Stefan Rocktäschel

A Branch-and-
Bound Algorithm
for Multiobjective
Mixed-integer Convex
Optimization
BestMasters
Mit „BestMasters“ zeichnet Springer die besten Masterarbeiten aus, die an renom­
mierten Hochschulen in Deutschland, Österreich und der Schweiz entstanden sind.
Die mit Höchstnote ausgezeichneten Arbeiten wurden durch Gutachter zur Veröf­
fentlichung empfohlen und behandeln aktuelle Themen aus unterschiedlichen
Fachgebieten der Naturwissenschaften, Psychologie, Technik und Wirtschaftswis­
senschaften. Die Reihe wendet sich an Praktiker und Wissenschaftler gleicherma­
ßen und soll insbesondere auch Nachwuchswissenschaftlern Orientierung geben.

Springer awards “BestMasters” to the best master’s theses which have been com­
pleted at renowned Universities in Germany, Austria, and Switzerland. The ­studies
received highest marks and were recommended for publication by supervisors.
They address current issues from various fields of research in natural sciences,
psychology, technology, and economics. The series addresses practitioners as well
as scientists and, in particular, offers guidance for early stage researchers.

More information about this series at http://www.springer.com/series/13198

Stefan Rocktäschel

A Branch-and-Bound
Algorithm for
Multiobjective
Mixed-integer Convex
Optimization
Stefan Rocktäschel
Ilmenau, Germany

ISSN 2625-3577 ISSN 2625-3615 (electronic)

BestMasters
ISBN 978-3-658-29148-8 ISBN 978-3-658-29149-5 (eBook)
https://doi.org/10.1007/978-3-658-29149-5

© Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2020

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with
regard to jurisdictional claims in published maps and institutional affiliations.

This Springer Spektrum imprint is published by the registered company Springer Fachmedien
Wiesbaden GmbH part of Springer Nature.
The registered company address is: Abraham-Lincoln-Str. 46, 65189 Wiesbaden, Germany
This book is dedicated to my family and friends.
Thank you for being there!
Contents

1 Introduction 1

2 Theoretical Basics 5
2.1 Basics of multiobjective optimization . . . . . . . . . . 5
2.2 The central multiobjective mixed-integer optimization
problem . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 A relaxation of (MOMICP) . . . . . . . . . . . . . . . 15

3 A basic Branch-and-Bound algorithm for (MOMICP) 17

3.1 The selection rule . . . . . . . . . . . . . . . . . . . . . 19
3.2 The bisection step . . . . . . . . . . . . . . . . . . . . 21
3.3 A necessary feasibility condition . . . . . . . . . . . . . 24
3.4 Determining lower bounds . . . . . . . . . . . . . . . . 27
3.5 Determining upper bounds . . . . . . . . . . . . . . . . 30
3.6 The discarding test and termination rule . . . . . . . . 34

4 Enhancing Algorithm 1 41
4.1 Preinitialization . . . . . . . . . . . . . . . . . . . . . . 41
4.2 Elimination step . . . . . . . . . . . . . . . . . . . . . 45
4.3 Decrease box width . . . . . . . . . . . . . . . . . . . . 46
4.4 Enhanced algorithm and theoretical results . . . . . . 47

5 Test instances and numerical results 51

VIII Contents

6 Outlook and further possible improvements 61

7 Conclusion 63

Bibliography 65

Plots referring to the numerical tests 67

1 Introduction

Mixed-integer optimization problems (MIP) appear in a variety of

applications like in economics or engineering. One example is the
uncapacitated facility location problem studied by Günlük, Lee, Weis-
mantel [9], where integer variables are used to model the decision
for a facility, whether it should be built or not. Additionally, there
are continuous variables which state the percentage of the respective
customers’ demands which is met by any given facility. The objective
hereby is to decide which facilities to build in order to minimize costs.
Mixed-integer optimization problems have been studied for example
in [13] and [2]. There are already some solvers for these optimization
problems [10], [1] .
Another class of optimization problems that are of interest in many
applications are multiobjective optimization problems (MOP). Hereby,
multiple objective functions have to be minimized simultaneously. In
general, there is no point, i.e. choice of variables that minimizes all
objective functions at the same time. As a result, there is another
concept of optimality used for this class of optimization problems
than we know from scalar optimization. Multiobjective optimization
problems have been studied in [11] and [6] for example and there are
already solvers for these problems [7], [16].
In this book, we use techniques from both of the above classes of
optimization problems and study multiobjective mixed-integer convex

© Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2020

S. Rocktäschel, A Branch-and-Bound Algorithm for Multiobjective Mixed-integer
Convex Optimization, BestMasters, https://doi.org/10.1007/978-3-658-29149-5_1
2 1 Introduction

optimization problems (MOMICP). This is a very important class of

optimization problems since it generalizes the concepts of both of the
above classes. Considering (MOMICP), we are able to add additional
objective functions, which we want to minimize, for example to the
uncapacitated facility location problem. One could derive an addi-
tional objective function for this case by introducing a parameter for
each facility that measures the negative impact on the environment
that occurs, if the respective facility is built. Our additional objective
could be to minimize the total negative impact on the environment
with our building plan for the facilities.
For solving (MOMICP) we are interested in finding the set of efficient
points. These are points for which there exists no other feasible point
that is better or equally good in all objectives. We are also interested
in finding nondominated points, which are the images of the efficient
points. This is one main difficulty of (MOMICP) in comparison to
(MIP). For (MIP) it is possible, at least theoretically, to completely
enumerate all combinations of integer variables and solve a continuous
optimization problem for them. Afterwards, one can compare all
obtained minima and the smallest solves the (MIP). Despite this is
not a ’good’ approach for solving (MIP), it is even worse for (MO-
MICP). The reason is that we would have to enumerate and solve
multiobjective optimization problems that are much harder to solve,
because we get a whole set of nondominated points for each of these
optimization problems instead of a unique minimal function value and
then we would have to compare all these nondominated sets. So we
note that this naive approach seems not practicable at all. There are
algorithms for solving (MOMICP) as can be seen in [5]. However, the
approach introduced in that paper is only heuristic. Thus, our aim is
to develop an algorithm that obtains efficient and nondominated points
 3

of the optimization problem (MOMICP). Hereby, we use concepts and

techniques that are known from mixed-integer optimization and from
global multiobjective optimization.
2 Theoretical Basics

In this chapter, we introduce the basic concepts of multiobjective

optimization. We introduce basic definitions and derive a concept of
optimality for multiobjective optimization problems. Based on this, we
formulate the central optimization problem that we study throughout
this book and introduce a relaxed optimization problem that we use
in order to solve the central optimization problem. Throughout this
book N denotes the set of natural numbers, Z the set of integers, R
the set of real numbers and R` the set of nonnegative real numbers.

2.1 Basics of multiobjective optimization

In this section, we set the basis for comparing vectors in Rp . Therefore,
we use the pointed convex cone K “ Rp` , which induces a partial order
relation ď on Rp defined by

z 1 ď z 2 :ô z 2 ´ z 1 P Rp`

for elements z 1 , z 2 P Rp , also called the componentwise ordering in Rp .

Obviously, it holds

z 1 ď z 2 ô zi1 ď zi2 for all i P t1, ..., pu.

© Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2020

S. Rocktäschel, A Branch-and-Bound Algorithm for Multiobjective Mixed-integer
Convex Optimization, BestMasters, https://doi.org/10.1007/978-3-658-29149-5_2
6 2 Theoretical Basics

Note that the partial order ď on Rp is not a total order relation for
p ě 2. Additionally, we use the following notations:

• z 1 ň z 2 :ô z 2 ´ z 1 P Rp` zt0p u and

• z 1 ă z 2 :ô z 2 ´ z 1 P intpRp` q,

where intpRp` q is the interior of the set Rp` .

The following definition allows us to consider projections of subsets
Z Ď Rp .

Definition 2.1. Let the set Z Ď Rp be nonempty and j P t1, ..., pu.
We define the j-projection of Z as

Zj :“ tzj P R | Dz̃ P Z : z̃j “ zj u.

Remark 2.2. We recall that for nonempty and compact set Z Ď Rp

all projections Z1 , ..., Zp of Z are nonempty and compact.

The following definition generalizes the concepts of minimality and

maximality from scalar optimization to multiobjective optimization
and introduces the ideal- and the anti-ideal point of a set Z Ď Rp .

Definition 2.3. Let Z Ď Rp be a given set. An element z ˚ P Z is

called

• a minimal element of Z, if Z X ptz ˚ u ´ Rp` q “ tz ˚ u and

• a maximal element of Z, if Z X ptz ˚ u ` Rp` q “ tz ˚ u.

Additionally,

• the set of minimal elements of Z is given by

minpZq :“ tz ˚ P Z | Z X ptz ˚ u ´ Rp` q “ tz ˚ uu

2.1 Basics of multiobjective optimization 7

and

• the set of maximal elements of Z is given by

maxpZq :“ tz ˚ P Z | Z X ptz ˚ u ` Rp` q “ tz ˚ uu.

Furthermore, if Z is nonempty and compact, we define

• the ideal point minpZq of Z as the element z ˚ P Rp with

zj˚ “ minpZj q “ mintzj P R | z P Zu for all j P t1, ..., pu and

• the anti-ideal point maxpZq of Z as the element z ˚ P Rp with

zj˚ “ maxpZj q “ maxtzj P R | z P Zu for all j P t1, ..., pu.

Thereby, minpZj q and maxpZj q for j P t1, ..., pu refer to the scalar
minimum and maximum of the set Zj .

Remark 2.4. The above definitions of minpZq and maxpZq are well
defined, because the respective minima and maxima exist due to the
compactness and non emptiness of the set Z and hence, of the projec-
tions.
Furthermore, it is easy to see that tz P Z | z ď z ˚ u “ tz ˚ u holds for
all z ˚ P minpZq. Analogously, tz P Z | z ˚ ď zu “ tz ˚ u holds for all
z ˚ P maxpZq.
Additionally,

minp´Zq “ tz P Rp | p´Zq X ptzu ´ Rp` q “ tzuu

“ t´z P Rp | p´Zq X pt´zu ´ Rp` q “ t´zuu
“ t´z P Rp | Z X ptzu ` Rp` q “ tzuu
“ ´tz P Rp | Z X ptzu ` Rp` q “ tzuu
“ ´maxpZq.
8 2 Theoretical Basics

This equation can be used to reformulate maximizing problems as

minimizing problems and vice versa.

The following example illustrates the above definitions.

Example 2.5. We consider the set Z “ tz P R2 | ||z||2 ď 1u.

The minimum of Z is given by minpZq “ tz P R2 | ||z||2 “ 1, z1 ď
0, z2 ď 0u and the maximum of Z is given by maxpZq “ tz P R2 |
||z||2 “ 1, z1 ě 0, z2 ě 0u. Furthermore, it holds minpZq “ p´1, ´1qJ
and maxpZq “ p1, 1qJ . The sets and points are illustrated in Figure
2.1.
z2

1 maxpZq

maxpZq

Z
0 1 z1

minpZq

minpZq

Figure 2.1: the sets and points from Example 2.5

2.2 The central multiobjective mixed-integer optimization problem 9

2.2 The central multiobjective mixed-integer

optimization problem
In this section, we introduce basic notations and definitions that allow
us to formulate the central optimization problem of this book. At first,
we recall the definition of intervals and generalize this concept for the
multidimensional case.
At first, we introduce intervals and boxes [15].

Definition 2.6.

(i) For b, b P R with b ď b we define the nonempty, compact interval

B by
B “ rb, bs :“ tb P R | b ď b ď bu.

(ii) We denote the set of all of these intervals by IR :“ trb, bs Ď R |

b, b P R, b ď bu.

Now, we introduce the definition of boxes that are a generalization

of intervals for the multidimensional case.

Definition 2.7.

(i) For b, b P Rr with b ď b we define the nonempty, compact,

r-dimensional Box B by
r
ą
B “ rb, bs :“ tb P Rr | b ď b ď bu “ rbi , bi s.
i“1

(ii) We denote the set of all of these r-dimensional boxes in Rr by

IRr .
10 2 Theoretical Basics

(iii) Let B, B̃ P IRr be boxes with B̃ Ď B. Then we call B̃ a subbox

of B.

We introduce notations for certain parameters of boxes.

Definition 2.8. For B “ rb, bs P IRr we define

(i) the infimum of B as infpBq :“ b,

(ii) the supremum of B as suppBq :“ b,

b`b
(iii) the midpoint of B as midpBq :“ 2 and

(iv) the width of B as widpBq :“ max pbi ´ bi q.

iPt1,...,ru

Furthermore, we call elements b P B with bi P tbi , bi u for all i P t1, ..., ru

vertices of B.

Obviously, boxes are convex sets.

Moreover, we will use the following notations. Let B Ď Rr be a set
and f : B Ñ R a function that is twice continuously differentiable on
an open superset B of B. We define f pB̃q :“ tf pbq | b P B̃u for B̃ Ď B.
Furthermore, for an element b P B we denote the gradient of f in b
as ∇f pbq and the Hessian of f in b as Hf pbq. In addition, λminf pbq
denotes the smallest eigenvalue of Hf pbq. Since f is twice continuously
differentiable λminf pbq is well defined, because Hf pbq is symmetric for
all b P B and hence, every eigenvalue of Hf pbq is real.
The following assumptions will be used to define the central optimiza-
tion problem of this book and to prove associated theoretical results.

Assumption 2.9. Let X “ rx, xs P IRm , Y “ ry, ys P IRn be boxes for

m, n P N and B :“ X ˆ Y P IRr with r :“ m ` n. Additionally, we
assume y, y P Zn .
2.2 The central multiobjective mixed-integer optimization problem11

Furthermore, let the functions f : B Ñ Rp and g : B Ñ Rq for p, q P N,

p ě 2 be convex and twice continuously differentiable on an open
superset B of B.

Remark 2.10. For our purposes, the assumption y, y P Zn can be

made without loss of generality, since we could simply round these vec-
tors up or down respectively and would obtain equivalent optimization
problems.
We will also make an additional assumption on regularity of certain
optimization problems in Assumption 3.9 later on.

Using the notation of Assumption 2.9, we introduce notations for

subsets of B that will be used later on.

Definition 2.11. Let Assumption 2.9 be fulfilled and let B̃ “ X̃ ˆ Ỹ Ď

B with X̃ P IRm and Ỹ P IRn be a given subbox of B. We define

B̃ g :“ tpx, yq P X̃ ˆ Ỹ | gpx, yq ď 0q u and

B̃ Z :“ tpx, yq P X̃ ˆ Ỹ | y P Zn u, as well as

B̃ g,Z :“ B̃ g X B̃ Z .

Now, we are able to formulate the central optimization problem

that we are going to study in this book. Under Assumption 2.9 we
are going to study the multiobjective mixed-integer convex problem

min f px, yq
s.t. gpx, yq ď 0q
(MOMICP)
px, yq P X ˆ Y “ B
y P Zn .
12 2 Theoretical Basics

Using the notations from above, we can also write (MOMICP) as

min f pbq
s.t. b P B g,Z .

Remark 2.12. Although (MOMICP) is a generalization of multi-

objective convex optimization problems, this approach cannot be used
in order to solve the interesting class of optimization problems defined
by
min y J f˜pxq
s.t. y J g̃pxq ď 0q
pPy q
xPX
y P t0, 1un

with X P IRm , f˜: X Ñ Rnˆp and g̃ : X Ñ Rnˆq , where f˜i,j and g̃i,k
are convex and twice continuously differentiable on an open superset
X of X for all i P t1, ..., nu, j P t1, ..., pu, k P t1, ..., qu.
This is due to the fact that we demand convexity of f and g in As-
sumption 2.9. With Y :“ r0, 1sn , B :“ X ˆ Y , f : B Ñ Rp defined
by f px, yq :“ y J f˜pxq and g : B Ñ Rq defined by gpx, yq :“ y J g̃pxq we
can rewrite pPy q in the form of (MOMICP). However, the functions f
and g are not necessarily convex in this case, even not if f˜ and g̃ are
linear. We illustrate this in Example 2.13.

Example 2.13. Consider X “ r´1, 1s, n “ 2, q “ 1 and g̃ : X Ñ R2

with g̃pxq :“ px, ´xqJ in pPy q. Note that g̃ is linear and hence convex.
Then we define g : X ˆ r0, 1s2 Ñ R by gpx, yq :“ y J g̃pxq.
For px, yq “ p´1, p0, 0qJ q P X ˆ r0, 1s2 , px1 , y 1 q “ p1, p0, 1qJ q P X ˆ
2.2 The central multiobjective mixed-integer optimization problem13

1
r0, 1s2 and λ “ 2 P r0, 1s we obtain

1 1 1
gpλpx, yq ` p1 ´ λqpx1 , y 1 qq “ gpp´ , p0, 0qJ q ` p , p0, qJ qq
2 2 2
1 J
“ gp0, p0, q q
2
1
“ p0, q ¨ g̃p0q
2
1
“ p0, q ¨ p0, 0qJ “ 0,
2
but

λgpx, yq ` p1 ´ λqgpx1 , y 1 q “ λp0, 0q ¨ g̃p´1q ` p1 ´ λqp0, 1q ¨ g̃p1q

1 1
“ p0, 0q ¨ p´1, 1qJ ` p0, 1q ¨ p1, ´1qJ
2 2
1 1
“ 0 ´ “ ´ ă 0.
2 2
Therefore, g is not convex.

Note that we could also find an example where the objective func-
tions f of pPy q are nonconvex in an analog way.
At least, there are possibilities that allow us to neglect the assumption
of convexity of f for (MOMICP). We will outline one approach for
this in Section 6.
Now, in order to derive a concept of optimality for the central multiob-
jective optimization problem (MOMICP), we introduce the concept
efficiency for multiobjective optimization problems

min f pbq
(MOP)
s.t. b P B,

where B Ď Rr is a given set with r P N and f : B Ñ Rp is a given

14 2 Theoretical Basics

function with p P N. In the following definitions we introduce the

concepts of efficiency and nondominated points [11], [8].

Definition 2.14.

(i) A point b˚ P B is called efficient for (MOP), if there is no b P B

with
f pbq ň f pb˚ q.

(ii) Let b, b˚ P B with

f pbq ň f pb˚ q.

Then we say b dominates b˚ .

(iii) The set of efficient points for (MOP) is called the efficient set
of (MOP).

(iv) A set L Ď PpRr q is called a cover of the efficient set E of

Ť Ť
(MOP), if E Ď L :“ B̃PL B̃ holds.

Remark 2.15. Referring to Definition 2.3, we observe that a feasible

point b˚ P B is efficient for (MOP) if, and only if, f pb˚ q is a minimal
element of f pBq or in other words f pb˚ q P minpf pBqq.

The above definitions consider points in the pre-image space of f .

Additionally, we introduce definitions for points in the image space of
f.

Definition 2.16.

(i) A point z ˚ “ f pb˚ q is called nondominated for (MOP), if b˚ P B

is efficient for (MOP).

(ii) The set of all nondominated points of (MOP) is called the

nondominated set of (MOP).
2.3 A relaxation of (MOMICP) 15

(iii) For elements z 1 , z 2 P Rp with z 1 ň z 2 we say z 1 dominates z 2 .

(iv) A set Z Ď Rp is called stable, if there are no z 1 , z 2 P Z with

z1 ň z2.

Remark 2.17. Referring to Definition 2.3, we observe that a point

z ˚ P Rp is nondominated if, and only if, z ˚ P minpf pBqq. Hence, the
nondominated set of (MOP) is equal to minpf pBqq.

2.3 A relaxation of (MOMICP)

In this section, deriving from the optimization problem (MOMICP),
we introduce a relaxed optimization problem pROPpB̃qq for a given
subbox B̃ of B. We use pROPpB̃qq to determine lower bounds for f on
B̃ g,Z . These lower bounds will be sets L Ď Rp of a ’simple’ structure
that fulfill f pB̃ g,Z q Ď L ` Rp` .
Under Assumption 2.9, the relaxed optimization problem we are
interested in is given by

min f pbq
pROPpB̃qq
s.t. b P B̃ g .

Obviously, the set of feasible points for (MOMICP) B g,Z is a subset of

the set of feasible points for pROPpBqq, which is B g . Because of this,
every feasible point for (MOMICP) is feasible for pROPpBqq and we
call the optimization problem pROPpBqq a relaxation of (MOMICP).
Since there are no integer constraints tied to the relaxation pROPpBqq,
it is a multi-objective (continuous) convex optimization problem. This
type of optimization problems has already been studied extensively.
3 A basic Branch-and-Bound
algorithm for (MOMICP)

In this chapter, we introduce a basic algorithm for computing a ’good’

cover of the efficient set of (MOMICP). The algorithm illustrates the
basic procedure that we use. The idea of this Branch-and-Bound
algorithm is to iteratively split the initial box B into smaller subboxes
and derive lower and upper bounds for respective subproblems. Using
these bounds, we can check certain criteria that allow us to discard a
subbox, if it cannot contain any efficient point for (MOMICP). We also
check for a termination criteria assuring that the boxes, which fulfill
it, are of interest for (MOMICP) (e.g. contain efficient points). If
neither of those criteria are fulfilled, the respective box will be further
investigated and split again later on. After checking any box regarding
these criteria, the next box to check is selected via a selection rule.
The selection rule, the bisection step, the discarding test and the
termination criteria, as well as the determination of lower and upper
bounds for the respective subproblems and their comparison are used
as black boxes for now. In the forthcoming chapters, we investigate
these black boxes and introduce possible options to replace them. We
will use the following notations.

• The list LW denotes the working list and contains boxes that still
have to be examined.

© Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2020

S. Rocktäschel, A Branch-and-Bound Algorithm for Multiobjective Mixed-integer
Convex Optimization, BestMasters, https://doi.org/10.1007/978-3-658-29149-5_3
18 3 A basic Branch-and-Bound algorithm for (MOMICP)

• The list LS 1 denotes the solution list and contains boxes that fulfill
the termination criteria.

• The list LN S 1 denotes the non-solution list and contains boxes that
can be discarded according to the discarding test or a necessary
feasibility condition.

• The list LPN S denotes the set of potentially nondominated points.

• The list LLUB denotes the set of local upper bounds w.r.t. LPN S .

How we manage LPN S and LLU B is not explicitly stated in Algorithm

1. We will go more into detail on how to manage these lists in Sections
3.4 and 3.5.
3.1 The selection rule 19

Algorithm 1 Basic Branch-and-Bound algorithm for (MOMICP)

INPUT: (MOMICP)
OUTPUT: LS 1 , LN S 1 , LPN S , LLU B
1: LS 1 Ð H, LN S 1 Ð H, LW Ð tBu
2: while LW ‰ H do
3: Select a box B̃ of LW via the selection rule.
4: Delete B̃ from LW .
5: Bisect B̃ into subboxes B̃ 1 and B̃ 2 by applying the bisection
step.
6: for j “ 1, 2 do
7: if B̃ j does not fulfill the necessary feasibility condition then
8: Discard B̃ j by adding it to LN S 1 .
9: else
10: Determine lower bounds of f on pB̃ j qg,Z and upper bounds
of f on B g,Z .
11: if B̃ j can be discarded via the discarding test then
12: Discard B̃ j by adding it to LN S 1 .
13: else
14: if B̃ j fulfills the termination criteria then
15: Add B̃ j to LS 1 .
16: else
17: Add B̃ j to LW .
18: end if
19: end if
20: end if
21: end for
22: end while

3.1 The selection rule

Referring to Algorithm 1, the selection rule determines, which box of
LW is going to be examined in the current iteration. Since we try to
find minimal elements of f pB g,Z q, it would make sense to choose a
20 3 A basic Branch-and-Bound algorithm for (MOMICP)

subbox B̃ of B that provides ’small’ function values in f pB̃ g,Z q. It is

not manageable to compare these image sets among all boxes on LW
pairwise in order to achieve this. Instead the ideal point minpf pB̃ g qq
for boxes B̃ P LW can easily be determined and is a relatively good
indicator for the function values of f on B̃ g and hence, can be used as
indicator for the function values of f on B̃ g,Z .
Using the definition of the ideal point (Definition 2.3), it is obvious
that we can obtain minpf pB̃ g qq for a box B̃ P LW by solving the scalar
convex optimization problems

min fi pbq
pROPi pB̃qq
s.t. b P B̃ g

with minimum value ϕi for all i P t1, ..., pu. The ideal point minpf pB̃ g qq
is then given by
minpf pB̃ g qq “ pϕ1 , ..., ϕp qJ .

Our selection rule is to select a box B̃ P LW with the lexicographic

smallest ideal point minpf pB̃ g qq. If there is more than one box with
the lexicographic smallest ideal point, we select the box, which we
obtained by bisection (see Section 3.2) at first. Although, there are
other tiebreaker rules one could apply here.
However, this selection rule is only a heuristic and does not guarantee
to select a ’good’ box. On the other hand, it is fairly easy to determine
these ideal points and since the lexicographical order on Rp is total,
there is exactly one ideal point that can be chosen.
Other approaches that could make sense would be to choose any ideal
point that is not dominated by others or calculate a weighted sum
of the components of the respective ideal points and choose one that
3.2 The bisection step 21

minimizes this sum. However, these approaches are only heuristics

as well and might not have a unique solution, what is the case for
our approach. As a consequence, these ideas would need additional
tiebreaker rules.
We can also observe that the lexicographical smallest ideal point is
always nondominated by others.

3.2 The bisection step

In this section, referring to Algorithm 1 we briefly discuss, how the
bisection step is executed. We therefore introduce the following defini-
tions.

Definition 3.1. Let y P R be a real number. We define

tyu :“ maxtc P Z | c ď yu,

rys :“ mintc P Z | y ď cu and
$
&rys if y ` 0.5 ě rys
rys :“ .
%tyu else

If y “ py1 , ..., yn qJ P Rn is a vector, we define

tyu :“ pty1 u, ..., tyn uqJ ,

rys :“ pry1 s, ..., ryn sqJ and
rys :“ pry1 s, ..., ryn sqJ .

Now, let Assumption 2.9 be fulfilled. We describe the bisection step

for a given subbox B̃ “ rb1 , b1 s “ X̃ ˆ Ỹ “ rx1 , x1 s ˆ ry 1 , y 1 s Ď B.
22 3 A basic Branch-and-Bound algorithm for (MOMICP)

• Determine the largest Index j P argmaxtpb1 i ´ b1 i q | i P t1, ..., ruu,

which is given by j “ maxtargmaxtpb1 i ´ b1 i q | i P t1, ..., ruuu.

• If j P tm ` 1, ..., ru, set

b 1 j ` b1 j 1
B̃ 1 :“ rb1 , pb1 1 , ..., b1 j´1 , t u, b j`1 , ..., b1 r qs and
2

b1 j ` b1 j 1
B̃ 2 :“ rpb1 1 , ..., b1 j´1 , r s, b j`1 , ..., b1 r q, b1 s,
2
if b1 j ` b1 j is odd and

1 1 b1 j ` b1 j 1
B̃ :“ rb , pb1 1 , ..., b1 j´1 , , b j`1 , ..., b1 r qJ s and
2

2 1 b1 j ` b1 j
1
B̃ :“ rpb 1 , ..., b j´1 , ` 1, b1 j`1 , ..., b1 r qJ , b1 s,
2
if b1 j ` b1 j is even.

• Else, set

b1 j ` b1 j 1
B̃ 1 :“ rb1 , pb1 1 , ..., b1 j´1 , , b j`1 , ..., b1 r qJ s and
2

b1 j ` b 1 j 1
B̃ 2 :“ rpb1 1 , ..., b1 j´1 , , b j`1 , ..., b1 r qJ , b1 s.
2
Note that in the bisection step, we choose the biggest index j P
argmaxtpb1 i ´ b1 i q | i P t1, ..., ruu for bisecting the box. In this way,
we split the box along a longest side and additionally at an integer
variable, if possible. In this case, the points in B̃zpB̃ 1 Y B̃ 2 q will not
be considered further.
Moreover, we assumed y i , y i P Z for all i P t1, ..., nu in Assumption 2.9.
3.2 The bisection step 23

Furthermore, for all subboxes B̃ obtained by using this bisection step it

obviously holds y 1 i , y 1 i P Z for all i P t1, ..., nu. Therefore, we are able
to check whether b1 j ` b1 j is odd or even in the case j P tm ` 1, ..., ru,
since b1 j , b1 j P Z holds then.
Also note that in the case j P tm ` 1, ..., ru it generally does not hold
B̃ 1 Y B̃ 2 “ B̃. This is intended, since obviously no b P B̃zpB̃ 1 Y B̃ 2 q
fulfills b P B̃ Z . Therefore, we do not ’loose’ any feasible points for
(MOMICP) due to our bisection method. The underlying idea is that
this might save some computational time and leads to the following
lemma and corollary.

Lemma 3.2. Let Assumption 2.9 be fulfilled and let the subbox B̃ P IRr
of B be derived from bisection steps. Then, it holds

B̃ g,Z Ď B̃ 1 Y B̃ 2 .

Corollary 3.3. Let Assumption 2.9 be fulfilled and let the subbox
B̃ P IRr of B be derived from bisection steps. Then the set of efficient
points for (MOMICP) in B̃ is a subset of B̃ 1 Y B̃ 2 .

The following example illustrates the bisection step.

Example 3.4. We consider the boxes X “ r0, 4s, Y “ r0, 8s and

B “ X ˆ Y . When bisecting B we obtain j “ maxtargmaxtpp4, 8qJ
i ´
p0, 0qJ 1 2
i q | i P t1, 2uuu “ 2 and hence, B “ r0, 4s ˆ r0, 4s and B “
r0, 4sˆr5, 8s. If we then bisect B 2 , we obtain j “ maxtargmaxtpp4, 8qJ
i ´
p0, 5qJ 2 1
i q | i P t1, 2uuu “ 1 and hence, pB q “ r0, 2s ˆ r5, 8s and
pB 2 q2 “ r2, 4s ˆ r5, 8s. When applying another bisection step to pB 2 q1
we obtain j “ maxtargmaxtpp2, 8qJ J
i ´ p0, 5qi q | i P t1, 2uuu “ 2 and
hence, ppB 2 q1 q1 “ r0, 2s ˆ r5, 6s and ppB 2 q1 q2 “ r0, 2s ˆ r7, 8s. The
resulting boxes that we could bisect from there on are illustrated in the
24 3 A basic Branch-and-Bound algorithm for (MOMICP)

following figure. The dotted area illustrates the points that we discard
due to the bisection step.

y
8
ppB 2 q1 q2
7
pB 2 q2
6
ppB 2 q1 q1
5

2 B1

0 1 2 3 4 x

Figure 3.1: some bisection steps

3.3 A necessary feasibility condition

Referring to Algorithm 1, we introduce a necessary condition that
a given subbox B̃ of B contains a feasible point for (MOMICP).
Furthermore, we show that this condition is also sufficient in a special
3.3 A necessary feasibility condition 25

case. The idea is, to solve the optimization problem

min γ
s.t. gpx, yq ď γ ¨ eq
pFOPpB̃qq
px, yq P B̃
γ P Γ,

where eq :“ p1, ..., 1qJ P Rq and Γ is the interval Γ :“ rγ, γs with

γ :“ min min gi px, yq

iPt1,...,qu px,yqPB

and
γ :“ max max gi px, yq.
iPt1,...,qu px,yqPB

Since B is a nonempty and compact set and gi is continuous for

all i P t1, ..., qu, the upper minima and maxima are well defined. We
observe that the inequalities

γ“ min min gi px, yq

iPt1,...,qu px,yqPB

ď min gj px, yq
px,yqPB

ď gj px1 , y 1 q (3.1)
ď max gj px, yq
px,yqPB

ď max max gi px, yq “ γ.

iPt1,...,qu px,yqPB

hold for all px1 , y 1 q P B and all j P t1, ..., qu. Hence, for every
px1 , y 1 q P B exists a γ 1 P Γ with gpx1 , y 1 q ď γ 1 ¨ eq . Therefore, for
every subbox B̃ of B and every px1 , y 1 q P B̃ there is a γ 1 P Γ so
26 3 A basic Branch-and-Bound algorithm for (MOMICP)

that px1 , y 1 , γ 1 q is feasible for pFOPpB̃qq. Hence, pFOPpB̃qq is an op-

timization problem with linear objective function, convex and twice
differentiable constraint functions on a nonempty and compact feasible
set. Therefore the minimum value γ ˚ exists and can easily be com-
puted. We observe that γ “ infpΓq can easily be obtained by solving
q scalar valued convex optimization problems. However, γ “ suppΓq
cannot be obtained easily, but since we are minimizing over γ P Γ
for pFOPpB̃qq and because of (3.1), it is equivalent to minimize over
γ ě γ instead of γ P Γ. This eliminates the necessity to compute γ in
order to solve pFOPpB̃qq.
The optimization problem pFOPpB̃qq is an often used tool in optim-
ization with convex constraints for finding feasible points. In the
following theorem we use the minimum value γ ˚ of pFOPpB̃qq in order
to formulate a necessary and sufficient condition for a subbox B̃ of B
to contain no feasible point for pROPpB̃qq or (MOMICP).

Theorem 3.5. Let Assumption 2.9 be fulfilled and let B̃ “ X̃ ˆ Ỹ Ď B

be a subbox and γ ˚ the minimum value of pFOPpB̃qq.

(i) It holds
γ ˚ ą 0 ô B̃ g “ H ñ B̃ g,Z “ H.

(ii) If Ỹ “ tỹu for some ỹ P Y X Zn (i.e. the integer variables in B̃

are fixed), it holds B̃ g “ B̃ g,Z and therefore

γ ˚ ą 0 ô B̃ g “ H ô B̃ g,Z “ H.

Proof. For the proof of piq, we observe that obviously B̃ g “ H ñ

B̃ g,Z “ H holds, because B̃ g,Z Ď B̃ g . Now, consider a minimal solution
px˚ , y ˚ , γ ˚ q P X̃ ˆ Ỹ ˆ Γ of pFOPpB̃qq.
3.4 Determining lower bounds 27

For the first case, suppose γ ˚ ą 0. Then for all px, yq P B̃ there is
a j P t1, ..., qu with gj px, yq ě γ ˚ ą 0. Else, a contradiction to the
minimality of γ ˚ would occur. Therefore, gpx, yq ď 0q holds for no
px, yq P B̃ and hence, B̃ g “ H.
If, on the other hand, γ ˚ ď 0 holds, then it is px˚ , y ˚ q P B̃ with
gpx˚ , y ˚ q ď γ ˚ ¨ eq ď 0 ¨ eq “ 0q . Hence, px˚ , y ˚ q P B̃ g holds and
B̃ g ‰ H.
For the proof of piiq, assume that Ỹ “ tỹu for some ỹ P Y X Zn . Then
it is easy to see that B̃ g,Z “ B̃ g holds. Using piq we are done.

In order to formulate a necessary feasibility condition we use the

following corollary.

Corollary 3.6. Let Assumption 2.9 be fulfilled and consider a subbox

B̃ of B. Furthermore let γ ˚ be the minimal value of pFOPpB̃qq. Then
it holds
B̃ g,Z ‰ H ñ γ ˚ ď 0.

Therefore, in line 7 of Algorithm 1, we check whether B̃ j does not

fulfill the necessary feasibility condition γ ˚ ď 0.

3.4 Determining lower bounds

In this section, considering (MOMICP) we introduce a method that
allows us to determine lower bounds of f on the set B̃ g,Z for a given
subbox B̃ of B. In detail, we determine a set L Ď Rp of ’simple’
structure that fulfills f pB̃ g,Z q Ď L ` Rp` . In order to achieve this, we
determine lower bounds of f on B̃ g . Using B̃ g,Z Ď B̃ g the Lemma 3.7
follows immediately.
28 3 A basic Branch-and-Bound algorithm for (MOMICP)

Lemma 3.7. Consider (MOMICP) and a given subbox B̃ of B. Then

it holds f pB̃ g,Z q Ď f pB̃ g q.

Now, we can observe that it is sufficient to determine a set L Ď Rp

with f pB̃ g q ` Rp` Ď L ` Rp` , because this then implies f pB̃ g,Z q Ď
f pB̃ g q ` Rp` Ď L ` Rp` . In order to determine such lower bounds, we
use the idea of Benson’s Algorithm [3], which is connected with the
concept of supporting hyperplanes.

Definition 3.8. Let Z Ď Rp be a nonempty set, ẑ P BZ and λ P

Rp zt0p u, where BZ is the boundary of the set Z. The Hyperplane

H λ,ẑ :“ th P Rp | λJ h “ λJ ẑu

is called supporting Hyperplane (of Z), if λJ z ě λJ ẑ holds for all

z P Z.

We can obtain supporting hyperplanes for f pB̃ g q by solving the

optimization problem

min t
s.t. f pbq ď z ` t ¨ ep
pHPOPz pB̃qq
b P B̃ g
tPR

for elements z P Rp .

Assumption 3.9. In addition to Assumption 2.9, we assume regular-

ity for the optimization problem pHPOPz pB̃qq for any z P Rp for which
we consider pHPOPz pB̃qq. This means any solution of pHPOPz pB̃qq
also induces a KKT point of pHPOPz pB̃qq.
3.4 Determining lower bounds 29

The following lemma allows us to construct supporting hyperplanes

of f pB̃ g q for a subbox B̃ of B.

Lemma 3.10. Consider (MOMICP), a subbox B̃ of B and an element

ẑ P Rp and let Assumption 3.9 for ẑ be fulfilled. If B̃ g is nonempty,
then the following statements hold.

(i) The optimization problem pHPOPẑ pB̃qq has a solution pb̂, t̂q P
B̃ g ˆ R with Lagrange multiplier λ̂ P Rp for the constraint
f pbq ď ẑ ` t ¨ ep .

(ii) The hyperplane H λ̂,p̂pẑq with p̂pẑq :“ ẑ ` t̂ ¨ e is a supporting

hyperplane of f pB̃ g q and it holds f pB̃ g q ` Rp` Ď H λ̂,p̂pẑq ` Rp` .

In line 10 of Algorithm 1 we solve pHPOPẑ pB̃qq for B̃ “ B̃ j and

certain elements ẑ P Z of a given nonempty and finite set Z Ď Rp . We
consider the set Z as given for now. In Section 3.5 we will go in detail
regarding the choice for the set Z.
We obtain supporting hyperplanes H λ̂,p̂pẑq with f pB̃ g q`Rp` Ď H λ̂,p̂pẑq `
Rp` for all ẑ P Z. Then we define
č
L :“ Bp pH λ̂,p̂pẑq ` Rp` qq,
ẑPZ

λ̂,p̂pẑq ` Rp` q.
Ş
i.e. as the boundary of the polyhedron ẑPZ pH
According to Lemma 3.10, it holds
č
f pB̃ g,Z q Ď f pB̃ g q ` Rp` Ď pH λ̂,p̂pẑq ` Rp` q “ L ` Rp` .
ẑPZ
30 3 A basic Branch-and-Bound algorithm for (MOMICP)

3.5 Determining upper bounds

In the previous section, considering (MOMICP) we investigated how
to derive lower bounds of f on B̃ g,Z for a given subbox B̃ of B. In this
section, we introduce a method to obtain upper bounds of f on B g,Z .
Later, we use these upper bounds and compare them with the lower
bounds of f on B̃ g,Z . We introduce the concept of local upper bounds
for this. This comparison will set the basis for us to check, whether B̃
can contain a efficient point for (MOMICP). This information will be
used for our discarding test.
In order to derive upper bounds of f on B g,Z , we try to find feasible
points b P B̃ g,Z and evaluate their respective function values. However,
finding such points is a difficult task and in general, there are no
methods known to achieve this within an adequate amount of time.
There are however heuristic approaches like the feasibility pump [4].
Additionally, there is the concept of granularity, which would need
additional assumptions, however.
For the sake of simplicity, we use a heuristic approach that simply
rounds points b “ px, yq P B̃ g in order to match the constraint y P Zn .
Afterwards, we check whether the obtained point still fulfills the con-
straints given by g.
Throughout proceeding Algorithm 1 we obtain points b P B̃ g . For
example, recall the necessary feasibility condition. After solving
pFOPpB̃qq for a subbox B̃ of B we obtain a minimal solution px˚ , y ˚ , γ ˚ q.
If B̃ fulfills the necessary feasibility condition γ ˚ ď 0, it holds
b˚ :“ px˚ , y ˚ q P B̃ g (Theorem 3.5).
Additionally, after solving pHPOPz pB̃qq we obtain a minimal solution
pb˚ , t˚ q P B̃ g ˆ R.
We are then able to apply the rounding step to such points b˚ and try
3.5 Determining upper bounds 31

to obtain a point b P B̃ g,Z in that way.

At first, we introduce the following definition.

Definition 3.11. Let Assumption 2.9 be fulfilled and consider a subbox

B̃ of B. For an element b “ px, yq P B̃ we define

bZ :“ px, rysq “ px, ry1 s, ..., ryn sq.

Lemma 3.12. Let Assumption 2.9 be fulfilled and consider a subbox

B̃ “ rx1 , x1 s ˆ ry 1 , y 1 s Ď B obtained by bisection steps starting with B.
Let b be an element b “ px, yq P B̃. It holds

bZ P B̃ Z .

Proof. Since we obtained B̃ by bisection steps, it holds y 1 i , y 1 i P Z for

all i P t1, ..., nu. Furthermore, it holds x1 ď x ď x1 and y 1 ď y ď y 1 ,
because px, yq P B̃. This implies

y 1 “ ry 1 s ď rys ď ry 1 s “ y 1 .

Therefore, bZ “ px, rysq P B̃ holds and since rys P Zn , it is bZ P B̃ Z .

Summarizing, for any element b “ px, yq P B̃ g Ď B̃ it holds bZ P B̃ Z .

However, it not necessarily holds gpbZ q ď 0q . Therefore, we cannot
assure bZ P B̃ g,Z , but it is easy to verify whether this holds.
Moreover, we are interested in finding many points b P B̃ g,Z with
function values that are nondominated by other known function values
of feasible points. Therefore, we introduce the set LPN S . At the
beginning of Algorithm 1, we initialize LPN S “ H. Then, we handle
LPN S as follows.
When considering a subbox B̃ of B we obtain a point b P B̃ Ď B after
32 3 A basic Branch-and-Bound algorithm for (MOMICP)

solving any of the optimization problems pROPpB̃qi q for i P t1, ..., pu,
pFOPpB̃qq and pHPOPz pB̃qq for some z P Rp and follow these steps.

Algorithm 2 Updating LPN S

INPUT: b P B, LPN S
OUTPUT: LPN S
1: Calculate bZ .
2: if bZ P B g,Z then
3: Calculate w :“ f pbZ q.
4: if Ew1 P LPN S : w1 ď w then
5: Set LPN S “ LPN S Y twu.
6: Set LPN S “ LPN S ztw1 P LPN S | w ň w1 u.
7: end if
8: end if

We observe that LPN S is a finite and stable set that contains

function values of points in B g,Z . We use these points in order to
derive so called local upper bounds which we introduce in the following
definition.

Definition 3.13. [12] Let Assumption 2.9 be fulfilled and let N Ď Rp

be a finite and stable set of points and Ẑ P IRp be a box with f pBq Ď
intpẐq. We define the search region

S :“ tz P intpẐq | w ę z for all w P N u.

Furthermore, a list L Ď Ẑ is called a local upper bound set with respect

to N , if

(i) @z P S Dw P L : z ă w,

(ii) @z P intpẐqzS @w P L : z ć w and

(iii) @w1 , w2 P L : w1 ę w2 or w1 “ w2 .
3.5 Determining upper bounds 33

Remark 3.14. Here, we consider the set Ẑ P IRp with f pBq Ď intpẐq
as given. One could determine such a box by using interval arithmetic
[15].

We want to make use of a local upper bound set with respect to

LPN S . Therefore, we introduce the set LLU B . We initialize LLU B “
tsuppẐqu at the beginning of Algorithm 1. Then, we have to update
LLU B everytime after Algorithm 2 has been executed and the respective
bZ was added to LPN S . Therefore, we execute Algorithm 3 from [12].
This algorithm assures that LLU B is a local upper bound set w.r.t. the
finite and stable set LPN S .
The following figure illustrates the local upper bound set LLU B with
respect to the set LPN S for the case p “ 2. Hereby, it is M 1 :“ suppẐq1
and M 2 :“ suppẐq2 .

f2

M2 LP N S
LLU B

f1
M1

Figure 3.2: LPN S and LLU B cf. [12], Figure 1

We will use the concept of local upper bounds to introduce the

discarding test in the next section.
34 3 A basic Branch-and-Bound algorithm for (MOMICP)

3.6 The discarding test and termination rule

In this section, under Assumption 2.9 we use the method for determin-
ing lower bounds of f on B̃ g,Z for a given subbox B̃ of B from Section
3.4 as well as the concept of local upper bounds from Section 3.5 in
order to introduce a discarding test. The following lemma is taken
from [16] and will be used for proving the fundamental theorem for
formulating our discarding test, which is inspired by the discarding
test proposed in [16].

Lemma 3.15. Let Assumption 2.9 be fulfilled and let L be a local upper
bound set with respect to a finite and stable set of points N Ď f pB g,Z q.
For every z P N and for every j P t1, ..., pu there is a z P L with
z j “ zj and z i ă zi for all i P t1, ..., puztju.

The following theorem will set the basis for the discarding test.

Theorem 3.16. Let Assumption 2.9 be fulfilled and consider a subbox

B̃ of B. Let LPN S Ď f pB g,Z q be a finite and stable set. Moreover, let
LLUB be the local upper bound set w.r.t. LPN S as described in Section
3.5. If
@z P LLU B : z R f pB̃ g q ` Rp` (3.2)

holds, then B̃ does not contain any efficient point for (MOMICP).

Proof. Assume that there is some efficient point x˚ P B̃ for (MOMICP).

This implies x˚ P B̃ g,Z . Since f pB̃ g,Z q ` Rp` Ď f pB̃ g q ` Rp` by Lemma
3.7, it holds z R f pB̃ g,Z q ` Rp` for all z P LLU B . This is implies

f px˚ q ę z for all z P LLU B . (3.3)

Because LLUB is a local upper bound set w.r.t. LPN S we conclude

with Definition 3.13 (i) f px˚ q R S. This implies that there exists a
3.6 The discarding test and termination rule 35

point q P LPN S with q ď f px˚ q. Since q P LPN S , it is q “ f px1 q

for some x1 P B g,Z . Because x˚ is efficient for (MOMICP), it follows
q “ f px1 q “ f px˚ q. Lemma 3.15 implies that there is a point z 1 P LLU B
with f px˚ q ď z 1 , which is a contradiction to (3.3). Thus, B̃ does not
contain any efficient point for (MOMICP).

Condition (3.2) can be tested by solving pHPOPz pB̃qq for the re-
spective local upper bounds z P LLU B and the box B̃. However,
in order to save some computational time, we do not simply solve
pHPOPz pB̃qq for every z P LLU B . Instead, we use Lemma 3.10 to
determine supporting hyperplanes H λ̂,p̂ of f pB̃ g q ` Rp` after we solved
pHPOPz pB̃qq for any z P LLU B . Before solving pHPOPz 1 pB̃qq for the
next local upper bound z 1 P LLU B , we check whether λ̂J z 1 ě λ̂J p̂ holds
for all determined hyperplanes. This is necessary for z 1 P f pB̃ g q ` Rp`
and only then we have to solve pHPOPz 1 pB̃qq.
We summarize some theoretical results on pHPOPz pB̃qq and how we
use them. Then, we formulate the discarding test in detail. There-
fore, we assume that we have a given accuracy parameter ε ą 0, a
local upper bound z P LLU B and the solution pb˚ , t˚ q P B̃ g ˆ R of
pHPOPz pB̃qq.

(1) If t˚ ď 0 holds, then it obviously is z P f pB̃ g q ` Rp` . Hence, we

cannot apply Theorem 3.16 and B̃ may contain efficient points
for (MOMICP). Thus, B̃ cannot be discarded and we distinguish
between the following two subcases.
(a) If t˚ ă ´ 2ε holds, then we stop the discarding test and bisect
B̃ later.
(b) If ´ 2ε ď t˚ ď 0 holds, then we construct a supporting hy-
perplane using Lemma 3.10 in order to improve the outer
36 3 A basic Branch-and-Bound algorithm for (MOMICP)

approximation of f pB̃ g q ` Rp` and consider the next local

upper bound.

(2) If t˚ ą 0 holds, then it obviously is z R f pB̃ g q ` Rp` . We then

construct a supporting hyperplane using Lemma 3.10 in order to
improve the outer approximation of f pB̃ g q ` Rp` and consider the
next local upper bound.

Thereby we apply the strategy from [16], which is motivated and

explained in more detail also in [16].
The following algorithm describes the discarding test in detail. There-
fore, we introduce a flag D for discarding a box and a flag B for bisecting
a box. In addition, we introduce a flag F that states, whether the lists
LPN S and LLUB are fixed, i.e. no further points will be added to them.
For the basic algorithm, we set F “ false for now. Furthermore, let
ui be the i-th unit vector in Rp for i P t1, ..., pu.
3.6 The discarding test and termination rule 37

Algorithm 3 Discarding test

INPUT: (MOMICP), a subbox B̃ of B, LPN S , LLUB , ε ą 0, F
OUTPUT: D, B, LPN S , LLU B
1: Set LPN S ˚ Ð LPN S , LLU B ˚ Ð LLU B .
2: Solve pROPpB̃qi q with solution bi P B̃ g for all i P t1, ..., pu.
3: if F then
4: Apply Algorithm 2 to update LPN S with bi for i P t1, ..., pu.
5: Update LLU B using Algorithm 3 from [12].
6: end if
7: Set a Ð pf1 pb1 q, ..., fp pbp qqJ .
i
8: Set H Ð tH u ,a | i P t1, ..., puu.
9: Set D Ð true, B Ð false.
10: for z P LLUB ˚ do
1
11: if @H λ,z P H : λJ z ě λJ z 1 then
12: Solve pHPOPz pB̃qq with solution pb˚ , t˚ q P B̃ ˆ R and Lag-
range multiplier λ̂ P Rp for the constraint f pbq ď z ` t ¨ ep .
13: if F then
14: Apply Algorithm 2 to update LPN S with b˚ .
15: Update LLU B using Algorithm 3 from [12].
16: end if
17: if t˚ ă ´ 2ε then
18: Set D Ð false, B Ð true.
19: break for ´ loop
20: else
˚ p
21: Set H Ð H Y tH λ̂,z`t ¨e u.
22: if t˚ ď 0 then
23: Set D Ð false.
24: end if
25: end if
26: end if
27: end for
38 3 A basic Branch-and-Bound algorithm for (MOMICP)

Referring to Algorithm 1, we discard a box B̃, if Algorithm 3 applied

to it returns D “ true.
Our termination rule for the basic Branch-and-Bound algorithm for a
given subbox B̃ is that Algorithm 3 returns D “ false and B “ false.
In this case, there is a local upper bound z P LLU B with z P f pB̃ g q`Rp`
and no local upper bound suggests a further bisection of B̃. This
suggests that f pB̃ g,Z q has reached a certain ’accuracy’ in f pB g,Z q
without getting discarded. Hence, this box is of interest for us, because
it might contain efficient points for (MOMICP). Later, we enhance
Algorithm 1 by further investigating the boxes on LS . For example,
we then assure a certain ’accuracy’ in the pre-image space as well by
bisecting these boxes until their box width is less or equal a given
accuracy parameter δ.
However, we want to prove the exactness of Algorithm 3 at first.
Therefore, we show that the discarding test returns D “ false for
every box that contains an efficient point for (MOMICP). Hence, we
do not discard such boxes throughout proceeding Algorithm 1.
Theorem 3.17. Let B̃ be a subbox of B that contains an efficient
point x˚ P B̃ g,Z of (MOMICP). Applying Algorithm 3 then returns
D “ false.
Proof. Assume that Algorithm 3 returns D “ true. This implies that
for all local upper bounds z P LLU B it holds t˚ ą 0, where pb˚ , t˚ q
is the solution of pHPOPz pB̃qq. Hence, z R f pB̃ g q ` Rp` holds for
all z P LLUB . Theorem 3.16 then implies that B̃ does not contain
any efficient point for (MOMICP). This is a contradiction to the
assumption that B̃ contains the efficient point x˚ .

We use this theorem to prove the exactness of Algorithm 1. This

means after applying Algorithm 1 the set of efficient points E of
3.6 The discarding test and termination rule 39

(MOMICP) is a subset of the union of all boxes of LS 1 , or in other

words LS 1 is a cover of E.

Theorem 3.18. Let E be the efficient set of (MOMICP). After ap-

plying Algorithm 1 the set LS 1 is a cover of E.

LS 1 Y LN S 1 Y BS,
Ť Ť
Proof. At first we observe that it holds B “
where BS is the set of all points that are discarded due to the bisection
step. Then it obviously holds E Ď LS 1 Y LN S 1 Y BS. Corollary
Ť Ť

3.3 implies E Ę BS and Theorem 3.17 and Corollary 3.6 imply E Ę

LN S 1 . Thus, E Ď LS 1 holds.
Ť Ť
4 Enhancing Algorithm 1

In this chapter, we introduce modifications that enhance the ba-

sic Branch-and-Bound algorithm for (MOMICP), we introduced in
Chapter 3. We follow different goals with these modifications. We
would like to reduce the amount of computational time, the algorithm
requires, as well as provide a ’better’ cover of the efficient set of (MO-
MICP). In order to create an understanding of the precision of a cover
of the efficient set of (MOMICP), we introduce the following definition.

Definition 4.1. Let L1 , L2 Ď PpRr q be covers of the efficient set of

(MOMICP). The cover L1 is called more precise than L2 , if L1 Ď
Ť
Ť 2
L holds.

4.1 Preinitialization
The first idea is to compute some imagepoints of f on B g,Z before
starting the basic algorithm. Using these points, we then initialize
LPN S and LLUB .
Therefore, let imgp P N be a given number. We want to distribute
evenly many points on every possible combination of integer variables
and then create an equidistant grid on the remaining continuous
variables in a way that we obtain about imgp points in B. We then
check for each of these points b P B, if b P B g,Z holds and in this case

© Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2020

S. Rocktäschel, A Branch-and-Bound Algorithm for Multiobjective Mixed-integer
Convex Optimization, BestMasters, https://doi.org/10.1007/978-3-658-29149-5_4
42 4 Enhancing Algorithm 1

update LPN S as well as LLU B .

The following algorithm describes this procedure in detail.

Algorithm 4 Preinitialization
INPUT: (MOMICP), imgp P N
OUTPUT: LPN S , LLU B
1: if imgp “ 0 _ m “ 0 then
2: Set LPN S Ð H, LLU B Ð tsuppẐqu.
3: else
xi ´xi 1
Set grd Ð p m
ř śn
4: 1 q ¨ p i“1 py i ´ y i ` 1qq .
m
i“1
m¨imgp m
5: Set Pos Ð 0 ¨ er and FV Ð H.
6: while x`grd¨pPos1 , ..., Posm qJ ď x^y `pPosm`1 , ..., Posr qJ ď
y do
7: Set b Ð px ` grd ¨ pPos1 , ..., Posm qJ , y ` pPosm`1 , ..., Posr qJ q.
8: if b P B g,Z then
9: Set FV Ð FV Y tbu.
10: end if
11: Set Pos1 Ð Pos1 ` 1.
12: for i “ 1, ..., r ´ 1 do
13: if pi ď m ^ xi ` grd ¨ Posi ą xi q _ pi ą m ^ y i´m ` Posi ą
y i´m q then
14: Set Posi Ð 0.
15: Set Posi`1 Ð Posi`1 ` 1.
16: end if
17: end for
18: end while
19: Set LPN S Ð H, LLU B Ð tsuppẐqu.
20: for b P FV do
21: Update LPN S using Algorithm 2 with input b.
22: Update LLU B using Algorithm 3 from [12].
23: end for
24: end if

Remark 4.2. Note that for imgp ą 0 the preinitialization can only
4.1 Preinitialization 43

be applied, if m ą 0 holds. Otherwise, the grid parameter grd is not

defined. However, since we only have to consider integer variables
for m “ 0, there is only a finite number of feasible points anyways.
Thus, a preinitialization might not be that useful regarding decreasing
computational time. Hence, Algorithm 4 returns LPN S “ H and
LLU B “ H in the case m “ 0 (lines 1,2).
Also note that we consider all y P Y X Zn for calculating points b P B
during the preinitialization due to the method of using the position
vector Pos. This implies that we consider all points px, yq for y P Y XZn
and hence at least |Y X Zn | points for imgp ą 0 and m ą 0 for the
preinitialization. This might be much more than the intended number
imgp of points that we want to consider. One should take this fact
into account when using the preinitialization.

In line 4, we calculate the grid parameter grd that is used for creating
an equidistant grid on the continuous variables. The idea behind this
formula is to consider roughly imgp points during the preinitialization.
If there were no integer variables and m “ 1, we would simply use
x1 ´x1
imgp as grid parameter, what is also done in Algorithm 4 in that
case. For m ą 1 and n “ 0 we take the mean value of interval lengths
řm xi ´xi 1
i“1 m into consideration and divide this by imgp m in order to
1
obtain roughly imgp m grid points on each Xi for i P t1, ..., mu and
1
hence pimgp m qm “ imgp grid points on X. If we have to consider
additional integer variables in the case n ą 0, we want to distribute
evenly many grid points on every possible y P Y X Zn . There are
śn n
i“1 py i ´ y i ` 1q possible y P Y X Z . Hence, we have to divide
the number of grid points on X by that number in order to still
obtain roughly imgp grid points in total on X ˆ Y . We achieve this
by multiplying the grid parameter from the case n “ 0, which is
44 4 Enhancing Algorithm 1

řm xi ´xi śn 1
, with p i“1 py i ´ y i ` 1qq . By doing that, the amount
1 m
i“1
m¨imgp m
1
imgp m
of grid points we consider on each Xi is now roughly śn 1 .
p m
i“1 py i ´y i `1qq
1
imgp m m “ ś imgp
Hence, we consider about p 1 q n grid
i“1 py i ´y i `1q
śn
p i“1 py i ´y i `1qq m
śn imgp ¨ ni“1 py i ´ y i ` 1q “ imgp
ś
points on X and thus py ´y `1q
grid
i“1 i i
points on X ˆ Y . This leads to the formula for the grid parameter
grd that we use in line 4 of Algorithm 4.
In line 5, we initialize the position vector Pos and the set of obtained
function values FV Ď f pB g,Z q. In line 6, we start going through the
box B using the position vector Pos until we obtain a point that is
not in B anymore. In line 7, we calculate the current point that has
to be considered. In the first iteration, this will be b “ px, yq and from
there on we use the position vector Pos to go through B. In lines
8-10, we check whether the current point b is in B g,Z and add it to
FV in that case. In lines 11-17, we update the position vector Pos to
determine the next point. Finally, in lines 19-23, we use the points
b P FV Ď B g,Z in order to initialize LPN S and LLU B .

We use the preinitialization in order to obtain some local upper

bounds and function values of f on B g,Z before applying Algorithm 1.
In that way, we might be able to discard certain boxes or verify the
termination rule after less iterations than we would need otherwise.
On the other hand, we might have to solve more optimization problems
in each iteration to apply the discarding test for example, because we
have more local upper bounds to consider for that. However, this is
just a heuristic approach.
Summarizing, we presume that we need less iterations but more
computational time per iteration, if we use the preinitialization before
4.2 Elimination step 45

applying Algorithm 1. Applying the enhanced algorithm to certain

test instances, we examine this relationship more precisely in Chapter
5.

4.2 Elimination step

In this section, we introduce a method that we can apply after Al-
gorithm 1 in order to obtain a more precise approximation of the
efficient set E of (MOMICP) than LS 1 . The idea is to consider all
boxes B̃ P LS 1 and check, if B̃ can be discarded using the set of local
upper bounds LLU B , which is fixed after applying Algorithm 1 and
contains more points then. Hence, the discarding test is more precise
now and might allow us to discard additional boxes.
In the following algorithm we describe this procedure in detail.

Algorithm 5 Elimination step

INPUT: (MOMICP), LS 1 ,LN S 1 , LPN S , LLU B
OUTPUT: LS 2 ,LN S 2
1: Set LS 2 Ð H, LN S 2 Ð LN S 1 .
2: Set F Ð true.
3: for B̃ P LS 1 do
4: Apply Algorithm 3 to B̃.
5: if D then
6: Set LN S 2 Ð LN S 2 Y tB̃u.
7: else
8: Set LS 2 Ð LS 2 Y tB̃u.
9: end if
10: end for
46 4 Enhancing Algorithm 1

4.3 Decrease box width

In this section, we introduce a method that we can apply after Al-
gorithm 5 in order to obtain a more precise approximation of the
efficient set E of (MOMICP) than LS 2 . The idea is to split the boxes
B̃ P LS 2 till it holds widpB̃q ď δ for all boxes B̃ for a given δ P R with
0 ă δ ă 1. Therefore, we bisect the boxes by applying the bisection
step and additionally check whether the respective subboxes can be
discarded. This procedure is similar to Algorithm 1.
In the following algorithm we describe this procedure in detail.
4.4 Enhanced algorithm and theoretical results 47

Algorithm 6 Decrease box width

INPUT: (MOMICP), LS 2 ,LN S 2 , LPN S , LLU B , δ P R with 0 ă δ ă 1
OUTPUT: LS 3 ,LN S 3
1: Set LS 3 Ð H, LN S 3 Ð LN S 2 .
2: Set F Ð true.
3: Set LW Ð LS 2 .
4: while LW ‰ H do
5: Select a box B̃ P LW via the selection rule.
6: Set LW Ð LW ztB̃u.
7: Bisect B̃ by applying the bisection step into B̃ 1 and B̃ 2 .
8: for j “ 1, 2 do
9: Apply Algorithm 3 to B̃ j .
10: if Dj _ B̃ j does not fulfill the necessary feasibility condition
then
11: Set LN S 3 Ð LN S 3 Y tB̃ j u.
12: else
13: if widpB̃ j q ď δ then
14: Set LS 3 Ð LS 3 Y tB̃ j u.
15: else
16: Set LW Ð LW Y tB̃ j u.
17: end if
18: end if
19: end for
20: end while

4.4 Enhanced algorithm and theoretical results

In this section, we introduce an enhanced algorithm that uses the ideas
of Algorithm 4, Algorithm 5, Algorithm 6 in addition to the basic
Branch-and-Bound algorithm for (MOMICP), referred to as Algorithm
1. Moreover, we prove some theoretical results regarding the enhanced
algorithm.
48 4 Enhancing Algorithm 1

Algorithm 7 Enhanced algorithm

INPUT: (MOMICP), imgp P N, ε ą 0, δ P R with 0 ă δ ă 1
OUTPUT: LS i , LN S i for i P t1, 2, 3u, LPN S , LLU B
1: Apply Algorithm 4.
2: Set F Ð false.
3: Apply Algorithm 1.
4: Set F Ð true.
5: Apply Algorithm 5.
6: Apply Algorithm 6.

Remark 4.3. Note that by setting F Ð true in line 4 of Algorithm 7

we do not update LPN S and LLU B during Algorithm 5 and Algorithm 6
anymore. The idea behind this is that we already obtained many points
b P LPN S Ď f pB g,Z q and hence many local upper bounds after applying
Algorithm 4 and Algorithm 1. We might save some computational time
by not adding even more points to LPN S . Moreover, the list LPN S
generated so far already contains enough information to approximate
the set of nondominated points of (MOMICP) sufficiently accurate.
Furthermore, we consider every box B̃ P LS 1 again in Algorithm 5
and Algorithm 6 and apply the discarding test to this box. Therefore,
we have to examine the respective hyperplanes for B̃. However, we
already obtained some hyperplanes for B̃ during the discarding test in
Algorithm 1. In order to save some computational time, we can save
these hyperplanes during Algorithm 1 and recall them later. In that
way we do not have to examine these hyperplanes again and only have
to solve the respective optimization problems for some new points that
were added to LLU B after B̃ was added to LS 1 .
The following theorem proves the exactness of Algorithm 7, i.e. all
LS i for i P t1, 2, 3u are covers of the efficient set of (MOMICP). In
addition, we prove a result regarding the precision of these covers.
4.4 Enhanced algorithm and theoretical results 49

Theorem 4.4. After applying Algorithm 7, the sets LS 1 , LS 2 , LS 3

are covers of the efficient set E of (MOMICP). Furthermore, LS 3 is
more precise than LS 2 and LS 2 is more precise than LS 1 . Moreover,
it holds y 1 “ y 1 for every box B̃ “ rx1 , x1 s ˆ ry 1 , y 1 s P LS 3 , i.e. the
integer variables of B̃ are fixed.

Proof. Considering Algorithm 5, it is obvious that LS 2 Ď LS 1

Ť Ť

holds. Moreover, it is easy to observe that LS 3 Ď LS 2 holds

Ť Ť

by considering Algorithm 6. Due to Theorem 3.18 it is E Ď LS 1 .

Ť

Because of Theorem 3.17 it also holds E Ď LS 2 . Moreover, Corollary

Ť

3.3, Theorem 3.17 and Corollary 3.6 imply E Ď LS 3 . So, the sets
Ť

LS 1 , LS 2 , LS 3 are covers of the efficient set E of (MOMICP) and LS 3

is more precise than LS 2 and LS 2 is more precise than LS 1 .
Now, assume that there is a box B̃ “ rx1 , x1 s ˆ ry 1 , y 1 s P LS 3 with
y 1 ‰ y 1 . This implies that there is a i P t1, ..., nu with y 1 i ă y 1 i .
Due to our bisection method this implies y 1 i ď y 1 i ` 1. This implies
widpB̃q ě y 1 i ´ y 1 i ě 1 ą δ. This is a contradiction, because every box
B 1 P LS 3 fulfills widpB 1 q ď δ. Hence, y 1 “ y 1 holds.
5 Test instances and numerical
results

We have implemented Algorithm 7 in MATLAB for p “ 2. All tests

have been run on an Intel(R) Core (TM) i7-6700K CPU @ 4.00GHz
with 32GB RAM (2x DDR4-2399/16GB) on the operating system
Microsoft Windows 10 Pro version 10.0.17134 in MATLAB R2018a.
In the following we introduce the test instances we designed.

© Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2020

S. Rocktäschel, A Branch-and-Bound Algorithm for Multiobjective Mixed-integer
Convex Optimization, BestMasters, https://doi.org/10.1007/978-3-658-29149-5_5
 52
name X Y gpx, yq f px, yq
«˜ ¸ ˜ ¸ff ˜ ¸
´2 2 x1 ` y
balls1 , r´2, 2s x21 ` x22 ´ 1
´2 2 x2 ´ y
«˜ ¸ ˜ ¸ff ˜ ¸
´2 2 x1 ` y
balls2 , r´2, 2s x21 ` x22 ´ 1
´2 2 x2 ` expp´yq
˜ ¸
x`y
parabolas1 r´2, 2s r´2, 2s 0
x2 ´ y
˜ ¸
x`y
parabolas2 r´2, 2s r´2, 2s px ` 2q2 ` py ` 2q2 ´ 16
x2 ´ y
˜ ¸
x`y
parabolas3 r´2, 2s r´2, 2s px ´ 2q2 ` py ´ 2q2 ´ 16
x2 ´ y
˜ ¸
x2
points1 r´2, 2s r´2, 2s 0
x2 ` py ´ 12 q2
˜ ¸
x
points2 r´2, 2s r0, 1s x2 ` py ´ 12 q2 ´ 14
x2
«˜ ¸ ˜ ¸ff ˜ ¸ ˜ ¸
´2 2 ´x2 ` 34 x1 ´ 12 x1 ´ y
triangles1 , r´2, 2s 2
´2 2 x2 ´ 43 x1 ´ 3 x2 ` y
«˜ ¸ ˜ ¸ff ˜ ¸ ˜ ¸
´2 2 ´x2 ´ 3x1 ´ 4 x1 ´ y
triangles2 , r´2, 2s 4
´2 2 ´x2 ´ 13 x1 ´ 3 x2 ` y
5 Test instances and numerical results
 ¨ 1
˛
«˜ ¸ ˜ ¸ff ´x2 ` 34 x1 ´ ˜ ¸
2
´2 2 ˚ ‹ x1 ´ y
triangles3 , r´2, 2s ˝ x2 ´ 43 x1 ´ 23 ‚
´2 2 x2 ` y
´x2 ´ x1 ´ 3 ´ y
¨ ˛
«˜ ¸ ˜ ¸ff ´x2 ` 34 x1 ´ 12 ˜ ¸
´2 2 ˚ ‹ x1 ´ y
triangles4 , r´2, 2s ˝ x2 ´ 43 x1 ´ 23 ‚
´2 2 x2 ` y
´x2 ´ x1 ´ 3 ´ 2y
¨ ˛
«˜ ¸ ˜ ¸ff ´x2 ` 34 x1 ´ 12 ˜ ¸
´2 2 ˚ ‹ x1 ´ 2y
triangles5 , r´2, 2s ˝ x2 ´ 43 x1 ´ 23 ‚
´2 2 x2 ` 2y
´x ´ x1 ´ 3 ´ 2y
¨ ř2 ˛
´ 8i“1 yi
˚ ‹
˚ x1 ´ x2 ‹
«˜ ¸ ˜ ¸ff ˚ ‹ ˜ ř3 ¸
´1 1 ˚ y ´y ‹
˚ 1 4 ‹ x21 ´ x2 ` yi
highdim , r´2e8 , 2e8 s ˚ ‹ ři“1
6
´1 1 ˚ y5 ´ y2 ‹
˚ ‹ x22 ´ x1 ` i“4 yi
˚ y ´y ‹
˝ 3 6 ‚
y7 ´ y8
Table 5.1: test instances
53
54 5 Test instances and numerical results

Furthermore, we have set ε “ δ “ 0.1 for all instances and we have

used different choices for imgp for which we compare the performance
of the algorithm later in this section. Also note that all instances
fulfill Assumption 2.9. Despite that, certain properties do not hold for
some instances due to the mixed-integer constraints. For example the
feasible sets B g,Z are nonconvex and even not connected. Furthermore,
the image sets f pB g,Z q are not necessarily connected and f pB g,Z q`Rp`
is not necessarily convex. These are properties that are often assumed
to hold in multiobjective optimization.
The idea was to design test instances with efficient sets and nondom-
inated sets that are easy to determine analytically but still illustrate
these ’difficulties’ that arise in multiobjective mixed-integer optimiza-
tion. The underlying idea of the test instance ’highdim’ was to test
the algorithm on an instance with a high dimension of the pre-image
space r “ 10.
Plots that illustrate the points in LPN S after the termination of the
algorithm with imgp “ 0 colored black and image points in f pB g,Z q
colored grey can be found in Appendix 7. We can see that we obtain a
good approximation of the nondominated set using Algorithm 7. Note
that it is not easy to illustrate the obtained covers LS 3 and the efficient
sets E due to the integer constraints and the fact that r ą 2 holds
for many of our test instances. Therefore, we will not compare them
here. In the following we compare how different choices of imgp for
the preinitialization affect the performance of the algorithm. There-
fore, we consider iteration counts and computational time needed.
Table 5.2 lists the iteration counts. Hereby N denotes the number of
points in f pB g,Z q we obtained during the preinitialization, J denotes
the iteration count of Algorithm 1, K denotes the iteration count of
Algorithm 5 and L denotes the iteration count of Algorithm 6.
 55

name imgp N J K L
0 0 319 280 810
balls1 3000 120 301 246 811
5000 195 276 222 842
0 0 257 215 675
balls2 3000 120 239 190 690
5000 195 239 188 691
0 0 85 78 194
parabolas1 3000 3005 8 9 281
5000 5005 8 9 279
0 0 74 65 157
parabolas2 3000 2100 10 8 228
5000 3500 10 8 228
0 0 45 41 150
parabolas3 3000 2100 6 7 193
5000 3500 6 7 191
0 0 1 2 100
points1 3000 3005 1 2 100
5000 5005 1 2 100
0 0 3 3 22
points2 3000 2 3 3 22
5000 2 3 3 22
0 0 12 5 41
triangles1 3000 150 12 5 41
5000 280 12 5 41
0 0 134 75 1351
triangles2 3000 410 121 73 1351
5000 785 123 74 1352
56 5 Test instances and numerical results

0 0 26 12 233
triangles3 3000 138 24 13 218
5000 262 24 14 219
0 0 27 6 148
triangles4 3000 126 26 5 149
5000 243 26 6 150
0 0 43 16 283
triangles5 3000 126 49 19 276
5000 243 46 18 281
Table 5.2: iteration counts

We observe that as expected, we need less iterations J and K for

most test instances when using the preinitialization. However, we then
still have to decrease the box width in Algorithm 6 and therefore need
more iterations for L in most cases. When investigating the values for
’parabolas1’, ’parabolas2’ and ’parabolas3’ we can observe this relation
very good. Summarizing it seems that we tend to have less overall
iterations when using the preinitialization as expected. However, we
might need more computational time for each iteration when using the
preinitialization. Hence, we consider the computational times in the
following table. Hereby, tinit denotes the time needed for Algorithm
4, t1 denotes the time needed for Algorithm 1, t2 denotes the time
needed for Algorithm 5, t3 denotes the time needed for Algorithm 6,
tvisual denotes the time needed for plots and ttotal denotes the total
time needed for executing Algorithm 7.
name imgp N tinit {s t1 {s t2 {s t3 {s tvisual {s ttotal {s
0 0 0.05 13.56 1.81 7.82 0.52 23.77
balls1 3000 120 0.06 11.60 1.63 8.12 0.29 21.71
5000 195 0.04 10.39 1.54 8.53 0.27 20.78
0 0 0.06 9.83 1.32 6.37 0.31 17.89
balls2 3000 120 0.05 9.27 1.30 6.83 0.30 17.75
5000 195 0.05 8.99 1.24 6.71 0.28 17.26
0 0 0.04 3.15 2.38 1.74 0.98 8.29
parabolas1 3000 3005 1.31 21.60 18.31 58.65 1.00 100.88
5000 5005 3.62 57.70 41.43 122.55 1.22 226.52
0 0 0.04 2.97 1.88 1.67 1.06 7.62
parabolas2 3000 2100 0.96 17.21 14.85 49.28 0.91 83.21
5000 3500 2.45 45.55 33.02 100.56 0.92 182.51
0 0 0.03 2.18 1.31 1.80 0.87 6.19
parabolas3 3000 2100 0.54 7.17 6.97 25.64 0.83 41.16
5000 3500 1.35 16.02 14.47 48.64 0.84 81.32
0 0 0.03 0.28 0.29 0.81 0.68 2.10
points1 3000 3005 0.17 0.32 0.25 0.76 0.68 2.18
5000 5005 0.23 0.28 0.26 0.81 0.68 2.26
0 0 0.04 0.36 0.26 0.39 0.54 1.58
points2 3000 2 0.04 0.36 0.36 0.37 0.55 1.68
5000 2 0.03 0.36 0.26 0.37 0.55 1.56
57
 58

0 0 0.05 0.67 0.02 0.48 0.30 1.53

triangles1 3000 150 0.05 0.61 0.02 0.59 0.29 1.55
5000 280 0.03 0.62 0.02 0.64 0.29 1.60
0 0 0.04 6.54 1.40 24.81 0.27 33.06
triangles2 3000 410 0.05 6.09 1.47 24.88 0.28 32.78
5000 785 0.05 6.05 1.37 24.48 0.30 32.25
0 0 0.04 1.56 0.14 4.36 0.28 6.38
triangles3 3000 138 0.05 1.42 0.13 4.17 0.28 6.05
5000 262 0.04 1.41 0.12 4.04 0.34 5.94
0 0 0.04 1.50 0.06 2.75 0.28 4.64
triangles4 3000 126 0.05 1.47 0.04 2.86 0.27 4.69
5000 243 0.04 1.53 0.06 2.85 0.28 4.75
0 0 0.04 2.48 0.17 5.27 0.27 8.24
triangles5 3000 126 0.05 2.79 0.20 5.07 0.29 8.39
5000 243 0.04 2.66 0.20 5.31 0.27 8.48
Table 5.3: computational times
5 Test instances and numerical results
 59

We observe that the preinitialization saves some computational time

for some test instances like ’balls1’ and ’balls2’, but this is neglectable
when considering the total times needed for ’parabolas1’, ’parabolas2’ and
’parabolas3’. We need extensively more computational time to solve these
problems when using the preinitialization despite less iterations needed. This
suggests that a preinitialization is not necessary for Algorithm 7 and even
hindering, because we might need much more computational time to solve
an optimization problem than without the preinitialization.
Note that we did not obtain results for the test instance ’highdim’, because
we stopped the test on this instance after 24 hours. So we can see that the
algorithm solves these low dimensional test instances rather quickly and
without a need of a preinitialization, but it definitely has its limits regarding
higher dimensional pre-image spaces.
6 Outlook and further possible
improvements
In this Chapter, we discuss an extension of the proposed algorithm to the
nonconvex case. Therefore, we introduce the concept of convex underestim-
ators. As we have seen in Example 2.13, the assumption of convexity of f
and g for (MOMICP) in Assumption 2.9 can be very restricting. However,
the concept of convex underestimators will allow us to drop the assump-
tion of convexity of f . Hence, we are then able to consider a larger class
of optimization problems. In the following, we outline how the concept
of convex underestimators allows us to drop the assumption of convexity
of f in Assumption 2.9. At first, we introduce the definition of convex
underestimators.

Definition 6.1. Let B P IRr be a box for r P N and h : B Ñ R a given

function. A function h : B Ñ R is called a convex underestimator of h (on
B), if h is convex and hpbq ď f phq holds for all b P B.

Considering (MOMICP) with an objective function f that is not necessarily

convex, we observe that the optimization problems pHPOPz pB̃qq for z P
LLU B that we solve during Algorithm 7 are not easy to solve anymore. We
would need a global optimization approach in order to solve these problems
due to f not being convex. However, solving such problems can take some
time and since we have to solve many of these optimization problems this
might result in a huge increase in computational time overall.
Instead, our idea is to determine a convex underestimator f i of fi for all
P t1, ..., pu on the subbox B̃ of B we are considering for pHPOPz pB̃qq. Since

© Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2020

S. Rocktäschel, A Branch-and-Bound Algorithm for Multiobjective Mixed-integer
Convex Optimization, BestMasters, https://doi.org/10.1007/978-3-658-29149-5_6
62 6 Outlook and further possible improvements

fi is twice continuously differentiable this can easily be done as shown in

[14]. Then we replace the constraint f pbq ď z ` t ¨ e of pHPOPz pB̃qq with
f pbq ď z ` t ¨ e which makes this problem easier solvable since f is convex.
It can be shown that we still obtain lower bounds for f pB̃ g q in that way [16].
These can then be used as lower bounds for f pB̃ g,Z q.
The drawback of this method is that the lower bounds are less tight. However,
with decreasing box width widpB̃q the bounds get tighter since f is a better
approximation for f then, see [14].
Summarizing, we are able to drop the assumption of convexity of f in
Assumption 2.9, but since we obtain bounds that are not very tight for
nonconvex f , we assume that solving (MOMICP) would need more iterations
and hence more computational time then. However, this would allow us
to solve optimization problems like pPy q for example, if we do not have to
consider any constraints (g̃ ” 0).
Note that the theoretical results of this book also hold for p ą 2. Hence, a
further improvement could be to revisit the implementation of Algorithm 7
in order to generalize it to being able to solve (MOMICP) for all p ě 2.
7 Conclusion
In this book, we have considered multiobjective mixed-integer convex op-
timization problems. We introduced basic definitions and concepts of mul-
tiobjective optimization. We derived a basic Branch-and-Bound algorithm
for solving (MOMICP) by using approaches from global multiobjective
(continuous) optimization [16]. We have described the main steps of this
algorithm and proven exactness. We have enhanced this basic algorithm
by introducing modifications that can save computational time or return a
more precise cover of the efficient set E of (MOMICP). The final algorithm
was implemented in MATLAB. We have tested the algorithm on several
instances which have been designed by us. We have discussed the impact
of the modifications on computational time and precision of the cover of E.
Finally, we have outlined further steps that could be done in order to save
more computational time, obtain a more precise cover of E and generalize
our theoretical results and the final algorithm in order to handle optimization
problems (MOMICP) with not necessary convex objective functions f .

© Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2020

S. Rocktäschel, A Branch-and-Bound Algorithm for Multiobjective Mixed-integer
Convex Optimization, BestMasters, https://doi.org/10.1007/978-3-658-29149-5_7
Bibliography
[1] Abhishek, K. ; Leyffer, S. ; Linderoth, J. : FilMINT: An outer
approximation-based solver for convex mixed-integer nonlinear pro-
grams. In: INFORMS Journal on computing 22 (2010), Nr. 4, S.
555–567
[2] Belotti, P. ; Kirches, C. ; Leyffer, S. ; Linderoth, J. ; Luedtke,
J. ; Mahajan, A. : Mixed-integer nonlinear optimization. In: Acta
Numerica 22 (2013), S. 1–131
[3] Benson, H. P.: An outer approximation algorithm for generating all
efficient extreme points in the outcome set of a multiple objective linear
programming problem. In: Journal of Global Optimization 13 (1998),
Nr. 1, S. 1–24
[4] Bonami, P. ; Cornuéjols, G. ; Lodi, A. ; Margot, F. : A feas-
ibility pump for mixed integer nonlinear programs. In: Mathematical
Programming 119 (2009), Nr. 2, S. 331–352
[5] Cacchiani, V. ; DâĂŹAmbrosio, C. : A branch-and-bound based
heuristic algorithm for convex multi-objective MINLPs. In: European
Journal of Operational Research 260 (2017), Nr. 3, S. 920–933
[6] Ehrgott, M. : Multicriteria Optimization. Springer, 2005
[7] Eichfelder, G. : Adaptive Scalarization Methods in Multiobjective
Optimization. Springer, 2008
[8] Fernández, J. ; Tóth, B. : Obtaining the efficient set of nonlinear
biobjective optimization problems via interval branch-and-bound meth-
ods. In: Computational Optimization and Applications 42 (2009), Nr.
3, S. 393–419

© Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2020

S. Rocktäschel, A Branch-and-Bound Algorithm for Multiobjective Mixed-integer
Convex Optimization, BestMasters, https://doi.org/10.1007/978-3-658-29149-5
66 BIBLIOGRAPHY

[9] Günlük, O. ; Lee, J. ; Weismantel, R. : MINLP strengthening

for separable convex quadratic transportation-cost UFL. In: IBM Res.
Report (2007), S. 1–16

[10] Gurobi Optimization, L. : Gurobi Optimizer Reference Manual.

http://www.gurobi.com. Version: 2018

[11] Jahn, J. : Vector optimization. Springer, 2009

[12] Klamroth, K. ; Lacour, R. ; Vanderpooten, D. : On the represent-

ation of the search region in multi-objective optimization. In: European
Journal of Operational Research 245 (2015), Nr. 3, S. 767–778

[13] Lee, J. ; Leyffer, S. : Mixed integer nonlinear programming. Bd. 154.

Springer Science & Business Media, 2011

[14] Maranas, C. D. ; Floudas, C. A.: Global minimum potential energy

conformations of small molecules. In: Journal of Global Optimization 4
(1994), Nr. 2, S. 135–170

[15] Neumaier, A. : Interval methods for systems of equations. Bd. 37.

Cambridge university press, 1990

[16] Niebling, J. ; Eichfelder, G. : A Branch-and-Bound based Algorithm

for Nonconvex Multiobjective Optimization. In: Preprint-Series of the
Institute for Mathematics (2018)
Plots referring to the numerical
tests

Figure 1: image set and LPN S of Figure 2: image set and LPN S of
balls1 balls2

© Springer Fachmedien Wiesbaden GmbH, part of Springer Nature 2020

S. Rocktäschel, A Branch-and-Bound Algorithm for Multiobjective Mixed-integer
Convex Optimization, BestMasters, https://doi.org/10.1007/978-3-658-29149-5
68 Plots referring to the numerical tests

Figure 3: image set and LPN S of Figure 4: image set and LPN S of
parabolas1 parabolas2

Figure 5: image set and LPN S of Figure 6: image set and LPN S of
parabolas3 points1
 69

Figure 7: image set and LPN S of Figure 8: image set and LPN S of
points2 triangles1

Figure 9: image set and LPN S of Figure 10: image set and LPN S of
triangles2 triangles3
70 Plots referring to the numerical tests

Figure 11: image set and LPN S of Figure 12: image set and LPN S of
triangles4 triangles5