3 , Bg,3 ).
We define the rectangular box corner points B̂g,g+1 and B̂g,g+1 in Cartesian
coordinate as follows:
B̂g,g+1 = (b̂xg,g+1 , b̂yg,g+1 ),
(3.61)
B̂g,g+1 = (ˆbxg,g+1 , ˆbyg,g+1 ).
Since the starting and end points are fixed the constraints (3.59) for the
g th segment can be rewritten as follows:
bxg,2 ≤ pxg,2 ≤ bxg,2 ∀g ∈ {1, 2, .....m} ,

byg,2 ≤ pyg,2 ≤ byg,2 ∀g ∈ {1, 2, .....m} ,
(3.62)
b̂xg,g+1 ≤ p̂xg,g+1 ≤ ˆbxg,g+1 ∀g ∈ {1, 2, .....m − 1} ,
b̂yg,g+1 ≤ p̂yg,g+1 ≤ ˆbyg,g+1 ∀g ∈ {1, 2, .....m − 1} .
The radius of curvature associated with each intersection point has a

minimum value. If the radius of curvature is too small (or zero) then a
horizontal alignment might get a sharp turn. Let Rmin be the minimum
radius of curvature. For the radius of curvatures rg,1 , rg,2 , and rg,3 of the
g th segment, we have the following constraints:
rg,1 ≥ Rmin ,
rg,2 ≥ Rmin , (3.63)
rg,3 ≥ Rmin .
For the starting and end points of the alignment, the radius of curvatures
are zero (i.e. r1,1 = 0 and rm,3 = 0). Since for two adjacent segments
r̂g,g+1 = rg,3 = rg+1,1 , we can rewrite the radius of curvature constraints as
follows:
r̂g,g+1 ≥ Rmin ∀g ∈ {1, 2, .....m − 1} ,
(3.64)
rg,2 ≥ Rmin ∀g ∈ {1, 2, .....m} .
39
3.3. Model Summary
3.3 Model Summary

The summary of the optimization model considering all of the geometric
specifications is stated as follows.
Objective function
if tlg,j ∈ [0, 1] ∀g ∈ IG , j ∈ ISG ,

CVA (X)
f (X) =
∞ otherwise. (3.65)
Continuity constraints
kPg,2 − Pg,1 k ≥ kPg,1 − Fg,1 k + kPg,2 − Eg,2 k ∀g ∈ {1, 2, .....m} , (3.66)
kPg,2 − Pg,3 k ≥ kPg,3 − Eg,3 k + kPg,2 − Fg,2 k ∀g ∈ {1, 2, .....m} , (3.67)
IP bound constraints
bxg,2 ≤ pxg,2 ≤ bxg,2 ∀g ∈ {1, 2, .....m} ,

byg,2 ≤ pyg,2 ≤ byg,2 ∀g ∈ {1, 2, .....m} ,
(3.68)
b̂xg,g+1 ≤ p̂xg,g+1 ≤ ˆbxg,g+1 ∀g ∈ {1, 2, .....m − 1} ,
b̂yg,g+1 ≤ p̂yg,g+1 ≤ ˆbyg,g+1 ∀g ∈ {1, 2, .....m − 1} ,
Minimum radius of curvature constraints
r̂g,g+1 ≥ Rmin ∀g ∈ {1, 2, .....m − 1} ,

(3.69)
rg,2 ≥ Rmin . ∀g ∈ {1, 2, .....m} .
40
Chapter 4
Numerical results
In this chapter, we present our experimental data and discuss the model
performance to solve the real-world problems. We solve the test problem
set using the two different derivative free optimization solvers: NOMAD
[ALT09] and HOPSPACK [Pla09]. Finally, we compare the results of the
two solvers.
4.1 Experimental setup

We performed numerical experiments on five different roads listed in
Table 4.1. The road profile data are given by our industry partner Soft-
ree Technical System Inc. Note that Road D is a backtracking road and
the other four roads i.e., Road A, Road B, Road C, and Road E are non-
backtracking roads.
Table 4.1: Specifications of the test problems
Road Name No. of stations No. of IPs

Road A 73 8
Road B 361 5
Road C 258 14
Road D 118 22
Road E 74 9
All of the experiments were carried out in a Dell workstation with an

Intel(R) Xenon(R) 2.40 GHz (2 cores) processor, 24 GB of RAM and a
64-bit Windows 7 Enterprise operating system. We used two derivative
free optimization solvers, NOMAD [ALT09] (version 3.5, available at http:
//www.gerad.ca/nomad) and HOPSPACK [Pla09] (version 2.0.2, available
at http://www.sandia.gov/hopspack) to solve the test problems. The
optimization model was implemented in C++ using Microsoft Visual Studio
2010 Professional Edition.
41
4.2. The NOMAD and HOPSPACK solvers
4.2 The NOMAD and HOPSPACK solvers

The NOMAD and HOPSPACK solvers use two different derivative free
optimization algorithms of the same category. Both of the solvers use the
pattern search to find an optimal solution. The NOMAD and HOPSPACK
solvers use the Mesh Adaptive Direct Search (MADS) algorithm [AD06]
and the Asynchronous Parallel Pattern Search (APPS) algorithm [Kol05,
GK06], respectively. Both algorithms convergence to a locally optimal point
[AD06], [Kol05]. When the mesh size of the MADS algorithm goes to zero,
it converges to a local minimum [AD06]. Similarly, if the step length of
the APPS algorithm goes to zero, it converges to a local minimum [Kol05].
It is worth mentioning that both MADS and APPS algorithm are globally
convergent to a locally optimum point and a solution (i.e., a minimizer)
obtained by the algorithms depends on the starting point.
As a stopping condition of the algorithms, we use minimum mesh size
and minimum step length for the MADS and APPS algorithms, respectively.
Since we are interested in a solution close to a local minimum, we set both
the minimum mesh size and minimum step length to 0.1. The input data of
our model are given in meters. The final scaling (see [ALT09] and [Pla09])
of the variables of our model goes down below 10 cm, which means a local
minimum should exist in less than 10 cm distance. However, we can set
these parameter values of the algorithms to a value less than 0.1 for a bet-
ter precision, but most of the agents (including our industry partner) only
require a 10 cm precision in the optimal solution.
Both solvers, NOMAD and HOPSPACK need an initial starting point
to run the algorithm. We used the baseline alignment of a corridor as an
initial starting point for both solvers. However, the NOMAD solver gives a
deterministic solution (i.e., different independent runs of the algorithm yield
the same solution), while the HOPSPACK solver gives a non-deterministic
solution (i.e., different independent runs of the algorithm might yield dif-
ferent solutions). So first, we solved the test problems using the NOMAD
solver and then compare with the HOPSPACK solver solutions obtained by
different independent runs.
4.3 Results for the test problems

Our model works well for both backtracking and non-backtracking align-
ments. Figure 4.1 and Figure 4.2 illustrate an initial alignment and an
optimum alignment for a non-backtracking road and a backtracking road,
42
4.3. Results for the test problems
respectively.
Table 4.2 shows the cost improvement of the objective functions, the
number of black-box evaluation and wall-clock time required to solve the
test problems using the NOMAD solver.
Table 4.2: Cost improvement, no. of black-box evaluations and wall-clock

time required to solve the test problems using the NOMAD solver.
No. of Wall-
Cost
Initial Optimized Black- clock
Road Improv-
alignment alignment box time
Name ement
cost cost evalua- (sec-
(%)
tions onds)
Road A 1,897 1,361 28% 2,073 1,445
Road B 17,036 15,198 11% 2,528 1,770
Road C 87,829 69,621 21% 37,165 45,647
Road D 31,031 14,418 54% 90,535 47,613
Road E 8,054 6,498 19% 11,101 5,588
Now we compare the HOPSPACK solver results with the results ob-
tained by the NOMAD solver. We solved each test problem five times inde-
pendently using the HOPSPACK solver. Table 4.3 lists the optimum values
of the objective functions obtained by five independent executions of the
HOPSPACK solver for each test problem. The differences in the optimum
objective function values are calculated with respect to the value obtained
by the NOMAD solver. So in Table 4.3, a “ + ” value in the Difference in
the optimum costs column indicates the HOPSPACK solver yields a better
solution than the NOMAD solver and a “ − ” value indicates the opposite.
Combining the results obtained for the different roads listed in Table 4.3,
we make an overall comparison between the two solvers. We observed that
the HOPSPACK solver might yield a better or a worse solution than the
solution obtained by the NOMAD solver. Thus, considering the tolerance
of the difference in the optimum objective values obtained by two solver, we
count the number of times a solver wins with respect to the other solver.
Table 4.4 shows the comparison of the solvers for different tolerance val-
ues of the difference in the optimum objective values. The x% tolerance of
the difference in optimum objective values means if the optimum objective
values obtained by the two solvers are in between −x% to +x%, then the
solvers yield the same solution (i.e., the two solvers tie), otherwise a posi-
tive percentage value indicates the HOPSPACK solver wins and a negative
43
4950
Initial Alignment (Cost: $ 1898.09)
Optimized Alignment: NOMAD (Cost: $ 1361.77)
Optimized Alignment: HOPSPACK (Cost: $ 1278.5)
4900
4850
Y
4800
4750
4700
9500 9550 9600 9650 9700 9750 9800 9850 9900
X
Figure 4.1: A non-backtracking alignment (the test problem associated with

Road A) showing the initial alignment and the optimized alignments ob-
tained by the NOMAD and HOPSPACK solvers.
percentage value indicates the NOMAD solver wins.

In Table 4.4, we see that if the tolerance of difference in the optimum
objective value is ±3% or above, then the two solver tie for more than 50%
test runs (i.e., more than 13 test runs among 25 test runs). We can also
observe that for any case of the tolerance change in the optimum objective
function value, the difference in the number of times the NOMAD solver
wins and the number of times the HOPSPACK solver wins is at most 2.
So in terms of optimum objective values obtained by the two solvers, the
performance of both solvers are roughly equivalent.
We also recorded the number of black-box evaluations required for the
HOPSPACK and NOMAD solvers to get the optimum solutions. We cal-
culated the difference in number of black-box evaluations with respect to
the number of black-box evaluations required by the NOMAD solver. So
in Table 4.5, a “ + ” value the Difference in no. of black-box evaluations
column indicates the HOPSPACK solver required less black-box evaluations
than the NOMAD solver and a “ − ” value indicates the opposite.
44
5
x 10
6.289
Initial Alignment (Cost: $ 31253)
Optimized Alignment−NOMAD (Cost: $ 14418)
Optimized Alignment−HOPSPACK (Cost: $ 13190.9)
6.2885
6.288
6.2875
6.287
Y
6.2865
6.286
6.2855
6.285
8.592 8.593 8.594 8.595 8.596 8.597 8.598 8.599 8.6
X x 10
5
Figure 4.2: A backtracking alignment (the test problem associated with

Road D) showing the initial alignment and the optimized alignments ob-
tained by the NOMAD and HOPSPACK solvers.
In Table 4.5, we can see that for all of the 25 test runs, the HOPSPACK
solver required less black-box evaluation than the NOMAD solver. For this
particular problem set, on average, the HOPSPACKS solver took 78% less
blackbox evaluation than the NOMAD solver. So the HOPSPACK solver is
roughly five time faster than the NOMAD solver to compute the optimum
solution.
45
Table 4.3: Comparison of optimum objective function values obtained by

the HOPSPACK and NOMAD solvers to solve the test problems.
Optimum Optimum
Difference
cost cost
Road Execution in the
function function
Name No. optimum
value- value-
costs (%)
NOMAD HOPSPACK
Test run 1 1,361 1,291 +5.2%
Test run 2 1,361 1,423 -4.6%
Road A Test run 3 1,361 1,418 -4.2%
Test run 4 1,361 1,278 +6.1%
Test run 5 1,361 1,486 -9.2%
Test run 1 15,198 15,510 -2.1%
Test run 2 15,198 15,141 +0.4%
Road B Test run 3 15,198 15,128 +0.5%
Test run 4 15,198 15,172 +0.2%
Test run 5 15,198 15,529 -2.2%
Test run 1 69,621 70,161 -0.8%
Test run 2 69,621 70,378 -1.1%
Road C Test run 3 69,621 69,995 -0.5%
Test run 4 69,621 67,301 +3.3%
Test run 5 69,621 67,045 +3.7%
Test run 1 14,418 13,190 +8.5%
Test run 2 14,418 15,154 -7.6%
Road D Test run 3 14,418 14,155 +1.8%
Test run 4 14,418 14,016 +2.8%
Test run 5 14,418 15,384 -6.7%
Test run 1 6,497 6,524 -0.4%
Test run 2 6,497 6,475 +0.3%
Road E Test run 3 6,497 6,497 0.0%
Test run 4 6,497 6,502 -0.1%
Test run 5 6,497 6,476 +0.3%
46
Table 4.4: Overall comparison of the HOPSPACK solver and the NOMAD
solver with the optimum objective function values.
Tolerance of No. of No. of No. of

the difference times the times the times the
in the NOMAD HOPSPACK two solvers
optimum costs solver wins solver wins ties
± 1% 8 7 10
± 2% 7 6 12
± 3% 5 5 15
± 4% 5 3 17
± 5% 3 3 19
± 6% 3 2 20
± 7% 2 1 22
± 8% 1 1 23
± 9% 1 0 24
± 10% 0 0 25
47
Table 4.5: Comparison of the no. of black-box evaluations required for the
HOPSPACK and NOMAD solvers to solve the test problems.
Difference
No. of No. of
in no. of
Road Execution black-box Black-box
black-box
Name No. evaluations- evaluations-
evaluations
NOMAD HOPSPACK
(%)
Test run 1 2,073 325 +84.3%
Test run 2 2,073 336 +83.8%
Road A Test run 3 2,073 697 +66.4%
Test run 4 2,073 486 +76.6%
Test run 5 2,073 665 +67.9%
Test run 1 2,528 316 +87.5%
Test run 2 2,528 309 +87.8%
Road B Test run 3 2,528 286 +88.8%
Test run 4 2,528 392 +88.5%
Test run 5 2,528 485 +80.8%
Test run 1 37,165 7,547 +79.7%
Test run 2 37,165 31,418 +15.5%
Road C Test run 3 37,165 3,392 +90.9%
Test run 4 37,165 3,213 +91.4%
Test run 5 37,165 4,661 +87.5%
Test run 1 90,535 15,852 +82.5%
Test run 2 90,535 19,997 +77.9%
Road D Test run 3 90,535 21,194 +76.6%
Test run 4 90,535 17,816 +80.3%
Test run 5 90,535 21,339 +76.4%
Test run 1 11,101 2,019 +81.8%
Test run 2 11,101 4,332 +60.1%
Road E Test run 3 11,101 1,996 +82.0%
Test run 4 11,101 2,501 +77.5%
Test run 5 11,101 3,120 +71.9%
48
4.4. Summary of the result
4.4 Summary of the result

Using our advanced model developed in Chapter 3, the NOMAD solver
and the HOPSPACK solver can solve all of the test problems. In our ad-
vanced model, the number of variables in the optimization problem depends
on the number of intersection points rather than the number of stations. For
a small non-backtracking road of 73 stations with 8 intersection points, the
NOMAD solver and the HOPSPACK solver required 2,073 and 325 function
calls, respectively. In the meanwhile, for a backtracking road of 118 stations
with 22 intersection points, the NOMAD solver and the HOPSPACK solver
required 90535 and 15852 function calls, respectively, to solve the problem.
So our advanced model can solve a reasonably large problem in a reasonable
number of function calls.
49
Chapter 5
Conclusion
5.1 Contributions
Geometric specifications of a horizontal alignment is one of the most im-
portant considerations in the road design process. In our model, a horizontal
alignment is represented using the geometric specifications which are used
by engineers in practice. Thus the solution of the optimization model yields
a practical horizontal alignment which satisfies geometric specifications and
engineering requirements.
In this research, we pursued a new approach to address the horizontal
alignment optimization problem. While most of the studies in the liter-
ature used heuristic based methods, we used derivative-free optimization
approach. The rationale to use the derivative-free optimization approach is,
it converges to a locally optimum solution. Thus our model always gives us
a mathematically proven local optimum solution.
It is well known that the backward bends of a horizontal alignment (i.e.,
in the case of backtracking roads) might give rise to some difficulties in
the optimization process [Nic73, page 123, Chapter 5], [JSJ06, page 21,
Section 2.4.3], [Par77]. Our model effectively handles backward bends in
a horizontal alignment. So our model can generate both backtracking and
non-backtracking alignments.
Our optimization model is a bi-level optimization problem. Our model
integrates the vertical alignment optimization problem and the horizontal
alignment optimization problem together. In our model, the cost of a hori-
zontal alignment is the cost of the optimized vertical alignment which corre-
sponds to that specific horizontal alignment. So our model yields a solution
which has not only a locally optimum horizontal alignment, but also the
corresponding optimum vertical alignment.
50
5.2. Recommendations for future research
5.2 Recommendations for future research

Although our proposed model works well for solving practical horizon-
tal alignment optimization problems, it can be improved further for better
precision and performance.
In our model formulation, we considered that the cross sections in a
corridor are fixed, which are taken corresponding to the baseline alignment.
However, cross sections should be always perpendicular to a horizontal align-
ment. When a horizontal alignment is significantly different from the base-
line alignment, a set of new cross sections should be generated to increase
the precision before calculating the corresponding vertical road profile.
A surrogate cost function is an approximation of the original cost func-
tion which is cheaper to compute. Two types of surrogates can be developed:
adaptive and non-adaptive surrogates. The surrogate function of an original
cost function might reduce the solution time required to get a solution. For
the development of an adaptive surrogate cost function for our original cost
function, we are motivated by the research work done in [BDF+ 99]. They
developed an adaptive surrogate for solving the helicopter rotor blade design
problem. The framework for optimization of expensive functions proposed in
[BDF+ 99] can be applied to our problem effectively. We could also develop a
non-adaptive surrogate cost function to accelerate the optimization process.
The NOMAD solver can exploit the usage of a non-adaptive surrogate.
In our model, we only consider the construction costs to formulate our
cost function. In the future, land acquisition costs could be incorporated
by considering the unit cost of a piece of land corresponding to two consec-
utive cross sections in a corridor. We can also include pavement costs by
considering the unit pavement cost between two consecutive stations.
We mentioned in Chapter 4 that both solvers need an initial starting
point. Since solutions obtained by both solvers are locally optimum, we
can use multiple starting points (i.e.; multiple initial alignments) to start
the algorithms to obtain a better solution quickly. How to choose multiple
alternative good alignments in a specified corridor can be a potential future
research direction.
The HOPSPACK solver can use a parallel computing environment ef-
fectively. The resulting optimization model can be solved using the parallel
version of the HOPSPACK solver to reduce the solution time.
During the optimization process, at each iteration, both of the deriva-
tive free optimization solvers solve a large number of vertical alignment
optimization problems (i.e., a large scale mixed linear programming (MILP)
problem). At the earlier stage of the optimization process (i.e., when the
51
5.2. Recommendations for future research
mesh size is coarse) we can relax some of the parameters of the vertical align-
ment optimization problem to get an approximation cost to go forward and
then at the later stage (i.e., when the mesh size becomes relatively small) we
can again tight the parameters to get the accurate costs. This policy might
reduce the solution time significantly. We can also use a warm start of the
vertical alignment optimization problem when the horizontal alignments are
close to each other to accelerate the vertical alignment optimization process.
So the interconnection between the derivative free optimization solver and
the MILP solver can be a potential way to reduce the solution time.
52
Bibliography
[AAS04] AASHTO. A policy on geometric design of highways and
streets, 2004. American Association of State Highway and
Transportation Officials. → pages 5
[AC73] G. C. Athanassoulis and V. Calogero. Optimal location of a

new highway from A to B: a computer technique for route
planning. In PTRC Seminar Proceedings on Cost Models &
Optimization in Highway, 1973. → pages 9
[AD06] C. Audet and J. E. Dennis, Jr. Mesh adaptive direct search

algorithms for constrained optimization. SIAM Journal on
Optimization, 17(1):188–217, 2006. → pages 16, 42
[Aka06] A. Akay. Minimizing total costs of forest roads with comput-

eraided design model. Sadhana, 31(5):621–633, 2006. → pages
10
[ALT09] C. Audet, S. Le Digabel, and C. Tribes. NOMAD user guide.

Technical Report G-2009-37, Les cahiers du GERAD, 2009. →
pages 41, 42
[Aru05] K. Aruga. Tabu search optimization of horizontal and vertical

alignments of forest roads. Journal of Forest Research, 10:275–
284, 2005. → pages 10
[BDE10] BDE. Bureau of design and environment manual, 2010. Illinois

Department of Transportation, USA. → pages 4
[BDF+ 99] A. J. Booker, Jr. Dennis, J. E., P. D. Frank, D. B. Serafini,

V. Torczon, and M. W. Trosset. A rigorous framework for
optimization of expensive functions by surrogates. Structural
optimization, 17(1):1–13, 1999. → pages 51
[Byr47] O. Byrne. The First Six Books of the Elements of Euclid.

William Pickering, London,UK, 1847. → pages 23
53
Bibliography
[CGF89] E. P. Chew, C. J. Goh, and T. F. Fwa. Simultaneous opti-

mization of horizontal and vertical alignments for highways.
Transportation Research Part B: Methodological, 23(5):315–
329, 1989. → pages 3, 10, 11
[CL06] J. Cheng and Y. Lee. Model for three-dimensional high-
way alignment. Journal of Transportation Engineering,
132(12):913–920, 2006. → pages 11
[DAAMP06] J. Dunlavy, A. Akuoko-Asibey, R. Masse, and D. Pilon. An
analysis of the transportation industry in 2005. Technical re-
port, Statistics Canada, 2006. → pages 1
[Eas88] S. M. Easa. Selection of roadway grades that minimize earth-
work cost using linear programming. Transportation Research
Part A: General, 22(2):121 – 136, 1988. → pages 8
[Fwa89] T. F. Fwa. Highway vertical alignment analysis by dy-
namic programming. Transportation Research Record, 1239:1–
9, 1989. → pages 7
[GCF88] C. J. Goh, E. P. Chew, and T. F. Fwa. Discrete and con-
tinuous models for computation of optimal vertical highway
alignment. Transportation Research Part B: Methodological,
22(6):399–409, 1988. → pages 7
[GK06] G. A. Gray and T. G. Kolda. Algorithm 856: APPSPACK
4.0: Asynchronous parallel pattern search for derivative-free
optimization. ACM Transactions on Mathematical Software,
32(3):485–507, September 2006. → pages 42
[GLA05] A. B. Goktepe, A. H. Lav, and S. Altun. Dynamic optimization
algorithm for vertical alignment of highways. Mathematical
and Computational Applications, 10:341–350, 2005. → pages
7
[Hay70] R. W. Hayman. Optimization of vertical alignment for high-
ways through mathematical programming. Highway Research
Record, pages 1–9, 1970. → pages 7
[HBS68] B. E. Howard, Z. Bramnick, and J. F. Shaw. Optimum curva-
ture principle in highway routing. In Proceeding of the Amer-
ican Society of Civil Engineers, volume 94, pages 61–82, 1968.
→ pages 9
54
Bibliography
[HHLM11] D. Hare, W. Hare, Y. Lucet, and S. Mehrotra. An arc-flow

model for earthwork operations. Technical report, (undis-
closed), 2011. → pages 8
[HHLR14] W. Hare, S. Hossain, Y. Lucet, and F. Rahman. Models and

strategies for efficiently determining an optimal vertical align-
ment of roads. Computers & Operations Research, 44:161 –
173, 2014. → pages 8, 16, 27
[HKL11] W. L. Hare, V. R. Koch, and Y. Lucet. Models and algorithms

to improve earthwork operations in road design using mixed
integer linear programming. European Journal of Operational
Research, 215(2):470 – 480, 2011. → pages 8
[Hog73] J. D. Hogan. Exprience with OPTLOC - optimum location

of highways by computer. In PTRC Seminer Proceddings on
Cost Models and Optimisation in Highways, 1973. → pages 10
[Hua73] J. P. Huarte. Opygar: Optimization and automatic design

of highway profiles. In PTRC Seminar Proceedings on Cost
Models and Optimization in Highways, 1973. → pages 7
[Jha03] M. K. Jha. Criteria-based decision support system for selecting

highway alignments. Journal of Transportation Engineering,
129(1):33–41, 2003. → pages 11
[JJS00] J.-C. Jong, M. K. Jha, and P. Schonfeld. Preliminary high-

way design with genetic algorithms and geographic informa-
tion system. Computer Aided Civil Infrastructure Engineering,
15(4):261–271, 2000. → pages 10
[JK06] M. K. Jha and E. Kim. Highway route optimization based

on accessibility, proximity, and land-use changes. Journal of
Transportation Engineering, 132:435–439, 2006. → pages 11
[Jon98] J.-C. Jong. Optimizing highway alignments with genetic al-

gorithms. PhD thesis, University of Maryland, College Park,
1998. → pages 10
[JS03] J.-C. Jong and P. Schonfeld. An evolutionary model for si-

multaneously optimizing 3d highway alignments. Tranporta-
tion Research, Part B: Methodology, 37B(2):107–128, 2003. →
pages 11
55
Bibliography
[JSJ06] M. K. Jha, P. Schonfeld, and J.-C. Jong. Intelligent Road

Design, volume 19. WIT Press, 2006. → pages 2, 3, 7, 8, 9,
10, 50
[KL10] V. R. Koch and Y. Lucet. A note on: Spline technique for
modeling roadway profile to minimize earthwork cost. Journal
of Industrial and Management Optimization, 6(2):393 – 400,
2010. → pages 8
[Kol05] T. G. Kolda. Revisiting asynchronous parallel pattern search
for nonlinear optimization. SIAM Journal on Optimization,
16(2):563–586, 2005. → pages 42
[LD11] S. Le Digabel. Algorithm 909: Nomad: Nonlinear optimization
with the mads algorithm. ACM Transactions on Mathematical
Software, 37(4):44:1–44:15, 2011. → pages 16, 19
[LTL09] Y. Lee, Y. Tsou, and H. Liu. Optimization method for high-
way horizontal alignment design. Journal of Transportation
Engineering, 135(4):217–224, 2009. → pages 10
[MA04] A. A. Moreb and M. S. Aljohani. Quadratic representation
for roadway profile that minimizes earthwork cost. Journal
of Systems Science and Systems Engineering, 13(2):245–252,
2004. → pages 8
[Mor96] A. A. Moreb. Linear programming model for finding optimal
roadway grades that minimize earthwork cost. European Jour-
nal of Operational Research, 93(1):148 – 154, 1996. → pages
8
[Mor09] A. A. Moreb. Spline technique for modeling roadway profile
to minimize earthwork cost. Journal of Industrial and Man-
agement Optimization (JIMO), 5(2):275–283, 2009. → pages
8
[MS81] R. H. Mayer and R. M. Stark. Earthmoving logistics. Journal
of the Construction Division, 107(2):297–312, 1981. → pages
7
[Mur73] J. D. Murchland. Methods of vertical profile optimisation for
an improvement to an existing road. In PTRC Seminar Pro-
ceedings on Cost Models and Optimisation in Highways, 1973.
→ pages 7
56
Bibliography
[NEW76] A. J. Nicholson, D.G. Elms, and A. Williman. A variational

approach to optimal route location. Highway Engineers, 23:22–
25, 1976. → pages 10
[Nic73] A. J. Nicholson. A variational approch to optimal route loca-

tion. PhD thesis, University of Canterbury, 1973. → pages 10,
50
[OEC73] OECD. Optimisation of road alignment by the use of com-

puters, 1973. Organisation for Economic Co-operation and
Development (Road Research Group), Paris. → pages 3, 10
[Par77] N. A. Parker. Rural highway route corridor selection. Trans-

portation Planning and Technology, 3(4):247–256, 1977. →
pages 9, 10, 50
[Pea73] A. D. Pearman. Investigation of the objective function in prob-

lems of optimising highway vertical alignment. In PTRC Sem-
inar Proceedings Cost Models and Optimisation in Highways,
1973. → pages 7
[Pla09] T. D. Plantenga. HOPSPACK 2.0 user manual. Technical

Report SAND2009-6265, Sandia National Laboratories, Al-
buqerque, NM and Livermore, CA, October 2009. → pages
41, 42
[Rah12] F. Rahman. Optimizing the vertical alignment under earth-

work block removal constraints in road construction. Master’s
thesis, The University of British Columbia, 2012. → pages 7,
8
[Rob73] R. Robinson. Automatic design of the road vertical alignment.

In PTRC Seminar Proceedings on Cost Models and Optimiza-
tion in Highways, 1973. → pages 7
[RS13] L. M. Rios and N. V. Sahinidis. Derivative-free optimization: a

review of algorithms and comparison of software implementa-
tions. Journal of Global Optimization, 56(3):1247–1293, 2013.
→ pages 16
[SH81] J. F. B. Shaw and B. E. Howard. Comparison of two integra-

tion methods in transportation routing. Transportation Re-
search Record, pages 8–3, 1981. → pages 9
57
Bibliography
[SH82] J. F. B. Shaw and B. E. Howard. Expressway route optimiza-

tion by OCP. Transportation Engineering Journal, 108:227–
243, 1982. → pages 9
[TM71] A. K. Turner and R. D. Miles. The gcars system: A computer

assisted method of regional route location. Highway Research
Record, pages 1–15, 1971. → pages 9
[Tri87a] D. Trietsch. Comprehensive design of highway networks. Tran-

portation Science, 21(1):1, 1987. → pages 9
[Tri87b] D. Trietsch. A family of methods for preliminary highway

alignment. Transportation Science, 21(1):17–25, 1987. →
pages 9
[TT03] C. W. Tat and F. Tao. Using GIS and genetic algorithm in

highway alignment optimization. In Intelligent Transportation
Systems, volume 2, pages 1563–1567, 2003. → pages 10
[Tur78] A. K. Turner. A decade of experience in computer aided route

selection. Photogrammetric Engineering and Remote Sensing,
pages 1561–1576, 1978. → pages 9
[Wan95] F. Y. M. Wan. Introduction to the calculus of variations and

its applications. Chapman & Hal, 1995. → pages 8
58
Appendix
59
Appendix A
Tables
A.1 Results for basic model
The numerical experiments were performed on the five different roads.
Tables A.1, A.2, A.3, A.4, and A.5 list the computational data for five
different input files (i.e., different roads). “ * ” indicates the problem cannot
be solved in 3 hours.
Table A.1: Computational Experience (Hart Rd Small.csv)
# of sta- # of # of itr. Wall- Memory

tion Fun. call clock
Time
(in Min) (in Mb)
5 509 53 1.8 0.10
10 1,256 70 5.6 0.35
15 1,998 81 11.9 0.73
20 3,517 1,119 21.0 1.56
25 4,734 131 29.7 2.48
30 6,702 175 53.4 4.09
35 8,011 173 57.3 5.56
40 11,213 211 91.7 8.57
45 16,413 285 137.9 13.86
50 16,182 257 149.8 14.78
60
A.1. Results for basic model
Table A.2: Computational Experience (Diamond Road align-1.csv)

Time
(in Min) (in Mb)
5 600 73 2.5 0.11
10 1,318 81 6.4 0.37
15 2,169 89 12.4 0.79
20 3,119 105 20.4 1.39
25 4,286 111 34.5 2.24
30 13,391 307 111.5 8.29
35 7,329 147 63.3 5.02
40 * * * *
45 * * * *
50 * * * *
Table A.3: Computational Experience (Diamond Road align-2.csv)

Time
(in Min) (in Mb)
5 2,810 359 13.8 0.57
10 1,787 107 7.7 0.50
15 2,715 117 16.0 0.98
20 3,690 115 28.0 1.60
25 5,424 147 42.6 2.87
30 20,032 447 177.3 12.60
35 * * * *
40 10,146 100.7 6,047 7.70
45 * * * *
50 * * * *
61
A.2. Optimized alignments of the test problems
Table A.4: Computational Experience (bluff road.csv)

Time
(in Min) (in Mb)
5 718 359 3.4 0.14
10 10,980 107 66.2 3.17
15 * * * *
20 6,207 207 38.9 2.74
25 9,281 275 66.6 4.87
30 17,141 407 146.3 10.60
35 * * * *
40 * * * *
45 * * * *
50 * * * *
Table A.5: Computational Experience (spur 3 demo.csv)

Time
(in Min) (in Mb)
5 728 83 3.4 0.14
10 1,727 101 8.2 0.48
15 3,099 127 19.3 1.12
20 19,268 627 144.1 8.79
25 6,987 189 50.0 3.61
30 10,609 245 85.5 6.41
35 13,625 285 121.21 9.34
40 * * * *
45 * * * *
50 * * * *
A.2 Optimized alignments of the test problems

The minimizer of the optimization problems ( developed using the model
described in Chapter 3) are listed in Tables A.6, A.7, A.8, A.9, and A.10.
62
Table A.6: The optimal alignments of the Road A obtained by the NOMAD and HOPSPACK solvers.
Opt. align.- Opt. align.- Opt. align.- Opt. align.- Opt. align.-
Initial Opt. align.-
Spec. HOPSPACK HOPSPACK HOPSPACK HOPSPACK HOPSPACK
alignment NOMAD
(run 1) (run 2) (run 3) (run 4) (run 5)
IP1-x 9,747.36 9,747.36 9,747.36 9,747.36 9,747.36 9,747.36 9,747.36
IP1-y 4,737.74 4,737.74 4,737.74 4,737.74 4,737.74 4,737.74 4,737.74
IP1-r 0.00 0.00 0.00 0.00 0.00 0.00 0.00
IP2-x 9,552.88 9,564.89 9,558.89 9,557.24 9,562.79 9,555.89 9,557.09
IP2-y 4,739.69 4,742.69 4,738.19 4,739.69 4,739.69 4,738.19 4,738.49
IP2-r 50.00 50.00 62.50 50.00 150.00 62.50 55.00
IP3-x 9,578.33 9,578.42 9,587.33 9,578.33 9,584.33 9,579.83 9,583.13
IP3-y 4,747.11 4,747.11 4,747.11 4,744.11 4,741.86 4,741.86 4,745.91
IP3-r 75.00 75.00 75.00 50.00 50.00 50.00 87.50
IP4-x 9,644.02 9,644.03 9,650.03 9,644.03 9,638.63 9,634.43 9,647.03
IP4-y 4,797.37 4,797.47 4,803.38 4,803.38 4,803.38 4,809.38 4,806.23
IP4-r 100.00 100.00 50.00 50.00 50.00 50.00 62.50
IP5-x 9,712.34 9,706.34 9,712.34 9,712.34 9,706.34 9,706.34 9,703.34
IP5-y 4,804.99 4,808.00 4,805.00 4,803.38 4,802.00 4,796.00 4,805.60
IP5-r 60.00 60.00 50.00 50.00 50.00 50.00 50.00
IP6-x 9,770.36 9,758.46 9,764.37 9,764.37 9,770.37 9,764.37 9,764.37
IP6-y 4,890.32 4,884.32 4,905.32 4,905.32 4,908.92 4,894.82 4,899.32
IP6-r 100.00 100.00 100.00 50.00 50.00 50.00 50.00
IP7-x 9,822.99 9,793.00 9,823.00 9,820.00 9,811.00 9,811.00 9,821.20
IP7-y 4,918.13 4,924.13 4,930.13 4,915.13 4,918.13 4,936.13 4,915.13
IP7-r 75.00 75.00 50.00 50.00 135.00 50.00 50.00
IP8-x 9,835.23 9,835.23 9,835.23 9,835.23 9,835.23 9,835.23 9,835.23
IP8-y 4,931.89 4,931.89 4,931.89 4,931.89 4,931.89 4,931.89 4,931.89
IP8-r 0.00 0.00 0.00 0.00 0.00 0.00 0.00
63
Table A.7: The optimal alignments of the Road B obtained by the NOMAD and HOPSPACK solvers.
alignment NOMAD
IP1-x 392,201.60 392,201.60 392,201.60 392,201.60 392,201.60 392,201.60 392,201.60
IP1-y 6,983,091.87 6,983,091.87 6,983,091.87 6,983,091.87 6,983,091.87 6,983,091.87 6,983,091.87
IP1-r 0.00 0.00 0.00 0.00 0.00 0.00 0.00
IP2-x 392,282.65 392,314.00 392,313.00 392,325.00 392,313.00 392,283.00 392,313.00
IP2-y 6,983,568.99 6,983,620.00 6,983,680.00 6,983,630.00 6,983,610.00 6,983,530.00 6,983,680.00
IP2-r 200.00 4,751.10 3,800.00 2,800.00 4,650.00 3,200.00 3,800.00
IP3-x 392,923.43 392,891.00 392,923.00 392,890.00 392,893.00 392,893.00 392,923.00
IP3-y 6,984,840.69 6,984,860.00 6,984,840.00 6,984,860.00 6,984,860.00 6,984,870.00 6,984,840.00
IP3-r 2,000.00 583.74 1,700.00 900.00 1,200.00 500.00 2,500.00
IP4-x 393,378.71 393,370.00 393,409.00 393,409.00 393,409.00 393,379.00 393,379.00
IP4-y 6,985,977.54 6,985,840.00 6,986,010.00 6,985,960.00 6,985,960.00 6,985,870.00 6,985,940.00
IP4-r 2,000.00 4,967.77 4,500.00 3,600.00 3,500.00 4,650.00 4,450.00
IP5-x 393,495.89 393,495.89 393,495.89 393,495.89 393,495.89 393,495.89 393,495.89
IP5-y 6,986,433.99 6,986,433.99 6,986,433.99 6,986,433.99 6,986,433.99 6,986,433.99 6,986,433.99
IP5-r 0.00 0.00 0.00 0.00 0.00 0.00 0.00
64
Table A.8: The optimal alignments of the Road C obtained by the NOMAD and HOPSPACK solvers.
alignment NOMAD
IP1-x 405,181.86 405,181.86 405,181.86 405,181.86 405,181.86 405,181.86 405,181.86
IP1-y 7,016,733.12 7,016,733.12 7,016,733.12 7,016,733.12 7,016,733.12 7,016,733.12 7,016,733.12
IP1-r 0.00 0.00 0.00 0.00 0.00 0.00 0.00
IP2-x 405,446.69 405,494.00 405,484.00 405,466.00 405,447.00 405,454.00 405,472.00
IP2-y 7,016,896.66 7,016,930.00 7,016,930.00 7,016,910.00 7,016,90.00 7,016,900.00 7,016,920.00
IP2-r 500.00 200.00 200.00 200.00 200.00 200.00 200.00
IP3-x 405,721.39 405,716.00 405,723.00 405,751.00 405,706.00 405,706.00 405,714.00
IP3-y 7,017,215.30 7,017,190.00 7,017,200.00 7,017,250.00 7,017,200.00 7,017,160.00 7,017,170.00
IP3-r 500.00 200.00 6,680.00 24,130.00 11,600.00 3,355.00 3,895.00
IP4-x 405,941.14 405,939.00 405,977.00 405,949.00 405,956.00 405,971.00 405,956.00
IP4-y 7,017,573.02 7,017,510.00 7,017,540.00 7,017,480.00 7,017,480.00 7,017,550.00 7,017,530.00
IP4-r 500.00 200.00 200.00 1,915.00 1,840.00 200.00 700.00
IP5-x 406,182.87 406,194.00 406,153.00 406,147.00 406,150.00 406,153.00 406,153.00
IP5-y 7,017,925.85 7,017,890.00 7,017,910.00 7,017,890.00 7,017,910.00 7,017,900.00 7,017,900.00
IP5-r 500.00 200.00 480.00 380.00 455.00 450.00 450.00
IP6-x 406,363.56 406,354.00 406,358.00 406,361.00 406,358.00 406,364.00 406,364.00
IP6-y 7,018,063.80 7,018,020.00 7,018,020.00 7,018,030.00 7,018,020.00 7,018,030.00 7,018,030.00
IP6-r 500.00 200.00 200.00 200.00 200.00 200.00 200.00
IP7-x 406,498.87 406,470.00 406,466.00 406,469.00 406,469.00 406,466.00 406,469.00
IP7-y 7,018,292.85 7,018,310.00 7,018,310.00 7,018,320.00 7,018,320.00 7,018,310.00 7,018,320.00
IP7-r 200.00 200.00 200.00 200.00 200.00 200.00 200.00
IP8-x 406,356.24 406,362.00 406,359.00 406,359.00 406,359.00 406,359.00 406,362.00
IP8-y 7,018,576.57 7,018,550.00 7,018,550.00 7,018,550.00 7,018,550.00 7,018,550.00 7,018,540.00
IP8-r 500.00 200.00 200.00 200.00 200.00 200.00 200.00
IP9-x 406,282.98 406,325.00 406,328.00 406,326.00 406,328.00 406,326.00 406,326.00
IP9-y 7,018,963.58 7,018,970.00 7,018,970.00 7,018,970.00 7,018,970.00 7,018,980.00 7,018,970.00
IP9-r 200.00 200.00 200.00 200.00 200.00 200.00 200.00
IP10-x 406,390.42 406,383.00 406,383.00 406,387.00 406,383.00 406,360.00 406,371.00
IP10-y 7,019,157.70 7,019,170.00 7,019,170.00 7,019,180.00 7,019,160.00 7,019,070.00 7,019,120.00
IP10-r 500.00 200.00 250.00 240.00 200.00 320.00 200.00
IP11-x 406,407.51 406,351.00 406,390.00 406,379.00 406,348.00 406,376.00 406,364.00
IP11-y 7,019,372.57 7,019,390.00 7,019,370.00 7,019,290.00 7,019,380.00 7,019,460.00 7,019,450.00
IP11-r 500.00 200.00 780.00 310.00 400.00 260.00 330.00
IP12-x 406,639.48 406,700.00 406,687.00 406,678.00 406,669.00 406,719.00 406,719.00
IP12-y 7,020,074.57 7,020,020.00 7,020,090.00 7,020,100.00 7,020,140.00 7,019,940.00 7,019,920.00

IP12-r 500.00 500.00 200.00 200.00 200.00 1,100.00 1,075.00
IP13-x 406,581.99 406,582.00 406,582.00 406,582.00 406,582.00 406,582.00 406,582.00
IP13-y 7,020,629.05 7,020,620.00 7,020,620.00 7,020,620.00 7,020,620.00 7,020,630.00 7,020,620.00
IP13-r 200.00 200.00 200.00 200.00 200.00 200.00 200.00
IP14-x 406,997.27 406,997.27 406,997.27 406,997.27 406,997.27 406,997.27 406,997.27
IP14-y 7,021,135.64 7,021,135.64 7,021,135.64 7,021,135.64 7,021,135.64 7,021,135.64 7,021,135.64
IP14-r 0.00 0.00 0.00 0.00 0.00 0.00 0.00
65
Table A.9: The optimal alignments of the Road D obtained by the NOMAD and HOPSPACK solvers.
alignment NOMAD
IP1-x 859,962.73 859,962.73 859,962.73 405,181.86 405,181.86 405,181.86 405,181.86
IP1-y 628,627.91 628,627.91 628,627.91 7,016,733.12 7,016,733.12 7,016,733.12 7,016,733.12
IP1-r 0.00 0.00 0.00 0.00 0.00 0.00 0.00
IP2-x 859,930.76 859,937.00 859,937.00 859,961.00 859,918.00 859,927.00 859,961.00
IP2-y 628,612.74 628,643.00 628,643.00 628,605.00 628,633.00 628,637.00 628,604.00
IP2-r 25.00 25.00 25.00 25.00 25.00 25.00 25.00
IP3-x 859,874.81 859,866.00 859,867.00 859,858.00 859,872.00 859,868.00 859,849.00
IP3-y 628,605.96 628,576.00 628,578.00 628,599.00 628,576.00 628,576.00 628,601.00
IP3-r 90.00 25.00 25.00 25.00 25.00 25.00 45.00
IP4-x 859,822.25 859,831.00 859,822.00 859,820.00 859,827.00 859,826.00 859,822.00
IP4-y 628,629.40 628,638.00 628,649.00 628,641.00 628,637.00 628,641.00 628,642.00
IP4-r 60.00 25.00 25.00 25.00 25.00 25.00 25.00
IP5-x 859,773.55 859,782.00 859,787.00 859,789.00 859,790.00 859,780.00 859,788.00
IP5-y 628,618.10 628,636.00 628,625.00 628,624.00 628,630.00 628,633.00 628,625.00
IP5-r 60.00 60.00 25.00 25.00 48.00 25.00 25.00
IP6-x 859,696.78 859,676.00 859,698.00 859,702.00 859,699.00 859,674.00 859,701.00
IP6-y 628,659.71 628,650.00 628,681.00 628,690.00 628,681.00 628,651.00 628,690.00
IP6-r 90.00 25.00 92.50 99.00 115.00 25.00 101.50
IP7-x 859,627.65 859,631.00 859,632.00 859,629.00 859,629.00 859,629.00 859,629.00
IP7-y 628,620.61 628,612.00 628,613.00 628,614.00 628,612.00 628,612.00 628,613.00
IP7-r 25.00 25.00 25.00 25.00 25.00 25.00 25.00
IP8-x 859,505.79 859,503.00 859,500.00 859,499.00 859,500.00 859,499.00 859,499.00
IP8-y 628,614.11 628,607.00 628,608.00 628,609.00 628,608.00 628,608.00 628,609.00
IP8-r 25.00 25.00 25.00 25.00 25.00 25.00 25.00
IP9-x 859,466.24 859,463.00 859,468.00 859,470.00 859,468.00 859,470.00 859,471.00
IP9-y 628579.36 628,574.00 628,574.00 628,576.00 628,573.00 628,575.00 628,577.00
IP9-r 25.00 25.00 96.25 93.75 97.50 90.00 90.00
IP10-x 859,404.57 859,412.00 859,414.00 859,412.00 859,414.00 859,414.00 859,412.00
IP10-y 628,565.11 628,563.00 628,565.00 628,565.00 628,565.00 628,565.00 628,565.00
IP10-r 25.00 25.00 25.00 25.00 27.50 27.50 25.00
IP11-x 859,343.95 859,346.00 859,351.00 859,360.00 859,347.00 859,348.00 859,359.00
IP11-y 628553.75 628,534.00 628,537.00 628,543.00 628,535.00 628,535.00 628,543.00
IP11-r 25.00 25.00 25.00 25.00 25.00 25.00 25.00
IP12-x 859,305.28 859,315.00 859,312.00 859,277.00 859,316.00 859,313.00 859,281.00
IP12-y 628,557.69 628,556.00 628,556.00 628,560.00 628,556.00 628,558.00 628,559.00
IP12-r 25.00 25.00 25.00 25.00 25.00 28.75 25.00

IP13-x 859,273.63 859,267.00 859,270.00 859,265.00 859,269.00 859,268.00 859,266.00
IP13-y 628573.70 628,550.00 628,550.00 628,571.00 628,551.00 628,548.00 628,566.00
IP13-r 25.00 25.00 25.00 25.00 25.00 25.00 25.00
IP14-x 859,255.50 859,263.00 859,262.00 859,262.00 859,262.00 859,263.00 859,263.00
IP14-y 628,585.89 628,598.00 628,590.00 628,601.00 628,595.00 628,590.00 628,587.00
IP14-r 25.00 25.00 55.00 25.00 38.75 25.00 42.50
continue to next page
66
continue from previous page
alignment NOMAD
IP15-x 859,258.87 859272.00 859,277.00 859,271.00 859,271.00 859,271.00 859,266.00
IP15-y 628,622.27 628,613.00 628,634.00 628,612.00 628,610.00 628,624.00 628,604.00
IP15-r 25.00 25.00 62.50 68.75 25.00 25.00 32.50
IP16-x 859,312.40 859,292.00 859,300.00 859,294.00 859,289.00 859,294.00 859,295.00
IP16-y 628,630.18 628,648.00 628,649.00 628,650.00 628,648.00 628,648.00 628,650.00
IP16-r 25.00 25.00 25.00 37.50 37.50 38.75 35.00
IP17-x 859,370.11 859,368.00 859,361.00 859,358.00 859,362.00 859,371.00 859,361.00
IP17-y 628,640.65 628,656.00 628,655.00 628,653.00 628,655.00 628,657.00 628,653.00
IP17-r 90.00 25.00 25.00 25.00 25.00 25.00 25.00
IP18-x 859,400.45 859,405.00 859,392.00 859,391.00 859,393.00 859,397.00 859,396.00
IP18-y 628,690.04 628,720.00 628,694.00 628,695.00 628,699.00 628,720.00 628,697.00
IP18-r 25.00 25.00 56.25 36.25 25.00 203.12 561.25
IP19-x 859,440.98 859,469.00 859,429.00 859,430.00 859,425.00 859,452.00 859,429.00
IP19-y 628,757.15 628,772.00 628,757.00 628,759.00 628,758.00 628,763.00 628,756.00
IP19-r 25.00 25.00 25.00 25.00 25.00 25.00 25.00
IP20-x 859,472.96 859,449.00 859,489.00 859,496.00 859,461.00 859,473.00 859,488.00
IP20-y 628,811.20 628,799.00 628,782.00 628,781.00 628,800.00 628,807.00 628,781.00
IP20-r 50.00 25.00 25.00 25.00 50.00 32.50 25.00
IP21-x 859,418.43 859,410.00 859,414.00 859,404.00 859,412.00 859,415.00 859,415.00
IP21-y 628,848.14 628,832.00 628,827.00 628,830.00 628,837.00 628,822.00 628,826.00
IP21-r 25.00 25.00 25.00 63.75 25.00 25.00 25.00
IP22-x 859,404.20 859,404.20 859,404.20 859,404.20 859,404.20 859,404.20 859,404.20
IP22-y 628,868.81 628,868.81 628,868.81 628,868.81 628,868.81 628,868.81 628,868.81
IP22-r 0.00 0.00 0.00 0.00 0.00 0.00 0.00
67
Table A.10: The optimal alignments of the Road E obtained by the NOMAD and HOPSPACK solvers.
alignment NOMAD
IP1-x 860,254.98 859,962.73 859,962.73 859,962.73 859,962.73 859,962.73 859,962.73
IP1-y 628,612.83 628,627.91 628,627.91 628,627.91 628,627.91 628,627.91 628,627.91
IP1-r 0.00 0.00 0.00 0.00 0.00 0.00 0.00
IP2-x 860,219.54 860,207.00 860,202.00 860,203.00 860,202.00 860,202.00 860,202.00
IP2-y 628,747.01 628,777.00 628,788.00 628,788.00 628,788.00 628,788.00 628,788.00
IP2-r 50.00 50.00 50.00 50.00 50.00 50.00 50.00
IP3-x 860,179.37 860,179.00 859,866.00 860,179.00 860,179.00 860,179.00 860,179.00
IP3-y 628,809.75 628,817.00 628,811.00 628,813.00 628,812.00 628,811.00 628,812.00
IP3-r 50.00 50.00 50.00 50.00 50.00 50.00 50.00
IP4-x 860177.93 860,180.00 860,178.00 860,178.00 860,178.00 860,178.00 860,178.00
IP4-y 628,861.00 628,856.00 628,849.00 628,858.00 628,854.00 628,849.00 628,865.00
IP4-r 50.00 50.00 117.50 170.00 100.00 102.50 110.0
IP5-x 860,181.57 860,190.00 860,182.00 860,188.00 860,188.00 860,182.00 860,188.00
IP5-y 628,896.38 628,897.00 628,878.00 628,893.00 628,896.00 628,878.00 628,892.00
IP5-r 50.00 50.00 220.00 112.50 50.00 235.00 122.50
IP6-x 860199.29 860,204.00 860,199.00 860,205.00 860,199.00 860,199.00 860,205.00
IP6-y 628,959.03 628,967.00 628,935.00 628,971.00 628,935.00 628,935.00 628,973.00
IP6-r 50.00 50.00 330.00 198.75 320.00 312.50 205.0
IP7-x 860,181.57 860,203.00 860,206.00 860,202.00 860206.00 860206.00 860203.00
IP7-y 629059.67 629,050.00 629,030.00 629,028.00 629,036.00 629,039.00 629,029.00
IP7-r 50.00 50.00 665.00 237.50 412.50 340.0 247.50
IP8-x 860,225.87 860,230.00 860,226.00 860,227.00 860,227.00 860,227.00 860,227.00
IP8-y 629,164.10 629,159.00 629,164.00 629,163.00 629,163.00 629,163.00 629,163.00
IP8-r 50.00 50.00 50.00 50.00 50.00 50.00 50.00
IP9-x 860,146.12 860,146.12 860,146.12 860,146.12 860,146.12 860,146.12 860,146.12
IP9-y 629,280.56 629,280.56 629,280.56 629,280.56 629,280.56 629,280.56 629,280.56
IP9-r 0.00 0.00 0.00 0.00 0.00 0.00 0.00
68
Appendix B
Figures
6
x 10
6.9865
Optimized Alignment−NOMAD (Cost: $ 15198.8)
Optimized Alignment−HOPSPACK (Cost: $ 15128)
6.986
6.9855
6.985
Y
6.9845
6.984
6.9835
6.983
3.92 3.922 3.924 3.926 3.928 3.93 3.932 3.934 3.936
X 5
x 10
Figure B.1: Optimum alignments of the Road B obtained by the NOMAD

solver and the HOPSPACK solver.
69
Appendix B. Figures
6
x 10
7.0215
Optimized Alignment−NOMAD (Cost: $ 69621)
7.021
7.0205
7.02
7.0195
7.019
Y
7.0185
7.018
7.0175
7.017
7.0165
4.05 4.055 4.06 4.065 4.07 4.075
X 5
x 10
Figure B.2: Optimum alignments of the Road C obtained by the NOMAD

70
Appendix B. Figures
5
x 10
6.294
Optimized Alignment−NOMAD (Cost: $ 6497.96)
6.293
6.292
6.291
6.29
Y
6.289
6.288
6.287
6.286
8.601 8.6015 8.602 8.6025 8.603 8.6035
X 5
x 10
Figure B.3: Optimum alignments of the Road E obtained by the NOMAD

71

Softree 3

Uploaded by

Copyright:

Available Formats

Softree 3

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Softree 3

Uploaded by

Copyright:

Available Formats

Horizontal Alignment Optimization in

B.Sc., Rajshahi University of Engineering & Technology, 2010

A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF

THE COLLEGE OF GRADUATE STUDIES

(Interdisciplinary Studies - Optimization)

THE UNIVERSITY OF BRITISH COLUMBIA

Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

Chapter 2: A basic approach to solve the horizontal align-

Chapter 3: Horizontal alignment optimization model . . . . . 20

Chapter 4: Numerical results . . . . . . . . . . . . . . . . . . . 41

Table 1.1 Classification of the construction costs. . . . . . . . . . 4

Table 4.1 Specifications of the test problems . . . . . . . . . . . 41

Table A.1 Computational Experience (Hart Rd Small.csv) . . . . 60

Figure 1.1 A three dimensional alignment (blue curve) showing

Figure 2.1 Corridor of a horizontal alignment. Base data points

Figure 3.1 Geometric representation of horizontal alignment. The

Figure 3.4 Horizontal alignment in a specified corridor. The red

Figure 4.1 A non-backtracking alignment (the test problem as-

Figure B.1 Optimum alignments of the Road B obtained by the

Dedicated to underprivileged talents of the world.

In the literature, many studies proposed computerized models to find the

1.2 Road design optimization

– planning and administrative cost,

– user cost, and

– social and environment cost.

Planning and administrative costs are not considered in the alignment

Table 1.1: Classification of the construction costs.

Cost components Contributions ( %)

1.2.1 Earthwork optimization

Cut area Fill area

Ground profile Road profile

Figure 1.2: Example of cut and fill in road construction.

1.2.2 Vertical alignment optimization

1.2.3 Horizontal alignment optimization

Figure 1.3: Two potential vertical alignments of a fixed horizontal alignment.

Figure 1.4: Some potential horizontal alignments in a specified corridor.

A horizontal alignment consists of some tangential line segments followed

the safety requirement.

1.3 Background and literature review

1.3.1 Earthwork and vertical alignment optimization model

Easa [Eas88] developed a linear programming model considering the

1.3.2 Horizontal alignment optimization model

A basic shortcoming of the network optimization models is that the

1.3.3 Three dimensional alignment optimization model

ment of forest roads.

1.4 Our research approach

1.5 Organization of the Thesis

A basic approach to solve

Base - offset Base Base + offset

2.2 Problem Formulation

Figure 2.2: Three dimensional alignment.

min f (X) = CVA (X)

alignment for a horizontal offset X. So the objective function of the pro-

2.3 Solution Approach

The algorithm takes a three dimensional alignment data as the input

Figure 2.4: Flowchart of the solution approach.

2.4 Numerical Results

Figure 2.6: Growth of no. of function calls with respect to number of

3.1 Geometric representation of horizontal