Models and strategies for efficiently

optimizing the vertical alignment of

roads for multimaterial
by

Shahadat Hossain

B.Sc., Bangladesh University of Engineering and Technology, 2009

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 2013
c Shahadat Hossain, 2013
Abstract
Selecting optimal vertical alignment while satisfying safety and design
constraints is an important task during road construction. The amount of
earthwork operations depends on the design of a vertical alignment, so a
good vertical alignment can have a profound impact on construction costs.
In “Optimizing the vertical alignment under earthwork block removal con-
straints in road construction” M.Sc. thesis 2012, F. Rahman proposed two
models for vertical alignment optimization. However, the models were lim-
ited to a single material and the volume of material needed to be a convex
function of the distance from the ground profile. Consequently, these models
cannot be used in commercial road design softwares. In this research, we
remove the previous limitations so the models can now be used for road de-
sign software. Both models use a piecewise quadratic curve to minimize the
cost for earthwork operations. The models consider several new features as
well as all the features of the previous models that are described in previous
research by F. Rahman. Moreover, we improve the solution time, so that
it is possible to make interactive design tools. We provide numerical tests
that validate our results.

ii
Preface
Models and strategies for efficiently optimizing the vertical alignment
of roads for single material part of this thesis has been adapted from a
manuscript co-authored with Dr. Warren Hare, Dr. Yves Lucet, and Faisal
Rahman at the University of British Columbia, Okanagan campus. The
manuscript has been submitted for publication to the Journal of Computers
and Operations Research [HHLR13]. The paper included models and tech-
niques for single material only. A US provisional patent was also filed based
on the results contained in the present thesis.

iii
Table of Contents

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . iv

List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . ix

Dedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x

Chapter 1: Introduction . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Road design problem . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Horizontal alignment . . . . . . . . . . . . . . . . . . . 2
1.2.2 Vertical alignment . . . . . . . . . . . . . . . . . . . . 4
1.2.3 3D alignment . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Our research focus . . . . . . . . . . . . . . . . . . . . . . . . 7

Chapter 2: The quadratic model for vertical alignment . . . . 10

2.1 Model variables and parameters . . . . . . . . . . . . . . . . . 10
2.2 The quadratic model . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.1 The quadratic model for single material . . . . . . . . 16
2.2.2 The quadratic model for multimaterial . . . . . . . . . 16

Chapter 3: Quasi network flow model for vertical alignment . 23

3.1 Quasi network flow for single material . . . . . . . . . . . . . 23
3.2 Quasi network flow for multimaterial . . . . . . . . . . . . . . 23

iv
 TABLE OF CONTENTS

Chapter 4: Speedup techniques . . . . . . . . . . . . . . . . . . 31

4.1 Techniques with offset levels . . . . . . . . . . . . . . . . . . . 31
4.1.1 Reducing binary variables for offset levels . . . . . . . 31
4.1.2 Modeling volume constraints with SOS2 variables . . . 33
4.2 Techniques with blocks . . . . . . . . . . . . . . . . . . . . . . 34
4.2.1 Block constraints with SOS1 variables . . . . . . . . . 34
4.2.2 Reducing the number of binary variables for blocks . . 35

Chapter 5: Numerical results . . . . . . . . . . . . . . . . . . . 40

5.1 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . 40
5.2 Performance measurement . . . . . . . . . . . . . . . . . . . . 42
5.3 Experimental results for single material . . . . . . . . . . . . 42
5.3.1 Results for problems with blocks . . . . . . . . . . . . 43
5.3.2 Results for problems with no blocks . . . . . . . . . . 47
5.3.3 Summary of experimental result for single material . . 47
5.4 Experimental results for multimaterial . . . . . . . . . . . . . 51
5.4.1 Results for problems with blocks . . . . . . . . . . . . 51
5.4.2 Results for problems without blocks . . . . . . . . . . 52
5.4.3 Summary of results for mutlimaterial . . . . . . . . . . 54

Chapter 6: Conclusion . . . . . . . . . . . . . . . . . . . . . . . . 55
6.1 Future research . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.1.1 Improvement in the quasi network flow model . . . . . 55
6.1.2 More options for selecting material . . . . . . . . . . . 56
6.1.3 Horizontal alignment . . . . . . . . . . . . . . . . . . . 56

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Appendix A: Tables . . . . . . . . . . . . . . . . . . . . . . . . . . 63

v
List of Tables

Table 2.1 Cut and fill volume of material . . . . . . . . . . . . . 14

Table 5.1 List of strategies for single material with blocks . . . . 41

Table 5.2 List of strategies for single material without blocks . . 47
Table 5.3 Precedence of techniques for CPLEX . . . . . . . . . . 50
Table 5.4 Precedence of techniques for CBC . . . . . . . . . . . . 50
Table 5.5 List of strategies without blocks for multimaterial test
cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Table 5.6 List of strategies with blocks for multimaterial test cases 53

Table A.1 Solution time by CPLEX for single material without

blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Table A.2 Solution time by CBC for single material without blocks 64
Table A.3 Solution time by CPLEX for single material with blocks 65
Table A.4 Solution time by CBC for single material with blocks . 66
Table A.5 Solution time by CPLEX for multimaterial with blocks 67
Table A.6 Solution time by CPLEX for multimaterial without
blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Table A.7 Test road configuration . . . . . . . . . . . . . . . . . . 69
Table A.8 Test configurations in terms of number . . . . . . . . . 70
Table A.9 Test configuration: (pattern of blocks, access roads) . 72
Table A.10 Sample input table (for multi-material) . . . . . . . . . 73

vi
List of Figures

Figure 1.1 Horizontal alignment . . . . . . . . . . . . . . . . . . 3

Figure 1.2 Vertical alignment . . . . . . . . . . . . . . . . . . . . 5

Figure 2.1 Side view of ground profile and road profile . . . . . . 10

Figure 2.2 Spline segment of vertical alignment. . . . . . . . . . 12
Figure 2.3 Cross-section of a road for a cut . . . . . . . . . . . . 13
Figure 2.4 Cross-section of a road for a fill . . . . . . . . . . . . 13
Figure 2.5 Approximation of side-slopes for a cut . . . . . . . . . 13
Figure 2.6 Approximation of side-slopes for a fill . . . . . . . . . 14
Figure 2.7 Volume approximation of side-slope . . . . . . . . . . 15
Figure 2.8 Stockpiling scenario during cut and fill . . . . . . . . 20

Figure 3.1 A typical section in quasi network flow model . . . . 24

Figure 3.2 A typical section in quasi network flow model with
blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

Figure 5.1 Performance profile for solver CPLEX (with blocks) . 43

Figure 5.2 Performance profile for solver CBC (with blocks) . . . 44
Figure 5.3 Performance profile of quadratic model for CPLEX
(with blocks) . . . . . . . . . . . . . . . . . . . . . . . 45
Figure 5.4 Performance profile of quadratic model for CBC (with
blocks) . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Figure 5.5 Performance profile of quasi network flow model for
CPLEX (with blocks) . . . . . . . . . . . . . . . . . . 46
Figure 5.6 Performance profile of quasi network flow model for
CBC (with blocks) . . . . . . . . . . . . . . . . . . . . 46
Figure 5.7 Performance profile for solver CPLEX (without blocks) 48
Figure 5.8 Performance profile for solver CBC (without blocks) . 48
Figure 5.9 Performance profile for solver CPLEX for multi-material
(with blocks) . . . . . . . . . . . . . . . . . . . . . . . 52

vii
 LIST OF FIGURES

Figure 5.10 Performance profile for solver CPLEX for multi-material

(without blocks) . . . . . . . . . . . . . . . . . . . . . 53

viii
Acknowledgements
I would like to express my sincere gratitude to my two supervisors, Dr.
Yves Lucet and Dr. Warren Hare, for their continuous help and support in
all stages of this research. I would like to thank them for being open and
encouraging to new ideas, listening to my ideas, and helping me to organize
those in proper way. Their knowledge, guidance, and patience were essential
for this thesis. It was a great privilege to have such friendly supervisors that
helped me to overcome many challenges and enjoy my graduate life.
A very special thanks goes to our industrial partner (Softree Technical
Systems Inc.) for providing us all the help that we need. In particular
David Mills, Craig Speirs, and Alexis Guigue were always very supportive
and helpful to give technical information for this thesis. The week that I
worked with them in Vancouver helped me a lot to organize my thesis.
I would like to thank everyone who directly or indirectly helped me in
my research, course works, and in other activities during my graduate life.
I owe my deepest gratitude to my parents, and family members for their
prayers and support throughout my life. Without their selfless and contin-
uous help I would not be able to come here.
I also like to thank all of my friends who helped me to pass some good
times in Kelowna, took me to camping, and gave me rides to several places.
This research is partially funded by an Collaborative Research and Devel-
opment (CRD) Grants from the Natural Sciences and Engineering Research
Council (NSERC) of Canada, MITACS, and Softree Technical Systems Inc.

ix
Dedication

To Rusdania Bushra Asha.

x
Chapter 1

Introduction
In this chapter, we describe our motivation for research in road design,
define the road design problem, and discuss some previous works in road
design.

1.1 Motivation
The state of the road network system is one of the strongest indica-
tor of a country’s level of development. So every year all countries spend
a significant amount of their budget on construction and maintenance of
roads and highways. For example, the Canadian government in 2011 allo-
cated $32,000,000,000 for roadway transportation [Min12a]. A significant
amount of this budget is spent on building new roads and maintaining ex-
isting roads. The cost of constructing a road without considering the cost of
engineering design or property acquisition, can be in the range of $2,100,000
to $5,000,000 per kilometer for two lanes [Min12b] . As a result, the gov-
ernment and the companies who build and maintain roads and highways
are interested to find a way to minimize the cost of road construction and
maintenance.
While constructing roads and highways, cost minimization is not the
only issue the road and highway engineers have to deal with. Every year
thousands of road collisions and casualties happen. These accidents cost
a lot, not only in terms of money but also in terms of life. So it is very
important to consider safety requirements, rules, and regulations while de-
signing roads to minimize accidents. Furthermore, road engineers also need
to consider the environmental impacts [Jha03] and the socioeconomic effects
[DJ09].
Designing roads or highways to minimize costs during the construction
process while satisfying safety requirements, rules, and regulations consid-
ering all socioeconomic and environmental impacts is a very difficult task.
Thus automating the road design process to the optimal design could save
a lot of money. Unfortunately there is no road design software in existence
today that has an automated and interactive way for computing optimal

1
 1.2. Road design problem

design of roads or highways considering multiple materials and all practical

design issues. In this research we work to address these issues and take a
step towards building a software tool for computing a mathematically proven
optimal design of a road.

1.2 Road design problem

Road design refers to the problem of connecting some given end points
by selecting an economical alignment while satisfying various design spec-
ifications, safety constraints, and considering environmental and socioeco-
nomic impacts [JS03], [JSJ06]. Minimizing the cost of road construction
spans a wide range of road types: from highways to mountain roads. It is a
sub-problem in forest management [CMW98], and in the design of efficient
highway networks [ACRV13].
The problem of designing roads can be split into three interrelated sub-
problems [AAS04]. First, the horizontal alignment provides the road trajec-
tory from a satellite’s eye view considering political and social issues [KSJ07].
Then, the vertical alignment is a modification of the ground profile that
minimizes the cost of construction while satisfying safety and regulation
constraints [Mor09]. Finally, a solution to the earthwork problem describes
how material is rearranged optimally to build the desired vertical alignment
at minimal cost [HKL11]. For computational efficiency, many researchers
combine the last two stages together to design optimal vertical alignments
while minimizing earthwork costs.
Since two given points can be connected in numerous possible ways,
selecting the best alignment quickly from many potential alignments is a
difficult task. Manual processes of designing roads take hours of skilled
labor and there is no guarantee on the optimality of the final design. In this
research we propose models and strategies to calculate the vertical alignment
(incorporating the earthwork problem) that can be solved by deterministic
optimization algorithms.

1.2.1 Horizontal alignment

A horizontal alignment is a route from source to destination; it follows a
topographical map and satisfies various design code requirements [AAS04].
It should also consider several socioeconomic and political issues [KSJ07]
, e.g., passing through specific town and staying away from hospitals and
schools. The cost of designing a horizontal alignment depends on several

2
 1.2. Road design problem

factors such as land cost, travel cost, and on the vertical alignment cost dis-
cussed next. A sample horizontal alignment is shown in Figure 1.1. Horizon-
tal alignment optimization is usually done using one of the three techniques:
calculus of variation, dynamic programming, and network optimization.

Figure 1.1: Horizontal alignment as visualized in RoadEng road design soft-

ware.

In calculus of variation, one tries to find a curve connecting two end

points in space which minimizes the integral of a cost function ([JSJ06],
[Wan95]). The horizontal alignment problem is similar to the typical prob-
lem in calculus of variation to some extent. This method is used to determine
horizontal alignment in [SH81], [SH82], and [HBS68]. This methods needs
to determine continuous local cost function. Since the cost depends on var-
ious factors, some approximation and assumptions are required to calculate

3
 1.2. Road design problem

the cost function. However, this method produces a continuous road, and a
global optimum is guaranteed.
Dynamic programming is a well known method for optimization. Mostly
this technique is used for discrete problems. Few studies ([OEC73], [Tri87b],
[JSJ06]) showed the optimization of a horizontal alignment using a dynamic
programming method. The problem with this approach is that it takes a
long time to solve large problems and the output alignment is non-smooth.
Network optimization method formulate the horizontal alignment prob-
lem as a network problem such as a shortest path problem, a min cost flow
problem, etc. Some early papers [TM70], [Tur78], [Tri87a] used network
optimization methods to optimize a horizontal alignment. This method is
simple and easy to understand, but formulating the cost function for each
edge of the network considering the cost of vertical alignment is hard and
requires significant approximations.

1.2.2 Vertical alignment

After computing a horizontal alignment, one can draw a curve between
the two given points and get the vertical ground profile along this curve
([HKL11]). An example of a vertical alignment is shown in Figure 1.2. In
vertical alignment optimization, a road profile is fitted to the ground profile
minimizing the cost of construction while satisfying safety and regulation
constraints. It determines which hill to cut or ravine to fill to construct a
road. The costs associated with the vertical alignment are the cut and fill
costs of materials. Once a vertical alignment is determined, the earthwork
optimization minimizes the cost of moving materials to build the desired
vertical alignment.
Vertical alignment optimization [Eas88], [GCF88], [Mor96], [LC01],
[FCS02], [MA04], [GLA05], [GAA09], [GLA09], [Mor09], [KL10], [HKL11] of
roads has been modeled by several researchers considering many constraints.
As the earthwork cost depends on the design of the vertical alignment, it is
common to combine the two stages together to design a vertical alignment
while considering the earthwork.
Nondeterministic methods to compute a vertical alignment were con-
sidered. Lee and Cheng [LC01] proposed a three layered heuristic method
for vertical alignment optimization. They used several layers to convert a
nonlinear mixed integer linear program to a mixed integer linear program.
Their model was nonlinear at first because of the use of crest and sag curve
equations to write the constraints on slopes limit. Fwa et al. [FCS02] and
Goktepe et al. [GLA09] presented genetic algorithms formulation of the ver-

4
 1.2. Road design problem

Height (meter)

Road length (meter)

Figure 1.2: Vertical alignment as visualized in RoadEng road design soft-

ware. The solid line is the ground profile and the dotted line is the road
profile.

tical alignment problem. A detailed model solved with a genetic algorithm

can be found in [JSJ06].
Several deterministic algorithms were proposed that take different con-
straints into account. Easa [Eas88] proposed a model for vertical alignment
considering earthwork costs. It enumerates all possible combinations of fea-
sible grades, and then for each grade it calculates the cut and fill using a
linear program to minimize earthwork operations. It is a trial and error pro-
cedure. Goh et al. [GCF88] modeled the optimal vertical alignment using
dynamic programming considering several costs such as cut, fill, pavement,
and land acquisition costs.
Moreb [Mor96] proposed a model combining vertical alignment and
earthwork operation together in a single linear program. However, it gives
a piecewise linear segments for the vertical alignment, which is not desir-
able. Moreb and Aljohani [MA04] improves the previous work of Moreb
[Mor96] by representing the road profile as a quadratic spline. Goktepe
et al. [GLA05] proposed a dynamic programming model to solve the ver-
tical alignment problem and later in [GAA09] they extended the previous
model by combining it with the Weighted Ground Line Method for cut-
fill balancing and earthwork minimization. Moreb ([Mor09]) also proposed
an improved linear program that solved the problem of sharp connectivity
of piecewise linear segments by adding more constraints. This model gen-
eralized the technique with a piecewise polynomial of any degree. Koch

5
 1.2. Road design problem

and Lucet ([KL10]) improved Moreb’s model ([Mor09]) by changing two

constraints to avoid unnecessary errors. They also showed that up to a
quadratic spline it is possible to maintain the linear structure of the model.
But they did not consider side slopes and blocks into the model.
Hare, Koch, and Lucet ([HKL11]) proposed a mixed integer linear pro-
gramming model for earthwork optimization considering blocks. They were
the first to introduce the block concept during earthwork optimization and
determined an optimal schedule for material movement. A block is an ob-
stacle such as mountains, rivers, etc. that needs to be dealt with before
materials can be moved over the location of the block. In [HLR13], Hare,
Lucet, and Rahman proposed a mixed integer linear programming model
for vertical alignment considering side slopes and blocks. Side slopes refer
to the gradual increase or decrease of the height from the road profile to
the ground profile. When blocks are not considered, Ahuja et al. [AMO93,
p. 12-13] gave a network flow model for earthwork optimization.
Rahman [Rah12] proposed two models for vertical alignment while min-
imizing the earthwork cost. He extended the work of Moreb [Mor09], Koch
and Lucet [KL10], and Hare et al. [HKL11]. The two models are called the
quadratic model and the quasi network flow model. The quadratic model
is a mixed integer linear program, and the quasi network flow model is a
modification of a minimum cost network flow problem that turns it into an-
other mixed integer linear program with significantly less variables. These
two models are the most recent and constitute the state-of-the-art for solv-
ing vertical alignment problems using deterministic optimization algorithms.
Sides slopes and blocks are considered in both models but only when a single
material is considered, which is not realistic and prevents the model from
being able to calculate a good approximation of the real construction cost.

1.2.3 3D alignment
In the ideal case, designing a three dimensional alignment of roads by
considering the above mentioned horizontal, vertical alignment, and earth-
work sub problems together would be the most efficient strategy. However,
calculating 3D alignment of a road, considering all of these simultaneously,
is a very difficult problem. Most of the researchers use heuristic based algo-
rithms to calculate 3D alignments.
Chew et al. [CGF89] proposed a three dimensional highway alignment
optimization model that determines the horizontal and vertical alignments
simultaneously. They applied the concept of optimal control theory and
used cubic spline polynomial interpolation to build a smooth vertical align-

6
 1.3. Our research focus

ment. The model is a general constrained non-linear optimization problem

and cannot guarantee global optimality. Tat and Tao [TT03] developed a
model for 3D alignment using genetic algorithms. It considers horizontal
and vertical curvature, and gradient constraints simultaneously using GIS
data. It also uses a cost model which depends on location, volume, length,
etc. Akay [Aka06] proposed a method to design 3D alignments for forest
roads using a combination of simulated annealing algorithm and linear pro-
gramming. This model considers various soil types along the road and takes
into account the safety of the drivers. But it manually selects an initial
trial route. Cheng and Lee [CL06] proposed an iterative process splitting
the whole process into three stages. In each iteration, a new horizontal
alignment is obtained by slightly adjusting the current alignment. A cor-
responding vertical alignment is determined by solving a series of mixed
integer linear program. Aruga [Aru05] uses a tabu search method using a
high-resolution digital elevation model and later Aruga et al. [ATSM06]
improved the model using Dijkstra’s shortest path algorithm and a cubic
spline function.
Models of the 3D alignment of highways are subject to nonlinear con-
straints. They have to deal with multiple design alternatives and various
design parameters [GLA05]. As a result, most researchers first determine
horizontal alignment and then compute a more precise vertical alignment.

1.3 Our research focus

Our research is focused on vertical alignment optimization for road de-
sign. We have already discussed above some research works for solving the
vertical alignment problem. Several research works emphasized different
issues and tried to model them accordingly. We worked closely with our
industry partner Softree Technical Systems Inc.1 who develops road design
software. We have consulted with professional road engineers and learned
what needs to be considered for a practical vertical alignment problem. The
following requirements were found most relevant for modeling the vertical
alignment problem:

− Multi-material: A typical road consists of several materials. The cost

of road construction depends on these various types of materials. So
we need to build a model which can consider more than one material.
1
http://www.softree.com

7
 1.3. Our research focus

Multi-material can be defined by both different quality of the same

material and different materials.

− Side slopes: Side-slopes needs to be considered to obtain an accept-

able material volume approximation. Side slopes refer to the gradual
increase or decrease of the height from the road profile to the ground
profile. Discarding them could introduces an unacceptable error in
cost calculation.

− Cut and fill at the same section: Cut and fill refers to the excavation
and embankment of materials to make a smooth surface for the road
construction. As the earth is uneven in shape, there could be cut
and fill at the same section of the same material. Section refers to
the smallest unit of a road, and there is no cost for moving materials
within it.

− Blocks: There could be blocks such as rivers, rocky mountains that

needs to be considered, because it is not possible to move materials
across or in between those blocks before we deal with them.

Therefore to make a practical design of a vertical alignment, one has to take

all of these constraints along with other common design code constraints
[AAS04] into account. In addition, the challenge is to make an interactive
software tool for designing optimal vertical alignment. So the model needs
to give an accurate solution with acceptable computational time.
Moreb’s model ([Mor09]) can be easily extended to accomodate multi-
materials. But it does not consider side slopes and blocks. Hare et al.
([HKL11]) introduces blocks for earthwork optimization only. Rahman’s
models ([Rah12]) incorporate blocks, side slopes. But it is only solvable for
single material. It needs the input volume to be convex with respect to the
height from the ground level. It also cannot consider cut and fill at the same
section: it needs strictly either a cut or a fill at a given section.
In our research, we have improved and combined the works done by
Moreb ([Mor09]), Koch and Lucet ([KL10]), Hare et al. ([HKL11]) and
Rahman ([Rah12]). We propose two models similar to Rahman’s models.
Both models incorporates multimaterial, side slopes, cut and fill, and blocks.
Our models can also consider convex, and nonconvex input volumes. It can
handle cut and fill of the same material at the same section. So we fulfill
all requirements for computing a practical vertical alignment. Moreover,
we devised some techniques to reduce the solution time in order to work
towards the target of developing an interactive design tool.

8
 1.3. Our research focus

In Chapter 2, we define the variables for both models, and then we

describe and improve the previously designed mixed integer linear program-
ming model for vertical alignment optimization. In Chapter 3, we modify the
quasi network flow model for vertical alignment optimization. We describe
several techniques to improve the solution time in Chapter 4. In Chapter 5,
we test the techniques from Chapter 4, show numerical results and deduce
the most efficient techniques. Chapter 6 concludes with directions for future
research.

9
Chapter 2

The quadratic model for

vertical alignment
In this chapter we describe the variables, parameters, and sets required
to model the vertical alignment of roads. Then we describe the quadratic
model for vertical alignment optimization for a known horizontal alignment.

2.1 Model variables and parameters

The road profile is designed as a quadratic spline. The idea of approx-
imating the road profile by a spline was described in [MA04] and [Mor09].
Later in [KL10], it was shown that up to a quadratic spline, the linearity of
the model can be maintained. In Figure 2.1, a ground profile and a vertical
alignment of a sample road are shown.

cross-section

𝑢1 𝑢5
Elevation (𝑧)

𝑢2 u3 𝑢4
𝑢0

Ground profile
ℎ1
ℎ2 ℎ4 Road profile (spline)

𝑑0 𝑑2 𝑑4 ℎ5
𝑑1 𝑑3 𝑑5

𝑠0 𝑠1 𝑠2 𝑠3 𝑠4 𝑠5
Station

Figure 2.1: Side view of ground profile and road profile. Here, di is the
effective length of a section, hi is the ground elevation, ui is the offset from
the ground profile towards the road profile. Cross-section areas are given at
each station si .

10
 2.1. Model variables and parameters

In order to model the vertical alignment, we split the road into several
sections. Let S = {1, 2, ..., n} be the index set for the sections. Every section
i (∈ S) has a station number si and an effective length di . The difference
between a station and a section is that a station is a point, whereas a
section is a piece of ground stretched by an associated effective length. At
every station, a cross section area of the ground profile is taken. So we can
determine the cross section area of each material at each station point. The
effective length is multiplied with the cross section area to approximate the
volume of materials. A section is associated with the volume of materials,
whereas a station is associated with the cross section areas at a particular
point.
The height of the ground profile of Section i is defined as hi . The decision
variable ui is the vertical offset from the ground profile at Section i. We
assume that throughout the entire length of the section, the road has a fixed
height and a fixed cross section. In reality the heights and cross sections
vary throughout the section, so this provides only an approximate value.
The approximation error can be reduced by decreasing the section length,
i.e., increasing the number of sections. However, if section lengths are small,
the number of section increases and the problem size becomes large.
The entire road profile is represented as a quadratic spline (Figure 2.2).
The sections are grouped into segments. In every segment there may be
several sections, and each segment is associated with a quadratic function.
Let G = {1, 2, . . . , ḡ} be the index set of the segments and Sg = {1, 2, ..., ng }
be the index set of ng sections in spline segment g. The function δ : G ×Sg →
S maps the index in the spline segment to the actual section index, i.e., if
δ(g, j) = i, then si = sδ(g,j) , for all g ∈ G, j ∈ Sg , i ∈ S.
Every two consecutive segments have one common station between them
(Figure 2.2). Therefore, the last station of a spline segment is equal to
the first station of the next spline segment, i.e., sδ(g,ng ) = sδ(g+1,1) , ∀g ∈
{1, . . . , ḡ − 1}. The quadratic function for each segment g ∈ G is defined as

Pg (s) = ag,1 + ag,2 s + ag,3 s2 ,

where s is the distance from the current station to the beginning of the road
along the alignment. The derivative of Pg (s) is written

Pg0 (s) = ag,2 + 2ag,3 s.

11
 2.1. Model variables and parameters

Slopes of two consecutive

segments are equal

Segment 1 Segment 2 Segment 3

Elevation (z)

𝑑0 𝑑2 𝑑4 h5 𝑑6
𝑑1 𝑑3 𝑑5

𝑠0 𝑠1 𝑠2 𝑠3 𝑠4 𝑠5 𝑠6

Section
Road profile (spline) Ground profile

Figure 2.2: Spline segment of vertical alignment.

In general, we can represent the spline function as follows



 P1 (s) if sδ(1,1) ≤ s ≤ sδ(1,n1 ) ,

P2 (s) if sδ(2,1) ≤ s ≤ sδ(2,n ) ,

2
P (s) = .

 .
.


Pḡ (s) if sδ(ḡ,1) ≤ s ≤ sδ(1,nḡ ) .


Similarly, we will use the notation P 0 (s) to represent the derivative of P (s).
Since every two adjacent sections have one common section between them,
to get a smooth curve, the height and slope of the beginning section of one
segment should be equal to the height and slope of the ending section of the
previous segment.
Remark 2.1. For a long road, the value of s in Pg (s) could be large. Since
Pg (s) has a square term of s, numerical errors can happen for large values
of s. To eliminate these numerical errors from the implementation of the
models, each value of s is calculated with respect to the beginning of its
segment, i.e., s − sδ(g,1) ∀g ∈ G. Since this does not change the models, for
the ease of describing the models, we use s instead of s − sδ(g,1) in Pg (s).
Side-slope refers to the gradual decrease and increase in height from the

12
 2.1. Model variables and parameters

road profile to the ground profile of a cut and fill section, respectively. Side-
slopes are very important from the civil engineering point of view because
they help to make the road durable. Therefore, to model the vertical align-
ment problem accurately, we must take side-slopes into consideration. In
Figures 2.3 and 2.4 the cross-section of a road with and without side-slopes
for a cut and a fill are shown, respectively.
Ground profile

Cut Cut
area area

Road profile
Without side-slopes With side-slopes

Figure 2.3: The cross-section of a road with and without side-slopes (for
cut).

Road profile

Fill Fill
area area

Ground profile
Without side-slopes With side-slopes

Figure 2.4: The cross-section of a road with and without side-slopes (for
fill).

Li,5
Li,4 M1
Cut
Li,3 ui M2
area
Li,2 M3
Li,1

Before approximation After approximation

Figure 2.5: Approximation of side-slopes (for a cut). Here, M1 , M2 , M3 are

three kinds of arbitrary materials shown for simple illustration.

To model the side slopes, the approximation can be made with trapezoid
shaped cross-sections with several rectangles (Figure 2.5). We make an
approximation of the volumes of material at different offset levels from the
ground profile. The volumes of material for cut and fill at every section are

13
 2.1. Model variables and parameters

Li,1
Li,2
Fill Li,3 ui M1
area Li,4 M2
Li,5
M3

Before After approximation

Figure 2.6: Approximation of side-slopes (for a fill).

the input data for the model. These volumes are approximated for each
offset level from the ground profile. Suppose for each section i ∈ S, Li =
{1, 2, ..., nl } is the index set for the different offset levels. The parameter
Li,l ∈ R represents the altitude of cut or fill for level l ∈ Li at section
i. Let K = {1, 2, ..., m} be the index set of the material types. Sample
+
input for a section is shown in Table 2.1. Let the parameter Ri,l,m represent
−
the given cut volume and Ri,l,m represent the given fill volume of material,
respectively, for each section i ∈ S, each level l ∈ Li , and for each material
type m ∈ K.

Table 2.1: Cut and fill volume of material

Level Offset Cut Volume Fill Volume
+ −
l Li,l Ri,l,m Ri,l,m
1 -4 262.06 0
2 -2 168.44 0
3 0 82.82 0
4 2 5.28 0.08
5 4 0 70.06

The relation between cut and fill volume with the height of the offset can
be understood clearly from Figure 2.7. For the vertical alignment problem,
we need to find out the optimal offset (ui ) from the ground profile and
+
the corresponding cut and fill volume for that offset. Let the variable Vi,m
−
(i ∈ S, m ∈ K), and Vi,m (i ∈ S, m ∈ K) represent the total cut and fill
volume of material, respectively.
During the construction of a road, some materials may need to be
dumped, and some materials may need to be brought in. Borrow pits
are the external sections from which materials can be borrowed, and waste
pits are the external sections in which materials can be dumped. Suppose
B = {1, 2, ..., nβ } and W = {1, 2, ..., nω } are the two sets of indices that are

14
 2.1. Model variables and parameters

300

250

200

Volume
150
(Ri,l ) Cut Volume
100 Fill Volume

50

0
-4 -2 0 2 4
Height of levels (Li,l)

Figure 2.7: Volume approximation of side-slope

used to index borrow and waste pits, respectively. The function ϑ : B → S

maps a borrow pit index to the section index to which it is attached, and
the function ϕ : W → S maps the same section for a waste pit. There are
also access roads from which the construction of the road can be initiated.
These access roads are also attached to a section and act as a gateway to
the road. Let R = {1, 2, ..., nr } be the set of indices that is used to index the
access roads. We assume that there is a borrow and a waste pit attached
to each access road with infinite capacity for dumping and borrowing ma-
terials. Suppose the function % : R → S maps an access road index to a
section index to which it is attached, and the set N = {S ∪ B ∪ W} is used
to represent sections, borrow pits, and waste pits collectively.
In practice, it happens that some sections of land cannot be accessed
before removing the obstacles. In [HKL11], the block concept was first
introduced for earthwork operation in road design. We use this concept for
vertical alignment modeling of road design. The block scenario is similar
to the situation where engineers have to build a bridge or tunnel before
they can move across a river or through a mountain, respectively. Suppose
I = {1, 2, ..., nb } is the set of indices that is used to index blocks. We define
time steps t to indicate when blocks are removed. It can be shown that
at most (nb + 1) number of time steps are needed to remove all the blocks
[HKL11]. Let T = {0, 1, 2, ..., nb } be the set to represent nb + 1 number of
time steps needed to remove nb blocks, and let the function ς : I → S map
a block index to a section index. To keep track of when a block is removed,
we introduce binary variables ykt for each block k ∈ I and t ∈ T . If a block

15
 2.2. The quadratic model

is removed at time step t, then movement across the block is possible from
the next time step t + 1.

2.2 The quadratic model

In this section the quadratic model is described for obtaining an optimal
vertical alignment. This model is referred to as the quadratic model because
the required number of material movement variables is O(n2 ), where n is
the number of sections.

2.2.1 The quadratic model for single material

In [Rah12], the quadratic model for single material was described. That
model had several features, for example, it incorporated blocks and sides
slopes into the vertical alignment optimization. However, it had several
limitations. That model was not capable to incorporate more than one
material. The input volumes needed to be convex with respect to the height
from the ground level. It could not consider cut and fill at the same section.
The reason of not being able to incorporate all of the above features
into the model was the way it defined the volume constraints. It required
that volume for cut and fill must be zero at the ground level, which is not
the case in practice because of the uneven shape of the earth. Therefore,
[Rah12] approximated the input to apply their models.
In this research, we change the volume constraint to accommodate all the
features into a new modified model which is described in the next section.

2.2.2 The quadratic model for multimaterial

In this section we describe the quadratic model for multimaterial by
modifying the volume constraint. We also incorporate all the features that
are required. For ease of explanation, first we introduce variables for material
movement without considering blocks. Later we will introduce movement
variables considering blocks. To track material movement, we define variable
xij , which represents the amount of material moved from section i to section
j. For each i ∈ N , the index set N→ i consists of all indices j such that x
ij
is a permitted move,
 
 j ∈ S ∪ W, if i ∈ S 
i
N→ = j : j∈S if i ∈ B .
∅ if i ∈ W
 

16
 2.2. The quadratic model

Similarly, for each j ∈ N the index set N←j which consists of all indices i
such that xij is a permitted move,
 
 i ∈ S ∪ B, if j ∈ S 
j
N← = i : ∅ if j ∈ B .
i∈S if j ∈ W
 

Note that for any i, j ∈ N , j ∈ N→ i if, and only if, i ∈ N j . Finally, N 2

←
consists of all index pairs (i, j) such that xij is a permitted move,

N 2 = {(i, j) : j ∈ N→
i
}

Note that N 2 6= N × N , as N 2 omits index pairs (i, j) that correspond to

illogical moves, such as movement from a road section into a borrow pit.
Now we introduce variables for material movements considering blocks.
For tracking material movement at every time step, we define the variable
xijt , which represents the amount of material moved from section i to section
j at time step t. Let Ī 2 = {(k1 , k2 ) ∈ I × I : k1 < k2 and k1 , k2 are two
consecutive blocks with no access road in between}, Ī← = {k ∈ I : the set
of blocks k before which there is no access road}, and Ī→ = {k ∈ I : the set
of blocks k after which there is no access road}.
For all k ∈ I, the set Nbk consists of all pairs of sections (i, j) ∈ N 2 , such
that movement between section i and j requires the removal of block k. The
set Nbk1,k2 , for all (k1 , k2 ) ∈ Ī 2 , consists of all pairs of sections (i, j) ∈ N 2 ,
such that sections i and j are between blocks k1 and k2 . For all k ∈ Ī← , we
set Nb←,k = {(i, j) ∈ N 2 : section i and j are before block k}, and for all
k ∈ Ī→ , we define Nbk,→ = {(i, j) ∈ N 2 : section i and j are after block k}.
The cost of a vertical alignment depends on the embankment, excava-
tion and hauling of materials from section to section. Let pi represent the
excavation cost per unit of volume and qi represent the embankment cost
per unit of volume for section i(∈ N ). Let rij be the hauling cost per unit
of volume from section i to section j.
The objective of the model is to minimize the total excavation costs,
embankment costs, and hauling costs for moving materials between sections,
borrow pits, and waste pits over all time steps. The objective function can
be represented as follows,
 
 X +
X
−
X 
min 
 p i V i,m + qi Vi,m + rij xijt 
,
i∈S∪B i∈S∪W (i,j)∈N 2
m∈K m∈K t∈T

17
 2.2. The quadratic model

+ −
where Vi,m (for i ∈ S ∪B, m ∈ K), and Vi,m (for i ∈ S ∪W, m ∈ K) represent
the total cut and fill volume of material respectively.
The total volume of material moved from a section to other sections at
all time steps must be equal to the cut volume of that section (Constraint
(2.1) below). Similarly, the total volume of material moved to a section
from other sections at all time steps must be equal to the fill volume of that
section (Constraint (2.2) below). We call these the balance constraints:

X
+
xijt,m = Vi,m , ∀ i ∈ S ∪ B, m ∈ K. (2.1)
i
j∈N→
t∈T
m∈K
X
−
xjit = Vi,m , ∀ i ∈ S ∪ W, m ∈ K. (2.2)
i
j∈N←
t∈T
m∈K

The offset value ui is the difference between the road profile, which is
represented by the quadratic spline P (si ), and the given ground profile (hi )
(Constraint (2.3)).

P (si ) − hi = ui , ∀i ∈ S (2.3)

There could be a cut and fill of the same material at the same section
according to the input volume table (see Table 2.1). Usually the cut volume
decreases as the height of the offset level Li,l increases (see Table 2.1); and
the fill volume increases along with the height of the offset level. That means
the relationship between volume and offset level is convex. In general, if we
have only a convex relationship for piecewise approximation in linear pro-
gramming and we need minimization, then we can model that relationship
without using binary variables (Constraints (2.4) and (2.5)).

+ +
+ +
Ri,l+1 − Ri,l,m
Vi,m ≥ Ri,l,m + (ui − Li,l ), ∀i ∈ S, l ∈ Li \ {nl }, m ∈ K
Li,l+1 − Li,l
(2.4)
− −
− −
Ri,l+1 − Ri,l,m
Vi,m ≥ Ri,l,m + (ui − Li,l ), ∀i ∈ S, l ∈ Li \ {nl }, m ∈ K.
Li,l+1 − Li,l
(2.5)

18
 2.2. The quadratic model

However, in the vertical alignment model, we are not minimizing only

the cut and fill cost of material, but also minimizing the movement cost of
material. As a result, if the cost of movement becomes larger than the cost
of cut and fill, a stockpiling pattern will appear. Therefore, at certain offset
level ui , the cut and fill volume could be larger than what we expect from
the input volume table (Table 2.1). These extra volume of materials would
cause stockpiling on the road.
For example, in Figure 2.8, a stockpiling scenario is shown. There is
a waste pit attached with Road Section 3. The dead haul distance of the
waste pit from Road Section 3 is 10m. Suppose the offset heights of the
Road Section 1, Section 2, and Section 3 are 0, 2 and 0 m respectively. The
cut volumes of these three sections are 82.82 m3 , 5.28 m3 , and 82.82 m3 , and
the fill volumes are 0 m3 , 0.08 m3 and 0 m3 respectively. In this instance,
Road Section 2 needs a small amount of fill. So all the cut volumes must be
hauled and dumped to the waste pit. However, if the fill cost is less than
the hauling cost of material to the waste pit, then instead of carrying the
material to the waste pit, the material will be dumped to a nearby station.
In Figure 2.8, materials are dumped from Road Section 1 to Road Section
2, and from Road Section 2 to Road Section 1, causing the fill volume to
get large. These dumpings of material to nearby road sections are valid
according to Constraints 2.4 and 2.5, because the constraints only ensure
the lower bound of cut and fill.
Therefore, we need to introduce binary variables to ensure that we get
the correct cut and fill volumes at optimal offset ui for each section i ∈ S
(Constraints (2.6) and (2.7)). The binary variable νi,l (∀i ∈ S, l ∈ Li \ nl ) is
used to indicate which interval (between Li,l and Li,l+1 ) of the offset levels
is selected. The parameter Mi is the largest volume possible to be cut or
filled at Section i. But for simplicity, we use M everywhere instead of Mi .
At any section, only one of the interval offsets can be selected (Constraint
(2.8)).

+ +
+ +
Ri,l+1 − Ri,l,m
Vi,m ≤ Ri,l,m + (ui − Li,l ) + M (1 − νi,l ),
Li,l+1 − Li,l
∀i ∈ S, l ∈ Li \ {nl }, m ∈ K, (2.6)

− −
− −
Ri,l+1 − Ri,l,m
Vi,m ≤ Ri,l,m + (ui − Li,l ) + M (1 − νi,l ),
Li,l+1 − Li,l
∀i ∈ S, l ∈ Li \ {nl }, , m ∈ K, (2.7)

19
 2.2. The quadratic model

Cut material of section

2 moves to section 1
Waste
𝑉1− = 5.28
pit

Cut material of section

1 moves to section 2 Cut material of
section 3 dumped at
𝑉2− = 82.9

10 m
waste pit

Section Section Section

1 2 3

20 m 40 m

𝑢1 = 0 𝑢2 =2 𝑢3 = 0
𝑉1+ ≥ 82.82 𝑉2+ ≥ 5.28 𝑉3+ ≥ 82.82
𝑉1− ≥ 0 𝑉2− ≥ 0.08 𝑉3− ≥ 0

Figure 2.8: Stockpiling scenario when Constraints (2.6) and (2.7) are not
considered.
X
νi,l = 1, ∀i ∈ S. (2.8)
l∈Li \{nl }

If there is a block between two sections, then material movement is not

possible between those two sections until the block is removed (Constraint
(2.9)). If there is no access road before a block, then material cannot be
moved among sections before that block until the block is removed (Con-
straint (2.10)). Similarly, material movement is not possible if there is no
access road after a block (Constraint (2.11)). Furthermore, if there is no
access road between two blocks, then material movement is not possible in
between those two blocks until one of them is removed (Constraint (2.12)).

xijt,m ≤ M yk,(t−1) , ∀t ∈ T \ {0}, k ∈ I,

(i, j) ∈ Nbk , m ∈ K, (2.9)
xijt,m ≤ M yk,(t−1) , ∀t ∈ T \ {0}, k ∈ Ī← ,
i, j ∈ Nb←,k , m ∈ K, (2.10)
xijt,m ≤ M yk,(t−1) , ∀t ∈ T \ {0}, k ∈ Ī→ ,
i, j ∈ Nbk,→ , m ∈ K, (2.11)

20
 2.2. The quadratic model

xijt,m ≤ M yk1 ,(t−1) + M yk2 ,(t−1) , ∀t ∈ T \ {0}, (k1 , k2 ) ∈ Ī 2 ,

i, j ∈ Nbk1,k2 , m ∈ K. (2.12)

The block removal indicator constraints (2.13, 2.14) ensure that when
the required amount of materials are cut and filled for any section containing
a block, then the block is considered removed from that section.

u
X X
+
xς(k)jt,m + M (1 − yku ) ≥ Vς(k),m , ∀k ∈ I, u ∈ T, m ∈ K,
t=0 j∈N ς(k)
→

(2.13)
u
X X
−
xjς(k)t,m + M (1 − yku ) ≥ Vς(k),m , ∀k ∈ I, u ∈ T, m ∈ K.
t=0 j∈N ς(k)
←

(2.14)

By the end of each time step t ∈ T , we want to remove at least the

number of blocks equal to the number of time steps, so that we do not
need more than nb + 1 time steps. This is guaranteed with block removal
enforcement constraint (2.15). The monotonicity constraint (2.16), states
that once a block is removed it remains removed.
u X
X
ykt ≥ t + 1, ∀u ∈ T \ {nb }, (2.15)
t=0 k∈I
yk,t ≥ yk,t−1 , ∀k ∈ I, t ∈ T \ {0}. (2.16)

The smoothness constraints (2.17) and (2.18) ensure that the transi-
tions from one spline segment to the next are smooth, i.e., same grade and
elevation (see Figure 2.2).

Pg−1 (sδ(g,1) ) = Pg (sδ(g,1) ), ∀g ∈ G \ {1}, (2.17)

0
Pg−1 (sδ(g,1) ) = Pg0 (sδ(g,1) ) ∀g ∈ G \ {1}. (2.18)

The maximum allowable speed depends on the grade of the road profile.
The grade constraint (2.19) makes sure that the grade of the spline segments
are within the allowed maximum (GU ) and the minimum (GL ) value, so that
it can maintain the safety parameters.

GL ≤ Pg0 (sδ(g,1) ) ≤ GU , ∀g ∈ G \ {1}. (2.19)

21
 2.2. The quadratic model

Usually, the elevation and the grade of the starting point and the ending
point of a road are defined explicitly. In addition to those points, the road
designer can set the elevation and grade value for some fixed control points
on the road (for example at an intersection). Let H be the index set of
all control points. The fixed point constraints (2.20), (2.21), (2.22) set the
grade and elevation on specific control points including the starting and the
ending point. The parameters yA , yB are the elevations of the starting and
the ending stations. Similarly, ȳA , ȳB are the slopes of the starting and the
ending spline segments. The parameters Hi is the elevation of the station
i ∈ H.

P (s1 ) = yA , P (sn ) = yB , (2.20)

0 0
P (s1 ) = ȳA , P (sn ) = ȳB , (2.21)
P (si ) = Hi , ∀i ∈ H. (2.22)

Finally, Constraints (2.23), (2.24), and (2.25) are bound constraints. The
parameter Mi+ , and Mi− are the largest volume possible to be cut and filled
from a section.

xijt,m ≥ 0, ∀(i, j) ∈ N 2 , t ∈ T , m ∈ K (2.23)

+
0≤ Vi,m ≤ Mi+ , ∀i ∈ S ∪ B, m ∈ K, (2.24)
−
0≤ Vi,m ≤ Mi− , ∀i ∈ S ∪ W, m ∈ K. (2.25)

In this chapter we have discussed the quadratic model for calculating

optimal vertical alignment of roads. We have considered several features
such as multimaterials, cut and fill at the same section of the same material,
blocks, slide slopes, etc. However, the problem size increases quadratically
with the number of sections. Consequently it takes a long time to solve
the model for long roads. Another limitation is that we still need convex
relationship between volumes of materials and the offset value. In Chapter
3, we discuss the quasi network flow model that reduces the size of the
problem, and the computation time.

22
Chapter 3

Quasi network flow model for

vertical alignment
In this chapter the quasi network flow model is described for computing
an optimal vertical alignment. This model is a modification of the traditional
network flow model [AMO93, p. 12-13]. It combines the ideas of network flow
and scheduling models. This model schedules the movement of materials
among different sections in discrete time steps because of the presence of
blocks in roads. In this model, the required number of material movement
variable is O(n), where n is the number of sections.

3.1 Quasi network flow for single material

In [AMO93, p. 12-13], a network flow model is described for earthwork
optimization. Rahman [Rah12] modeled the vertical alignment problem by
modifying the network flow model and considering blocks and side slopes.
Same as the quadratic model described in the previous chapter, this quasi
network flow model has the same limitation. However, by modifying the
volume constraint and without altering the input volume of materials we
are able to use it for multimaterial.

3.2 Quasi network flow for multimaterial

In the quasi network flow model, sections and pits are considered as
nodes, and the feasible moves are represented as arcs. This quasi network
flow model consists of a series of nodes, which are connected to their adjacent
nodes (Figure 3.1).
The materials can flow from section to section towards the East and
West directions. In every section we have to decide whether we want to fill
the section with materials or transfer it to the next section, or cut materials
to add more to the flow. To represent these flows towards the East and
West directions, at each section virtual nodes are introduced called transit

23
 3.2. Quasi network flow for multimaterial

West transit flow

West 𝑟 West 𝑟 West

𝑓𝑖,𝑖−1,𝑡,𝑚 𝑓𝑖+1,𝑖,𝑡,𝑚
transit transit transit
node node node
𝑖−1 𝒊 𝑖+1
𝑙
𝑢
𝑓𝑖,𝑖−1,𝑡,𝑚 𝑓𝑖+1,𝑖,𝑡,𝑚

𝑢 𝑙
𝑓𝑖,𝑖,𝑡,𝑚 Section 𝒊 𝑓𝑖,𝑖,𝑡,𝑚

𝑢
𝑙
𝑓𝑖−1,𝑖,𝑡,𝑚 𝑓𝑖,𝑖+1,𝑡,𝑚

East East East

transit transit transit
node 𝑟 node 𝑟 node
𝑖−1 𝑓𝑖−1,𝑖,𝑡,𝑚 𝒊 𝑓𝑖,𝑖+1,𝑡,𝑚 𝑖+1

East transit flow

Figure 3.1: A typical section i at time-step t.

nodes for both directions (Figure 3.1). If materials moved from one section
to another towards the West direction, then the material goes through the
West transit nodes, and this flow of material is referred as West transit flow.
Similarly, the flow towards the East direction goes through East transit
nodes, and is referred as East transit flow. Let the variables fi,i−1,t,m r and
r
fi,i+1,t,m (∀i ∈ S, t ∈ T , m ∈ K) be the flow of materials from one transit
node to the next in the West and East directions, respectively.
To represent the cut and fill volume of materials at each section, unload
and load flows are introduced for each section. Note that, cut and fill can
happen within the same section. The cut materials can be transferred to
the East and West transit nodes, and within the same section. Similarly,
the fill materials can be brought from the East and West transit nodes, and
within the same section. Let the variables fi−1,i,t,m l l
, fi+1,i,t,m l
, and fi,i,t,m
(∀i ∈ S, t ∈ T , m ∈ K) represent the load (fill) flows of materials to the East
and West transit nodes, and within the same section, respectively. Similarly,
u
the variables fi,i−1,t,m u
, fi,i+1,t,m u
, and fi,i,t,m (∀i ∈ S, t ∈ T , m ∈ K) represent
the unload (cut) flows of materials from the West and East transit nodes,
and within the same section, respectively.
The borrow and waste pits are attached to their associated sections.

24
 3.2. Quasi network flow for multimaterial

Each borrow pit can only have unload flows to both transit nodes. These un-
load flows represent the cut volume from the borrow pit. The flows from bor-
b
row pits are called borrow flows. Let the variables fj,ϑ(j)−1,t b
and fj,ϑ(j)+1,t,m
(∀j ∈ B, t ∈ T , m ∈ K) be the flows of materials to the West and East
transit nodes, respectively, from the borrow pit. Similarly, for each waste
pit there are load flows from both directions. These load flows represent
the fill volume of the waste pit and are called waste flows. Let the vari-
w
ables fϕ(j)−1,j,t,m w
and fϕ(j)+1,j,t,m (∀j ∈ W, t ∈ T , m ∈ K) be the flows of
materials from the East and West transit nodes to waste pit, respectively.

West transit flow

West 𝑟 West 𝑟 West

𝑓𝑖,𝑖−1,𝑡,𝑚 𝑓𝑖+1,𝑖,𝑡,𝑚
transit transit transit
node node node
𝑖−1 𝒊 𝑖+1
𝑙
𝑢
𝑓𝑖,𝑖−1,𝑡,𝑚 𝑓𝑖+1,𝑖,𝑡,𝑚

Borrow pit 𝑓 𝑢 𝑙
𝑖,𝑖,𝑡,𝑚 Section 𝒊 𝑓𝑖,𝑖,𝑡,𝑚 Waste pit
𝜗(𝑗) 𝜑(𝑘)

𝑢
𝑙
𝑓𝑖−1,𝑖,𝑡,𝑚 𝑓𝑖,𝑖+1,𝑡,𝑚

East East East

transit transit transit
node 𝑟 node 𝑟 node
𝑖−1 𝑓𝑖−1,𝑖,𝑡,𝑚 𝒊 𝑓𝑖,𝑖+1,𝑡,𝑚 𝑖+1

East transit flow

Figure 3.2: A typical section i at time-step t with borrow and waste pit
associated with it (where ϑ(j) = ϕ(k) = i). The function ϑ : B → S maps
a borrow pit index to the section index to which it is attached, and the
function ϕ : W → S maps the same section for a waste pit.

The cost of hauling materials from Section i to Section i−1 (respectively

i + 1) is represented by cri,i−1 = cdi,i−1 (respectively cri,i+1 = cdi,i+1 ), where c
represents the cost of moving one unit volume of material per unit distance,
and dij is the distance between section i and j.
The objective function of the quasi network flow model is also to mini-
mize the total cost in excavation, embankment and hauling materials, which

25
 3.2. Quasi network flow for multimaterial

can be represented as follows,

X X X
+ −
cri,i−1 fi,i−1,t,m
r
+ cri,i+1 fi,i+1,t,m
r


min pi,m Vi,m + qi,m Vi,m +
i∈S∪B i∈S∪W i∈S
m∈K m∈K t∈T
m∈K
X  
+ cd¯j fj,ϑ(j)−1,t,m
b b
+ fj,ϑ(j)+1,t,m
j∈B
t∈T
m∈K
X  
+ cd¯j fϕ(j)−1,j,t,m + fϕ(j)+1,j,t,m
w
.
j∈W
t∈T
m∈K

Notice that the objective function in this model is more strictly defined
than in the quadratic model described earlier. In particular, the hauling
cost of material must be a linear function of the distance traveled. This is
unrealistic for long distances, but reasonably accurate for short hauls.
The transit nodes are virtual nodes that provide transits for all associated
flows. Therefore, the sum of the flows (demand) coming to a transit node
must be equal to the sum of the flows leaving (supply) the node. The flow
constraints for the East transit node (Constraint (3.1)), and for the West
transit node (Constraint (3.2)) ensure that the supply and demand at both
transit nodes are equal.

X X
r u b r l w
fi−1,i,t,m + fi,i+1,t,m + fj,i+1,t = fi,i+1,t,m + fi−1,i,t,m + fi−1,j,t,m ,
j∈B j∈W
ϑ(j)=i ϑ(j)=i

∀t ∈ T , i ∈ S, m ∈ K, (3.1)

X X
r u b r l w
fi+1,i,t,m +fi,i−1,t,m + fj,i−1,t,m = fi,i−1,t,m +fi+1,i,t,m + fi+1,j,t,m ,
j∈B j∈W
ϑ(j)=i ϑ(j)=i

∀t ∈ T , i ∈ S, m ∈ K. (3.2)

The balance constraints ensure that the total unload flows from a section
is equal to the cut volume of that section (Constraint (3.3)), and the total
load flows to a section is equal to the fill volume (Constraint (3.4)). Borrow

26
 3.2. Quasi network flow for multimaterial

(Constraint (3.5)) and waste (Constraint (3.6)) pits constraints are written
similarly.
X
u u u +
fi,i−1,t,m + fi,i+1,t,m + fi,i,t,m = Vi,m , ∀i ∈ S, m ∈ K, (3.3)
t∈T
X
l l l −
fi−1,i,t,m + fi+1,i,t,m + fi,i,t,m = Vi,m , ∀i ∈ S, m ∈ K, (3.4)
t∈T
X
b b +
fj,ϑ(j)−1,t,m + fj,ϑ(j)+1,t,m = Vj,m , ∀j ∈ B, m ∈ K, (3.5)
t∈T
X
w −
fϕ(j)−1,j,t + fϕ(j)+1,j,t,m = Vj,m , ∀j ∈ W, m ∈ K. (3.6)
t∈T

The volume constraints (2.4), (2.5), (2.6), (2.7), (2.8), smoothness con-
straints (2.17), (2.18), grade constraints (2.19) and fixed point constraints
(2.20), (2.21), (2.22) are the same for this quasi network flow model as de-
scribed in the quadratic model.
r
If section i is a block, then for the East transit node equations fi−1,i,t,m =
l r u
fi−1,i,t,m and fi,i+1,t,m = fi,i+1,t,m (∀t ∈ T , m ∈ K) ensure that no material
can transfer across section i to the East direction. Because fi−1,i,t,m r =
l
fi−1,i,t,m implies that when materials are transferred from section i − 1 to
r
section i, all the materials are loaded (filled) to section i. And fi,i+1,t,m =
u
fi,i+1,t,m implies that when materials are transferred from section i to section
i + 1, all the materials are unloaded (cut) from Section i to Section i +
1. Similarly, in the case of West transit nodes, fi+1,i,t,mr l
= fi+1,i,t and
r u
fi,i−1,t,m = fi,i−1,t,m ensure that no material can transfer across Section i to
the West direction.
From the above reasoning, we can ensure that no material movement
can occur over a block to the East direction, until that block is removed by
Constraints (3.7), (3.8), (3.9), and (3.10). Similarly, for the West direction,
Constraints (3.11), (3.12), (3.13), (3.14) ensure no movement over a block.
r l
−(fς(k)−1,ς(k),t,m − fς(k)−1,ς(k),t,m ) ≤ M yk,t−1 , ∀k ∈ I, t ∈ T , m ∈ K,
(3.7)
r l
fς(k)−1,ς(k),t,m − fς(k)−1,ς(k),t,m ≤ M yk,t−1 , ∀k ∈ I, t ∈ T , m ∈ K,
(3.8)
r u
−(fς(k),ς(k)+1,t,m − fς(k),ς(k+1),t,m ) ≤ M yk,t−1 , ∀k ∈ I, t ∈ T , m ∈ K,
(3.9)
r u
fς(k),ς(k)+1,t,m − fς(k),ς(k+1),t,m ≤ M yk,t−1 , ∀k ∈ I, t ∈ T , m ∈ K.
(3.10)

27
 3.2. Quasi network flow for multimaterial

r l
−(fς(k)+1,ς(k),t,m − fς(k)+1,ς(k),t,m ) ≤ M yk,t−1 , ∀k ∈ I, t ∈ T , m ∈ K,
(3.11)
r l
fς(k)+1,ς(k),t,m − fς(k)+1,ς(k),t,m ≤ M yk,t−1 , ∀k ∈ I, t ∈ T , m ∈ K,
(3.12)
r u
−(fς(k),ς(k)−1,t,m − fς(k),ς(k)−1,t,m ) ≤ M yk,t−1 , ∀k ∈ I, t ∈ T , m ∈ K,
(3.13)
r u
fς(k),ς(k)−1,t,m − fς(k),ς(k)−1,t,m ≤ M yk,t−1 , ∀k ∈ I, t ∈ T , m ∈ K.
(3.14)

The block constraints must also ensure that no earth movement can
occur among sections, borrow pits, and waste pits: in between two blocks,
before the first block, and after the last block with no access roads, until
blocks are removed. Therefore, there is no transit flow (Constraints (3.15),
(3.16)), borrow flow (Constraints (3.17), (3.18)), or waste flow (Constraints
(3.19), (3.20)) between two blocks with no access roads until one of the
blocks is removed.

r
fi,i+1,t,m ≤ M (yk1 ,t−1 + yk2 ,t−1 ), ∀t ∈ T , (k1 , k2 ) ∈ Ī 2 , m ∈ K,
ς(k1 ) ≤ i, i + 1 ≤ ς(k2 ), (3.15)
r 2
fi+1,i,t,m ≤ M (yk1 ,t−1 + yk2 ,t−1 ), ∀t ∈ T , (k1 , k2 ) ∈ Ī , m ∈ K,
ς(k1 ) ≤ i, i + 1 ≤ ς(k2 ), (3.16)
b 2
fj,ϑ(j)+1,t,m ≤ M (yk1 ,t−1 + yk2 ,t−1 ), ∀t ∈ T , (k1 , k2 ) ∈ Ī , m ∈ K,
ς(k1 ) ≤ ϑ(j) − 1, ϑ(j),
ϑ(j) + 1 ≤ ς(k2 ), (3.17)
b 2
fj,ϑ(j)−1,t,m ≤ M (yk1 ,t−1 + yk2 ,t−1 ), ∀t ∈ T , (k1 , k2 ) ∈ Ī , m ∈ K,
ς(k1 ) ≤ ϑ(j) − 1, ϑ(j),
ϑ(j) + 1 ≤ ς(k2 ), (3.18)
w 2
fϕ(j)+1,j,t,m ≤ M (yk1 ,t−1 + yk2 ,t−1 ), ∀t ∈ T , (k1 , k2 ) ∈ Ī , m ∈ K,
ς(k1 ) ≤ ϕ(j) − 1, ϕ(j),
ϕ(j) + 1 ≤ ς(k2 ), (3.19)
w
fϕ(j)−1,j,t,m ≤ M (yk1 ,t−1 + yk2 ,t−1 ), ∀t ∈ T , (k1 , k2 ) ∈ Ī 2 , m ∈ K,
ς(k1 ) ≤ ϕ(j) − 1, ϕ(j),
ϕ(j) + 1 ≤ ς(k2 ). (3.20)

28
 3.2. Quasi network flow for multimaterial

There will also be no transit flow (Constraints (3.21) and (3.22)), borrow
flow (Constraints 3.23 and 3.24), or waste flow (Constraints 3.25 and 3.26)
after the last block with no access road until the block is removed.
r
fi,i+1,t,m ≤ M yk1 ,t−1 , ∀t ∈ T , ς(k) ≤ i, i + 1 ≤ n, k ∈ Ī→ , m ∈ K,
(3.21)
r
fi+1,i,t,m ≤ M yk1 ,t−1 , ∀t ∈ T , ς(k) ≤ i, i + 1 ≤ n, k ∈ Ī→ , m ∈ K
(3.22)
b
fj,ϑ(j)+1,t,m ≤ M yk1 ,t−1 , ∀t ∈ T , ς(k) ≤ ϑ(j) − 1, ϑ(j), ϑ(j) + 1 ≤ n,
k ∈ Ī→ , m ∈ K, (3.23)
b
fj,ϑ(j)−1,t,m ≤ M yk1 ,t−1 , ∀t ∈ T , ς(k) ≤ ϑ(j) − 1, ϑ(j), ϑ(j) + 1 ≤ n,
k ∈ Ī→ , m ∈ K (3.24)
w
fϕ(j)+1,j,t,m ≤ M yk1 ,t−1 , ∀t ∈ T , ς(k) ≤ ϕ(j) − 1, ϕ(j), ϕ(j) + 1 ≤ n,
k ∈ Ī→ , m ∈ K, (3.25)
w
fϕ(j)−1,j,t,m ≤ M yk1 ,t−1 , ∀t ∈ T , ς(k) ≤ ϕ(j) − 1, ϕ(j), ϕ(j) + 1 ≤ n,
k ∈ Ī→ , m ∈ K. (3.26)

Similarly, there will also be no transit flow (Constraints (3.27) and

(3.28)), borrow flow (Constraints (3.29) and (3.30)), or waste flow (Con-
straints (3.31) and (3.32)) before the first block with no access road until
the block is removed.
r
fi,i+1,t,m ≤ M yk1 ,t−1 , ∀t ∈ T , 1 ≤ i, i + 1 ≤ ς(k), k ∈ Ī← , m ∈ K,
(3.27)
r
fi+1,i,t,m ≤ M yk1 ,t−1 , ∀t ∈ T , 1 ≤ i, i + 1 ≤ ς(k), k ∈ Ī← , m ∈ K,
(3.28)
b
fj,ϑ(j)+1,t,m ≤ M yk1 ,t−1 , ∀t ∈ T , 1 ≤ ϑ(j) − 1, ϑ(j), ϑ(j) + 1 ≤ ς(k),
k ∈ Ī← , m ∈ K, (3.29)
b
fj,ϑ(j)−1,t,m ≤ M yk1 ,t−1 , ∀t ∈ T , 1 ≤ ϑ(j) − 1, ϑ(j), ϑ(j) + 1 ≤ ς(k),
k ∈ Ī← , m ∈ K, (3.30)
w
fϕ(j)+1,j,t,m ≤ M yk1 ,t−1 , ∀t ∈ T , 1 ≤ ϕ(j) − 1, ϕ(j), ϕ(j) + 1 ≤ ς(k),
k ∈ Ī← , m ∈ K, (3.31)
w
fϕ(j)−1,j,t,m ≤ M yk1 ,t−1 , ∀t ∈ T , 1 ≤ ϕ(j) − 1, ϕ(j), ϕ(j) + 1 ≤ ς(k),
k ∈ Ī← , m ∈ K. (3.32)

29
 3.2. Quasi network flow for multimaterial

The block removal indicator constraints ensure that when the required
amount of earth is excavated from or embanked to a section with a block,
then the block is considered removed (Constraints (3.33) and (3.34)).
u 
X 
u u +
fς(k),ς(k)−1,t,m + fς(k),ς(k)+1,t,m + M (1 − yku ) ≥ Vς(k),m ,
t=0
∀k ∈ I, u ∈ T, m ∈ K, (3.33)
u 
X 
l l −
fς(k)−1,ς(k),t,m + fς(k)+1,ς(k),t,m + M (1 − yku ) ≥ Vς(k),m ,
t=0
∀k ∈ I, u ∈ T, m ∈ K. (3.34)

The block removal enforcement Constraint (2.15) and monotonicity Con-

straint (2.16) are the same for the quasi network flow model as described
for the quadratic model. The extra bound constraints set the lower bounds
of the variables for this model.
r r
fi,i−1,t,m ≥ 0, fi,i+1,t,m ≥ 0, ∀i ∈ S, t ∈ T , m ∈ K, (3.35)
l l
fi−1,i,t,m ≥ 0, fi+1,i,t ≥ 0, ∀i ∈ S, t ∈ T , m ∈ K, (3.36)
u u
fi,i−1,t,m ≥ 0, fi,i+1,t ≥ 0, ∀i ∈ S, t ∈ T , m ∈ K, (3.37)
b b
fj,ϑ(j)−1,t,m ≥ 0, fj,ϑ(j)+1,t,m ≥ 0, ∀j ∈ B, t ∈ T , m ∈ K, (3.38)
w w
fϕ(j)−1,j,t,m ≥ 0, fϕ(j)+1,j,t,m ≥ 0, ∀j ∈ W, t ∈ T , m ∈ K. (3.39)

In this chapter we have discussed the quasi network flow model for the
vertical alignment of roads. This model reduces the size of the problem
from the quadratic model (see Chapter 2) from quadratic to linear with
respect to the number of sections. However, it has one drawback: the cost
function must be linear with respect to the hauling distance of materials.
And we still need a convex relationship between volumes of materials and
the offset value. In Chapter 4, we discuss some techniques to improve the
solution time for both models and remove the dependency on the convex
relationship between volume and offset levels.

30
Chapter 4

Speedup techniques
One of our goals in this research is to determine methods to solve the
vertical alignment problem in the most efficient manner. The models we
described in the previous two chapters can solve problems according to the
full specification except one limitation, i.e., the volume of each material must
be convex with respect to the height of the offset level. In this chapter we
also remove that limitation of not able to handle non-convexity. To make
an interactive design tools it is necessary to compute the optimal solution
quickly. In this chapter we describe four strategies that we used to decrease
the time to find a solution. In Chapter 5 we numerically test and compare
these methods.

4.1 Techniques with offset levels

As input for both models (the quadratic and the quasi network flow),
we get volume information at different offset levels. We select an optimal
height among those input offset levels by approximating the volumes of
materials as a piecewise linear function (see Constriants (2.4), (2.5), (2.6),
(2.7), and (2.8)). We have binary variables for each offset levels for all
sections. In total we need n(nl − 1) binary variables, where n is the number
of sections and nl is the number of offset levels. This large number of binary
variables increases the solution time significantly. In the next two subsection
we describe two techniques to accelerate the computation by modifying the
constraints involving with offset levels. These techniques are applicable for
both models. However, the second technique will remove the limitation of
non-convexity discussed above.

4.1.1 Reducing binary variables for offset levels

In the volume constraints (Constraints (2.4), (2.5), (2.6), and (2.7)) of
the quadratic and quasi network flow models, there are binary variables
for all the intervals of the input offsets (Li,l ) for each section i ∈ S. This

31
 4.1. Techniques with offset levels

requires n(nl − 1) binary variables, where n is the number of sections and

(nl − 1) is the number of intervals of the input offsets.
For simplicity, we now assume that every section i ∈ S has the same num-
ber of input offsets (Li,l ) numbered from 0 to nl −1. Let L = {0, 1, ..., nl −2}
be the index set for the intervals of the input offset levels of all sections. A
number from 0 to nl − 2 can be represented using dlog2 (nl − 2)e binary
variables as follows:
blog2 (nl −2)c
X
l= blk · 2k ,
k=0

where blk ∈ {0, 1} for each l ∈ L. Let Bk1 = {l ∈ L : blk = 1} and

Bk0 = {l ∈ L : blk = 0}.
Define the set L0 = {0, 1, 2, ..., blog2 (nl −2)c} to help representing (nl −1)
number of intervals of offsets in binary format. Let αi,k (∀k ∈ L0 , i ∈ S)
be the binary variable which represents the selected interval of offsets for
Section i.
Note that indicator variables are still needed to write the volume con-
straints (2.6, 2.7). Therefore, we introduce extra continuous variables Γi,l ∈
[0, 1] (for all i ∈ S, l ∈ L) to indicate which interval of the offset levels is
selected. The new constraints are written in such a way that Γi,l can only
take the value 0 or 1. The variable Γi,l plays the role of the binary variable
νi,l of the previous models. The new constraints (4.1) and (4.2) ensure that
only one of Γi,l (∀l ∈ L) can equal 1 for each section i ∈ S.

X
Γi,l ≤ αik , ∀i ∈ S, k ∈ L0 , (4.1)
l∈Bk1
X
Γi,l ≤ (1 − αik ), ∀i ∈ S, k ∈ L0 . (4.2)
l∈Bk0

In applying the above, we also note that Constraint (2.8) is no longer re-
quired because for a given combination of αik (∀k ∈ L0 ), for each section
i ∈ S, only one of the Γi,l (∀l ∈ L) can be one.

Example 4.1. Consider the case of 5 input offset levels. Then we have 4
intervals and the set L0 = {0, 1}. Thus for each section i ∈ S, we get the

32
 4.1. Techniques with offset levels

following inequalities:

Γi,1 + Γi,3 ≤ αi0 ,

Γi,0 + Γi,2 ≤ 1 − αi0 ,
Γi,2 + Γi,3 ≤ αi1 ,
Γi,0 + Γi,1 ≤ 1 − αi1 .

Therefore, if we choose (αi1 , αi0 ) = (1, 0) that represents interval 2 of offset

levels, then we get the following concrete inequalities from where it is easily
noticeable that only Γi,2 equals to 1.

Γi,1 + Γi,3 ≤ 0,
Γi,0 + Γi,2 ≤ 1,
Γi,2 + Γi,3 ≤ 1,
Γi,0 + Γi,1 ≤ 0.

Now, if we replace the binary variable νi,l with the continuous indicator
variable Γi,l in the volume constraints (2.6, 2.7), and add new constraints
(4.1) and (4.2), then we will get the same result as the original model,
but the binary variables are reduced from n(nl − 1) to ndlog2 (nl − 1)e and n
constraints (2.8) are discarded. However, the number of continuous variables
is increased by n(nl − 1), and 2ndlog2 (nl − 1)e new constraints are added.

4.1.2 Modeling volume constraints with SOS2 variables

The volume constraints in the quadratic and quasi network flow models
work only when the input volume maintains the convex relationship with
the height. But if the convexity does not hold, as often occurs in practice,
then we need to select two adjacent binary variables to indicate an interval.
However, instead of writing complicated volume constraints and binary vari-
able selection constraints, we can use SOS2 variables to model this piecewise
relationship between volume and offset height. Recall that SOS2 variables,
or special ordered set variables of type 2, are variables with the property
that at most two variables can be nonzero and they must be adjacent to
each other [KdFJN04],[KdFN06],[MMM06],[Tom81]. Using this technique
we are able to handle non-convex relation of volumes with respect to the
heights of offset levels.
Let λi,l (∀l ∈ Li ) be the set of SOS2 variables for each section i ∈ S.
These variables denote non-negative weights for all levels such that their
sum is 1 (Constraint (4.3)). The offset variable ui can be any point in

33
 4.2. Techniques with blocks

between two input offset levels (l and l + 1) that can be calculated by the
weighted sum of the two adjacent input offset heights (Constraint (4.4)).
Similarly, the cut and fill volumes can be calculated using the weighted sum
of input volumes of two adjacent offset levels (Constraints (4.5) and (4.6)).
In this formulation, we do not have to calculate slopes for each interval.

X
λi,l = 1, ∀i ∈ S, (4.3)
l∈Li
X
λi,l Li,l = ui , ∀i ∈ S, (4.4)
l∈Li
X
+
λi,l Ri,l = Vi+ , ∀i ∈ S, (4.5)
l∈Li
X
−
λi,l Ri,l = Vi− , ∀i ∈ S. (4.6)
l∈Li

4.2 Techniques with blocks

Considering blocks during earthwork operation is a great idea. How-
ever, blocks make the problem harder to solve. We need to find an optimal
schedule for material movement during vertical alignment optimization. To
indicate the removal of blocks we define binary variables in Chapter 2. We
need nb (nb + 1) binary variables for blocks in both models (Quadratic and
Quasi network flow), where nb is the number of blocks. Moreover, it in-
creases the number of movement (quadratic) or flow (quasi network flow)
variables by a factor of (nb + 1) to associate time steps into them.

4.2.1 Block constraints with SOS1 variables

In the basic quadratic and quasi network flow models, the block con-
straints have monotonicity constraint (2.16). The monotonicity constraint
ensures that once a block is removed it remains removed by setting 1 to the
binary variables ykt , for all k ∈ I, t ∈ T , upto the last time step. However,
the block k is removed when 1 is set in ykt for the first time. From this
reasoning, it follows that we can model the block constraints with SOS1
variable. Recall that SOS1 variables have the property that at most one
variable can be 1 from that set. The SOS1 variables can be treated very
efficiently by most Mixed Integer Linear Programming (MILP) solvers.
For each block k ∈ I, we introduce a set of SOS1 variables γkt for all
t ∈ T . As a result, the monotonicity constraints are no longer needed.

34
 4.2. Techniques with blocks

However we have to rewrite all the block constraints for both basic models
as a summation of γkt instead of just one ykt per constraint. In particular,
the block Constraints (2.9), (2.10), (2.11),(2.12), (2.13), (2.14), and (2.15)
of the quadratic model are rewritten using SOS1 variables as Constraints
(4.7), (4.8), (4.9), (4.10), (4.9), and (4.10).

t−1
X
xijt,m ≤ M γkw , ∀t ∈ T \ {0}, k ∈ I,
w=0
(i, j) ∈ Nbk , m ∈ K, (4.7)
t−1
X
xijt,m ≤ M γkw , ∀t ∈ T \ {0}, k ∈ Ī← ,
w=0
i, j ∈ Nb←,k , m ∈ K, (4.8)
t−1
X
xijt,m ≤ M γkw , ∀t ∈ T \ {0}, k ∈ Ī→ ,
w=0
i, j ∈ Nbk,→ , m ∈ K, (4.9)
t−1
X
xijt,m ≤ M (γk1 ,w + γk2 ,w ) , ∀t ∈ T \ {0}, (k1 , k2 ) ∈ Ī 2 ,
w=0
i, j ∈ Nbk1,k2 , m ∈ K, (4.10)

u u
!
X X X
+
xς(k)jt,m + M 1− γkw ≥ Vς(k) , ∀k ∈ I, u ∈ T, m ∈ K,
t=0 j∈N ς(k) w=0
→
m∈K
(4.11)
u u
!
X X X
−
xjς(k)t + M 1− γkw ≥ Vς(k) , ∀k ∈ I, u ∈ T, m ∈ K.
t=0 j∈N ς(k) w=0
←
m∈K
(4.12)
The block constraints for the basic quasi network flow model are rewrit-
ten in an analogous manner.

4.2.2 Reducing the number of binary variables for blocks

An alternative to SOS1 constraints for working with blocks, is to rewrite
the block variables using log n binary variables. This modeling method uses

35
 4.2. Techniques with blocks

similar techniques as those discussed in Subsection 4.1.1 for offset levels. We

can represent each time step t ∈ T in binary format as follows:
blog2 nb c
X
t= btj · 2j ,
j=0

where btj ∈ {0, 1}. Define Bt1 = {j : btj = 1} and Bt0 = {j : btj = 0}, for
each t ∈ T . Let T 0 = {0, 1, 2, ..., blog2 nb c} be a set that helps to represent
(nb + 1) number of time steps in binary format. Let vkj (∀j ∈ T 0 , k ∈ I) be
the set of binary variables which represent the time step at which the block
k is removed.
We then introduce extra continuous variables βkt ∈ [0, 1] (for all t ∈
T , k ∈ I) to indicate time step t when the block k is removed. The new
constraints (4.13) and (4.14) ensure that there will always be consecutive
ones in βkt (∀t ∈ T ) for each k ∈ I upto the last time step t. This represents
that once a block is removed it remains removed.

t−1
X X
βku ≤ t (1 − vkj ), ∀k ∈ I, t ∈ T \ {0, nb }, (4.13)
u=0 j∈Bt1
nb blog2 nb c
X X
βku ≥ (nb + 1) − 2j · vkj , ∀k ∈ I. (4.14)
u=0 j=0

Therefore, the monotonicity constraint (2.16) is not needed any more. The
other block constraints for both (quadratic and quasi network flow) models
remain the same. However, one needs to replace the binary variables ykt
with the continuous variables βkt .
Example 4.2. If we have I = {1, 2 . . . , 7} and T = {0, 1, 2 . . . , 7} then for
each k ∈ I we get the following inequalities from the Constraint (4.13):
βk0 ≤ [(1 − vk0 )] · 1,
βk0 + βk1 ≤ [(1 − vk1 )] · 2,
βk0 + βk1 + βk2 ≤ [(1 − vk1 ) + (1 − vk0 )] · 3,
βk0 + βk1 + βk2 + βk3 ≤ [(1 − vk2 )] · 4,
βk0 + βk1 + βk2 + βk3 + βk4 ≤ [(1 − vk2 ) + (1 − vk0 )] · 5,
βk0 + βk1 + βk2 + βk3 + βk4 + βk5 ≤ [(1 − vk2 ) + (1 − vk1 )] · 6,
βk0 + βk1 + βk2 + βk3 + βk4 + βk5 + βk6 ≤ [(1 − vk2 ) + (1 − vk1 ) +
(1 − vk0 )] · 7.

36
 4.2. Techniques with blocks

From Constraint (4.14), for each k ∈ I we get the following inequality:

βk0 + βk1 + βk2 + βk3 + βk4 + βk5 + βk6 + βk7 ≥ 8 − [22 vk2 + 2vk1 + vk0 ].

Lemma 4.3. Assume constraints (4.13) and (4.14) hold. If block k ∈ I is

removed at time step t ∈ T , then

i. βk0 = βk1 = ... = βk(t−1) = 0,

ii. βkt = βk(t+1) = ... = βknb = 1.

Proof. (Part i): We know, the binary representation of the tth time step
can be given as,
blog2 nb c
X
t= btj · 2j .
j=0

If the tth time step is selected for a block k ∈ I, then we can write t =
Pblog2 nb c
j=0 vkj · 2j . So, btj = vkj for all index j for the tth time step of block
k. Therefore, Bt1 = {j : btj = vkj = 1}. Thus,P in Constraint (4.13), the
P t−1
j∈Bt1 (1 − v kj ) term becomes zero. As a result, i=0 βki = 0. This proves
the first part.
(Part ii): If we have nb blocks, then we know we have nb + 1 time steps
for each block k ∈ I. Also notice, time step starts from 0 and ends at nb . If
Pblog n c
the tth time step is selected for block k, then t = j=02 b vkj · 2j for that
block. In the first part of this theorem, we prove that, the first t continuous
indicator variables (βk0 , βk1 , ..., βk(t−1) ) are zero. From Constraint (4.14), we
know, the sum of all the continuous variables should be at least (nb + 1) − t.
The continuous indicator variable can have value at most 1. So, to make the
sum (nb +1)−t, we need at least (nb +1)−t continuous indicator variables to
be 1. If we rewrite Constraint (4.14), then we get the following expression,

Sum of first t continuous indicator variables + Sum of rest of the

continuous indicator variables ≥ (nb + 1) − t
⇒ [βk0 + βk1 + ... + βk(t−1) ] + [βkt + βk(t+1) + ... + βknb ] ≥ (nb + 1) − t
⇒ [0] + [sum of rest of the (nb + 1 − t) continuous indicator variables]
≥ (nb + 1) − t.

To satisfy the last line, the continuous indicator variables from βkt to βknb
should be one.

37
 4.2. Techniques with blocks

Remark 4.4. (Feasibility) If block k ∈ I is removed at time step t ∈ T , then

the last (nb − t) inequalities of the first constraint (4.13) will always have
at least t0 ∈ T , where t0 > t, on the right hand side. Also, on the left hand
side, there are t0 continuous indicator variables, which makes the solution
feasible to always get consecutive ones up to the end.
Remark 4.5. We no longer require the monotonicity constraint βk(t+1) ≥ βkt
for each t ∈ T ∩ {1, 2, ..., nb } under this formulation, because Lemma 4.3
proves that these properties will be satisfied for any selection of t using
binary variables vkj , where j ∈ T 0 .
Theorem 4.6. The solution of the quadratic model is (xijtm , ykt ), where
ykt is binary variable if and only if (xijtm , βkt ) is the solution to the current
model, where βkt is continuous variable.
Proof. Let (xijtm , ykt ) be the solution for the quadratic model. In the cur-
rent model, we have the same constraint sets as the quadratic model except
for the monotonicity constraints (2.16), which is not required anymore. Ad-
ditionally, in the current model we have constraints (4.13) and (4.14) that
ensure the monotonicity of βkt (Lemma 4.3). So, if we have monotonicity
in the quadratic model, then the solution should also satisfy the extra two
constraints (4.13) and (4.14) of the current model. Therefore (xijtm , βkt ) is
the solution to the current model.
Let (xijtm , βkt ) be the solution to the current model. In the quadratic
model, ykt is a binary variable with the same constraint satisfying Con-
straints (2.16), but without the extra two constraints (4.13) and (4.14) of
the current model. These two extra constraints (4.13) and (4.14) ensure
monotonicity in the current model. So in the solution βkt will have consecu-
tive ones till the last time step. Therefore, it should satisfy the monotonicity
constraint (2.16) in the quadratic model. Hence, (xijtm , ykt ) is the solution
to the quadratic model.

Proposition 4.7. Representing nb + 1 time steps for each block k ∈ I in

binary format, can be done using nb (blog2 nb c + 1) binary variables.
Proof. Let, t be an integer number. We know, t can be represented in binary
format using the following formulation,
blog2 tc
X
t= b · 2j ,
j=0

where b ∈ {0, 1}. So, to represent time step nb in binary format, we can use
blog2 nb c + 1 binary variables. For each block k ∈ I, we need to have a time

38
 4.2. Techniques with blocks

step which will indicate when block k is removed. There are nb number of
blocks that need to be removed, and time steps are ranged from 0 to nb .
Therefore, one can use nb (blog2 nb c + 1) binary variables.

This approach reduces the number of binary variables from nb (nb + 1)

to nb (blog2 nb c + 1), but increases the number of continuous variable by
nb (nb + 1) and constraints by nb .

39
Chapter 5

Numerical results
In this chapter we discuss the performance of the models and techniques
that are described in earlier chapters.

5.1 Experimental setup

The experiments were setup with 60 different test problems with various
section lengths, numbers of sections, blocks, access roads, and offset levels.
These 60 test problems were created with help from our industry partner.
We have used 7 roads that are listed in Table A.7. The configuration of all
60 test problems is listed in Tables A.8 and A.9. A sample input file for
multi-material case is shown in Table A.10.
All experiments were performed on a Dell workstation with an Intel(R)
Xeon(R) 3.20 GHz (4 cores) processors, a 24 GB of RAM and a 64-bit Win-
dows 7 Enterprise operating system. We used two different MILP solvers
to solve all the test problems. First, an academic edition of the IBM ILOG
CPLEX Optimizer v12.4 (http://www.cplex.com) and second, an open
source MILP solver CBC v2.7.7 (https://projects.coin-or.org/Cbc).
The models were programmed in C++ using MS Visual Studio 2010 Pro-
fessional edition.
We have two basic models: a quadratic model and a quasi network flow
model (see Chapters 2 and 3). In Chapter 4, we outlined several strategies
to modify both models to speedup the solution time. For offset levels, in
addition to the basic method, we use a reduced binary variable technique
and SOS2 variable technique (see Subsections 4.1.1 and 4.1.2). For the
cases of blocks, in addition to the basic way of dealing with blocks, we use
SOS1 variable technique and a reduced n log n technique (see Subsections
4.2.1 and 4.2.2). In sum, we have 2 models (quadratic, quasi network flow)
with 3 level modeling techniques (basic, reduced binary, SOS2) and 3 block
modeling techniques (basic, n log n, SOS1), creating 18 different strategies
and two solvers. These 18 strategies are shown in Table 5.1. All the test
problems are solved with a 1% MIP gap tolerance and a maximum 4 minutes
of CPU time.

40
 5.1. Experimental setup

Table 5.1: List of strategies for single material with blocks

Strategy Model Level Block

Technique Technique
QBB Quadratic Basic Basic
QBN Quadratic Basic n log n
QBS Quadratic Basic SOS1
QNB Quadratic Binary reduction Basic
QNN Quadratic Binary reduction n log n
QNS Quadratic Binary reduction SOS1
QSB Quadratic SOS2 Basic
QSN Quadratic SOS2 n log n
QSS Quadratic SOS2 SOS1
NBB Quasi network flow Basic Basic
NBN Quasi network flow Basic n log n
NBS Quasi network flow Basic SOS1
NNB Quasi network flow Binary reduction Basic
NNN Quasi network flow Binary reduction n log n
NNS Quasi network flow Binary reduction SOS1
NSB Quasi network flow SOS2 Basic
NSN Quasi network flow SOS2 n log n
NSS Quasi network flow SOS2 SOS1

41
 5.2. Performance measurement

5.2 Performance measurement

In order to compare the performance of different strategies, we use per-
formance profiles as developed in [DM02]. Performance profiles are designed
to graphically compare both speed and robustness of strategies across a test
set. This is done by plotting, for each strategy, the percentage of problems
that are solved within a factor of the best solve time. Let P be the set of all
problems, and S be the set of strategies. We first compute the performance
ratio
tp,s
rp,s = ,
min{tp,s : s ∈ S}
where p ∈ P and tp,s is the solving time of strategy s for problem p. The
percentage of problems that are solved within a factor τ ∈ R are computed
in logarithmic scale by

|{p ∈ P : log2 (rp,s ) ≤ τ }|

ρs (τ ) = × 100%.
|P|

The percentage ρs of all problems solved by strategy s at τ gives us an

overall assessment of the performance of strategy s. High values of ρs near
τ = 1 represents fast solve times. High values of ρs when τ is large represents
high success rates. Therefore, for a good algorithm, the plot of ρ(τ ) should
be found above the other algorithms. For a more detailed description of
performance profiles we refer the reader to [DM02].

5.3 Experimental results for single material

Out of the 60 test problems, 46 test problems contain blocks and 14 test
problems are setup without blocks. Some of the test problems are created
for stress testing. For example, Test 8 (see Table A.9) is created with 40
blocks in just 50 sections that cannot be solved by any strategies within
time limits. In reality, almost all problems have at most 5 and extremely
rarely 10 blocks. But we include harder problems with extreme conditions
to learn the limit of the models.
In Tables A.3 and A.4 the solution times required by solvers CPLEX
and CBC, respectively, are listed for 18 different strategies for the 46 test
problems with blocks. In Tables A.1 and A.2 the solution times are listed
for the 14 test problems with no blocks for 6 different techniques (see Table
5.2) for both MIP solvers.

42
 5.3. Experimental results for single material

5.3.1 Results for problems with blocks

The performance profiles of 18 different strategies with solver CPLEX
are shown in Figure 5.1. We have used MATLAB to draw the performance
profiles. Note that when MATLAB plots vertical lines the value of the
function ρs is equal to the largest values of the graph at that point. We
see that the performances of Strategies NSB, NSN and NSS are very close
to each other. These 3 strategies are all quasi network flow models with
the SOS2 level technique and three different block techniques. These three
strategies solved almost all of the problems in the problem set. Although
Strategy NSB is better than the other two strategies, strategy NSN, which is
the quasi network flow model with the SOS2 level technique and the n log n
block technique, is very close. However, strategy NSN solves one problem
less than strategy NSB.
1
QBB
QBN
QBS
0.9 QNB
QNN
QNS
QSB
0.8 QSN
QSS
NBB
NBN
0.7 NBS
NNB
NNN
NNS
0.6 NSB
NSN
NSS
ρs(log2(τ))

0.5

0.4

0.3

0.2

0.1

0
0 2 4 6 8 10 12
log2(τ)

Figure 5.1: Performance profile for solver CPLEX (with blocks)

Similarly, in the case of CBC, the performance profiles of 18 different

strategies are shown in Figure 5.2. Strategy NSS, which is a quasi network
flow model with the SOS2 level technique and the SOS1 block technique,
outperforms all the other strategies. It solves almost 90% of the problem
set. Strategy NSB comes next: it can also solve 90% of the problem set but

43
 5.3. Experimental results for single material

1
QBB
QBN
QBS
0.9 QNB
QNN
QNS
QSB
0.8 QSN
QSS
NBB
NBN
0.7 NBS
NNB
NNN
NNS
0.6 NSB
NSN
NSS
ρs(log2(τ))

0.5

0.4

0.3

0.2

0.1

0
0 1 2 3 4 5 6 7 8 9 10
log2(τ)

Figure 5.2: Performance profile for solver CBC (with blocks)

takes a little bit more time.

From Figures 5.1 and 5.2, we observe that the quasi network flow model
performs better than the quadratic model for both solvers for solving prob-
lems with blocks. However, for the block technique (Section 4.2.1 and 4.2.2),
we see that the SOS1 technique works better than the other techniques for
CBC, whereas the basic technique works well for CPLEX.
In Figures 5.3, 5.4, 5.5, and 5.6 the performance profiles are shown with
various techniques separately for the quadratic and quasi network flow mod-
els for each MIP solver. For CBC (Figures 5.4, 5.6), the SOS2 level tech-
nique and the SOS1 block technique work better than the other strategies
by a slight margin. However, for CPLEX (Figures 5.3, 5.5), the SOS2 level
technique combined with the basic block technique works better than other
strategies.

44
 5.3. Experimental results for single material

1
QBB
QBN
QBS
0.9 QNB
QNN
QNS
QSB
0.8 QSN
QSS

0.7

0.6
ρs(log2(τ))

0.5

0.4

0.3

0.2

0.1

0
0 1 2 3 4 5 6 7 8 9
log2(τ)

Figure 5.3: Performance profile of different strategies of the quadratic model

for solver CPLEX (with blocks)

1
QBB
QBN
QBS
0.9 QNB
QNN
QNS
QSB
0.8 QSN
QSS

0.7

0.6
ρs(log2(τ))

0.5

0.4

0.3

0.2

0.1

0
0 1 2 3 4 5 6 7 8
log2(τ)

Figure 5.4: Performance profile of different strategies of the quadratic model

for solver CBC (with blocks)

45
 5.3. Experimental results for single material

1
NBB
NBN
NBS
0.9 NNB
NNN
NNS
NSB
0.8 NSN
NSS

0.7

0.6
ρs(log2(τ))

0.5

0.4

0.3

0.2

0.1

0
0 2 4 6 8 10
log2(τ)

Figure 5.5: Performance profile of different strategies of the quasi network

flow model for solver CPLEX (with blocks)

1
NBB
NBN
NBS
0.9 NNB
NNN
NNS
NSB
0.8 NSN
NSS

0.7

0.6
ρs(log2(τ))

0.5

0.4

0.3

0.2

0.1

0
0 1 2 3 4 5 6 7 8 9 10
log2(τ)

Figure 5.6: Performance profile of different strategies of the quasi network

flow model for solver CBC (with blocks)

46
 5.3. Experimental results for single material

5.3.2 Results for problems with no blocks

For problems without blocks, the list of strategies is shown in Table 5.2.
The run times for these 6 strategies in the case of both solvers are listed
in Tables A.1 and A.2. The performance profiles for CPLEX and CBC
for these 6 strategies are shown in Figures 5.7 and 5.8, respectively. From
both performance profiles, we see that the SOS2 level technique makes a
significant difference. Both quasi network flow and quadratic models with
SOS2 can solve all of the problems, and as before, the quasi network flow
model outperforms the quadratic model.

Table 5.2: List of strategies for single material without blocks

Strategy Model Level Trick

QB Quadratic Basic
QN Quadratic Binary reduction
QS Quadratic SOS2
NB Quasi network flow Basic
NN Quasi network flow Binary reduction
NS Quasi network flow SOS2

5.3.3 Summary of experimental result for single material

The basic quadratic model (QBB) with the basic level and block tech-
niques can solve 38% of the problem set with blocks using CPLEX and 20%
using CBC. Using the quasi network flow model (NSB), with the SOS2 level
technique and the basic block technique, we improve this to 95% for CPLEX
and 88% for CBC.
To compare the improvement in solution time, we consider the common
problems that can be solved by all strategies. The improvement in solution
time for the problems with blocks are listed below.

− For the solver CPLEX, we found the following results.

– The quadratic model with the basic level and block technique can
solve 17 problems out of 46 test problems taking 1197 seconds.
Whereas the quadratic model with the SOS2 level technique and
the basic block technique can solve those 17 problems in 151
seconds. The improvement in speed is a factor of 8.

47
 5.3. Experimental results for single material

1
QB
QN
QS
0.9 NB
NN
NS

0.8

0.7

0.6
ρs(log2(τ))

0.5

0.4

0.3

0.2

0.1

0
0 2 4 6 8 10 12
log2(τ)

Figure 5.7: Performance profile for solver CPLEX (without blocks)

1
QB
QN
QS
0.9 NB
NN
NS

0.8

0.7

0.6
ρs(log2(τ))

0.5

0.4

0.3

0.2

0.1

0
0 2 4 6 8 10 12
log2(τ)

Figure 5.8: Performance profile for solver CBC (without blocks)

48
 5.3. Experimental results for single material

– The quasi network flow model with the basic level and block tech-
nique can solve 41 test problems taking 1755 seconds. Whereas
the quasi network flow model with the SOS2 level technique and
the basic block technique takes 175 seconds to solve those prob-
lems. In this case the improvement in speed is a factor of 10.
– The improvement from the basic quadratic model to quasi net-
work flow model introduced in [Rah12] with SOS2 level technique
is a factor of 173.

− For the solver CBC, we found the following results.

– The basic strategy QBB (quadratic model with basic level and
block technique) takes 785 seconds to solve 9 problems out of
46 problems. Whereas the Strategy QSB takes 50 seconds and
Strategy QSS takes 205 seconds to solve those 9 problems. The
performance improvement in solution time from QBB to QSB is
a factor 16.
– On the other hand, for the quasi network flow model, the basic
strategy NBB takes 2128 seconds to solve 36 problems out of 46.
Whereas the strategy NSS (quasi network flow with SOS2 level
and SOS1 block technique) takes 623 seconds. The improvement
is a factor of 3. The performance improvement from QBB to NSS
is a factor of 142.

For no blocks, the basic quadratic model (QB) with basic level technique
can solve 42% of the problem set using CPLEX and 50% using CBC. Using
both the quadratic (QS) and the quasi network flow model(NS) with SOS2
level technique, we improve this to 100% for both solvers. For problems
without blocks, the performance improvement in solution time for different
strategies are listed below.

− The solution time is improved from basic quadratic model (QB) to

quadratic model with SOS2 level technique (QS) by a factor of 244.2.

− The solution time is improved from basic quasi network flow model
(NB) to quasi network flow model with SOS2 level technique (NS) by
a factor of 94.1.

− The solution time is improved from basic quadratic model (QB) to

quasi network flow model with SOS2 level technique (NS) by a factor
of 1498.3. Similar results can be found for CBC.

49
 5.3. Experimental results for single material

The precedence of all techniques are shown in Tables 5.3, 5.4 for CPLEX
and CBC respectively in terms of number of problems solved and solution
speed.
Although the performance of the quasi network flow model is better than
the quadratic model, the quasi network flow model cannot have a non linear
cost function with respect to hauling distance. The quadratic model is more
flexible as it can handle a nonlinear cost function.

Table 5.3: Precedence of techniques for CPLEX in terms of number of prob-

lems solved and solution speed. Precedence is decreasing from left to right
in each row.
CPLEX
Model Quasi network flow Quadratic
Level SOS2 Basic Binary SOS2 Basic Binary
Technique Reduction Reduction
Block Basic SOS1 n log n Basic SOS1 n log n
Technique

Table 5.4: Precedence of techniques for CBC in terms of number of problems

solved and solution speed. Precedence is decreasing from left to right in each
row.
CBC
Model Quasi network flow Quadratic
Level SOS2 Basic Binary SOS2 Basic Binary
Technique Reduction Reduction
Block SOS1 Basic n log n SOS1 Basic n log n
Technique

50
 5.4. Experimental results for multimaterial

5.4 Experimental results for multimaterial

In this section, we show the experimental results for the multimaterial
vertical alignment optimization. We use the same 60 test problems that
we have used for single material, but this time we consider four materials
instead of one single material, and give ten minutes CPU time to solve a
problem. We also use the same configuration for the computer and software
(see Section 5.1).
In multimaterial test cases, the volumes of materials are mostly non-
convex in nature with respect to the height of the offset levels. So we do
not use the basic and binary reduction techniques for offset levels since
they requires that the volumes are convex in nature. We apply only SOS2
techniques which can handle non-convex volumes with respect to heights.

5.4.1 Results for problems with blocks

The strategies that we have used for numerical testing multimaterial
models with blocks are listed in Table 5.5. In Table A.5 the solution times
required by CPLEX are listed for the 46 test problems with blocks. CBC
cannot solve any of the test problem of mulimaterial.

Table 5.5: List of strategies without blocks for multimaterial test cases

Strategy Model Level Block

Technique Technique
QSB Quadratic SOS2 Basic
QSN Quadratic SOS2 n log n
QSS Quadratic SOS2 SOS1
NSB Quasi network flow SOS2 Basic
NSN Quasi network flow SOS2 n log n
NSS Quasi network flow SOS2 SOS1

The performance profiles of these six strategies is shown in Figure 5.9. In

this case we observe the same behavior that we have seen for single material.
NSB and NSS solve almost all the problems from the test set, although NSB
has better solution time. NSN solves 88 percent problem. Among the three
quasi network flow strategies, NSB is the best in terms of solution speed
and robustness.
In the case of strategies with the quadratic model, the performance of
the QSN and QSB are very close and significantly better than the other

51
 5.4. Experimental results for multimaterial

strategy QSS. Among the three strategy QSN is the best method in terms
of number of problem solved. The reason for this kind of performance profile
for the quadratic model is that the problems size grows very fast with the
number of materials and blocks. Since QSN reduces the number of binary
variables, the size of the problem is less than the other two strategies.
Overall, the performance profile with blocks for multimaterial test is
similar to the performance of single material.

1
QSB
QSN
QSS
0.9 NSB
NSN
NSS

0.8

0.7

0.6
ρs(log2(τ))

0.5

0.4

0.3

0.2

0.1

0
0 1 2 3 4 5 6 7
log2(τ)

Figure 5.9: Performance profile for solver CPLEX for multi-material (with
blocks)

5.4.2 Results for problems without blocks

For the case of without blocks, we now have only two strategies that are
listed in Table 5.2. The solution times required by CPLEX solver are listed
in Table A.6 for the 14 test cases without blocks. The performance profile is
shown in Figure 5.10. In this case, it is obvious that the quasi network flow
strategy with SOS2 level technique is better than the quadratic strategy
with SOS2.

52
 5.4. Experimental results for multimaterial

Table 5.6: List of strategies with blocks for multimaterial test cases

Strategy Model Level

Technique
QS Quadratic SOS2
NS Quasi network flow SOS2

1
QS
NS

0.9

0.8

0.7

0.6
ρs(log2(τ))

0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
log2(τ)

Figure 5.10: Performance profile for solver CPLEX for multi-material (with-
out blocks)

53
 5.4. Experimental results for multimaterial

5.4.3 Summary of results for mutlimaterial

We observe almost the same behavior in performance for the multima-
terial case as we have seen in single material in terms of number of problem
solved. For the mutimaterial case, the problem size increases as the number
of material increases. Solver CPLEX can cope with the increase of problem
size, but solver CBC is not able to solve multimaterial problems.
In the case of multimaterial, we have to consider nonconvexity in volume
with respect to the height offset level. In general, all strategies with quasi
network flow is significantly better than the strategies with quadratic model.
In order to measure the performance in solution time explicitly, we consider
the problems that are solved by both strategies. QSB can solve 27 problems
with blocks out of 46 problems taking 3675 seconds in total. NSB takes
148 seconds for the same problem set. So performance increase in solution
time is a factor of 25. Without block problems, the performance increase
in solution time is 11 from QS to NS. This performance increase is caused
mostly by the quasi network flow model, while SOS2 techniques help to
increase performance for both models and accommodate the nonconvexity
relation between volume and offset level.

54
Chapter 6

Conclusion
We started this thesis with the practical requirements that need to be
considered for vertical alignment design in Chapter 1. We also discussed the
shortcomings of previous research works. We developed two models based
on the previous works of [Mor09], [HKL11], [Rah12]. The quadratic and
quasi network flow models are an enhancement of the previous work done
in [Rah12] and are described in Chapters 2 and 3. These two models are
capable of considering all requirements for optimal design of the vertical
alignment. Both models incorporate features such as multimaterial, sides
slopes and physical blocks.
We further introduced six potential techniques to improve the solution
speed that are applicable for both models in Chapter 4. Numerical tests
shows that using the best combination of techniques can consistantly gen-
erate major improvements in solution time, regardless of which LP solver
is used. For example, the quasi network flow model introduced in [Rah12]
and further improved by the best speed techniques discussed in this research
improves solution speed from the basic quadratic model by a factor of 173.
The techniques developed in this research have made it possible to solve
a vertical alignment design for a very long road within a reasonable amount
of time. Using these improvement techniques, our industry partner (Softree
Technical Systems Inc.) has developed an interactive road design software.

6.1 Future research

6.1.1 Improvement in the quasi network flow model
The quasi network flow model has a limitation in the structure of its cost
function. The hauling cost is linear with distance traveled by the volume of
materials. In practice, for hauling materials, it may be enough to consider
only two kinds of haul such as short haul and long haul. Short haul refers
to moving material in short distance. Long haul refers to moving materials
in long distance using big trucks. There could be a fixed cost for loading
materials in long haul which is much higher than hauling materials in short

55
 6.1. Future research

distance. However, once the materials are put in a large truck, a long travel-
ing distance does not increase the cost significantly. To accommodate these
two hauls in our quasi network flow model, we can define another pair of
East and West transit flow. Then with appropriate cost values set for the
short and long haul it is probably possible to have a nonlinear cost function.

6.1.2 More options for selecting material

In our existing models, we cut and fill various sections of roads with
the materials according to the given input volume tables (Table 2.1). We
assume that materials are not replaceable by any other type of materials.
But it can be modeled and implemented to have the alternate material
options. However, it will increase the size of the problem, as we have to
define all possible movement variables as xijt,m1 m2 , where m1 and m2 are
the alternate materials pair such that low quality material m1 is replaceable
with high quality material m2 .

6.1.3 Horizontal alignment

The ultimate goal of automating the road design process is to calculate
the horizontal alignment while considering the sub problem of vertical align-
ment. First, road engineers select a corridor for building a road using the
topographical analysis of the ground profile. After fixing a corridor, road
engineers select a route for the horizontal alignment within that corridor
minimizing the cost while satisfying the safety, other design constraints, and
considering other socioeconomic impacts. While computing the horizontal
alignment, the cost is heavily dependent on the cost of the vertical align-
ment. Certainly, during the horizontal alignment optimization, our research
will help to find the optimal solution.

56
Bibliography
[AAS04] AASHTO, editor. A policy on geometric design of highways and
streets. AASHTO, 5 edition, 2004. → pages 2, 8

[ACRV13] E. Angulo, W. Castillo, R. G. Rdenas, and J. S. Vizcano. A

continuous bi-level model for the expansion of highway networks.
Computers & Operations Research, 2013. → pages 2

[Aka06] A. Akay. Minimizing total costs of forest roads with computer-

aided design model. Sadhana, 31:621–633, 2006. → pages 7

[AMO93] R. K. Ahuja, T. L. Magnanti, and J. B. B. Orlin. Network Flows:

Theory, Algorithms, and Applications. Prentice Hall, Englewood
Cliffs, NJ, 1993. → pages 6, 23

[Aru05] K. Aruga. Tabu search optimization of horizontal and vertical

alignments of forest roads. Journal of Forest Research, 10:275–
284, 2005. → pages 7

[ATSM06] K. Aruga, T. Tasaka, J. Sessions, and S. Miyata. Tabu search op-

timization of forest road alignments combined with shortest paths
and cubic splines. Croatian Journal of Forest Engineering, 27:37–
47, 2006. → pages 7

[CGF89] E. P. Chew, C. J. Goh, and T. F. Fwa. Simultaneous optimization

of horizontal and vertical alignments for highways. Transportation
Research Part B: Methodological, 23(5):315–329, 1989. → pages 6

[CL06] J. F. Cheng and Y. Lee. Model for three-dimensional highway

alignment. Journal of Transportation Engineering, 132(12):913–
920, 2006. → pages 7

[CMW98] R. L. Church, A. T. Murray, and A. Weintraub. Locational issues

in forest management. Location Science, 6(14):137 – 153, 1998. →
pages 2

57
 Bibliography

[DJ09] C. Davis and M. K. Jha. Modeling the effects of socioeconomic

factors in highway construction and expansion. Journal of Trans-
portation Engineering, 135(12):990–998, 2009. → pages 1

[DM02] E. D. Dolan and J. J. Moré. Benchmarking optimization software

with performance profiles. Mathematical Programming, 91(2):201–
213, 2002. → pages 42

[Eas88] S. M. Easa. Selection of roadway grades that minimize earthwork

cost using linear programming. Transportation Research Part A:
General, 22(2):121 – 136, 1988. → pages 4, 5

[FCS02] T. F. Fwa, W. T. Chan, and Y. P. Sim. Optimal vertical alignment

analysis for highway design. Journal of Transportation Engineer-
ing, 128(5):395–402, 2002. → pages 4

[GAA09] A. B. Göktepe, S. Altun, and P. Ahmedzade. Optimization of

vertical alignment of highways utilizing discrete dynamic program-
ming and weighted ground line. Turkish Journal Of Engineering
& Environmental Sciences, 10:105–116, 2009. → pages 4, 5

[GCF88] C. J. Goh, E. P. Chew, and T. F. Fwa. Discrete and continuous

models for computation of optimal vertical highway alignment.
Transportation Research Part B: Methodological, 22(6):399 – 409,
1988. → pages 4, 5

[GLA05] A. B. Goktepe, A. H. Lav, and S. Altun. Dynamic optimization

algorithm for vertical alignment of highways. Mathematical and
Computational Applications, 10(3):341–350, 2005. → pages 4, 5,
7

[GLA09] A. B. Goktepe, A. H. Lav, and S. Altun. Method for optimal

vertical alignment of highways. Proceedings of The Institution
of Civil Engineers-transport, 162(4):177–188, December 2009. →
pages 4

[HBS68] B. E. Howard, Z. Bramnick, and J. F. Shaw. Optimum curvature

principle in highway routing. Journal of the Highway Division,
94:61–82, 1968. → pages 3

[HHLR13] W. L. Hare, S. Hossain, Y. Lucet, and F. Rahman. Models and

strategies for efficiently determining an optimal vertical alignment
of roads. Technical report, Under submission, 2013. → pages iii

58
 Bibliography

[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 2, 4, 6, 8, 15, 55

[HLR13] W. Hare, Y. Lucet, and F. Rahman. A mixed-integer linear pro-

gramming model to optimize the vertical alignment considering
blocks and side-slopes in road construction. under submission.
Technical report, Under submission, 2013. → pages 6

[Jha03] M. K. Jha. Criteria-based decision support system for select-

ing highway alignments. Journal of Transportation Engineering,
129(1):33–41, 2003. → pages 1

[JS03] J. C. Jong and P. Schonfeld. An evolutionary model for simulta-

neously optimizing three-dimensional highway alignments. Trans-
portation Research Part B: Methodological, 37(2):107 – 128, 2003.
→ pages 2

[JSJ06] M. K. Jha, P. Schonfeld, and J. C. Jong. Intelligent Road Design,

volume 19. WIT Press, 2006. → pages 2, 3, 4, 5

[KdFJN04] A. B. Keha, I. R. de Farias Jr., and G. L. Nemhauser. Models for

representing piecewise linear cost functions. Operations Research
Letters, 32(1):44 – 48, 2004. → pages 33

[KdFN06] A. B. Keha, I. R. de Farias, and G. L. Nemhauser. A branch-and-

cut algorithm without binary variables for nonconvex piecewise
linear optimization. Operations Research, 54(5):847–858, Septem-
ber 2006. → pages 33

[KL10] V. R. Koch and Y. Lucet. A note on: Spline technique for model-
ing roadway profile to minimize earthwork cost. Journal of Indus-
trial and Management Optimization, 6(2):393 – 400, May 2010.
→ pages 4, 6, 8, 10

[KSJ07] M. W. Kang, P. Schonfeld, and J. C. Jong. Highway alignment

optimization through feasible gates. Journal of Advanced Trans-
portation, 41(2):115–144, 2007. → pages 2

[LC01] Y. S. Lee and J. F. Cheng. . Journal of Transportation

Engineering-ASCE, 127(4):303–310, Jul-Aug 2001. → pages 4

59
 Bibliography

[LV92] M.G. Lay and J.E. Vance. Ways of the World: A History of the
World’s Roads and of the Vehicles That Used Them. RUTGERS
University Press, 1992. → pages
[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, April 2004. →
pages 4, 5, 10
[Min12a] Minister of Public Works and Government Services. Transporta-
tion in canada 2011 - comprehensive review. Transport Canada,
Cat. No. T1-23A/2011E-PDF:15, 2012. → pages 1
[Min12b] Ministry of Transportation and Infrastructure. Construction and
rehabilitation cost guide, July 2012. → pages 1
[MMM06] A. Martin, M. Mller, and S. Moritz. Mixed integer models for
the stationary case of gas network optimization. Mathematical
Programming, 105:563–582, 2006. → pages 33
[Mor96] A. A. Moreb. Linear programming model for finding optimal road-
way grades that minimize earthwork cost. European Journal of
Operational Research, 93(1):148 – 154, 1996. → pages 4, 5
[Mor09] A. A. Moreb. Spline technique for modeling roadway profile to
minimize earthwork cost. Journal of Industrial and Management
Optimization (JIMO), 5(2):275–283, 2009. → pages 2, 4, 5, 6, 8,
10, 55
[OEC73] OECD. Optimisation of road alignment by the use of computers.
Technical report, Organization of Economic Co-operation and De-
velopment, Paris, 1973. → pages 4
[Rah12] F. Rahman. Optimizing the vertical alignment under earthwork
block removal constraints in road construction. Master’s thesis,
UBC, 2012. → pages 6, 8, 16, 23, 49, 55
[SH81] J. F. B. Shaw and B. E. Howard. Comparison of two integra-
tion methods in transportation routing. Transportation Research
Record, 806:8–13, 1981. → pages 3
[SH82] J. F. B. Shaw and B. E. Howard. Expressway route optimization
by ocp. Transportation Engineering Journal, 108:227–243, 1982.
→ pages 3

60
 Bibliography

[TM70] A.K. Turner and R. D. Miles. The GCARS system: a computer-

assisted method of regional route location. Technical Report
FHWA/IN/JHRP-70/27, Purdue University, Joint Highway Re-
search Project, 1970. → pages 4

[Tom81] J. A. Tomlin. A suggested extension of special ordered sets to non-

separable non-convex programming problems. Studies on Graphs
and Discrete Programming, 11:359–370, 1981. → pages 33

[Tri87a] D. Trietsch. Comprehensive design of highway networks. Trans-

portation Science, 21(1):26–35, 1987. → pages 4

[Tri87b] D. Trietsch. A family of methods for preliminary highway align-

ment. Transportation Science, 21(1):17–25, February 1987. →
pages 4

[TT03] C. W. Tat and F. Tao. Using gis and genetic algorithm in highway
alignment optimization. In Intelligent Transportation Systems,
2003. Proceedings. 2003 IEEE, volume 2, pages 1563 – 1567 vol.2,
oct. 2003. → pages 7

[Tur78] A. K. Turner. A decade of experience in computer aided route

selection. Photogrammetric Engineering and Remote Sensing,
44:1561–1576, 1978. → pages 4

[Wan95] F. Y. M. Wan. Introduction to the calculus of variations and its

applications. Chapman & Hall, 1995. → pages 3

61
Appendix

62
Appendix A

Tables

Table A.1: Solution times for CPLEX with 6 strategies for 14 test problems
without blocks
Test # QB QN QS NB NN NS
1 1.485 NaN 0.032 1.181 NaN 0.021
2 1.134 NaN 0.027 1.119 NaN 0.018
3 NaN NaN 0.028 0.647 NaN 0.023
4 31.377 136.647 0.06 1.005 203.027 0.031
5 NaN NaN 0.065 1.474 NaN 0.035
6 74.593 NaN 0.128 2.514 NaN 0.046
7 NaN NaN 0.122 3.144 NaN 0.05
8 NaN NaN 0.125 4.97 NaN 0.044
9 199.839 NaN 0.176 5.671 NaN 0.056
10 NaN NaN 0.155 4.062 NaN 0.047
11 NaN NaN 1.605 18.932 NaN 0.147
12 NaN NaN 1.312 22.998 NaN 0.189
13 NaN NaN 1.322 12.813 NaN 0.119
14 156.044 NaN 1.479 10.17 86.527 0.138

63
 Appendix A. Tables

Table A.2: Solution times for CBC with 6 strategies for 14 test problems
without blocks
Test # QB QN QS NB NN NS
1 45.716 NaN 0.057 101.07 NaN 0.04
2 15.417 NaN 0.034 38.33 NaN 0.03
3 31.591 NaN 0.047 55.593 NaN 0.043
4 12.985 NaN 0.214 36.97 NaN 0.084
5 35.614 NaN 0.181 23.706 NaN 0.091
6 NaN NaN 0.359 NaN NaN 0.115
7 NaN NaN 0.281 112.62 NaN 0.123
8 NaN NaN 0.333 45.162 NaN 0.149
9 160.272 NaN 0.602 15.58 NaN 0.192
10 48.029 NaN 0.541 9.091 NaN 0.164
11 NaN NaN 5.05 NaN NaN 0.664
12 NaN NaN 4.999 NaN NaN 0.8
13 NaN NaN 5.392 161.062 NaN 0.61
14 NaN NaN 5.32 8.943 89.347 0.564

64
 Table A.3: Solution time for CPLEX with 18 strategies for 48 test problems with blocks
Test # QBB QBN QBS QNB QNN QNS QSB QSN QSS NBB NBN NBS NNB NNN NNS NSB NSN NSS

1 1.45 3.14 2.45 11.69 16.11 10.95 0.27 0.28 1.23 0.70 0.97 0.94 4.28 12.23 4.77 0.05 0.05 0.08
2 3.03 3.46 3.57 70.85 133.41 123.73 0.40 0.38 1.60 1.62 1.56 1.40 11.50 19.18 9.02 0.05 0.05 0.06
3 9.55 26.27 24.89 NaN NaN NaN 1.46 1.43 9.06 2.26 2.28 1.92 33.45 63.37 52.82 0.11 0.10 0.16
4 NaN NaN NaN NaN NaN NaN 20.88 34.66 223.09 8.91 14.00 52.68 102.79 NaN 207.15 0.37 0.43 0.75
5 43.44 23.92 49.00 NaN NaN NaN 2.53 2.42 18.64 3.48 2.28 4.13 83.02 78.67 240.34 0.18 0.17 0.42
6 54.05 43.39 127.68 NaN NaN NaN 4.05 3.29 42.56 4.53 8.71 4.98 48.24 64.71 41.54 0.16 0.30 0.36
7 89.68 175.70 NaN NaN NaN NaN 14.19 12.39 102.07 17.15 25.66 52.08 NaN NaN NaN 1.41 6.28 10.40
8 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
9 NaN NaN NaN NaN NaN NaN 28.58 20.61 114.49 25.56 17.55 49.82 NaN NaN NaN 1.43 1.34 2.74
10 111.67 NaN NaN NaN NaN NaN 16.52 6.47 75.03 39.76 15.58 21.86 NaN NaN NaN 0.95 0.78 0.90
11 NaN NaN NaN NaN NaN NaN 50.46 78.99 NaN 40.62 46.36 27.03 NaN NaN NaN 1.34 1.18 1.18
12 NaN NaN NaN NaN NaN NaN 78.59 114.20 NaN 77.60 58.30 164.51 NaN NaN NaN 3.84 4.55 1.10
13 57.78 56.80 92.90 NaN NaN NaN 4.54 4.12 23.42 7.96 7.49 13.95 39.06 NaN 91.41 0.30 0.38 0.36
14 43.98 92.03 51.95 NaN NaN NaN 1.15 1.19 6.17 15.62 13.07 14.79 NaN 34.16 130.92 0.28 0.11 0.12
15 60.40 NaN NaN NaN NaN NaN 12.46 11.75 191.66 11.28 22.65 23.36 NaN NaN NaN 0.52 0.42 0.81
16 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
17 NaN NaN NaN NaN NaN NaN 33.28 28.50 132.78 34.84 18.33 34.44 NaN NaN NaN 2.12 1.93 2.17
18 NaN NaN NaN NaN NaN NaN 115.21 101.36 NaN 16.40 22.56 28.85 NaN NaN NaN 2.87 4.89 5.68
19 231.87 NaN NaN NaN NaN NaN 47.39 43.10 NaN 14.62 24.40 13.07 NaN 203.81 NaN 1.36 1.33 1.37
20 NaN NaN NaN NaN NaN NaN 122.71 89.45 NaN 21.00 16.15 65.18 NaN NaN NaN 1.60 1.70 6.48
21 NaN NaN NaN NaN NaN NaN NaN NaN NaN 70.54 100.87 206.28 NaN NaN NaN 23.45 14.16 11.26
22 NaN NaN NaN NaN NaN NaN NaN NaN NaN 210.73 NaN NaN NaN NaN NaN 91.53 225.44 131.60
23 150.65 NaN NaN NaN NaN NaN 25.07 22.78 NaN 7.82 30.89 18.62 NaN NaN NaN 0.36 0.38 1.09
24 27.61 56.59 110.77 NaN NaN NaN 3.76 3.18 50.44 2.20 3.46 3.60 15.62 33.19 37.86 0.13 0.13 0.13
25 NaN NaN NaN NaN NaN NaN 14.62 10.13 NaN 13.57 15.09 45.34 NaN NaN NaN 0.22 0.22 0.41
26 NaN NaN NaN NaN NaN NaN 109.03 NaN NaN 58.77 231.60 220.83 NaN NaN NaN 0.75 0.77 2.48
27 109.58 182.50 NaN NaN NaN NaN 6.57 6.12 68.05 29.95 26.31 40.93 NaN NaN NaN 0.16 0.24 0.20
Appendix A. Tables

28 NaN NaN NaN NaN NaN NaN 10.33 6.37 200.25 39.88 60.48 62.07 NaN NaN NaN 0.20 0.17 0.19
29 70.76 41.68 99.71 NaN NaN NaN 2.25 1.92 24.03 35.54 25.09 35.96 NaN NaN NaN 0.11 0.11 0.20
30 NaN NaN NaN NaN NaN NaN 40.41 27.70 NaN 198.18 20.70 11.30 NaN NaN NaN 0.44 0.64 0.62
31 54.89 127.81 70.25 NaN NaN NaN 6.99 6.77 33.75 28.95 11.31 11.08 NaN NaN NaN 0.16 0.14 0.17
32 NaN NaN NaN NaN NaN NaN 54.91 45.04 NaN 19.65 19.08 27.24 NaN NaN NaN 0.33 0.34 1.10
33 NaN NaN NaN NaN NaN NaN 36.86 21.53 NaN 21.19 26.75 26.01 NaN NaN NaN 0.28 0.29 0.45
34 NaN NaN NaN NaN NaN NaN 51.57 36.25 NaN 33.64 21.29 44.98 NaN NaN NaN 0.64 0.58 2.80
35 NaN NaN NaN NaN NaN NaN 65.12 40.16 NaN 15.32 17.66 20.64 NaN NaN NaN 0.48 0.41 0.58
36 NaN NaN NaN NaN NaN NaN 186.31 209.07 NaN 136.10 50.84 165.77 NaN NaN NaN 1.42 1.54 5.24
37 NaN NaN NaN NaN NaN NaN 212.62 NaN NaN 50.16 155.81 22.39 NaN NaN NaN 2.59 2.50 2.01
38 76.63 76.14 71.28 NaN NaN NaN 1.44 0.59 0.50 9.66 9.27 9.16 NaN NaN NaN 0.66 0.11 0.53
39 NaN NaN NaN NaN NaN NaN 49.95 34.84 NaN 18.57 24.80 27.78 NaN NaN NaN 0.36 0.45 0.76
40 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN 7.39 NaN 135.91
41 NaN NaN NaN NaN NaN NaN NaN NaN NaN 170.95 NaN NaN NaN NaN NaN 24.57 22.40 107.02
42 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN 216.44 NaN NaN NaN NaN 5.66 7.47 5.07
43 NaN NaN NaN NaN NaN NaN NaN NaN NaN 133.79 65.11 228.36 NaN NaN NaN 4.01 5.28 4.68
44 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN 2.75 3.74 5.73

65
45 NaN NaN NaN NaN NaN NaN NaN NaN NaN 95.16 144.01 104.32 NaN NaN NaN 2.36 2.11 2.73
46 NaN NaN NaN NaN NaN NaN NaN NaN NaN 10.64 12.03 13.62 49.16 69.87 49.84 1.23 1.42 2.01
 Table A.4: Solution time for CBC with 18 strategies for 48 test problems with blocks
Test # QBB QBN QBS QNB QNN QNS QSB QSN QSS NBB NBN NBS NNB NNN NNS NSB NSN NSS

1 7.94 14.92 12.48 NaN NaN NaN 1.78 5.44 0.53 3.22 4.95 4.57 NaN NaN NaN 0.47 1.61 0.17
2 84.62 29.96 30.63 NaN NaN NaN 0.49 127.72 1.00 21.67 6.74 3.85 NaN NaN 217.71 0.29 10.10 0.25
3 106.41 89.66 108.05 NaN NaN NaN 25.03 NaN 7.28 45.20 58.22 4.96 NaN NaN NaN 3.70 19.47 1.99
4 NaN NaN NaN NaN NaN NaN 199.37 NaN 191.11 102.06 207.69 40.13 NaN NaN NaN 20.91 NaN 2.94
5 NaN NaN NaN NaN NaN NaN 162.16 NaN 17.63 27.07 86.68 59.04 NaN NaN NaN 7.68 27.18 3.84
6 86.93 NaN NaN NaN NaN NaN 5.75 NaN 153.05 15.15 91.81 19.47 NaN NaN NaN 1.26 21.18 0.69
7 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
8 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
9 NaN NaN NaN NaN NaN NaN NaN NaN NaN 74.79 NaN 50.99 NaN NaN NaN 22.06 24.67 38.00
10 NaN NaN NaN NaN NaN NaN 89.99 NaN NaN 44.77 21.55 28.32 NaN NaN NaN 2.32 18.69 9.74
11 NaN NaN NaN NaN NaN NaN NaN NaN NaN 100.17 184.09 149.99 NaN NaN NaN 36.60 59.15 22.68
12 NaN NaN NaN NaN NaN NaN NaN NaN NaN 55.86 231.40 87.56 NaN NaN NaN 80.08 NaN 27.12
13 NaN 167.28 NaN NaN NaN NaN 12.29 NaN 22.33 14.79 11.78 15.24 NaN NaN NaN 2.28 9.18 1.83
14 31.92 59.55 30.01 NaN NaN NaN 1.33 38.83 1.53 40.16 8.86 47.30 NaN NaN NaN 0.74 0.70 0.36
15 NaN NaN NaN NaN NaN NaN NaN NaN NaN 92.78 172.95 19.83 NaN NaN NaN 26.32 30.48 24.78
16 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
17 NaN NaN NaN NaN NaN NaN 17.16 NaN 11.57 107.00 78.94 106.88 NaN NaN NaN 1.71 32.75 0.99
18 NaN NaN NaN NaN NaN NaN NaN NaN NaN 113.88 NaN NaN NaN NaN NaN 62.15 94.80 153.02
19 NaN NaN NaN NaN NaN NaN 25.78 NaN 55.81 139.14 144.08 36.36 NaN NaN NaN 3.61 62.36 1.07
20 NaN NaN NaN NaN NaN NaN 175.26 NaN NaN NaN NaN 121.82 NaN NaN NaN 7.34 58.31 57.15
21 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN 222.24 NaN 10.72
22 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
23 NaN NaN NaN NaN NaN NaN 32.30 NaN 26.36 46.26 192.51 86.85 NaN NaN NaN 2.78 49.74 2.46
24 144.72 NaN NaN NaN NaN NaN 4.49 NaN 9.16 8.78 5.43 27.27 228.01 151.99 NaN 0.59 6.39 0.67
25 NaN NaN NaN NaN NaN NaN NaN NaN 112.01 56.22 76.39 52.62 NaN NaN NaN 17.29 37.68 5.51
26 NaN NaN NaN NaN NaN NaN NaN NaN NaN 102.68 NaN 89.12 NaN NaN NaN 97.88 96.15 93.68
27 NaN NaN NaN NaN NaN NaN 24.09 NaN 21.63 37.81 13.90 33.54 NaN NaN NaN 1.81 11.62 1.84
Appendix A. Tables

28 NaN NaN NaN NaN NaN NaN 41.81 NaN 8.86 18.80 107.69 22.53 NaN NaN NaN 0.84 10.58 0.89
29 125.53 129.79 114.93 NaN NaN NaN 4.34 NaN 4.40 23.74 17.82 25.01 NaN NaN NaN 0.47 1.66 0.43
30 NaN NaN NaN NaN NaN NaN NaN NaN NaN 114.71 117.30 25.19 NaN NaN NaN 28.48 45.83 11.26
31 137.33 134.24 178.72 NaN NaN NaN 5.34 NaN 5.97 34.70 17.44 16.66 NaN NaN NaN 0.50 0.50 0.45
32 NaN NaN NaN NaN NaN NaN 35.37 NaN 58.96 44.59 67.66 41.11 NaN NaN NaN 2.64 65.39 3.54
33 NaN NaN NaN NaN NaN NaN NaN NaN 99.76 29.52 19.02 37.11 NaN NaN NaN 14.04 23.46 7.82
34 NaN NaN NaN NaN NaN NaN NaN NaN NaN 216.00 160.99 135.16 NaN NaN NaN 14.73 86.14 27.83
35 NaN NaN NaN NaN NaN NaN NaN NaN NaN 65.16 149.84 80.59 NaN NaN NaN 27.01 56.27 9.78
36 NaN NaN NaN NaN NaN NaN NaN NaN NaN 122.57 NaN NaN NaN NaN NaN 80.26 134.80 96.22
37 NaN NaN NaN NaN NaN NaN NaN NaN NaN 147.59 NaN 230.79 NaN NaN NaN 110.57 NaN 60.43
38 59.80 94.50 81.45 NaN NaN NaN 1.78 1.65 1.77 11.87 17.77 13.85 NaN NaN NaN 1.53 0.44 0.53
39 NaN NaN NaN NaN NaN NaN NaN NaN 93.20 20.81 120.71 50.20 NaN NaN 157.72 20.95 35.08 5.34
40 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
41 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN
42 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN 127.43 NaN NaN NaN 208.73 NaN 111.51
43 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN 101.28 195.48 91.70
44 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN 92.31 221.35 21.92

66
45 NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN NaN 211.04 NaN NaN NaN 15.44 116.83 8.25
46 NaN NaN NaN NaN NaN NaN NaN NaN NaN 28.65 31.34 28.30 47.49 201.13 163.63 5.12 45.18 4.43
 Appendix A. Tables

Table A.5: Solution time for CPLEX with 6 strategies for 48

test problems with blocks for multimaterial

Test # QSB QSN QSS NSB NSN NSS

1 1.611 2.551 2.658 0.54 0.491 0.641
2 1.733 1.729 3.252 0.315 0.37 0.476
3 11.101 12.924 27.801 1.166 1.15 1.82
4 144.51 187.665 NaN 7.579 21.74 10.832
5 15.376 15.218 43.646 1.449 1.412 2.73
6 NaN 175.466 NaN 7.028 5.507 6.925
7 NaN NaN NaN 42.762 NaN 113.876
8 NaN NaN NaN 125.299 NaN 241.943
9 32.174 98.9 135.887 3.771 3.05 11.003
10 16.287 16.69 86.277 1.269 1.085 1.674
11 208.916 214.335 NaN 4.24 6.295 4.961
12 509.851 577.067 NaN 5.358 10.068 6.342
13 10.091 9.078 29.145 0.909 0.993 1.637
14 8.547 2.881 8.911 0.506 0.403 0.62
15 77.049 47.665 238.032 2.523 3.176 3.601
16 NaN NaN NaN 19.82 202.195 17.395
17 21.439 31.844 143.385 1.988 1.79 2.249
18 318.855 370.849 NaN 20.114 29.77 45.046
19 NaN 430.451 NaN 4.022 6.429 5.086
20 265.369 330.761 NaN 18.988 25.7 60.412
21 NaN NaN NaN 28.216 40.785 45.003
22 NaN NaN NaN 26.515 92.717 74.644
23 197.132 220.926 596.321 12.628 19.769 32.604
24 8.105 6.591 61.082 0.992 0.855 0.779
25 90.885 126.161 469.719 3.972 4.225 6.601
26 NaN NaN NaN 23.249 46.813 37.171
27 37.006 52.907 106.244 3.054 2.324 3.749
28 38.6 88.74 240.214 2.301 2.148 2.583
29 13.463 13.805 36.571 1.842 1.853 2.149
30 223.456 282.791 NaN 4.706 6.533 5.817
31 80.338 108.861 106.847 7.014 7.347 10.067
32 168.785 162.139 565.35 10.863 11.766 18.709
33 237.192 278.53 NaN 7.697 9.289 11.594
34 NaN NaN NaN 20.627 19.715 54.811
Continued on next page

67
 Appendix A. Tables

Table A.5 – continued from previous page

Test # QSB QSN QSS NSB NSN NSS
35 248.307 418.303 NaN 7.933 9.46 9.372
36 NaN NaN NaN 34.481 44.471 96.462
37 NaN NaN NaN 25.11 54.156 33.075
38 17.273 17.247 17.38 2.749 2.817 2.827
39 100.066 130.871 NaN 3.013 4.826 7.016
40 NaN NaN NaN NaN NaN NaN
41 NaN NaN NaN 304.881 NaN 511.828
42 NaN NaN NaN 62.56 391.583 84.689
43 NaN NaN NaN 74.586 135.362 100.502
44 NaN NaN NaN NaN NaN NaN
45 NaN NaN NaN 28.381 36.46 35.489
46 572.097 NaN NaN 8.617 8.713 12.912

Table A.6: Solution time for CPLEX with two strategies for 16 test problems
without blocks for multimaterial
Test # QS NS
1 0.307 0.16
2 0.685 0.271
3 9.62 4.192
4 0.686 0.301
5 2.63 1.78
6 9.117 1.454
7 25.994 3.433
8 17.785 3.108
9 65.313 7.668
10 16.316 2.54
11 NaN NaN
12 NaN NaN
13 270.647 15.958
14 44.071 1.401

68
 Appendix A. Tables

Table A.7: Test road configuration

Road Section Length Section Count Size in kilometre

A 20 50 1
B 100 50 5
C 20 100 2
D 20 150 3
E 100 150 15
F 100 200 20
G 20 450 9

69
 Appendix A. Tables

Table A.8: Test configuration of 60 test problems in terms of

number of access roads, blocks, and offset levels

Test # Road Access Rd Blocks Offset Levels

1 A 1 3 10
2 A 2 5 10
3 A 5 10 10
4 A 20 25 10
5 A 4 12 10
6 B 10 15 10
7 B 4 20 10
8 B 8 40 10
9 C 2 7 10
10 C 3 8 10
11 C 13 19 10
12 C 17 22 10
13 C 1 5 10
14 C 2 3 10
15 C 4 12 10
16 C 20 30 10
17 C 5 10 16
18 D 3 12 10
19 D 9 10 10
20 D 2 7 10
21 D 25 30 10
22 D 5 20 10
23 D 3 8 10
24 D 4 5 6
25 E 7 9 10
26 E 10 21 10
27 E 2 5 10
28 E 6 8 10
29 E 1 3 10
30 E 10 15 10
31 F 1 2 10
32 F 2 5 10
33 F 4 8 10
34 F 2 10 10
Continued on next page

70
 Appendix A. Tables

Table A.8 – continued from previous page

Test # Road Access Rd Blocks Offset Levels
35 F 8 12 10
36 F 5 15 10
37 F 10 20 10
38 F 1 1 10
39 F 4 7 6
40 G 25 40 20
41 G 15 30 10
42 G 20 22 10
43 G 10 15 10
44 G 5 10 20
45 G 7 9 10
46 G 3 4 4
47 A 10 0 40
48 A 1 0 30
49 B 1 0 40
50 C 1 0 16
51 C 1 0 24
52 D 1 0 16
53 D 1 0 22
54 E 1 0 24
55 F 1 0 16
56 F 1 0 10
57 G 1 0 16
58 G 1 0 30
59 G 1 0 10
60 G 1 0 4

71
 Table A.9: Test configuration for 60 test problems with patterns of blocks and access roads associated with sections
Test # Road Offset Levels Access Roads Blocks

1 A 10 1 10;25;37
2 A 10 15;40 10;20;30;43;45
3 A 10 1;10;19;30;50 2;3;7;13;17;24;25;40;41;45
4 A 10 1;3;5;7;9;11;15;17;20;23;26;29;33;35;38;41;43;46;48;50 2;4;6;8;10;12;13;14;16;18;19;21;22;24;27;28;30;31;32;36;40;42;44;47;49
5 A 10 15;30;40;50 4;5;9;10;20;21;22;23;24;35;45;46
6 B 10 1;5;10;15;20;25;30;35;40;50 2;3;7;13;18;23;28;33;34;38;43;44;45;46;49
7 B 10 1;25;30;50 4;5;6;9;12;14;15;16;19;20;21;26;27;33;34;38;40;45;46;47
8 B 10 4;10;15;21;28;36;42;48 1;2;3;5;6;7;8;9;11;12;13;14;16;17;18;19;20;22;23;24;25;26;27;29;30;31;32;33;34;35;37;39;40;41;43;44;45;46;47;49
9 C 10 1;100 20;30;40;41;55;60;70
10 C 10 30;50;80 20;25;40;60;61;70;85;90;
11 C 10 1;10;20;25;30;40;45;50;60;70;80;90;100 5;6;14;15;22;27;33;34;35;43;48;55;56;65;68;73;75;85;96
12 C 10 1;7;14;20;25;30;35;44;50;55;60;66;70;75;80;90;100 4;5;10;11;16;22;23;27;33;38;40;41;48;53;58;64;68;73;78;81;82;94
13 C 10 50 10;20;30;60;70
14 C 10 20;50 10;30;60
15 C 10 10;30;50;80 5;6;17;25;35;40;45;55;60;85;90;95
16 C 10 1;5;10;15;20;25;30;35;45;50;55;60;65;70;75;80;85;90;95;100 2;3;7;8;12;17;23;28;33;36;40;41;42;48;53;54;58;62;63;67;72;73;77;78;83;84;88;92;97;98
17 C 16 5;25;45;65;80 2;10;15;20;30;31;35;50;70;90
18 D 10 20;80;130 10;30;35;40;50;60;70;90;100;110;140;145
19 D 10 1;10;25;35;45;55;89;97;123 7;15;30;39;51;60;70;93;109;120
20 D 10 1;150 20;40;50;55;80;100;120
21 D 10 1;10;15;20;25;30;35;40;50;55;60;70;75;80;85;90;100;105;110;120;125;130;135;140;150 5;6;7;12;17;23;27;34;42;48;53;58;66;67;68;73;78;83;88;93;94;102;107;113;114;123;127;133;137;145
22 D 10 20;40;70;100;120 5;6;10;15;25;30;34;36;45;55;60;80;85;90;95;110;115;125;130;140
23 D 10 1;80;150 10;15;20;30;50;90;100;130
24 D 6 10;40;90;130 5;20;60;100;140
25 E 10 1;20;30;50;70;90;120 10;25;40;41;80;100;130;131;140
26 E 10 1;25;35;45;55;65;75;95;125;135 5;10;15;30;40;43;50;60;70;71;80;81;85;90;100;105;115;130;131;140;141
27 E 10 30;90 10;40;50;70;120
28 E 10 20;50;90;120;130;150 10;25;26;27;70;100;125;140
29 E 10 50 20;40;100
30 E 10 1;30;45;60;70;88;101;123;130;150 10;11;20;40;51;65;78;92;109;110;128;140;141;142;143
31 F 10 1 50;100
32 F 10 1;200 10;20;40;60;100
33 F 10 1;50;100;150 10;20;70;80;120;129;140;160
34 F 10 50;100 10;20;21;22;23;30;31;60;90;130
35 F 10 1;24;44;58;77;98;120;150 9;10;11;30;32;50;60;80;100;130;140;170
36 F 10 1;50;90;140;200 10;20;30;60;70;88;101;120;121;122;130;150;170;189;182
37 F 10 10;20;40;60;80;100;120;140;160;180 5;15;30;31;32;50;55;70;71;90;110;130;131;149;150;170;172;189;190;191
38 F 10 1 100
Appendix A. Tables

39 F 6 1;50;130;200 30;70;90;100;120;150;180
40 G 20 1;13;20;33;40;55;66;77;90;100;120;130;150;160;190;220;240;260;280;290;300;330;340;350;360 7;8;9;16;25;35;50;60;61;70;84;88;95;96;105;110;115;125;135;140;144;155;165;170;180;185;200;210;230;250;270;285;295;310;320;335;345;355;370;380
41 G 10 1;30;70;100;130;150;170;200;230;250;270;290;330;350;370 10;15;20;40;45;55;60;61;62;80;110;120;140;160;180;185;190;198;211;220;225;245;265;280;300;310;320;340;360;380
42 G 10 10;25;40;55;70;85;100;115;140;155;165;175;190;220;250;280;300;330;360;380 4;5;15;30;50;60;65;75;90;105;125;150;160;170;200;210;240;270;290;310;340;370
43 G 10 1;30;60;90;130;160;190;250;300;350 10;15;20;50;80;110;150;170;220;230;240;280;290;330;380
44 G 20 50;100;150;250;330 10;20;70;120;130;170;200;300;310;350
45 G 10 1;50;90;130;160;260;350 10;30;70;100;120;150;200;290;380
46 G 4 1;100;300 50;120;150;200;330
47 A 40 1;5;10;15;20;30;35;40;45;50 X
48 A 30 1 X
49 B 40 1 X
50 C 16 1 X
51 C 24 1 X
52 D 16 1 X
53 D 22 1 X
54 E 24 1 X
55 F 16 1 X
56 F 10 1 X
57 G 16 1 X
58 G 30 1 X
59 G 10 1 X
60 G 4 1 X

72
 Table A.10: A sample input table for multi-material
Station X Y h Offset M1-Cut M1-Fill M1-Type M2-Cut M2-Fill M2-Type M3-Cut M3-Fill M3-Type M4-Cut M4-Fill M4-Type

0 392201.6 6983092 311.88 -10 11.48 0 GR 32.93 0 HP 137.6 0 SR 0 0 OR

0 392201.6 6983092 311.88 -8 10.48 0 GR 29.93 0 HP 97.2 0 SR 0 0 OR
0 392201.6 6983092 311.88 -6 9.48 0 GR 26.93 0 HP 60.8 0 SR 0 0 OR
0 392201.6 6983092 311.88 -4 8.48 0 GR 23.93 0 HP 28.4 0 SR 0 0 OR
0 392201.6 6983092 311.88 -2 7.48 0 GR 20.93 0 HP 0 0 SR 0 0 OR
0 392201.6 6983092 311.88 0 0 0 GR 0 0 HP 0 0 SR 0 0 OR
0 392201.6 6983092 311.88 2 0 -32.4 GR 0 0 HP 0 0 SR 0 0 OR
0 392201.6 6983092 311.88 4 0 -76.8 GR 0 0 HP 0 0 SR 0 0 OR
0 392201.6 6983092 311.88 6 0 -133.2 GR 0 0 HP 0 0 SR 0 0 OR
0 392201.6 6983092 311.88 8 0 -201.6 GR 0 0 HP 0 0 SR 0 0 OR
0 392201.6 6983092 311.88 10 0 -282 GR 0 0 HP 0 0 SR 0 0 OR
20 392203.3 6983112 313.24 -10 11.48 0 GR 32.93 0 HP 137.6 0 SR 0 0 OR
20 392203.3 6983112 313.24 -8 10.48 0 GR 29.93 0 HP 97.2 0 SR 0 0 OR
20 392203.3 6983112 313.24 -6 9.48 0 GR 26.93 0 HP 60.8 0 SR 0 0 OR
20 392203.3 6983112 313.24 -4 8.48 0 GR 23.93 0 HP 28.4 0 SR 0 0 OR
20 392203.3 6983112 313.24 -2 7.48 0 GR 20.93 0 HP 0 0 SR 0 0 OR
20 392203.3 6983112 313.24 0 0 0 GR 0 0 HP 0 0 SR 0 0 OR
20 392203.3 6983112 313.24 2 0 -32.4 GR 0 0 HP 0 0 SR 0 0 OR
Appendix A. Tables

20 392203.3 6983112 313.24 4 0 -76.8 GR 0 0 HP 0 0 SR 0 0 OR

20 392203.3 6983112 313.24 6 0 -133.2 GR 0 0 HP 0 0 SR 0 0 OR
20 392203.3 6983112 313.24 8 0 -201.6 GR 0 0 HP 0 0 SR 0 0 OR
20 392203.3 6983112 313.24 10 0 -282 GR 0 0 HP 0 0 SR 0 0 OR
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . .

73