Optimization-based decision support systems

for planning problems in processing industries

G.D.H. Claassen
Thesis committee

Promotor

Prof. Dr J.G.A.J. van der Vorst

Professor of Logistics and Operations Research
Wageningen University

Other members

Prof. A.J.M. Beulens, Wageningen University

Prof. Dr G. Kant, Tilburg University
Prof. Dr J.H. Trienekens, Wageningen University
Prof. Dr I.F.A. Vis, University of Groningen

This research was conducted under the auspices of Wageningen School of Social
Sciences (WASS)
Optimization-based decision support systems for
planning problems in processing industries

G.D.H. Claassen

Thesis
submitted in fulfillment of the requirements for the degree of doctor
at Wageningen University
by the authority of Rector Magnificus
Prof. Dr M.J. Kropff,
in the presence of the
Thesis Committee appointed by the Academic Board
to be defended in public
on Monday 8 December 2014
at 11 a.m. in the Aula.
G.D.H. Claassen
Optimization-based decision support systems for planning problems in processing industries
172 pages.

PhD thesis, Wageningen University, Wageningen, NL (2014)

With references, and summaries in English and Dutch

ISBN 978-94-6257-208-9
Preface
More than a quarter of a century ago I discovered my appreciation for the theory and
application of mathematical models and techniques for decision support in practice.
Barely two years after my MSc-graduation I wrote a letter of application for an assistant
professorship at the department of mathematics. Hardly difficult to remind, after all, I just
wrote one successful letter in my entire life. Let’s see what that single letter spawned
and might bring into play ☺.

At the time of my application for employment I just finished my first contribution in

a collaborative approach to bridge the gap between theory and the scientific challenge
of its applicability in real life. We worked on a planning and scheduling problem. This
problem resulted in, and was deliberately chosen as, the first case study for this thesis in
Chapter 2. The study intends to demonstrate the anticipating value of case-based
research for daily practice. Meanwhile it is clear that the main architecture of currently
available (commercial) Advanced Planning Systems (APS) provide in what has been
proposed in (many) earlier case studies. The downside of that remarkable observation is
that, in spite of the tremendous progress that has been made in all these years, the real
problem in Chapter 2 has neither been solved (generically) in literature, nor in this thesis
(Chapter 3). But we should ask whether this is really bad. At least there is hope for the
applicability and relevance of case-based research outside a given application area (e.g.
the chapters 4 and 5).

A thesis like this calls for a day to finalise something that – according to present-
days standards – should have been finished many years earlier. The question remains:
“Should I regret my unrestrained search for professional satisfaction far beyond the
social significance of a PhD-degree?” Definitely not! Partly unconsciously following
personal motives, never satisfied with what has been reached, horribly abasing my own
results and incessantly pushing the bar to almost unreachable limits; it is simply me! Let
me give my readership some basic sense for the preceding character sketch. Once,
somebody in the academic community stated – I quote – : “You cannot be a good
teacher if you are not a good researcher”. Well, without that elementary formal degree in
research, I must be a horrible teacher. So, think as a good researcher and define the
next “research question” (of course to be answered after the 8th of December 2014):
“How to prove that the opposite holds?” Any future PhD-candidate who launches the
proposition “You cannot be a good researcher if you are not a good teacher” may
provoke an interesting debate.

In all those years of employment, my search for understanding, insight and

overview was, and still is, indispensable. As the years passed, teaching became one of
my prime activities. It provided significantly in my ceaseless quest for professional
satisfaction. From my point of view, students are the best potential ambassadors for any
applied (research) field in practice. So, I constantly ask myself: “How to get large
audiences of students daily in my lecture rooms, inspiring them to study the subjects
such that the added value of our profession will become a second nature for every
generation we deliver for a professional career?”. Meanwhile, I’m convinced that the
related mission of any university cannot be projected on, or captured by commonly
accepted tape measures.

I should emphasize that teachers and/or researchers have had teachers too, in
the broadest sense of the word. Many people contributed to my professional satisfaction
in one way or another. I would like to thank them all. Some of them need to be
mentioned in particular.

In the first place, I express my sincere gratitude to Jack van der Vorst. Not just
as a patient promotor but also as our valued head of the department. It’s amazing how
you manage to guide a complete corridor of different personalities with all their strengths
and weaknesses through rapidly changing environments. Our interpretation and
perception on research and education may not always be the same but you constantly
showed to be a great listener and extremely fast thinker. Always radiating patience,
confidence, and continuous support. Your ability to take the necessary distance and
keep the big picture in mind is indispensable to structure my activities. I always left your
room with new energy (particularly at difficult times) and valuable notes for the next
moves. I have to admit, sometimes I was crashed by the idea that I went in “for a blue
sweater” but in the end, implicitly, bought a “vacuum cleaner”. A great gift for any
manager ☺.

I’m also much obliged to Theo Hendriks, my former colleague and mentor who
patiently taught me, for instance, how to use the most simple and powerful tools in
educational settings. I remember one of my first educational experiences. At that time
we had two groups of students, all following the same course in dynamic programming.
After two weeks Theo visited my room in the mathematics building and asked with that
well-known expression of sympathy around his eyes: “How do you do?” I knew enough.
Apparently, the major part of my initial audience filled the window sills of his classroom.
In spite of that devastating experience, you always showed confidence and contributed
to my everlasting learning curve, i.e. to find a proper balance between mathematical
correctness and its applicability in real life. You simply knew, I would never give up.
Hence, you mined my skills and enthusiasm ceaselessly. Two chapters of this thesis are
based on research you initiated. Over the years I learned to accept that the outcome,
e.g. this thesis, will never be Theo-proof. The disillusionment might start already in the
first paragraph of each chapter ☺.

Joke and Eligius, I owe you my great debt of gratitude for all continuous support,
the contributions and involvement that allowed me to finish this piece of work! Your
completely different personal styles and skills floated the vessel successfully at times
there was something wrong with the engine. Eligius, I suggest we wait for the moment
that our mutual opinion regarding “research questions”, is widely approved ☺. Joke, in
spite of our preference for deterministic behaviour, several of your colleagues (including
myself) experienced the impact of stochastic, “superior forces”. I will make a splendid
picture for our audience and provide for your ultimate relief as a paranymph: only the
two of us know which question of the opposition will be yours to answer during my
defense ☺.

I should mention the continuous and mostly invisible support of all my

colleagues who never showed any hesitation in assisting or taking over parts of my work
and give substance to my responsibilities. Teaching activities, in the broad sense of the
word, seem to expand exponentially at our university. Yes indeed, a problem of luxury.
However, it is usually done at the expense of measured indicators. Dear Joke, Argyris,
Karin, Aleksander, Willem and Mehmet thanks for your dedication and involvement in
my activities.

Last and certainly not least, I want to thank all my current and former LDI-
colleagues, the members of the TIFN project team and all other (former) colleagues
within and outside our university for the close collaboration and their contributions to my
daily professional satisfaction!

The major part of this thesis had to be done outside regular working hours.
Finally the mess on my desk at home can be stored. It will certainly feel as a great relief,
not just for me. Within this context it seems almost impossible to find the right words for
the most important person in my life. In some way I am convinced she does not want to
be mentioned at all. It is reassuring that she does not need my words to communicate,
we simply look, feel and know…

Liefste Kim en Li-An, de aandachtige lezer mag ook iets van jullie leren! Wij allen
begrijpen nu waarom uitgevers gruwelen van het woord “kaft”. Dat wordt een “omslag”
genoemd; wat een mooi woord!

G.D.H. (Frits) Claassen Valburg, November 1, 2014

Table of contents

1 General introduction 1

1.1 Introduction 2
1.2 Concepts and relevance of model-based DSS 4
1.2.1 Concepts and origin of DSS 4
1.2.2 Classical architecture of DSS 5
1.2.3 Professional relevance of DSS 7
1.3 Decision support for industrial practice 8
1.3.1 Origin of advanced planning systems 8
1.3.2 Main architecture of APS systems 10
1.3.3 Vertical integration of production planning and scheduling 13
1.3.4 Horizontal integration of planning tasks 14
1.4 Research motivations, objective and questions 16
1.5 Research method and outline of the thesis 18

2 Planning and scheduling in food processing industry 21

Abstract 22

Part I Planning and Scheduling Packaging lines 23

2.1 Introduction 23
2.2 Problem analysis 24
2.2.1 The tactical planning level 25
2.2.2 Modelling 26
2.3 Results 32
2.4 Conclusions Part I 34

Part II Modelling approaches for planning and scheduling 36

2.5 Introduction 36
2.6 Small bucket approaches 39
2.6.1 DLSP: Discrete Lot-sizing and Scheduling Problem 39
2.6.2 Extensions of the DLSP 39
2.6.3 CSLP: Continuous Setup Lot-sizing Problem 40
 2.6.4 PLSP: Proportional Lot-sizing and Scheduling Problem 40
2.7 Big bucket approaches 40
2.7.1 CLSP: Capacitated Lot-Sizing Problem 40
2.7.2 GCLP: Generalized Capacitated Lot-sizing Problem 41
2.7.3 GLSP: General Lot-sizing and Scheduling Problem 41
2.7.4 Extensions of the CLSP 42
2.8 Product decay 42
2.9 Literature overview 43
2.10 Conclusions Part II 44

3 Vertical integration of lot-sizing and scheduling in food processing industry 47

Abstract 48
3.1 Introduction 49
3.2 Embedding in the literature 50
3.2.1 Small bucket approaches 51
3.2.2 Big bucket approaches 51
3.2.3 Product decay 52
3.3 Model formulation 53
3.3.1 Outline of the lot-sizing and scheduling problem 53
3.3.2 Notation and model formulation 54
3.3.3 Modelling product decay 56
3.4 Numerical illustrations and benchmark 56
3.4.1 Impact of (non-)triangular setup and product decay 57
3.4.2 Complexity discussion 59
3.4.3 Heuristic approach 59
3.5 Numerical analysis 61
3.6 Concluding remarks 62

4 Integrated planning between procurement and production 65

Abstract 66
4.1 Introduction 67
4.2 Problem description 70
4.3 Model formulation 70
 4.4 Solving the model 73
4.5 Discussion and concluding remarks 81
4.5.1 The pilot decision support system 81
4.5.2 Conclusions 83
Appendix 4.1 85

5 Mixed Integer (0-1) Fractional Programming in Paper Production Industry 89

Abstract 90
5.1 Introduction 91
5.2 Background 92
5.3 Problem description and model formulation 94
5.4 Solving mixed integer (0-1) fractional problems 99
5.5 Numerical results 104
5.6 Discussion 108
5.7 Conclusions 110

6 General discussion 113

6.1 Introduction 114

6.2 Reflection on research questions and main findings 115
6.2.1 Research question 1 115
6.2.2 Research question 2 119
6.2.3 Research question 3 120
6.2.4 Research question 4 123
6.3 Discussion of findings 126
6.3.1 Model building and solving 126
6.3.2 Professional relevance and applicability of DSS in practice 133
6.3.3 Advanced Planning Systems 136
6.4 Main conclusions and future research 137

Summary 142

Samenvatting 149

References 157
 Chapter 1

General introduction

A study of decision making for “stochastic, multi-objective fractional programming with conditional
terms, subject to a non-linear, fuzzy set of constraints in integer variables“ (title arbitrarily chosen)
may be of little value, unless the usefulness of such a study is demonstrated (Schaible and
Ibaraki 1983).
Chapter 1

1.1 Introduction
Nowadays manufacturing strategy is an important part of corporate strategy, particularly
in food processing industry (FPI). Due to global competition, the diversity of products
increased considerably in this branch of industry which forced manufacturers to
participate in an on-going trend towards increased variety (i.e. ingredients and flavours,
customised packaging, prints and/or labels) and new products. In this environment,
efficient production planning and scheduling is of vital importance and has become one
of the most challenging problems for decision support in practice. To keep up with global
competition and deal with developments in today’s society, management teams of
enterprises have to take all kinds of interrelated decisions on different levels and
timeslots within the organization. As a consequence, the need for computerized support
has increased substantially.

The tremendous progress in hard- and software of the past decades was an
important gateway for developing computerized systems that are able to support
decision-making on different levels within enterprises. The history of such systems
started in the late 1960s, and in 1971 the concept of Decision Support Systems (DSS)
emerged (Gorry and Morton 1971). Meanwhile, the field of DSS has evolved into a
broad variety of directions. DSS is not a homogenous field and over its history a number
of distinct subfields have emerged (Arnott and Pervan 2008). There are a number of
fundamentally different approaches to DSS and each has had a period of popularity in
both research and practice (Arnott and Pervan 2005). Due to its interdisciplinary context,
a unique framework for categorizing the different types of DSS does not exist. However,
based on the dominant architectural components providing the functionality of decision-
making, Power and Sharda (2007) identified five categories of DSS, i.e. model-driven,
communication-driven, data-driven, document-driven and knowledge-driven DSS.

According to Power and Sharda (2007), model-driven DSS emphasize access to

and manipulation of a quantitative model (e.g. accounting and financial models,
representation models, and/or optimization models). Hence, quantitative models are the
dominant component in the architecture that provides the functionality for the DSS.
Communication-driven DSS derive their functionality from communications and
information technologies that are used in the system to support shared-decision-making
(e.g. computer-based bulletin boards or group decision support systems). The
functionality of data-driven DSS results from access to and manipulation of large
databases of structured data (e.g. management report systems, data warehousing and
analysis systems or business intelligence systems). Document-driven DSS integrate a
variety of computer storage facilities and processing technologies in which a search
engine is a primary tool to provide sophisticated document retrieval and analysis to
support decision-makers. Finally, knowledge-driven DSS suggest or recommend actions
based upon knowledge that has been stored using Artificial Intelligence or statistical
tools like case-based reasoning, rules, frames and Bayesian networks. The knowledge
component, usually based on specialised problem-solving expertise, provides the
primary functionality for knowledge-based DSS.

2
 General Introduction

This thesis will explore DSS developments for current-days practice including its
added value for industrial practice in the future. We focus on the category of model-
driven DSS. In addition, the term model-driven DSS is further refined to modelling and
solving (production) planning problems.

Production planning is considered here as the planning of the acquisition of raw

materials, the planning of production activities required to transform raw materials into
intermediate and final products, and the coordination of production scheduling with
physical distribution of finished products to clients in order to meet customer demand in
the most efficient or economical way possible. In industrial environments, the problems
to be addressed in this field call for (interrelated) decisions with respect to the required
kind of raw materials, the types of production quantities to be manufactured, the lot-
sizes (or batch-sizes) of the different products to be processed and - last but not least -
the time at which the raw materials and production orders must be available.

Due to a high complexity in production structure and layout, process industries

show a distinctive role among the various industries, particularly with respect to
production planning and scheduling (Entrup 2005). Processing industries are
characterized by specific production operations like blending, milling, refining, heating
and/or cooling which in turn change and/or define the final properties of (intermediate)
products (Kallrath 2002; Günther and van Beek 2003). Usually, specific processes can
only be performed efficiently using large installations, which tend to be very expensive
(Fransoo and Rutten 1994). Moreover, margins are often relatively low in capital-
intensive process industries (e.g. pulp and paper production). Process industries often
obtain their raw materials from mining or agricultural industries. These materials have
natural variations in quality which often lead to variations in recipes and prices of
(alternative) ingredients (Fransoo and Rutten 1994).

Processing can take place in batches or by continuous flows and quite often
shared or multi-purpose equipment is used to produce a wide variety of products
(Kallrath 2002; Günther and van Beek 2003). As a consequence, sequence-dependent
changeover costs and/or times are often incurred (Soman, Van Donk et al. 2004a;
Stadtler 2005; Stadtler and Kilger 2008). For instance, for the sake of pureness and
safety regulations exhaustive cleaning operations may be prescribed in food processing
industry (Günther and van Beek 2003; Soman, Van Donk et al. 2004a). Both, between
different types of process industries as within a specific branch, product structures may
be completely different, i.e. converging (e.g. paper production industry) or diverging (e.g.
dairy industry). The complexity of lot-sizing and scheduling in food processing industry
may also be determined by an inevitable decline in quality of products or limited shelf
lives, finite intermediate storage facilities, the use of product specific storage devices,
no-wait production for certain types of products, and complex packaging facilities
(Günther and van Beek 2003; Soman, Van Donk et al. 2004a). Compared to, for
instance, discrete parts manufacturing, the specific characteristics of processing industry
complicate planning problems considerably which give rise to focus in this thesis on
model-driven (i.e. optimization-based) decision support in the domain of processing
industry.

3
Chapter 1

As the field of decision support concerns the process of choosing the most
attractive alternative, the underlying process of decision-making needs to be analysed to
a certain extent in Section 1.2. Section 1.3 will discuss the current state of the art for
industrial practice and provides the basis for a basic perception on the question “Which
decision-making processes in processing industry need to be supported and how?”
Section 1.4 will present the overall research objective including its translation into a
number of research questions to be addressed in the next chapters. An outline of the
thesis is presented in Section 1.5.

1.2 Concepts and relevance of model-based DSS

This section offers a brief overview of the main developments in DSS including its basic
principles. The origin and main concepts of DSS are described (Section 1.2.1) followed
by its general architecture (Section 1.2.2). After more than four decades of research in
DSS, Section 1.2.3 briefly summarizes the current state of affairs with respect to its
professional relevance. The described perception of DSS will be the starting point for
one of the research premises for this thesis.

1.2.1 Concepts and origin of DSS

Decision problems arise in many varieties. Some problems are simple while others are
extremely complex. Problems can be deterministic and/or contain stochastic elements.
Simon (1960) described decision problems as existing on a continuum from
programmed (routine, repetitive, well structured, and easy to solve) to non-programmed
(new, ill-structured, and difficult to solve). A programmable task can be captured in clear
rules, substituting the judgement of the decision-maker. For example, setting up a bill of
materials for material requirement planning is a programmable task. For these kinds of
programmable tasks, the decision-maker can be replaced by a computer program.
However, ambiguity, creativity, and ingenuity may also be involved in decision-making.
In these situations, human decision-makers cannot be simply replaced by computers.
According to Simon’s taxonomy of decision types, the process of decision-making
includes four phases, i.e. intelligence, design, choice and review. The phase of
intelligence comprises the search for problems i.e. which problems need a decision. The
design phase involves the development of alternatives i.e. finding possible courses of
action. The third phase consists of the selection of available courses of action. Finally,
past choices are evaluated in the review phase.

Anthony (1965) introduced one of the most generally accepted categories of

management activities or frameworks for planning problems which consists of three
decision levels i.e. strategic, tactical and operational control level. Anthony defines
strategic planning problems as “the process of deciding on the objectives of the
organization, on changes in these objectives, on the resources used to attain these
objectives, and on the policies that are to govern the acquisition, use and deposition of
these resources”. Strategic decisions are extremely important because they are, to a

4
 General Introduction

great extent, responsible for maintaining the competitive capabilities of a firm. Strategic
decisions determine the rate of growth, and eventually define the success or failure of
an enterprise. An essential characteristic of strategic decisions is that they have long-
lasting effects, thus forcing long planning horizons in their analysis (Hax and Candea
1984). Once strategic decisions have been made, the next problem to be resolved is the
effective allocation of resources on tactical planning level, also called management
control. Anthony defines tactical decisions as “the process by which managers assure
that resources are obtained and used effectively and efficiently in the accomplishment of
the organization’s objective”. Tactical decisions usually involve the consideration of a
medium-range time horizon, divided into several periods, and require significant
aggregation of the relevant managerial information. Typical tactical planning decisions
are purchasing of raw materials, utilization of regular and overtime workforce, allocation
of aggregate capacity resources to product families, maintenance planning, and order
acceptance strategy. After an aggregate allocation of resources, it is necessary to deal
with day-to-day (operational) decisions in a small-range time horizon. Anthony defines
operational decision-making as ”the process of assuring that specific tasks are carried
out effectively and efficiently on a day to day basis”. Typical decisions at this level are
the assignment of customer orders to individual machines, the scheduling of orders or
vehicle routing problems. For an overview of the major elements of Anthony’s
management activities, the interested reader is referred to Hax and Candea (1984).
Anthony’s framework and Simon’s description of decision problems are considered as
the cornerstones of Decision Support Systems (DSS).

The term and concept of DSS was introduced and defined by Gorry and Morton
(1971) who integrated Anthony’s categories of management activities and Simon’s
taxonomy of decision types. The authors defined DSS as “Computer systems that
support decision-making for problems that are at least at some stage semi-structured or
unstructured”. Computer systems could be developed to deal with the structured part of
a problem, but the judgement of a decision-maker is needed on the unstructured part,
hence constituting a human-machine problem-solving system. The concept of DSS
aimed to assist and make ill-structured, non-programmable tasks more tractable. Models
and computers proved to be very valuable for many (programmable) decision problems.
However, they can easily demonstrate their weakness too for decision-making in daily
practice; particularly with respect to model-based DSS. In the late 1980’s it became
clear that the added value and applicability of (mathematical) models and computers in
daily practice, needed a general architecture which will be discussed in the next section.

1.2.2 Classical architecture of DSS

Model-driven or optimization-based decision support is usually associated with the field
of Operations Research (OR). One of the main characteristics of OR (also called
Management Science) is the attempt to quantify aspects of decision problems with
abstract (mathematical) models. A model of a decision problem is always an abstract
description of reality. No model can capture all characteristics of an unstructured
decision problem, and is by default a simplification of reality (Ackoff 1977; Claassen and

5
Chapter 1

Hendriks 2007). In order to handle models and use them for generating solutions,
assumptions are necessary. Moreover, it is often hard or even impossible to quantify
certain aspects of a decision problem. Sometimes, these aspects are either disregarded
(Ackoff 1977) or artificially embedded into models as a compromise to the applied
technique (Claassen and Hendriks 2007). One (recognized) way to cope with the limits
of mathematical models and computers for daily practice is a profound architecture for
model-based DSS.

The most basic and classical architecture of a DSS was given by Sprague Jr
(1980). According to this scheme the software system of model-based DSS comprises
three components, i.e. the Model Base Management System (MBMS) or model-base,
the Data Base Management System (DBMS) or database and the user interface which
Sprague called the Dialog Generation Management System (DGMS). The components
are briefly discussed here.

For many OR scientists, the model base is the actual core of the system. It
contains abstract models and algorithms for generating high-quality plans to be used for
further analyses. Building models requires profound insight in the problem. Therefore,
the focus on the problem itself leads to better insight into the decision situation and part
of the problem may already be solved. The exercise of building models often reveals
relationships that are scarcely apparent to decision-makers. As a result, there is an
increase in insight and understanding of the object being modelled. On the other hand,
the gained insight and understanding of the underlying problem is often a prerequisite to
solve the generated problems in practice. Solving real-life problems by OR models
generally implies that the focus must be towards taking advantage of important problem
characteristics, i.e. to recognize them and to exploit special structures for solving the
generated problems efficiently. In addition, the motivation to solve real-life problems may
also provide new theoretical insight including the basis for new approaches in new
contexts that have an added value of their own, even outside the original application
area. The main principles for designing models and solution techniques are defined by
Little (1970) in his seminal paper “The concept of decision calculus”. The author
presents a set of guidelines along six issues (i.e. simplicity, robustness, ease of control,
adaptability, completeness and ease of communication) to bridge the gap between
mathematical theory and the scientific challenge of its applicability in real-life
enterprises.

The database can be seen as the facts of a decision situation. It contains all data
necessary to create problem instances for models in the model base (e.g. the type,
dimension or quality of raw materials, semi-finished products and end products,
inventory levels and demand figures of the products, machinery specification, work
force, available capacities and lay-out. Databases fulfil a crucial role in a DSS as they
are the key to separate the data from models (Carlsson and Turban 2002). A strict
separation of data and models can be considered as the gateway for the applicability of
model-based DSS in real-life practice.

6
 General Introduction

The software, managing the interface between the user and the system, is called
the user interface. The user-friendliness of this component is of extreme importance for
the acceptance of the DSS. The user interface carries all communication between the
end-user and the system in practice. Even if a DSS offers a wide range of functional
routines and delivers incredibly good solutions, it will hardly be accepted when the
underlying routines are hard to use or do not look like what the end-users expect.

Historically, OR-scientists consider the development of models and algorithms as

the dominant component of optimization-based DSS (Power and Sharda 2007).
However, the contributions and continuous development of other design issues like
databases, effective user interfaces and particularly tools to analyse the generated
solutions, may be even more important for the applicability of model-based DSS.
Kallrath (2004) confirmed the importance of the latter issue and described it as the final
stage in decision processes, which includes the delivery and analysis of the generated
solutions in a usable form to non-technical end-users.

1.2.3 Professional relevance of DSS

Carlsson and Turban (2002) mentioned in a special issue about the future of decision
support systems that the term DSS was seen less and less frequently, both in trade
journals and in vendor web sites. The paper mentioned the conception that DSS
matured to a point of losing its identity and may even disappear as a stand-alone field.
The authors stated the opposite and claimed that the developments of DSS will actually
thrive into the next decade because most of the challenges of DSS are still valid.
Moreover, the so-called second generation of Enterprise Systems recognized the need
for supporting not only transaction processes, but also analytical processing (Carlsson
and Turban 2002).

After four decades of DSS research, Arnott and Pervan (2008) reported on a
long-term project that critically analysed the academic field of DSS and showed that the
gap between research and practice, still exists. The authors analysed almost 1100
articles published in fourteen major journals and showed that almost half of the analysed
research was regarded as having low or no practical relevance while only ten per cent of
the research was regarded as having a (very) high relevance. The authors argue that
the practical contribution of DSS research faces a crisis of relevance due to a long-term
issue, i.e. the tension between academic rigor and professional relevance.

Framinan and Ruiz (2010) confirmed the gap between theory and practice in
their review on the development of customised and realistic manufacturing scheduling
systems. The authors stated that a vast amount of literature is available for
manufacturing and scheduling models including solution techniques. However, very little
has been written on how to bring these models and procedures into practice. The
evidenced trend regarding a lack of relevance and applicability of model-based DSS
was an important premise for this thesis:

7
Chapter 1

Research premise P1: Professional relevance and applicability

A professionally oriented academic area like DSS needs a reasonable balance between
development of theory and real-life applications since research and practice inform each
other (Arnott and Pervan 2008).

Since the field of DSS aims to be an application-oriented discipline, a logical step

is to focus in the next section on its elaboration in practice, i.e. in processing industry.

1.3 Decision support for industrial practice

Although Arnott and Pervan (2008) found a moderate to low impact for DSS in practice,
the basic concepts of model-based DSS did find their application in a subset of
commercial software suites called Advanced Planning Systems (APS) (Günther and
van Beek 2003; Stadtler 2005; Pochet and Wolsey 2006; Stadtler, Fleischmann et al.
2012). APS can be regarded as the latest offspring in the development of Enterprise
Systems (ES) like Material Requirement Planning (MRP) and its successors. APS
particularly aim to give substance to the lack of decision support in (prior) enterprise
systems (Entrup 2005). The next sections aim to describe the basic architecture of APS
including its relation with the field of DSS. The main architecture of APS constitutes the
basis for i) positioning the core of this study by defining additional research premises
and a further demarcation of the types of problems to focus on in the next chapters, ii)
exposing the needs for additional decision support in daily practice, particularly with
respect to processing industry.

1.3.1 Origin of advanced planning systems

In the 1960’s manufacturing strategies were mainly focused on inventory control (Umble,
Haft et al. 2003). In those days, companies could afford to keep lots of ‘‘just-in-case’’
inventory on hand to satisfy customer demand and stay competitive. In the late 1960’s it
became increasingly clear that companies could no longer afford the luxury of
maintaining large quantities of inventory. The conventional thrust of product-focused
manufacturing strategies based on high-volume production, cost minimization and
assuming stable economic conditions came to an end (Jacobs and Weston 2007). At
that time, and simultaneously with the emergence of DSS, software vendors recognized
the high potentials of available data and developed software, i.e. Enterprise Systems
(ES), to standardize and control production planning problems (Jacobs and Weston
2007). In those early days, the wave of real-world DSS applications was still in its
infancy. Nevertheless, the introduction of Material Requirement Planning (MRP) was a
major step forwards (Pochet and Wolsey 2006).

The name Material Requirement Planning (MRP) was coined in the late 1960s
through a joint effort between J.I. Case, a manufacturer of tractors and other
construction machinery, and IBM which resulted in one of the earliest software
applications for planning and scheduling materials for complex manufactured products

8
 General Introduction

(Jacobs and Weston 2007). Pochet and Wolsey (2006) state that the first serious efforts
to formulate mixed integer programming (MIP) models for planning problems of the type
that MRP systems are designed to tackle, date from the 1960s and 1970s. However, at
that time, MIP systems were only able to solve “toy” instances and efforts to solve these
problems mainly concentrated on simple and rapid heuristics. Powerful (personal)
computers with internal data storage facilities became available and the era for
computerized planning and control systems started. However, despite of the fact that
generic optimization-based Material Requirement Planning models were available for at
least discrete parts manufacturing systems, MRP and its successors were first, and
foremost transaction and information-oriented systems, necessary but not sufficient for
efficient planning and decision support at factory level or for planning problems of
complete enterprises (Pochet and Wolsey 2006).

The interested reader is referred to Pochet and Wolsey (2006) for a basic insight
in mathematical formulations of classical production planning models considered in
Enterprise Resource Planning (ERP) or MRP systems. This includes the drawbacks and
limitations of these systems due to the applied decomposition approach on product level
in order to solve these models. Pochet and Wolsey (2006) stated that the observed
limitations all relate to the MRP decomposition approach and planning process, and not
to the MRP model itself. Nevertheless, the concept of MRP systems can be considered
as the basic vein that would become the key for all subsequent developments on the
software market with respect to enterprise systems for industrial practice (Pochet and
Wolsey 2006). The authors stated that superior results can be obtained for production
planning problems if today’s transaction-oriented systems are changed into planning
systems for coordination and optimization.

Nowadays, large companies face the challenge of increasing competition,

expanding markets, and rising customer expectations. Software industry provides
software suites consisting of a number of interrelated modules each intended for specific
planning tasks (Umble, Haft et al. 2003). These so-called Advanced Planning Systems
(APS) share one major characteristic, namely extending transaction- and information-
oriented systems by optimization-based tools for decision support. APS incorporate
models and solution approaches attributed to Operations Research (Stadtler 2005;
Stadtler and Kilger 2008). The introduction of APS intended to shift the objective of
production planning for industrial practice from generating plans to solutions that are
subject to constraints and company-specific optimization criteria (Entrup 2005). APS are
either add-ons or direct integral components of enterprise resource planning (ERP)
systems, which create the support mechanism for planning and decision-making at the
strategic, tactical, and operational planning level (Møller 2005; Jonsson, Kjellsdotter et
al. 2007). Advanced Planning Systems particularly aim to support decision-making.
They do not intend to substitute MRP systems or their successors but can be regarded
as a top layer for these systems in order to support planners in making decisions on
different levels in organizations (Entrup 2005; Jonsson, Kjellsdotter et al. 2007; Stadtler
and Kilger 2008). Due to the added functionality of optimization-based decision support,

9
Chapter 1

industrial practice started to demand for APS. The historical development and market
penetration of computerized planning systems is depicted in Figure 1.1 (Entrup 2005).

Figure 1.1 Historical development of planning systems (Entrup 2005)

1.3.2 Main architecture of APS systems

It is well known that the strength of transactional systems like enterprise resource
planning (ERP) is not in the area of planning (Stadtler 2005). APS are developed to fill
this gap. The three main characteristics of an APS are integral planning (i.e.
coordination of the planning processes of an entire supply chain), a hierarchical planning
approach (i.e. the decomposition into planning modules, and their vertical and horizontal
coordination) and true optimization (Pochet and Wolsey 2006; Stadtler and Kilger 2008).
The MRP concept in nearly all ERP systems is a planning system restricted to the
procurement and production area. It does not optimize and in most cases even not
consider an objective function (Stadtler and Kilger 2008).

Rohde, Meyr et al. (2000) introduced the main structure or architecture of

planning processes in APS which is known as the Supply Chain Planning Matrix
(SCPM). Different variants of the SCPM exist but they all share the same basic principle,
i.e. to support the main planning tasks related to material flows in organizations along
two dimensions: the supply chain process and the planning horizon. The main focus
from a supply chain point of view is to support decision-making at different stages or
phases in the material flow, i.e. from procurement, production, distribution to sales
(horizontal-axis), within the framework of Anthony’s levels of aggregation (vertical-axis)
ranging from strategic (long-term) to operational (short-term) planning (Anthony 1965).
Figure 1.2 depicts a variant of the SCPM. The interested reader is referred to literature
for an extensive description of the SCPM (Stadtler and Kilger 2008; Stadtler,
Fleischmann et al. 2012). APS systems typically consist of different software modules,
each of them covering a certain range of planning tasks in the SCPM.

10
 General Introduction

Due to the earlier mentioned complexity of the production structure in process

industries, the impact of specific production operations on (intermediate) products, and
the need for an efficient use of expensive installations, this thesis will mainly focus on
medium- to short-term decision support at production phase, including its horizontal
integration with decision problems on procurement and distribution phase.

Monolithic models for all planning tasks in the SCPM will neither be solvable nor
accepted by various managers of specific tasks (Stadtler, Fleischmann et al. 2012).
Monolithic models will also require large amounts of up-to-date data, and revising data
will result in frequent replanning.

Figure 1.2 Supply Chain Planning Matrix; based on Rohde (2004). The dark shaded
parts refer to the focus of the research in this thesis.

Moreover, bottom-line managers will be reluctant to input their local knowledge

into an abstract model at the top of an organization’s hierarchy (Stadtler, Fleischmann et
al. 2012). Consequently, different models are proposed for individual building blocks
and/or between (adjacent) blocks in the SCPM. According to Stadtler and Kilger (2008)
it is not possible and not recommended to perform optimization on detailed data. In
order to reduce the complexity and the need for detailed data, the principles of
aggregation, decomposition and reformulation are important starting points for model
development in hierarchical planning approaches like APS.

Aggregation is particularly important on higher planning levels. Wijngaard (1982)

stated that aggregation can be achieved along four dimensions: aggregation over time,
i.e. the length of periods in the time horizon, aggregation over product types,
aggregation over capacities or resources and aggregation over product stages.
Aggregation is crucial on medium-term planning level because it reduces the size of
models (i.e. computational complexity). Moreover, detailed data (e.g. demand) may not
be available over the complete planning interval. However, aggregation is done at the
expense of accuracy which may imply that decision-makers in practice poorly support

11
Chapter 1

the generated solutions (Stadtler, Fleischmann et al. 2012). The latter gives rise to the
next research premise.

Research premise P2: Aggregation

Models for decision support should be based on adequate aggregation levels, carried by
decision-makers at crucial decision levels in practice.

Besides aggregation, the principle of decomposition is often applied in order to

reduce the computational complexity of solving (monolithic) models and the need for
detailed data. Decomposition may refer to the scope of the problem, the developed
model and/or the applied solution techniques. The SCPM in Figure 1.2 can be regarded
as an example of decomposition as it divides an “overall problem” for supply chains into
sub problems for each planning or decision unit in the SCPM (Stadtler, Fleischmann et
al. 2012). A disadvantage of decomposition approaches in hierarchical planning systems
is that separate mathematical models are used for every level. The major drawback of
decomposition in hierarchical systems is the risk of a weak linkage between different
models. Because of the separate models, some mechanism is needed for obtaining
solutions which are consistent across planning levels.

Reformulation is often needed for improving the (initial) formulation of a specific

problem. Many companies try to develop planning systems able to optimize productivity.
The first step is often to develop new planning models. The resulting large-size mixed
integer programming problems are typically much harder to solve to (near) optimality
than linear models. Nevertheless, it is often possible to (re)formulate these models such
that the solution time is drastically reduced. Unfortunately, some of these reformulation
techniques are not generic and depend on specific structures in the problem/model. The
identification of special structures in planning problems is important during model
construction, especially for the use of reformulation techniques (Pochet and Wolsey
2006). Kallrath (2002) confirmed that many problems in process industries lead to
complex MI(N)LP models. Moreover, solution efficiency strongly depends on the
individual problem and model formulation. However, for both problem types, MILP and
MINLP, it is recommended that the full mathematical structure of a problem is exploited,
appropriate reformulations of models are made and/or specific valid inequalities or cuts
are used (Kallrath 2000). Pochet and Wolsey (2006) confirmed that the identification of
special structures in production planning problems is important during model
construction, especially for the use of reformulation techniques.

The computational complexity of (monolithic) planning approaches force model

developers to apply principles of aggregation and decomposition. However, their
disadvantages should be considered at model construction. Taking advantage of
specific (domain-oriented) problem characteristics, identifying and exploiting special
(mathematical) structures, and applying favourable reformulation approaches may be
even more important to solve problems faster and to generate solutions that are actually
carried by decision-makers in practice. The latter gives rise to the next premise for this
research:

12
 General Introduction

Research premise P3 : Decomposition and reformulation

Decision support in practice requires decomposition and/or reformulation. These
principles may refer to the scope of the problem, the developed model and/or the
applied solution techniques. Characteristics of the problem domain including the
requirements of specific industries should be taken into account and exploited in order to
find effective decomposition schemes and/or reformulation approaches.

The next section will focus on two major aspects at production planning stage:
the aggregation level in time (i.e. segmentation of the planning horizon) on different
planning levels and the (related) decomposition of planning and scheduling at
production level, which are both of particular relevance for processing industry.

1.3.3 Vertical integration of production planning and scheduling

While master planning in Figure 1.2 particularly coordinates material flows between
locations, production planning and detailed scheduling is usually run on single locations
(Stadtler 2005). On short-term planning level, the aggregated master production plan
should be disaggregated to derive detailed plans for different plants and production
units. The aim of production planning and scheduling modules in APS should be to
generate detailed production schedules for the shop floor over a relatively short interval
of time (Stadtler and Kilger 2008). On medium-term planning level, the time horizon is
usually divided into (big) time-buckets of variable length. However, sequence-dependent
set-up costs and times on flow lines in processing industry cannot be represented
properly by big bucket models (Stadtler and Kilger 2008). Due to for instance sequence-
dependent set-up times, big-bucket-oriented production plans on a medium-term
planning level may lead to infeasible solutions after disaggregation on a lower planning
level. Reversely, the generated schedules on the shop floor often fail to realise
production targets because changeover losses are not correctly accounted for on a
higher planning level. As a consequence, the planning process has to be redone (with or
without over-time) and/or frequent rescheduling takes place in daily practice (Kreipl and
Pinedo 2004). One option on the higher planning level may be to reserve a certain
portion of available capacity for set-up times. However, the portion may be either too
large or too small (Stadtler and Kilger 2008).

The time horizon for production planning usually covers a period between weeks
and months with time buckets of days or weeks while a typical horizon for production
scheduling covers a period between hours and days (Entrup 2005). In order to reduce
the complexity of decision problems at production phase, Stadtler, Fleischmann et al.
(2012) propose to separate (i.e. decompose) the planning tasks into at least two levels:
production planning first and sequencing and scheduling second. In general practice,
lot-sizing and scheduling problems are usually solved separately in successive
hierarchical phases (Claassen and Beek van 1993; Drexl and Kimms 1997; Kreipl and
Pinedo 2004; Soman, Van Donk et al. 2004a; Soman, van Donk et al. 2007; Framinan
and Ruiz 2010; Stadtler, Fleischmann et al. 2012). First optimal lot sizes for given
product families are determined and afterwards production schedules of customer
orders are generated. Stadtler and Kilger (2008) suggested to take the industrial sector

13
Chapter 1

as a starting point to determine the planning interval including its segmentation. The
authors also stated that the production type at the shop floor should determine whether
production planning and scheduling are executed by a single planning level or by a, less
elegant, two-level planning hierarchy. Stadtler (2005) stated that if the loading of
resources and lot-size decisions are strongly affected by the sequence of jobs, which
often applies to the process industry, both production planning and scheduling should be
performed simultaneously.

Although most of the lot sizing literature is focused on discrete manufacturing,

there exists an increasing interest for other areas like processing industries (Günther
and van Beek 2003; Quadt and Kuhn 2008; Clark, Almada-Lobo et al. 2011). Nowadays
there is also a general consensus regarding a closer integration of planning and
scheduling (Meyr 2000; Jans and Degraeve 2008; Clark, Almada-Lobo et al. 2011).
However, today’s APS systems do not provide modules for simultaneous lot-sizing and
scheduling (Stadtler, Fleischmann et al. 2012). This lack of integration is of particular
relevance for processing industry. The utilization of flow lines in this branch of industry is
usually high and different products (lot-sizes) have to compete for scarce available
capacity. Solving models for the two types of problems simultaneously, usually takes a
large computational burden.

The complexity of (mixed) integer programming models to describe these kinds

of problems can easily exceed today’s hardware and algorithmic capabilities (Kallrath
2002). Although the computational complexity may increase, Soman, Van Donk et al.
(2004a) stated that the majority of research contributions do not address specific
characteristics of food processing industry in production phase, e.g. high capacity
utilisation, sequence-dependent set-ups and limited shelf life due to product decay. The
latter gives rise to the next research premise:

Research premise P4: Vertical integration in production phase

Due to specific characteristics of processing industry, decision support in the production
phase should include simultaneous planning and scheduling in which sequence-
dependent set-ups and product decay are considered.

Although the borders between the different building blocks of the SCPM may be
less strict than depicted in Figure 1.2 (e.g. between production planning and
scheduling), Stadtler (2005) stated that the general aim of APS is to achieve a better fit
between modules, planning tasks and decision-making. The next section will focus on
the importance of a close (horizontal) integration and coordination between building
blocks of the SCPM.

1.3.4 Horizontal integration of planning tasks

APS hierarchically decompose all planning tasks in a supply chain into partial planning
problems and solve them within single modules (Entrup 2005). A strong coordination
(i.e. the configuration of data flows and the division of planning tasks to modules) of APS
modules is a prerequisite to achieve consistent plans for the different planning phases

14
 General Introduction

and for each entity of the supply chain (Stadtler and Kilger 2008). The authors defined
the incremental update and major changes on master data in ERP systems as the key
for integration between APS and (transactional-oriented) ERP systems. Although
different APS modules can interact directly by sending messages, exchanging data and
information between different decision phases, coordination and integration is often
restricted to the exchange of data flows between different modules and/or the related IT
infrastructure (Stadtler and Kilger 2008). Literature on integrated modelling approaches
for separated planning issues in the SCPM, is relatively sparse (Kanyalkar and Adil
2005). Studies, in which issues of integration are considered, mostly refer to
simultaneously considering production and distribution planning (Kanyalkar and Adil
2005). Recently, Mula, Peidro et al. (2010) presented a review of mathematical
programming models for supply chain production and transport planning. The authors
found 44 studies within a time frame of 25 years that focussed on tactical and/or
operational decision levels and their possible combination with aspects of a strategic
nature. Kanyalkar and Adil (2005) developed a single model for consumer goods
industry integrating aggregated and detailed production planning with a detailed
distribution plan. In a follow-up study, the authors focussed on the missing link with
procurement (Kanyalkar and Adil 2007). Although planning issues between production
and distribution have been the concern of research, integrated modelling approaches
between other building blocks of the SCPM retrieved remarkably little attention.

Sourcing of (various) raw materials needs planning both on medium-term and

short-term planning level, particularly in processing industry. On a medium-term
planning level, decisions regarding which, how much, and when various raw materials
must be purchased and delivered at processing sites, are of major importance for
production environments that are characterized by i) (semi-) batch type production
processes (e.g. the choice of various raw materials to be processed in different batches
on shared or multi-purpose equipment), ii) decline in quality of raw materials (e.g. raw
milk in dairy industry), and iii) limited capacity of (special) storage facilities, both on
supply and processing level. For instance, if perishable raw materials are produced at a
constant level in a push-oriented supply chain (e.g. raw milk) and processing of different
end products is planned on shared resources at a limited number of discrete moments in
a planning horizon, sourcing and production planning decisions are interrelated and
complicated. In those situations, the collection schedules at supply phase are not
restricted to solving vehicle routing problems. On a short-term level, the routing problem
for collecting perishable raw materials (e.g. in dairy industry) is more complicated than
solving a classical vehicle routing problem (VRP) with a typical planning period of a
single day (Chao, Golden et al. 1995). If raw materials at supply level are collected at
various frequencies, the collection problem can be classified as a periodic vehicle
routing problem (PVRP). The PVRP extends the classical VRP from a single day to a
time horizon of T days in which each supplier must be visited at least once but some of
them may or must be visited several times during the T-days period (Chao, Golden et al.
1995; Cordeau, Gendreau et al. 1997). However, this class of problems either concerns
the construction of pickup routes or delivery routes, not both. For an integrated approach

15
Chapter 1

between procurement and production, both pickup and delivery conditions should be
considered simultaneously.

Another special feature in processing industry refers to specific processing

operations like blending, refining or heating (Günther and van Beek 2003). These
operations may have a variable impact on multi-component streams entering a
processing unit which in turn defines the final composition of the combined mass flow. If
flows of raw materials with different chemical or physical properties are treated by
processing units at variable technical settings, the impact of processing operations
refers both to the final properties of end products and (simultaneously) to the required
types of raw materials. This two-sided impact requires a close integration and
coordination between procurement and production.

Research premise P5: Horizontal coordination and integration

There is a need for integrated approaches in processing industry particularly on the
tangent plane between procurement and production.

1.4 Research motivations, objective and questions

Many resources emphasise the need for developing specific (integrated) decision
models for each production segment and planning step in APS (Günther and van Beek
2003; Entrup 2005; Stadtler and Kilger 2008; Stadtler, Fleischmann et al. 2012).
According to Entrup (2005), it is crucial to examine the requirements of specific
industries and develop industry-specific solutions. Given the need for specific
(integrated) decision support in process industries, the general research objective (RO)
of this study is:

To support medium- to short-term planning problems by optimization-based models and

solution techniques such that:

i) The applicability and added value of (prototype) systems is recognized and

carried by decision-makers in practice
ii) The proposed approaches contribute to knowledge, understanding and insights
from a model-building and -solving point of view.

A number of planning issues are studied to give substance to the research

objective. The translation of the general objective into concrete research questions
(RQ’s) is based on the research premises P1 to P5 in the Sections 1.2 and 1.3.

Due to the complexity of decision problems on production level in process

industries, both the first and second RQ focus on the relation and integration between
planning and scheduling.

In companies where APS are implemented, planning and scheduling decisions

are often transferred from the shop floor to the new APS. As a result, there can be a dis-

16
 General Introduction

agreement between the system and the shop floor which may lead to problematic use of
APS in practice (Wiers 2009). The reserved use of APS in daily practice corresponds
with the findings of earlier studies (Kreipl and Dickersbach 2008; Ivert and Jonsson
2011; Ivert 2012). As the field of decision support systems was initiated and aims to be
an application-oriented discipline and APS particularly intend to support decision-making
in practice, the first research question (RQ) is:

RQ1
How to apply aggregation, decomposition and reformulation in model-based DSS at
planning and scheduling level such that the aspect of decision support is recognized and
appreciated by decision-makers in practice, and which level of aggregation is needed to
integrate production planning (i.e. lot-sizing) and scheduling problems in a single model?

Literature shows that the boundaries between planning (i.e. lot-sizing) and scheduling
are fading, but further integration still constitutes a challenging research track (Jans and
Degraeve 2008; Quadt and Kuhn 2008). Both reviews showed that there is an on-going
research trend directed towards incorporating real-world issues and specificities of
simultaneous lot-sizing and scheduling. For instance, the problem of contamination is a
key aspect in animal food production. An ingredient needed for one type of animal can
be lethal for another (Wiers 2009). These kinds of typical characteristics in food
processing industry make it necessary to relax all assumptions with respect to
changeover matrices, particularly with respect to the so-called triangular set-up
conditions. Moreover, lot sizing and scheduling models in food processing industry
should include issues of deterioration due to perishability of inventory. In a recent
special issue on lot-sizing and scheduling Clark, Almada-Lobo et al. (2011) confirmed
and emphasized the need for more realistic and practical variants of models for
simultaneous lot-sizing and scheduling. Features like non-triangular set-ups,
perishability, and delivery time windows were explicitly labelled by the authors as open
research opportunities. Therefore, the next research question refers to both a vertical
integration of production planning and scheduling at production level, and a closer
coordination between production and physical distribution level in the SCPM:

RQ2
How to integrate production planning (i.e. lot-sizing) and scheduling problems in a single
model, such that common assumptions regarding the triangular set-up conditions are
relaxed and issues of product decay and limited shelf lives are taken into account?

Both production and distribution planning of (end) products are part of the APS
framework. However, coordination and integration issues should not be restricted to
these two phases in material flows. Comparable planning problems of integration may
manifest at procurement phase with a reverse impact on planning problems in the
production phase. In their literature review, Kanyalkar and Adil (2007) explicitly
concluded that issues of integration between distribution at procurement level with
production, are rarely addressed. The critical issue is to implement systems that
integrate organizational decision-making vertically (among strategic, tactical, and

17
Chapter 1

operational levels) and horizontally (among many functional fields at the same level) to
coordinate and manage conflicts among the various subunits of the organization (Eom
and Lee 1990). The next research question focuses on the observed need and missing
link between both phases procurement and production (horizontally) and the time
horizon (vertically) in the APS framework (see Figure 1.2). Special emphasis is directed
to integrated decision support across organizational borders. If the “strongest” partner is
in charge of a supply chain and dictates decision-making in the production phase, the
planning and distribution problem at the supply phase becomes even more complicated
and challenging for a “weaker” partner in the same supply chain. Therefore, the third
research question is:

RQ3
How to model and solve an integrated planning problem between procurement and
production, both on a medium-term and short-term planning level, in an inter-
organizational supply chain?

A special feature in processing industry with major impact on computational efficiency,

refers to specific processing operations like blending, refining or heating (Günther and
van Beek 2003). These operations may have a variable impact on multi-component
streams entering a processing unit which in turn define the final composition of the
combined mass flow. An almost classical example in OR is the non-linear “pooling or
fuel mixture” problem in refinery and other branches of process industry (Amos,
Ronnqvist et al. 1997). If these types of production planning problems are treated in the
context of mathematical optimization, they may lead to MINLP problems which are often
hard to solve (Kallrath 2002). Other (non-linear) problems may occur if flows of raw
materials with different chemical or physical properties are treated by processing units at
different technical settings which in turn determine the final properties of end products.
Production according to customer specifications requires interrelated decision-making
with respect to procurement of raw materials, assignment of available raw materials to
different end products including the technical setting of processing units. Depending on
changing production targets of final products, optimization-based decision support may
provide a way for selecting the right raw materials (on the market), to be processed at
various technical settings in available production units, and assign them to end products
that meet customer specifications.

RQ4
How to support decision-makers in practice if crucial properties of end products
simultaneously depend on (endogenous) types of raw materials with different chemical
or physical properties and (endogenous) technical settings of processing units?

1.5 Research method and outline of the thesis

The extensive analysis of DSS by (Arnott and Pervan 2008) showed that the gap
between research and practice of DSS is widening, which is confirmed by (Framinan

18
 General Introduction

and Ruiz 2010) in their study on the development of customised and realistic
manufacturing scheduling systems.

In a recent review, particularly devoted to mathematical programming models for

supply chain production and transport planning, Mula, Peidro et al. (2010) mentioned a
striking finding that more proposed models were validated by numerical examples than
by case studies applied to real supply chains. Arnott and Pervan (2008) proposed a
strategy for improving the relevance of DSS research by increasing the number of case
studies which automatically increases the commitment of all parties involved. A field that
is removed from practice needs case study work to ensure that the questions it is
addressing are both relevant and important (Arnott and Pervan 2008). The authors
stated that researchers need to select problems with a consideration for professional
relevance and interest, in addition to considering the recommendations of previous
academic research. According to Arnott and Pervan (2008), case studies are the
research papers with the highest proportional relevance scores and can illuminate areas
of contemporary practice in ways that experimental studies or surveys cannot. The
review of Mula, Peidro et al. (2010) affirms that applying planning models to real case
studies needs more attention. We take this statement as a starting point for the case-
based approach in (most of) the following chapters. All studies concern modelling and
solving (production) planning problems in process industries by optimization-based
decision support.

We introduced five research premises P1 to P5. Premise P1 refers to the first

part of the research objective (RO) in Section 1.4, while all other premises are related to
the second part of the RO, i.e. model-building and/or -solving. The relation between the
defined premises and research questions is given in Table 1.1. Each research question
RQ(n) refers to the Chapter (n+1).

Table 1.1 Relation between research questions and premises

Premise RQ1 RQ2 RQ3 RQ4

Professional relevance
P1
and applicability
( )

P2 Aggregation
Decomposition and/or
P3
reformulation
P4 Vertical integration

P5 Horizontal integration

Chapter 2 is based on a pilot DSS in dairy industry. The study aims to

demonstrate the validity and contribution of case-based DSS research in the past to the

19
Chapter 1

current framework of APS. The study focuses on a medium-term planning problem (i.e.
lot-sizing) and short-term scheduling problem. Chapter 3 focuses on complete
integration of production planning (i.e. lot-sizing) and scheduling in a single model. The
emphasis is to incorporate specific issues for food processing industry (i.e. non-
triangular set-ups, product decay and delivery time windows) into the model and to
demonstrate its impact on generated solutions. The case-study in Chapter 4 deals with
integrated decision support combining procurement and production in an inter-
organizational supply chain. The goal of the study is to demonstrate the importance of a
distribution level for decision-making between procurement and production in the SCPM.
Chapter 5 studies the impact of technical settings of production units on raw material
flows in processing industry. Moreover, the study demonstrates the impact of
continuously changing decision environments in practice for a real-life DSS, both from a
modelling and solving point of view. Chapter 6 presents a general discussion, an
overview of findings and points out some directions for further research.

20
 Chapter 2

Planning and scheduling in food

processing industry

The big problem with management science models is that managers practically never use them
(Little 1970)

This chapter is based on:

Claassen, G.D.H., and Beek, P. van, (1993)

Planning and Scheduling Packaging Lines in Food-Industry
European Journal of Operational Research
Vol. 70 (2), pp. 150 – 158.

Claassen, G.D.H. and Hendrix, E.M.T., (2014)

On Modelling Approaches for Planning and Scheduling in Food Processing Industry
In: ICCSA 2014, Part II, Lecture Notes in Computer Science 8580
Eds: B. Murgante et al.
Springer, pp. 47 – 59
Chapter 2

Abstract
This chapter consists of two parts. Part I concerns the development and implementation
of a pilot Decision Support System for the bottleneck packaging facilities of a large dairy
company. The planning and scheduling problem has been decomposed into two levels:
a tactical and operational control level. On the tactical level a feasible (daily) production
schedule of the orderbook is determined. A Mixed Integer Linear Programming model is
the basis for making this schedule. On the operational control level two sequencing sub
problems are solved. For the solution of these sub problems well-known heuristics have
been used.

As the case study is based on an earlier study, Part II consists of a literature

research on modelling developments for simultaneous lot-sizing and scheduling. We
consider developments in lot-sizing and scheduling, particularly relevant for problem
settings arising in food processing industry. Food processing industry (FPI) reveals
several specific characteristics which make integrated production planning and
scheduling a challenge. First of all, set-ups are usually sequence-dependent and may
include the so-called non-triangular set-up conditions. Secondly, planning problems in
FPI have to deal with product decay due to deterioration of inventory. We give an
overview of lot-sizing and scheduling models, and assess their suitability for addressing
sequence-dependent set-ups, non-triangular set-ups and product decay. We show that a
trend exists towards so-called big bucket models. However, the advantage of these
approaches may become a major obstacle in addressing the identified characteristics in
FPI.

22
 Planning and scheduling in food processing industry

Part I
Planning and Scheduling Packaging Lines

2.1 Introduction
This first part deals with an approach to solve a planning and scheduling problem
for the bottleneck packaging facilities of the cheese production division of a large dairy
company. This approach is mainly concerned with the development and evaluation of a
pilot Decision Support System (DSS) in order to generate and to display ‘high-quality’
schedules with a reasonable efficiency. The pilot DSS should combine the power of
human judgement and experience on the one hand with the accuracy and speed of the
computer on the other hand. Special attention should be paid to the development and
implementation of a user interface.

Figure 2.1 illustrates the goods flow of the cheese product division. The
divergent product structure can be partitioned as follows: during the first stage the
company produces about 300 different kinds of cheese-varieties. Next, the cheese has
to be stored in a large warehouse for the purpose of ripening. The duration of this so
called maturation period determines the taste of the final product. Consequently the
number of cheese varieties triples in this stage of the goods flow. Finally the cheese will
be transported to the packaging department. In this stage several treatments or
operations have to be performed in order to cope with the specific demand for
packaging requirements of each individual customer. As a result the total number of final
products increases dramatically to 2500. If we take the time horizon of the various
stages Figure 2.1 into account, then it will be obvious that the packaging department is
the bottleneck facility of the cheese product division.

Time

Number of 4 Weeks ·········· 2 Years 1-2 Days 0-3 Days

products

2500

2000

1500

1000

500

Milk Cheese Maturation Packaging Transport

supply production
Stage 1 Stage 2 Stage 3

Figure 2.1 Overview of the product increase in time

23
Chapter 2

In order to handle a throughput of five million pounds of cheese weekly, the

packaging department is equipped with ten packaging lines. Actually, in this case, a
packaging line consists of m machines arranged in series. Generally a job requires m
operations, each operation being performed on a different machine. As the machines of
a packaging line are physically connected, the processing time of each job depends on
the bottleneck operation to be performed. The flow of work is unidirectional: each job
has to pass each machine in a prescribed order. Moreover, it is not possible to interrupt
an operation on a machine before completion of the corresponding job (nonpreemptive
scheduling). In general a job can only be processed on one packaging line that is
specific for that job. However, there are also some jobs that can be processed on more,
not necessarily identical lines. All the products are strictly made to order and in general
no inventory of final products is carried (on open shop). According to Hax and Candea
(1984) the problem can be considered as a generalisation of an open nonpreemptive, n-
job m-machine flow shop problem. The job arrival process can be classified as a
deterministic dynamic shop with a time horizon of two weeks. This means that the new
jobs are periodically released to the shop floor and the processing times of the
intermittently arriving jobs are more or less known.

Several planners of the packaging department are in charge of drawing up a

working schedule for the next few days. This planning process is executed manually and
principles of developing a schedule are strictly based on experience. Moreover, the
transfer and reuse of the specialized expertise to other (often non expert) planner
seldom takes place. It can be a nasty task to obtain a feasible schedule and the
procedure does not lead to an optimal solution in general. Once the generation of a
schedule has been completed, it turns out to be extremely difficult to handle rush orders
by changing this schedule.

The management team of the cheese product division had gained the insight that
the manual planning and scheduling procedure was inadequate for future planning. An
increasing flexibility of the packaging department should get higher priority in order to
meet due dates, reduce lead times, optimize utilisation of resources and reduce
minimize changeover costs.

2.2 Problem analysis

The ultimate goal of the management team was twofold. At first the aim was to support
planners concerned with the planning and scheduling of the packaging lines by drawing
up an effective and efficient working schedule of the order book. Furthermore, the
system should also support the order entry process. If the sales manager possesses a
thorough overview of the working schedule of the packaging department, the order
acceptance department will be able to anticipate more adequately on the remaining
capacity.

The scheduling problem introduced before turns out to be disappointingly difficult

to solve (Hax and Candea 1984). Garey, Johnson et al. (1976) proved that the non-

24
 Planning and scheduling in food processing industry

preemptive scheduling flow shop problem belongs to the class of NP-hard problems. In
order to reduce the complexity of the stated problem, several researches proposed a
decomposition of this kind of problems into a number of control levels. Anthony (1965)
classifies decisions into three categories: decisions on strategic planning level, tactical
planning level and operational control level. Strategic planning is mainly concerned with
long-term decision-making, related to investment decisions, product development etc.
The emphasis of tactical planning is effective and efficient use of all resources. Having
allocated all resources, it is necessary to deal with day-to-day operational decisions.
This is called operational control. Van Wassenhove and Vanderhenst (1983) also
discuss a hierarchical framework for the development and implementation of a (similar)
planning problem for a set of production facilities of a large chemical firm. Our research
is only focussed on both the tactical and the operational control level.

2.2.1 The tactical planning level

On the tactical planning level a feasible (daily) ‘master packaging schedule’ of the
orderbook has to be determined. This notion will be explained in the sequel. The time
horizon is subdivided into ten working days (two weeks). The emphasis is on the
fulfilment of the due dates of the individual jobs. Early handling of orders is possible but
restricted to a small extent (due to the ongoing maturation of the new cheese).
Furthermore, it is also desirable (for reasons of capacity utilization) to minimize the total
changeover time on the packaging lines. For this purpose, the packaging department
created clusters of similar jobs. The jobs belonging to one specific cluster consist of
several similar operations and the changeover costs related to the jobs within one
cluster are negligible. However, for different clusters the jobs require different
operations. For that reason switching over from one cluster to another on a specific
packaging line implies a substantial changeover time. We distinguish about fifty different
clusters; forty of them can be processed on only one, not necessarily the same,
packaging line. The remaining clusters can be processed on several, mostly not
identical, lines working in parallel. Changeover time reduction is achieved by scheduling
the jobs daily to production lots consisting of one or more clusters, taking into account
the availability and due dates of the individual jobs. Moreover, clustering the jobs into
large clusters complies with the endeavour to minimize the remnants of the fixed lots in
the storage yard.

Most of the jobs can only be processed on one specific packaging line. However,
there are also jobs that can be processed on alternative, not necessarily identical, lines
with different processing times. In order to optimize the capacity utilization, the elapsed
time between the arrival and completion of the jobs on the shop floor (the mean flow
time) has to be minimized.

In general a crew of workers on the shop floor can operate only one specific
packaging line. The department can make use of a so-called special shift. This shift can
only be scheduled in the night and has the skills to process any cluster on any
packaging line. The special shift is not considered as overtime; it just fills up a shortage

25
Chapter 2

of capacity in a flexible manner. Moreover it is possible to increase the available

capacity by overtime of the regular labour. An important goal is to minimize the hours of
overtime and special shift.

2.2.2 Modelling
In this section we describe a Mixed Integer Linear Programming (MILP) model
for the tactical planning level. This model turns out to show a great similarity with the
‘capacitated facility location model’. For that reason we briefly review the latter model
before dealing with the tactical planning model.

Capacitated facility location model

The Capacitated Facility Location Model (CFLM) deals with the problem how to locate a
number of facilities (with finite capacity) which have to service a given set of customers,
at minimum cost. Mathematically, this problem can be formulated as a MILP model in
which the index i refers to I potential locations where facilities can be established and
the index j to J customers. Let us now formulate the CFLM, specified by the following
parameters:

Fi ~ The fixed costs associated with a facility at location i.

Ci,j ~ The transportation costs of supplying the demand of customer j from
facility i.
Si ~ The capacity of facility i (units per year).
Dj ~ The demand of customer j (units per year).

Furthermore, define the following decision variables:

Xi,j = The fraction of the total demand Dj of customer j that is supplied from
facility i
1 if facility i is opened,
Yi =
0 if facility i is closed.

Now the capacitated facility location model can be formulated:

 I I J 
∑
min  Fi Yi +
 i =1 i =1
∑∑ C j =1
i, j X i, j 

(1)

Subject to

∑X
i =1
i, j =1 ∀ j, (2)

∑D
j =1
j X i , j ≤ S i Yi ∀ i, (3)

26
 Planning and scheduling in food processing industry

X i, j ≥ 0 ∀ i, j, (4)

Yi ∈ {0, 1} ∀ i. (5)

Equations (2) ensure that the demand of every customer is satisfied. The equations (3)
are the capacity constraints, a facility at location i cannot handle more than Si units a
year. Moreover, these constraints imply that no customer can be supplied from a not
existing facility at location i (Yi =0). The conditions in (4) and (5) complete the set of
constraints.

If there is no restriction on the capacity of the facilities (Si is a very large

number), the solution of the LP relaxation represented by (1)-(4) supplemented with the
restriction

0 ≤ Yi ≤ 1, ∀i (6)

is mostly integer in the Yi ’s. We call this LP model the LP relaxation of problem (1)-(5).
Erlenkotter (1978) developed for this so-called uncapacitated facility location problem an
efficient solution procedure.

The capacitated facility location problem, however, is much harder to solve.

Substantial research has been focussed on this particular problem class. One approach
is directed to the definition of Variable Upper Bound (VUB) constraints. Adding the
(redundant) inequalities

X i , j ≤ Yi ∀i , j (7)

will enrich the model formulation (1)-(5) in such a way that the LP relaxation of this
problem, (1)-(4) together with (6) and (7), tends to generate integer Yi ’s (Schrage 1975).
In order to see that (7) is valid, note that Yi are binary variables. If Yi = 1, (7) is implied
I
by ∑X
i =1
i, j ≤ 1 see (2). If Yi = 0, (3) implies that Xi,j = 0 and in that case (7) is also true.

According to Vanroy (1986), these VUB constraints yield a much tighter LP relaxation
than the formulation without (7). The author reported several studies in which the
inclusion of the inequalities (7) gave very tight lower bounds and sparse search trees.
Cornuejols, Sridharan et al. (1991) described a comparison of several approaches which
are mainly based on heuristics and Lagrangean relaxation.

The tactical planning model

As described before, our main goal on the tactical planning level is to determine the
clusters and to assign clusters to each time period of the planning horizon in such a way
that the due dates of the individual jobs are met and total changeover time is minimized.
These kinds of production scheduling problems are closely related to the above-
mentioned ‘capacitated facility location problem’. In both cases the decision variables
can be subdivided into two classes: the (binary) location variables (Yi) and the
(continuous) allocation variables (Xi,,j).

27
Chapter 2

However, in or case the managerial decisions require the consideration of more goals:
optimizing (i) the capacity utilization, (ii) the hours of special shift and (iii) the hours of
overtime. Moreover, these goals are incommensurable with each other. For this purpose
we approached the problem partly as a goal programming model. The hours of overtime
as well as those of special shift are modelled as deviational variables which are a part of
both the capacity constraints (slack) and the objective function. By means of several
penalty and weighing coefficients in the objective function, we are able to include all the
decision criteria into the model and to assign weights to them.

Mathematically the problem can be formulated as a MILP model, in which the

index j refers to the individual jobs of the orderbook (j = 1, 2,…, J), l to the packaging
lines (l =1, 2,…, L) and i to the potential clusters I (i = 1, 2,…., I ). The index t denotes
the specific day within the planninghorizon T (t = 1, 2,…, T). Now we define the following
coefficients:

PST j,t ~ A penalty coefficient for the starting time t of each job j.
PPT j,l ~ A penalty coefficient for the processing time of each job j on packaging line l.
PFD l,t ~ A penalty coefficient for the forecasted demand on day t at each packaging
line l.
WSC i,t ~ A set-up cost coefficient for scheduling a cluster i on day t.
WSS l,t ~ A cost coefficient for hours of special shift scheduled on packaging line l on
day t.
WOTl,t ~ A cost coefficient for hours of overtime scheduled on packaging line l on day
t.
PTj,l ~ The processing time on job j on packaging line l.
RM l,t ~ The total hours of regular labour available on packaging line l on day t.
RZS t ~ The total hours of special shift labour available on day t.
RZO l ~ The total hours of overtime labour available on packaging line l.
JOB I ~ The set of jobs (collected at the start of the planning horizon) belonging to
cluster i.

Let the following variables be defined:

Xj,l,t = The fraction of job j to be processed on packaging line l on day t.

1 if cluster i will be scheduled on day t,
Yi ,t =
0 otherwise.
ZSl,t = The planned hours of special shift on packaging line l on day t.
ZOl,t = The planned hours of overtime on packaging line l on day t.

Now, the tactical planning model can be stated as follows:

28
 Planning and scheduling in food processing industry

 J
∑∑∑ (PST + PPT j ,l + PFD l,t ) X j ,l ,t +
L T I T L T
Min
 j =1 l =1 t =1
j ,t ∑∑ WSC
i =1 t =1
i ,t Yi ,t + ∑∑ WSS
l =1 t =1
l ,t ZS l ,t
(8)
L T

+ ∑∑ WOTl ,t ZO l ,t 
l =1 t =1 

L T

∑∑ X
l =1 t =1
j ,l ,t =1 ∀ j, (9)

∑ PT
j =1
j ,l X j ,l ,t − ZS l ,t − ZO l ,t ≤ RMl ,t ∀ l , t, (10)

∑ ZS
l =1
l ,t ≤ RZS t ∀ t, (11)

1
2
T −1

∑ ZO
t =1
l ,t ≤ RZO l ∀ l,

∑ ZO l ,t ≤ RZO l ∀ l, (12)
t = 21T

ZO l ,t + ZS l ,t ≤ RM l ,t ∀ l , t, (13)

∑X j ,l ,t − Yi ,t ≤ 0 ∀ i , t , ∀ j ∈ JOB i , JOB i ⊂ JOB := {1,2,...,J } (14)

l =1

Yi ,t ∈ {0, 1} ∀ i, t, (15)

X j ,l ,t ≥ 0 ∀ j , l , t, (16)

ZS l ,t , ZO l ,t ≥ 0 ∀ l, t. (17)

Equation (9) ensures that all the jobs are processed within the planning horizon
T. The capacity constraints are formulated in (10): A nine-hour working day can be
augmented by the available special shift (ZSl,t) or by overtime of the regular labour
(ZOl,t). The constraints (11) put a daily maximum on the total amount of special shift
labour, while the constraints in (12) restrict the total manhours of overtime to a weekly
limit. In addition, the combined manhours of overtime and special shift labour is
restricted to a certain extent, represented by (13). The constraints in (14) imply that a job
j can only be processed on day t if cluster i, to which job j belongs, will be scheduled on
that particular day. Moreover, these VUB constraints enrich the model formulation in a
way already discussed. Constraints (15)–(17) complete the set of restrictions.

29
Chapter 2

One important goal is to meet the due dates of the individual jobs. Early handling
of orders is possible but, regarding the maturation period of the cheese, restricted to a
small extent. So each job has to be scheduled somewhere on the time horizon between
its availability date and its due date. Within the specified period the emphasis is to match
the completion time of the jobs with the day preceding their due dates. For this purpose
the square of the deviation between the due date of each job and the planned arrival
time plus one is minimized. The corresponding penalty coefficient for the starting time t
of each job j (PSTj,t) has been defined as:

PST j ,t = c 1 {DD j − (t + 1)}

2
for AV j ≤ t ≤ DD j

in which
c1 ~ A weighing coefficient.
DDj ~ The due date of job j.
AVj ~ The availability date of job j.

A quadratic function has been preferred to a linear one because it reduces the number
of jobs which will be divided over more than one day within the planning horizon (see
Figure 2.2).

Penalty
(PST)

C1= ⅔

1 2 3 4 5 6 7 8 9 10

Availability Due date Time (t)

Figure 2.2. Penalty coefficients dependent on the starting time t of each job j

So most of the allocation variables Xj,l,t will get a binary value. As a result the LP
relaxation of the problem tends to give solutions which are integer in the Yi,t (see (14)).

30
 Planning and scheduling in food processing industry

As stated before, 30% of the jobs can be processed on alternative, not

necessarily identical, packaging lines with different processing times. In order to
optimize capacity utilization, the elapsed time between the arrival and completion of the
jobs on the shop floor (the mean flow time) has to be minimized. For this purpose the
objective function contains a penalty coefficient (PPTj,l), which has been defined as,

SI j
PPT j ,l = c 2
PS l

in which
c2 ~ A weighing coefficient.
SIj ~ The size of job j.
PSl ~ The processing speed of packaging line l.

In practice a scheduler has a restricted knowledge of the amount of intermittently

arriving jobs. However, he/she tries to create a smooth working schedule. In order to
anticipate in an adequate way the future demand, a packaging plan should not only be
based on the jobs in the orderbook. If a reliable estimation of the future demand is taken
into account, the system will be forced to smooth peak production which is mostly
related to some specific days within the planning horizon. In relation to this aspect we
assume that a rolling horizon of two weeks is adequate to anticipate on a short-term
trend in orderbooking. A forecast of the future demand is based on the order process of
the last two weeks and the inventory level in the storage yard. It is incorporated into the
model by a penalty coefficient (see Figure 2.3).

Penalty
(PDF)

C3 = 2

1 3 6 9 12 15 17
Forecasted
RMl,t demand

Figure 2.3. Penalty coefficients dependent on the forecasted demand on

line l at day t

31
Chapter 2

The coefficient PFDl,t has been defined as,

PFD l ,t = c 3 max {0, FD l ,t − RM l ,t }

in which
c3 = A weighing coefficient.
FDl,t = The forecasted demand on day t, on packaging line l.
RMl,t = The total hours of regular labour available on packaging line l, on day t.

Operational control
The tactical planning is concerned with the grouping of jobs into clusters and to
assign these clusters to a particular packaging line within a feasible timetable. In this
way the complexity of the problem can be reduced substantially. The remaining problem
on operational level can be partitioned into two sub problems. In both cases the problem
is a sequencing problem; however the measures of performance are different. At first
the sequence in which the clusters should be processed on each packaging line has to
be determined. This problem, with sequence-dependent set-up times, can be
considered as the well-known ‘asymmetric traveling salesman problem for whose
solution a satisfactory heuristic approach has been chosen based on a savings
algorithm.

Subsequently, the sequence in which the individual jobs within each cluster are
processed has to be determined. This sequence depends on several (logical) rules. The
related planning criteria in descending order of importance are:

• The orders within a cluster are grouped to production code and article number.
Together, these distinguishing marks make up an indication about the cheese
variety, ordered by the customers. As the various cheese varieties are stored in
fixed batches in a warehouse, a grouping into varieties within each cluster
prevents unnecessarily large remnants in the warehouse.
• By clustering the jobs to customer name, a large inventory level on the shipping
department can be avoided.
As the invoicing process cannot start previous to the completion of the packaging
process, the invoice department is served most by a processing sequence of the jobs in
the last (two) cluster to an increasing extent. This working method prevents excessive
activities on the invoice department at the end of the day.

2.3 Results
The development of the (pilot) interactive planning system was started in the spring of
1989. From the beginning it was obvious that a regular and intensive dialogue with the
planners of the company would be of crucial importance. In this way we gained both a
thorough insight into the planning process and substantial support from the planning

32
 Planning and scheduling in food processing industry

department. After about one year a first release of the pilot DSS was implemented on a
VAX-8600 main-frame computer. Some characteristics of the problem on the tactical
planning level are given in Table 2.1. Here, case 1 refers to a stand-alone packaging
line while the figures concerning a number of three parallel lines are summarized by the
second case.

Table 2.1 Some characteristics of the MIP problem matrix

Case 1 Case 2
Rows 1104 2324
Columns 539 1309
Density (%) 0.3 0.3
Binary variables 380 190

Several test runs showed that the selected decomposition and solution
techniques could solve the problem in a satisfactory way. In order to test the model it
had to be implemented and evaluated in a real-world environment. Within this context
special attention had to be given to the development of a user-friendly Man-Machine
Interface (MMI). The development of a pilot interactive planning system including a
direct interface to the local mainframe (IBM AS/400), on behalf of the daily data
collection, has been completed by the end of 1990. It has been implemented on a
powerful personal computer (IBM PS/2). The software package XPRESS-MP (DOS
extended) was used to handle the tactical planning level. The system uses two modules:
the model builder and the optimizer. The input consists of the relevant data files and a
model file. The model builder (or matrix generator) interprets the symbolic specification
statements of the model file and generates a problem matrix. This matrix file can be
read by the optimizer which performs the optimization part. Moreover, it produces a
readable (ASCII) representation of the solution, which in turn constitutes the input for the
modules of the operational control level and the man-machine interface.

For about two months the planning model has been run several times a day. In
spite of the intensive communication with the planners of the packaging department,
they showed substantial detachment in the beginning. This was mainly caused by the
competitive aspect of the system, the initial teething trouble as well as unacquaintance
with computers of the target group. However, after a couple of weeks the planners
appreciated the potential value of the system by its true merits. On the one side they
realized that every computer system has its own deficiencies. Only the human way of
reasoning and his/her experience can compensate for these deficiencies. On the other
hand, the DSS made it possible to generate schedules at any time and within a
reasonable amount of time (about two minutes for each collection of similar packaging
lines). The favourable performance of the MIP model can mainly be contributed to the
VUB constraints (14).

Without the DSS, a planner needs a few hours in order to finish the daily
planning and scheduling problem at the agreed time. Occasionally he has to start the

33
Chapter 2

scheduling process even before all the jobs for the next day have been blocked. This
working method implies that the remaining jobs, partly with a due date of only one day
ahead in the planning horizon, will never fit optimally into the schedule. Moreover,
coping with rush orders is an extremely difficult and frustrating task. With the help of the
DSS the planner can postpone the start of his scheduling task at least until all the orders
with a due date of the next day have been booked. The system has also shown to be
very powerful in generating alternatives or revised plans when unforeseen disturbances
occur; for example a breakdown of a packaging line or a sudden change in demand
(rush orders).

During the period of testing the planners were delighted with the speed in which
some time-consuming clerical actions and data processing were completed. Because a
planner always possesses more information than the system, we created the possibility
to review the generated schedules critically. In this connection the gain of time during
the total scheduling process is of great value. With the help of a menu-driven man-
machine interface the planner is able to modify the computed schedules in such a way
that the solution will be tuned to the actual and future situation on the packaging
department. In most cases the generated schedules have proven to be a good starting
point for the planners and they are on the average of better quality than those drawn up
by hand in the present situation. An additional advantage of the various utilities of the
MMI is the possibility to employ human judgement and experience optimally, in order to
improve the generated schedule.

The (daily) graphical presentation of the complete orderbook and the proposed
final working schedule to the sales manager at the order entry level has proven to be
very valuable. It enables the order acceptance department to make use of the remaining
capacity optimally. As a result the interactive planning system made for a better and
smoother working schedule for the packaging department.

2.4 Conclusions Part I

We described the development and implementation of a pilot Decision Support System
for the bottleneck packaging facilities of a large dairy company. Its major benefits was
the generation of packaging line schedules in a more efficient and effective way. The
quality of the final schedule turned out to be on the average of higher quality than the
solutions found by hand. The efficiency enables the decision-maker to ‘optimize’ his/her
own performance with respect to his/her planning mission.

The conceptual approach to the problem appears to be appropriate. The

quadratic penalty coefficient PSTj,t in the objective function of the MIP model prevents
an excessive split up of the jobs within the feasible part of the planning horizon.
Therefore, most of the allocation variables (Xj,l,t) will get a binary value automatically.
Fortunately only a limited part of the potential jobs possess a processing time which
exceeds the daily capacity. If the priority of the set-up cost coefficient WSCi,t is high, a

34
 Planning and scheduling in food processing industry

large number of jobs (Xj,l,t) will be allocated to each planned cluster (Yi,t). As a result the
LP solution tends to give answers which are integer in the Yi,t’s (see (14)).

The desired or ‘optimal’ working schedule turned out to be strongly dependent on

all kinds of unpredictable situations at the packaging department. Consequently an
appropriate and mutual adjustment of the penalty functions and weighing coefficients in
the objective function is very hard. Even restricted access to the relevant coefficients via
the interactive MMI was not always satisfactory. Hence the system should not be
considered as an optimizer but rather as a tool for generating high-quality schedules to
be used for further analyses. In this connection the various utilities of the user-friendly
and fully interactive MMI are essential.

Initially we planned to extend the heuristic approach on the operational control

level by a 2- or 3-OPT improvement algorithm. However, experiments showed that this
additional effort was hardly relevant.

Acknowledgements
This research has been supported by the Dutch Ministry of Economic Affairs. The
authors wish to express their thanks to the dairy company FRIESLAND FRICO DOMO
B.A. for the fruitful and constructive discussions. They commemorate Drs. Irene Goris
(member of the software house CAPGEMINI PANDATA) who was involved in the
development of the Decision Support System. She died early in 1992 due to a road
accident. The authors dedicate this publication to her.

35
Chapter 2

Part II
Modelling approaches for planning and
scheduling

2.5 Introduction
Adequate and efficient production planning and scheduling is one of the most
challenging problems for present-days enterprises. Especially scheduling and sizing of
production lots, is an area of increasing research attention within the wider field of
production planning and scheduling (Clark, Almada-Lobo et al. 2011). Although lot-
sizing problems have been studied extensively, most of the literature is focused on
discrete manufacturing. Moreover, there is an on-going research trend directed towards
incorporating real-world issues and specificities of simultaneous lot-sizing and
scheduling (Jans and Degraeve 2008; Quadt and Kuhn 2008).

Lot-sizing and scheduling in Food Processing Industry (FPI) is usually more

complex than in other continuous and discrete processing environments. This is
primarily due to issues like inevitable decline in quality of products, related quality
requirements and safety regulations of products, market-driven standards regarding
shelf life and variability of demand and prices. Secondly, the diversity of products in FPI
increased considerably in the past decades and global competition on the food market
has forced manufacturers to participate in an on-going trend towards increased variety
(i.e. ingredients and flavors, customized packaging, prints and/or labels) of (new)
products. Soman, Van Donk et al. (2004a) state that the majority of research
contributions do not address specific characteristics of food processing, e.g. high
capacity utilization, sequence-dependent set-ups and limited shelf life due to product
decay.

Production lines in FPI usually operate under tight capacity constraints. As

products take the same route, a production line may be planned as a single resource.
Changeovers between products that share the same line in a food processing
environment often imply that both changeover costs and times depend on the production
sequence of individual items. In order to avoid unnecessary changeovers and improve
efficient use of available production capacity, customer demand has to be pooled in
production orders (lots). When sequence-dependent set-up times are predominant,
available capacity for production depends on both the sequence and the size of the lots.
In such a situation, lot-sizing and scheduling should be applied simultaneously (Meyr
2000).

In general practice, lot-sizing and scheduling problems are solved separately in

successive hierarchical phases (Claassen and Beek 1993; Drexl and Kimms 1997;
Kreipl and Pinedo 2004; Soman, Van Donk et al. 2004a; Soman, van Donk et al. 2007).

36
 Planning and scheduling in food processing industry

First optimal lot-sizes for given product families are determined and afterwards
production schedules of customer orders are generated. The generated schedules on
the shop floor often fail to realize production targets because changeover losses are not
correctly accounted for on a higher planning level. As a consequence, the planning
process has to be redone (with or without over-time) and/or frequent rescheduling takes
place in daily practice (Kreipl and Pinedo 2004). Currently, there exists a general
consensus regarding a closer integration of lot-sizing and scheduling decisions (Meyr
2000; Gupta and Magnusson 2005; Almada-Lobo, Oliveira et al. 2008; Jans and
Degraeve 2008; Clark, Almada-Lobo et al. 2011; Menezes, Clark et al. 2011).

Although the survey of Drexl and Kimms (1997) already focused on the
integration of lot-sizing and scheduling, Jans and Degraeve (2008) conclude after
another decade in their review that the boundaries between lot-sizing and scheduling
are fading, but further integration still constitutes a challenging research track. The latter
may explain why even in present-days Advanced Planning and Scheduling (APS)
systems, the planning and scheduling modules are seen as unusable, or unable to
handle the complexity of the underlying capacitated planning problems (Pochet and
Wolsey 2006).

Planning (i.e. lot-sizing) models differ from scheduling models in a number of

ways. Kreipl and Pinedo (2004) give an extensive overview of practical issues for
planning and scheduling processes. In a recent special issue on lot-sizing and
scheduling, Clark, Almada-Lobo et al. (2011) confirm the need for more realistic and
practical variants of models for simultaneous lot-sizing and scheduling. Features such
as (i) non-triangular set-ups, (ii) perishability, and (iii) delivery time windows are labelled
by the authors as hot topics and open research opportunities. We focus on two
interrelated problem characteristics that argue the need for simultaneous planning and
scheduling, particularly in FPI:

(i) Sequence-dependent set-ups and non-triangular set-ups.

With respect to sequence-dependent set-up costs and times under tight capacity
constraints there is a complicating issue, referred to as the triangular set-up conditions
(Gupta and Magnusson 2005; Almada-Lobo, Oliveira et al. 2008; Clark, Almada-Lobo et
al. 2011), that holds for FPI too (Menezes, Clark et al. 2011). Due to processing
conditions of different product variants (e.g. several heating and/or cooling levels) and
other product specific requirements (e.g. flavors, addition of specific additives and/or the
danger of contamination between subsequent production runs), the common
assumption regarding the triangular set-up conditions often does not hold in FPI. If these
conditions don not hold, it implies that changeover costs and times between two
subsequent products i and j may become substantially less by processing another
product k between i and j. Consequently, applying models that assume triangular set-up
conditions may generate non-consistent solutions from a scheduling point of view.

(ii) Product decay.

The quality or value of perishable food products usually deteriorates rapidly after
production. Product decay may be delayed by conditioned storage, but quality depends

37
Chapter 2

on product age, and restricted shelf lives are inevitable. Considering product decay in
lot-sizing enforces smaller production quantities for perishable products. Consequently,
individual products are produced i.e. scheduled at higher frequency. This increases the
difficulty of sequencing.

The literature overview intends to contribute to the recognized need for more
realistic variants of models for simultaneous lot-sizing and scheduling under tight
capacity constraints (Almada-Lobo, Klabjan et al. 2007; Almada-Lobo, Oliveira et al.
2008; Jans and Degraeve 2008; Clark, Almada-Lobo et al. 2011; Menezes, Clark et al.
2011). We give an overview of model developments for simultaneous lot-sizing and
scheduling directive for a problem with the following characteristics: a multi-item, single
machine lot-sizing and scheduling problem for FPI with sequence-dependent set-up
costs and times and product decay. The set-up state of the machine should be
preserved over period boundaries including idle time (i.e. set-up carry-overs) and any
additional assumption with respect to the changeover matrices should be relaxed (e.g.
the triangular set-up conditions).

We assess the proposed models for addressing sequence-dependent set-ups

(including non-triangular set-ups) and product decay. The objective here is to focus on
modelling developments in time that are directive for the identified problem
characteristics, and to expose their shortcomings and disadvantages. For a general
overview of lot-sizing problems we refer to several reviews of the past (Kuik, Salomon et
al. 1994; Drexl and Kimms 1997; Karimi, Ghomi et al. 2003) and two more recent
overviews (Jans and Degraeve 2008; Quadt and Kuhn 2008).

The overview in this paper shows that a trend exists of preferred modelling
approaches. However, these approaches may i) disrupt a crucial balance between total
set-up costs and inventory-holdings costs and ii) hamper a further integration between
production and distribution planning. We state that crucial aspects for integrated
planning and scheduling may unfoundedly disappear from sight. One of the most
important features of models for lot-sizing and scheduling is the segmentation of the
planning horizon. From a modelling point of view it is convenient to distinguish two
general classes of models (Eppen and Martin 1987), i.e. small bucket (SB) and big (or
large) bucket (BB) modelling approaches. In SB models, the planning horizon is divided
into a finite number of small time periods such that in each period either at most two
products can be produced, or there will be no production at all. Conversely, in BB
approaches the planning horizon is divided into longer periods, usually of the same
length. In each period, multiple products may be produced. As a consequence, SB
models have been applied mostly over short time planning horizons and BB models are
usually associated with medium-term planning horizons.

Sections 2.6 and 2.7 provide an overview of model developments for SB and BB
approaches respectively. Section 2.8 describes the state of affairs regarding issues of
product decay for lot-sizing and scheduling. A summary can be found in Section 2.9 to
analyse the literature overview. Section 2.10 concludes.

38
 Planning and scheduling in food processing industry

2.6 Small bucket approaches

Crucial for small bucket modelling approaches is that at most one set-up may occur in a
period. In this class of models, the so-called ‘all-or-nothing’ assumption usually holds. In
most models only one item may be produced in a time interval and, if so, production
uses (in most cases) full capacity. In SB models, lot-sizes include the production of the
same product for one or several consecutive periods. Alternatively, if a set-up is
performed and when it comes to non-zero set-up times, both set-ups and production
runs comprise a number of time intervals. A lot includes the production of a single
product for one or several consecutive periods. Next, we discuss development in time of
SB-approaches.

2.6.1 DLSP: Discrete Lot-sizing and Scheduling Problem

The Discrete Lot-sizing and Scheduling Problem (DLSP) is a typical example within the
class of small bucket approaches. The basic DLSP includes (sequence-independent)
set-up costs and set-up carry-over but at zero set-up times (Fleischmann 1990).
Inclusion of set-up carry-over implies that set-up states of a machine are carried over
between period boundaries. Porkka, Vepsalainen et al. (2003) compare models with and
without set-up carry-overs. The authors show that substantial savings (regarding costs
and production time) can be derived from fundamentally different production plans
enforced by carry-overs. Comparable results are found by Sox and Gao (1999).
However, in the basic DLSP, set-up states are not preserved over idle time. Sequence-
dependent set-up costs and times are neither considered in the DLSP. Many extensions
of the (basic) DLSP have been described in literature. We refer to Drexl and Kimms
(1997) and Salomon, Kroon et al. (1991) for a broader view on variants of the DLSP.

2.6.2 Extensions of the DLSP

Fleischmann (1994) analyses the multi-item single machine DLSP with sequence-
dependent set-up costs. An artificial product (i=0) is introduced to represent idleness of
the machine. Salomon, Solomon et al. (1997) continue this work and reformulate the
DLSP to capture the characteristic of sequence-dependent times (DLSPSD). However,
the triangular set-up conditions are assumed to hold. Machine idleness is represented
by an artificial product. Jordan and Drexl (1998) present a comparable model in which
idleness is indicated by an artificial product too. It should be mentioned that for models
in which idleness is represented by an artificial product (i=0), the changeover matrix
must comply with strict conditions to cope with sequence-dependent set-up times. In all
other cases the set-up state of the machine is not correctly carried over across the
boundaries of idleness.

Wolsey (1997) extended the work of Constantino (1996) for problems with
sequence-independent set-ups to formulations with sequence-dependent set-up times
and costs. In this paper, the presented model will be referred to as (GSB), i.e. the

39
Chapter 2

general small bucket model. In the (GSB), idleness is not represented by an artificial
product (i=0). However, the triangular set-up conditions should hold.

2.6.3 CSLP: Continuous Set-up Lot-sizing Problem

An early paper in which sequence-dependent costs are modelled is due to Karmarkar
and Schrage (1985). Their model is called the Continuous Set-up Lot-sizing Problem
(CSLP). The CSLP is closely related to the DLSP. Main difference is that the CSLP
allows production of quantities less than the available production capacity in a time
period. Still, at most one product can be produced in each time interval.

2.6.4 PLSP: Proportional Lot-sizing and Scheduling Problem

The fundamental assumptions of the DLSP and the CSLP stimulated Drexl and Haase
(1995) to study a new type of model, the Proportional Lot-sizing and Scheduling
Problem (PLSP). The PLSP is based on a widening of the common “all-or-nothing”
production principle in SB models. The PLSP assumes that at most one set-up may
occur within a period. Hence, at most two products can be produced in a period. Main
difference between the PLSP and the DLSP is the possibility to compute continuous lot-
sizes and to preserve the set-up state over idle time. However, set-up costs and times of
(extended) PLSP formulations are considered to be sequence-independent (Suerie
2006).

2.7 Big bucket approaches

In contrast to small bucket models, the planning horizon of a big bucket (BB) model is
usually divided into longer periods, mostly of equal length. Time intervals in a BB model
may represent a time slot of one week (or more) in the real world (Drexl and Kimms
1997). In each period, multiple products can be manufactured. Relaxing the “all-or-
nothing” production principle of (most) SB models implies that a BB model includes the
possibility to determine continuous lot-sizes.

2.7.1 CLSP: Capacitated Lot-Sizing Problem

The Capacitated Lot-Sizing Problem (CLSP) is a typical example of a big bucket model.
It is closely related to the (small bucket) DLSP; decision variables, parameters and
objective function are the same for both problems (Drexl and Kimms 1997). However,
sequence-dependent set-up costs and times, or more in general scheduling decisions,
are not integrated into the CLSP. As a consequence, set-up carry-overs between period
boundaries are not included either. Suerie and Stadtler (2003) use the simple plant
location problem to obtain a tight and new model formulation for set-up carry-overs in
the CLSP with sequence-independent set-up costs and times.

40
 Planning and scheduling in food processing industry

2.7.2 GCLP: Generalized Capacitated Lot-sizing Problem

Sox and Gao (1999) introduce the Generalized Capacitated Lot-sizing Problem (GCLP).
The GCLP uses less binary variables for including set-up carry-overs in the CLSP with
sequence-independent set-up costs and no set-up times. Sequence-independent set-up
times can be included; probably at the expense of additional computational effort. The
authors also apply the network reformulation approach as proposed by Eppen and
Martin (1987) and compare the behavior of a set of models. The results demonstrate
that incorporating set-up carry-over has a significant effect on both costs and lot-sizes.

In all aforementioned BB approaches, the emphasis is directed towards

combining characteristics of a big bucket model like the CLSP (i.e. allow production of
more products per period without set-up carry-overs) with a small bucket model like the
DLSP (production of only one product per period with set-up carry-overs) in a single
framework. Still, sequence-dependent set-up costs and times are not considered in the
BB models above.

2.7.3 GLSP: General Lot-sizing and Scheduling Problem

Fleischmann and Meyr (1997) proposed a combination of CLSP and DLSP, i.e.
the General Lot-sizing and Scheduling Problem (GLSP). The GLSP is a big bucket
model in which the planning horizon is divided into T macro-periods. To obtain the
production sequence of the items, each macro-period is subdivided into a subset of
micro-periods of variable length. The GSLP assumes all-or-nothing production for micro-
periods. The number of micro-periods within each macro-period must be fixed in
advance in the MIP model. As a consequence, a lot (i.e. a sequence of micro-periods
assigned to the same item) may contain idle micro-periods. Sequence-dependent costs
are considered, but set-up times are disregarded in the (basic) GLSP. In order to cope
with cases in which the triangular set-up conditions of the cost matrix do not hold, the
authors introduce minimum lot-sizes. Meyr (2000) extended the GLSP with sequence-
dependent set-up times. Again, minimum lot-sizes are used to avoid a wrong evaluation
of set-up costs (and set-up time, respectively) if the set-up matrices do not satisfy the
triangular set-up conditions. It should be mentioned that the introduction of minimum lot-
sizes may have an impact on economical lot-sizes.

Transchel, Minner et al. (2011) present a tailored hybrid mixed-binary model

based on the GLSP for a practical problem from process industry and show that
minimum production quantities affect the MIP performance for real world test instances.
Ferreira, Morabito et al. (2009) present a GLSP-based model too that integrates
production lot-sizing and scheduling decisions for a Brazilian soft drink plant.

Block planning approaches can be considered as a practical variant of the GLSP

in which macro- (i.e. blocks) and micro periods are distinguished. An important
assumption in block planning approaches is a predefined production sequence of
(variable) batch-sizes (Entrup, Gunther et al. 2005; Bilgen and Günther 2010; Baumann
and Trautmann 2012). In other words, there is a unique period-block assignment and

41
Chapter 2

each product occurs at the same given position (micro-period) in each block. As a
consequence, within the planning horizon of T periods, each product i =1..N is
scheduled T times. The number of production lots in the schedule equals N*T. We refer
to Gunther, Grunow et al. (2006) for a complete description of block planning.

2.7.4 Extensions of the CLSP

A study to extend the CLSP was initiated by Gopalakrishnan, Miller et al. (1995). The
authors developed a modelling framework for the (single machine) CLSP with set-up
carry-overs. Set-up times and costs were assumed to be constant across all products
and time periods. This assumption was relaxed in a modified framework that included
product-dependent and sequence-independent set-up costs and times (Gopalakrishnan
2000).

Haase (1996) takes the CLSP as a starting point but extends the model with
sequence-dependent set-up costs. Moreover, the set-up state of the machine can be
preserved over idle times. The model formulation does not consider (sequence-
dependent) set-up times and it is assumed that the triangular set-up conditions for set-
up costs hold. Haase and Kimms (2000) consider both sequence-dependent set-up
costs and times. It is assumed that set-up times satisfy the triangular set-up conditions.
The authors formulate the problem by considering only efficient (predefined) production
sequences. Efficient sequences are found by solving a travelling salesman problem.

Gupta and Magnusson (2005) extend the framework of Gopalakrishnan (2000)

by including sequence-dependent set-up times and set-up costs. From a scheduling
point of view, the CLSP with sequence-dependent set-up times is closely related to the
travelling salesman problem (TSP). In every period a connected tour (or sequence)
between multiple products has to be determined. The distance matrix in the TSP
corresponds to the matrix of set-up costs in the (extended) CLSP. Almada-Lobo,
Oliveira et al. (2008) show that the model formulation as proposed by Gupta and
Magnusson (2005) does not eliminate disconnected sub tours. As a consequence, it
may generate infeasible solutions. Almada-Lobo, Klabjan et al. (2007) present two
correct model formulations for the identified problem characteristics, provided that the
triangular set-up conditions with respect to the set-up matrices (costs and times), hold.
In order to avoid disconnected sub tours, the authors add a polynomial set of sub tour
elimination constraints. Menezes, Clark et al. (2011) present an extension of the CLSP
which handles non-triangular set-up costs and times while enforcing minimum lot-sizes.

Next, we will focus on papers that discuss lot-sizing and scheduling of perishable
products.

2.8 Product decay

Although a vast body of literature exists on inventory management for perishable
products, surprisingly little has been done to include product decay in traditional lot-

42
 Planning and scheduling in food processing industry

sizing and scheduling models. One of the first contributions in this area is provided by
Soman, van Donk et al. (2004b). The authors focus on shelf-life considerations in the
economic lot scheduling problem (ELSP). Models of this class usually assume constant
demand, do not account for sequence-dependent set-up times and aim to generate
production cycles for several products on a single resource. Entrup, Gunther et al.
(2005) propose three MILP models that integrate shelf-life issues into production
planning and scheduling for an industrial case study of yoghurt production. The authors
apply a block planning approach (see Section 3.3) in which a block covers all products
based on the same recipe. Shelf-life aspects are taken into account by considering a
shelf-life-dependent pricing component that may also include inventory-holding costs.
Chen, Hsueh et al. (2009) and Kopanos, Puigjaner et al. (2012) argue the need to
develop models for better coordination between production scheduling and vehicle
routing for perishable food products. Lee and Yoon (2010) consider a coordinated
production-and-delivery scheduling problem that incorporates different inventory-holding
costs between production and delivery stages. The results may only apply to specific
situations but the study can be regarded as a first attempt to allow different (stage-
dependent) inventory-holding costs. Chen, Hsueh et al. (2009) conclude that papers
discussing production scheduling and/or distribution of perishable goods are relatively
rare. Amorim, Antunes et al. (2011) state that papers discussing simultaneous lot-sizing
and scheduling for perishable goods are even rarer. These authors deal with
simultaneous lot-sizing and scheduling of perishable products using a multi-objective
framework. The main idea is to separate economic production tangible costs from
intangible value of having fresher products in two conflicting objectives.

2.9 Literature overview

Table 2.2 summarizes the literature research and gives an overview of key publications
that are directive for the problem formulation in the next chapter. The basic problem can
be characterized as a multi-item, single machine lot-sizing and scheduling problem with
sequence-dependent set-up costs and times. The set-up state of the machine should be
preserved over period boundaries (including idle time) and any additional assumption
with respect to the changeover matrix (e.g. the triangular set-up conditions) is excluded.

Note that Table 2.2 only refers to the presented model formulations and not to
the proposed solution approaches. The (GSB) is a SB formulation for the specified
problem without product decay, provided that the triangular set-up conditions hold. If
these conditions do not hold, set-up state changes will occur without production
changes. Literature shows that there is a clear tendency to propose BB models for short
time horizons too. Moreover, both the survey of Quadt and Kuhn (2008) and the results
in Table 2.2 reveal an interesting trend in which BB approaches are preferred to SB
models.

43
Chapter 2

Table 2.2 Overview of key publications *

extension of Sequence- Non-

/or Set-up dependent triangular
Year Author Basic Model name SB BB carry-over costs times set-ups

DLSP − − − −
1985 Kamarkar & Schrage CSLP − − −
1994 Fleishmann DLSP − − −
1995 Drexl & Haase PLSP − − −
1997 Salomon et al. DLSP − − −
1997 Wolsey GSB −
CLSP − − − −
1996 Haase CLSP − −
1997 Fleishmann & Meyr GLSP − −
1999 Sox & Gao GCLP − − −
2000 Gopalakrishnan et al. CLSP − − −
2000 Meyr GLSP −
2003 Suerie & Stadtler CLSPL − − −
2005 Gupta & Magnusson CLSP − −
2007 Almada-Lobo et.al. CLSP −
2011 Menezes et al. CLSP

* A minus sign in Table 2.2 means that the issue is either not considered in the model formulation or
not adequately modelled.

2.10 Conclusions Part II

Purpose the second part of this chapter was to study how literature deals with practical
variants of models for simultaneous lot-sizing and scheduling models. We considered
developments in lot-sizing and scheduling particularly relevant for problem settings
arising in food processing industry (FPI) and focus on i) sequence-dependent set-ups
(including non-triangular set-ups), ii) product decay of inventory due to perishability, and
(iii) a desired tuning of modules for production planning with physical distribution
planning (i.e. delivery time windows).

Although Big Bucket (BB) models are usually associated with medium-term
planning horizons, literature reveals an interesting trend in which these models are
proposed for short-term planning horizons too. From a computational point of view,
models with large time intervals (i.e. a week) are preferred over Small Bucket (SB)
approaches. However, we argue that segmentation of the planning horizon is a key

44
 Planning and scheduling in food processing industry

issue for simultaneous lot-sizing and scheduling. The observed preference for
segmentation in BB approaches implies that the following crucial aspects may disappear
from sight:

• Main principle of optimality for lot-sizing models.

The general accepted principle of optimality for lot-sizing models is based on the best
compromise between total production costs on the one hand and total inventory-
holding costs on the other hand. Inventory costs in lot-sizing models are calculated
from inventory levels at the end of each period. As time intervals in BB models
represent long periods (e.g. a week) multiple batches can be produced in a single
period. As a consequence, inventory costs for batches manufactured at the start of
periods are assumed to be equal to inventory costs of batches produced at the end of
the same period. As a result, total inventory-holding costs are underestimated and the
crucial principle of optimality for lot-sizing problems may be disrupted. Segmentation
of the planning horizon is the key in modelling this balance correctly. SB models offer
the framework to calculate these costs more adequately.

• Decline of product quality and limited shelf lives.

In FPI, product decay is primarily associated with the “age” of products. Incorporating
perishability issues like product decay requires defining the moments of production for
manufactured products, unambiguously. Segmentation of the planning horizon is the
key to capture the age of manufactured products.

• Delivery time windows for physical distribution

Obviously, a close coordination of production scheduling and delivery planning will
become an important issue (Chen, Hsueh et al. 2009; Clark, Almada-Lobo et al.
2011). Products in FPI usually include highly perishable items that must be delivered
within allowable time frames. In order to contribute to improved logistical performance,
production planning and scheduling modules for FPI should have, at least to a certain
extent, the flexibility to take issues for physical distribution into consideration. In
contrast to BB models, SB approaches offer the timeframe to attune short-term
physical distribution planning to production planning and scheduling, e.g. by assigning
demand to specific time slots in a 24-hours production environment.

• Scheduling lots
A major advantage of small time intervals may be a better control of the sequence of
lots. Using large time intervals implies that sequencing the lots within each period may
become complex. Moreover, planned maintenance for production facilities can be
scheduled much easier and accurately by applying SB models. In each period of a BB
model, a sequencing problem (i.e. travelling salesman problem) has to be solved.
Incorporating this feature may become complex, especially in case the triangular set-
up conditions do not hold.

In spite of a larger number of time periods in the planning horizon, the strengths of SB
approaches will be used in the next chapter to develop models that (i) can handle

45
Chapter 2

sequence-dependent set-ups (including non-triangular set-ups), (ii) addresses product

decay by using age-dependent holding costs. Such models offer (iii) a starting point to
integrate solutions of a production planning with delivery time windows for physical
distribution of products to customers.

46
 Chapter 3

Vertical integration of lot-sizing and

scheduling in food processing industry

The optimal solution of a model is not an optimal solution of a problem unless the model is a perfect
representation of the problem (Ackoff 1977).

This chapter is based on:

Claassen, G.D.H., Gerdessen, J.C., Hendrix, E.M.T., and van der Vorst, J.G.A.J.,

On production planning and scheduling in food processing industry:

modelling non-triangular set-ups and product decay
Under review
Chapter 3

Abstract
Based on the conclusions in the preceding chapter we consider a complete vertical
integration of lot-sizing and scheduling problems which is particularly relevant for food
processing industry (FPI). Problem settings in FPI require to take specific characteristics
into account. First of all, set-ups are usually sequence-dependent and may include the so-
called non-triangular set-up conditions. These conditions make it necessary to relax all
assumptions with respect to the changeover matrices (both with respect to costs and
times). Secondly, lot sizing and scheduling models in FPI must take product decay into
consideration. We present an MILP model that handles these characteristics. We study its
behaviour and complexity and show that optimal production schedules become significantly
different when non-triangular set-ups and product decay are taken into account. Numerical
results are provided for small size instances and a time-based decomposition heuristic is
applied to solve larger problem instances.

48
 Vertical integration of lot-sizing and scheduling in food processing industry

3.1 Introduction
Adequate and efficient production planning and scheduling is one of the most challenging
problems for present-days enterprises. Lot-sizing and scheduling in Food Processing
Industry (FPI) is usually more complex than in other continuous and discrete processing
environments. First of all, planners have to deal with decline in quality of products, related
quality requirements and safety regulations of products, market-driven standards regarding
shelf life, and variability of demand and prices. Secondly, the diversity of products in FPI
increased considerably in the past decades and global competition on the food market has
forced manufacturers to participate in an on-going trend towards increased variety (e.g.
ingredients and flavours, customised packaging, prints and/or labels) of (new) products.
Soman, Van Donk et al. (2004a) state that the majority of research contributions do not
address specific characteristics of food processing, e.g. high capacity utilisation, sequence-
dependent set-ups and limited shelf life due to product decay.

In general practice, lot-sizing and scheduling problems are solved separately in

successive hierarchical phases (Claassen and Vanbeek 1993; Drexl and Kimms 1997;
Kreipl and Pinedo 2004; Soman, Van Donk et al. 2004a; Soman, van Donk et al. 2007).
First optimal lot-sizes for given product families are determined and afterwards production
schedules are generated. The generated schedules on the shop floor often fail to realise
production targets, because changeover losses are not correctly accounted for on a higher
planning level. As a consequence, the planning process has to be redone (with or without
over-time) and/or frequent rescheduling takes place in daily practice (Kreipl and Pinedo
2004). Currently, there exists a general consensus regarding a closer integration of lot-
sizing and scheduling decisions, see Meyr (2000), Gupta and Magnusson (2005), Jans and
Degraeve (2008), Almada-Lobo, Oliveira et al. (2008), Clark, Almada-Lobo et al. (2011),
and Menezes, Clark et al. (2011).

Planning (i.e. lot-sizing) models differ from scheduling models in a number of ways.
Kreipl and Pinedo (2004) give an extensive overview of practical issues for planning and
scheduling processes. In a special issue on lot-sizing and scheduling, (Clark, Almada-Lobo
et al. 2011) confirm the need for more realistic and practical variants of models for
simultaneous lot-sizing and scheduling. Features such as (i) non-triangular set-ups, (ii)
perishability, and (iii) delivery time windows are labelled by the authors as hot topics and
open research opportunities. The research question of this paper is how to include the first
two characteristics in models for simultaneous lot-sizing and scheduling.

(i) Sequence-dependent set-ups and non-triangular set-ups. There is a complicating

issue with respect to sequence-dependent set-up costs and times, commonly referred to as
the assumption of the triangular set-up. Menezes, Clark et al. (2011) confirm that non-
triangular set-ups may occur in FPI. Due to processing conditions of different product
variants (e.g. several heating and/or cooling levels) and other product specific requirements

49
Chapter 3

(e.g. flavours, addition of specific additives, the danger of contamination between

subsequent production runs), changeover costs and times between two subsequent
products i and j may become substantially less by processing another product k between i
and j. As a consequence, applying models that assume triangular set-up conditions may
generate non-consistent solutions from a scheduling point of view.

(ii) Product decay. In many FPI cases, the quality or value of perishable food products
deteriorates rapidly after production. Considering product decay in lot-sizing enforces
smaller production quantities. Consequently, individual products are produced at higher
frequency. This increases the difficulty of sequencing.

This paper investigates implementing the characteristics into models for simultaneous
lot-sizing and scheduling under tight capacity constraints. We present an MILP model that
includes the identified characteristics. Moreover, the approach offers a natural starting point
for integrating delivery time windows in lot-sizing and scheduling models as mentioned by
Clark, Almada-Lobo et al. (2011). Small scale examples demonstrate that optimal
production schedules become significantly different when including non-triangular set-ups
and product decay. Two model formulations are presented and compared with a known
approach from literature.

The remainder of the paper is organised as follows. Section 3.2 embeds the model in
existing approaches from literature. Section 3.3 presents two MILP models for the problem
under consideration. Section 3.4 provides small scale numerical examples to demonstrate
the impact of non-triangular set-ups and product decay. Moreover, the complexity of the
model is studied. Section 3.5 provides numerical results for medium size instances,
including a comparison with a straightforward MP-based heuristic. Concluding remarks and
suggestions for further research are given in Section 3.6.

3.2 Embedding in the literature

Models for lot-sizing and scheduling can be classified according to the segmentation of
the planning horizon. From a modelling point of view, it is convenient to distinguish two
general classes of models (Eppen and Martin 1987), i.e. small bucket (SB) and big (or
large) bucket (BB) modelling approaches. In SB models, the planning horizon is divided into
a finite number of small time periods such that in each period either at most two products
can be produced, or there will be no production at all. Conversely, in BB approaches the
planning horizon is divided into longer periods, usually of equal length. In each period,
multiple products may be produced. As a consequence, SB models are usually associated
with short-term planning horizons and BB models with medium-term planning horizons.

50
 Vertical integration of lot-sizing and scheduling in food processing industry

3.2.1 Small bucket approaches

A typical example of SB approaches is the Discrete Lot-sizing and Scheduling Problem
(DLSP). The basic DLSP includes (sequence-independent) set-up costs and set-up carry-
over at zero set-up time (Fleischmann 1990). Inclusion of set-up carry-over implies that set-
up states of a machine are carried over between period boundaries. Porkka, Vepsalainen et
al. (2003) compare models with and without set-up carry-overs. They show that substantial
savings in costs and production time can be achieved by fundamentally different production
plans enforced by carry-overs. Comparable results are found by (Sox and Gao 1999).
However, in the basic DLSP, set-up states are not preserved over idle time. Sequence-
dependent set-up costs and times are neither considered in the DLSP. Many extensions of
the (basic) DLSP have been described in literature (Salomon, Kroon et al. 1991; Drexl and
Kimms 1997).

Fleischmann (1994) analyses the multi-item single machine DLSP with sequence-
dependent set-up costs. An artificial product is introduced to represent idleness of the
machine. Salomon, Solomon et al. (1997) continue the latter study and reformulate a DLSP
that captures sequence-dependent set-up times (DLSPSD). The triangular set-up
conditions are assumed to hold. However, machine idleness is represented by an artificial
product. Jordan and Drexl (1998) present a comparable model in which idleness is
indicated by an artificial product too. It should be mentioned that if idleness is represented
by an artificial product, the changeover matrix must fulfil very strict conditions to cope with
sequence-dependent set-up times. Otherwise the set-up state of the machine is not
correctly carried over across the boundaries of idleness.

Wolsey (1997) extended the work of Constantino (1996) for problems with sequence-
independent set-ups to formulations with sequence-dependent set-up times and costs.
Idleness is not represented by an artificial product. However, the triangular set-up
conditions are assumed to hold. We will refer to Wolsey’s model as the general small
bucket model (GSB).

3.2.2 Big bucket approaches

In contrast to small bucket models, the planning horizon of a big bucket (BB) model is
usually divided into longer periods of equal length. Time intervals in a BB model may
represent a time slot of one week (or more) in practice (Drexl and Kimms 1997). In each
period, multiple products can be manufactured. Releasing the “all-or-nothing” production
principle of (most) SB models implies that a BB model includes the possibility to determine
continuous lot-sizes.

The Capacitated Lot-Sizing Problem (CLSP) is a typical example of a big bucket model.
It is closely related to the (small bucket) DLSP. Decision variables, parameters and
objective function are comparable in both problems (Drexl and Kimms 1997). However, the

51
Chapter 3

CLSP does not include sequence-dependent set-up costs and times. As a consequence,
set-up carry-over between period boundaries is not included either. Suerie and Stadtler
(2003) use the simple plant location problem to obtain a tight new model formulation for set-
up carry-over in the CLSP with sequence-independent set-up costs and times.

Sox and Gao (1999) introduce the Generalized Capacitated Lot-sizing Problem
(GCLP). The GCLP uses less binary variables for including set-up carry-over in the CLSP
with sequence-independent set-up costs and no set-up times. Sequence-independent set-
up times may be included; probably at the expense of additional computational effort. The
authors also apply the network reformulation approach as proposed by Eppen and Martin
(1987) and compare the behaviour of a set of models. The results demonstrate that
incorporating set-up carry-over has a significant effect on both costs and lot-sizes.

We observe a tendency in simultaneous lot-sizing and scheduling to incorporate

characteristics of small bucket models into big bucket models. For confirmation we refer to
proposed variants of the CLSP (Gopalakrishnan, Miller et al. 1995; Gopalakrishnan 2000;
Haase and Kimms 2000; Gupta and Magnusson 2005; Almada-Lobo, Klabjan et al. 2007;
Almada-Lobo, Oliveira et al. 2008; Menezes, Clark et al. 2011), variants of hybrid BB and
SB models like the General Lot-sizing and Scheduling Problem (Fleischmann and Meyr
1997; Meyr 2000; Ferreira, Morabito et al. 2009; Transchel, Minner et al. 2011), and
variants of block planning approaches, originally introduced by Gunther, Grunow et al.
(2006). The literature review on extensions of capacitated lot-sizing by Quadt and Kuhn
(2008) confirm the trend in which BB approaches are preferred to SB models.

3.2.3 Product decay

Although a vast body of literature exists on inventory management for perishable products,
surprisingly little has been done to include product decay in traditional lot-sizing and
scheduling models. One of the first contributions in this area is provided by Soman, van
Donk et al. (2004b). The paper studies shelf life considerations in the economic lot
scheduling problem (ELSP). Models of this class usually assume constant demand, do not
account for sequence-dependent set-up times and aim to generate production cycles for
several products on a single resource. Entrup, Gunther et al. (2005) propose three MILP
models that integrate shelf-life issues into production planning and scheduling for an
industrial case study of yoghurt production. The models use a block planning approach in
which a block covers all products based on the same recipe. Shelf-life aspects are taken
into account by considering a shelf-life-dependent pricing component that may also include
inventory-holding costs. Lee and Yoon (2010) consider a coordinated production-and-
delivery scheduling problem that incorporates different inventory-holding costs between
production and delivery stages. The results may be only applicable to limited situations but
the study can be regarded as a first attempt to allow different (stage-dependent) inventory-
holding costs. Chen, Hsueh et al. (2009) conclude that papers discussing production

52
 Vertical integration of lot-sizing and scheduling in food processing industry

scheduling and/or distribution of perishable goods are relatively rare. Amorim, Antunes et
al. (2011) state that papers discussing simultaneous lot-sizing and scheduling for perishable
goods are very rare.

3.3 Model formulation

This section presents two SB models and demonstrates the impact of (i) relaxing the
triangular set-up conditions and (ii) taking product decay into account. Section 3.3.1
describes the problem which is then modelled in Section 3.3.2 as a lot-sizing and
scheduling problem with non-triangular set-ups. In Section 3.3.3 we extend the model such
that it addresses product decay of inventory by including an age-dependent component in
the inventory-holdings costs.

3.3.1 Outline of the lot-sizing and scheduling problem

The complete problem under study is characterized as follows:

- Consider N products to be scheduled over a finite planning horizon of T periods.

Consider a small bucket problem i.e. in each time period at most one item can be
produced at full capacity, or there is no production at all.
- A lot or batch of item i is defined as an uninterrupted sequence of periods in which
production takes place for item i.
- Manufacturing items requires a common equipment or resource with limited capacity.
Without loss of generality, machine capacity is normalised to 1 unit per period.
- Demand is assumed to be varying and deterministic and expressed in the number of
required production periods.
- Demand must be satisfied (without backlogging) either by production in the same
period or from stock. In the latter case, an inventory carrying cost is incurred.
- It is assumed that the Wagner-Whitin cost condition holds, i.e. given the set-ups, it
always pays to produce as late as possible. This condition is also referred to as the
absence of speculative motives for early production.
- Unit production costs are assumed to be constant over periods and are therefore
ignored.
- A changeover between items incurs a loss of production capacity (non-zero set-up
times) and associated set-up and changeover costs. Both are sequence-dependent.
- Set-up costs are assumed to be proportional to changeover times.
- The triangular set-up conditions refer to the changeover matrix A (with respect to set-up
times), the changeover matrix S (with respect to set-up costs), or both:

a ij ≤ a ik + a kj and/or s ij ≤ s ik + s kj for all items i, j, k (1)

53
Chapter 3

- If changeover and idleness occur in subsequent time intervals, we follow common

practice and assume that periods of idleness are preceded by changeover time.
- In contrast to the DLSP, the set-up state of the machine should be preserved over idle
time. This assumption is made in many lot-sizing and scheduling models (Drexl and
Kimms 1997). Generally speaking, it is also valid in FPI.
- Starting inventory levels for product i are assumed to be zero, i.e. I i ,0 = 0 for all i .

The objective is to determine the production sequence and lot-sizes that minimise the
sum of set-up and inventory carrying costs over the complete planning horizon.

3.3.2 Notation and model formulation

The following notation is used to formulate the problem:

Parameters
N number of items or products i.e. i , j = 1...N
T number of time intervals (i.e. periods) in the planning horizon; t = 1...T
hi,t unit storage costs of item i at the end of period t
St fixed set-up costs in period t
di,t demand for item i in period t (expressed in units of required capacity)
ai,j changeover time between products i and j in units of lost capacity

Variables

Ii,t inventory level of item i at the end of period t

1 if product i is produced in period t 

Yi ,t =  
0 otherwise 
1 if the machine is changing over to product j in period t 
V j ,t =  
0 otherwise 
1 if the machine is idle in period t and the next item to produce is product j 
W j ,t =  
0 otherwise 

Now, the problem can be formulated as follows:

 
Min 

∑∑j t
St V j ,t + ∑∑ i t
hi I i ,t 

(2)

s.t.

I i , t − 1 + Yi , t − d i , t = I i , t ∀ i, t (3)

54
 Vertical integration of lot-sizing and scheduling in food processing industry

∑Y i
i,t + ∑V
j
j,t + ∑W j
j,t =1 ∀ t (4)

t +τ −1
Yi ,t − ∑ ∑Y k ,s + Y j ,t +τ ≤ 1 ∀ (i , j ) ∈ A : = {(i , j ) | a i , j > 0 } (5)
k :k ≠i ≠ j s =t +1

t = 1,...,T − τ , τ = 1,..., a i , j

t −1
V j ,( t −ai , j +τ ) ≥ Yi ,( t −ai , j −1) + Y j , t + W j , t − ∑ ∑Y k, l −1 ∀ (i , j ) ∈ A : = {(i , j ) | a i , j > 0 } (6)
k :k ≠ i ≠ j l = t −aij

t = (a i , j + 2),...,T , τ = 0,..., (a i , j − 1)

∑W j
j ,t + ∑V j
j , t +1 ≤1 ∀ t = 1,...,T − 1 (7)

Wi , t + ∑Y j ≠i
j , t +1 + ∑Wj ≠i
j , t +1 ≤1 ∀ i , t = 1,...,T − 1 (8)

Yi , t ,V j , t ∈ {0,1} ∀ i, j , t (9)

I i , t ,W j , t ≥ 0 ∀ i, j , t (10)

Objective function (2) minimises the sum of changeover and inventory-holding costs.
Constraints (3) represent the inventory balance equations and assures demand di,t for item i
in period t is fulfilled without backlogging. Equations (4), together with (9) and (10),
guarantee that in each time interval the machine is either producing item i at full capacity
     



∑Y i ,t

 j   j  ∑
= 1 , changing over  V j ,t = 1 , or idle  W j ,t = 1 before manufacturing the next ∑
i    
batch of an item. Constraints (5) assures that between two subsequent production batches i
and j, sufficient time (ai,j) is reserved for changeover. For positive change over time (ai,j>0),
inequalities (6) enforce the set-up variables to be non-zero between two subsequent
batches i and j; if item j is produced in period t (Y j ,t = 1) or the machine is idle in period t
(W j ,t = 1) before manufacturing item j in period t’ (t’>t), set-up variables V j ,( t −ai , j +τ ) should get
t −1

∑ ∑Y in (6) represents the production

a value of one for τ = 0,..., (ai , j − 1) . The term
k :k ≠ i ≠ j l = t − aij
k,l

of another item than i and j within time interval [ t − a , t − 1] . Constraints (7) assure that i, j

periods of idleness are scheduled after a changeover. The inequalities (8) prevent that
idleness in period t before manufacturing item i is followed by the production (or preliminary

55
Chapter 3

idleness) of another item j. Finally, constraints (9) and (10) define the integrality and non-
negativity requirements.

3.3.3 Modelling product decay

When a traditional objective function like (2) is used for perishable items, traditional linear
holding costs in lot-sizing models may disrupt a crucial balance between changeover costs
on the one hand and inventory-holding costs on the other hand. Product decay has an
impact on the remaining shelf life of products. This aspect is included by an age-dependent
component in the inventory-holding costs (Entrup, Gunther et al. 2005).

Product decay of inventory can be incorporated in a SB type model like (2)-(10) in

the following way. Let additional parameter pri ≥ 1 represent the perishability rate of item i.
Next, we redefine the inventory variables Ii,t by I i ,t ,q ≥ 0 (for all i, t, q in which q ≤ t ) as the
inventory level of item i at the end of period t, produced in period q. Now, objective function
(2) is replaced by (2b) in which (t - q) represents product age:

 t 
Min 

∑∑ j t
St V j ,t + ∑∑∑
i t q =1
hi pri( t −q )I i ,t ,q 

(2b)

Note that if the perishability rate is pri = 1 for all items i, then objectives (2) and (2b) are
equal. Replacing (3) by (3a)-(3e) describes the age dynamics of the inventory levels:

I i , t ,q = Yi , t − d i , t ∀ i, t = 1, q = 1 (3a)

I i , t ,q ≥ I i ,t −1,q − d i , t ∀ i , t = 2...T , q = 1 (3b)

(Y τ − d τ ) − I
t
I i , t ,q = ∑
τ =1
i, i, i ,t ,q −1 ∀ i , t = 2, q = 2 (3c)

q q −1
I i , t ,q ≥ ∑
τ =1
I i ,t −1,τ − d i ,τ − ∑
τ
I
=1
i ,t ,τ ∀ i , t = 3...T , q = 2...t − 1 (3d)

t −1

∑ (Y τ − d τ ) − ∑ I
t
I i , t ,q = i, i, i ,t ,τ ∀ i , t = 3...T , q = t (3e)
τ =1 τ =1

3.4 Numerical illustrations and benchmark

This section illustrates the impact of (i) (relaxing) the triangular set-up conditions and (ii)
incorporating product decay of inventory on optimal production schedules. We use small
numerical examples to compare the behaviour and characteristics of the model with the
general small bucket model GSB from literature (Wolsey 1997). Model formulation

56
 Vertical integration of lot-sizing and scheduling in food processing industry

(2)-(10) is referred to as SB1, and formulation (2b), (3a)-(3e), (4)-(10) is referred to as SB2.
We first consider a tailored problem for N=3 and T=10.

3.4.1 Impact of (non-)triangular set-up and product decay

The impact of (relaxing) the triangular set-up conditions and modelling product decay by
age-dependent holding costs is demonstrated by three illustrative problem instances.

Example 1. Triangular set-ups, no product decay

Consider inventory-holding costs of (hi ) = (30, 25, 23) . The (nonzero) values of demand d i ,t
and changeover times a i , j (expressed in required periods of capacity) given by:

Let the set-up costs St = 100 for all t and the processing time be unitary for all items i. Note
that the triangular set-up conditions (1) hold for matrix A1. However, the matrix is
asymmetric. Table 3.3 shows the optimal production schedule for this instance.

Model formulations GSB, SB1, and SB2 with pri = 1 all obtain the same solution in Table
3.3 with total costs of 767 units. The set-up state of the machine is preserved over idle time
in period 7.

57
Chapter 3

Example 2. Non-triangular set-up, no product decay.

Increasing the changeover time a3,1 between item 3 and 1 to a3,1 > 2 results into
 0 1 2
 
changeover matrix A2 =  1 0 2  for which one of the triangular set-up conditions (1)
3 1 0
 
does not hold, as a3,1 > a3,2 + a2,1 . The optimal production schedule obtained by model SB1
and SB2 with pri = 1 is presented in Table 3.4. Obviously, the optimal production schedules
in the Tables 3.3 and 3.4 are substantially different.

Model formulation GSB generates again the production schedule of Table 3.3.
Clearly, this solution is infeasible: a changeover from item i =3 (in period 3) to item j =1 (in
period 6) requires 3 time intervals. Apparently, using model formulation GSB for cases in
which the triangular set-up conditions (1) do not hold, implies that set-up state changes will
occur (a changeover from item i =3 to item j =2 in period 4 and from item i =2 to item j =1 in
period 5) without a production change, i.e. no associated production for item 2.

Example 3 Non-triangular set-up and product decay.

We include product decay of inventory by using age-dependent holding costs: a
perishability rate pri = 1.12 is used for all items and applied to model SB2. All other data
from Example 2 remain unchanged, including changeover matrix A2.

Model SB2 yields a completely different production schedule as presented in Table

3.5. Apparently, a small change in the balance between inventory holding- and changeover
costs has a major impact on the generated production schedules, see Tables 3.4 and 3.5.

58
 Vertical integration of lot-sizing and scheduling in food processing industry

3.4.2 Complexity discussion

The time required to solve a MIP by a branch-and-bound approach depends heavily on the
way in which problems are formulated (Pochet and Wolsey 2006). We conduct a model
benchmark for the problem size of the formulations GSB, SB1 and SB2. Table 3.6 provides
a general overview. The general problem size is expressed in a common notation, i.e. the
dominating term of the number of constraints and variables. Table 3.6 shows that adding
the functionality to cope with non-triangular set-ups for simultaneous lot-sizing and
scheduling requires more constraints, see column “Constraints” for the rows GSB and SB1.
On the other hand, formulation SB1 requires substantially less (binary) variables than model
GSB. Nevertheless, the models GSB, SB1, and SB2 are all potentially very large
formulations. The impact of the time horizon on problem sizes for an SB-approach becomes
most obvious for model SB2 (adding product decay).

Complexity considerations for (variants of) the DLSP are published in Salomon,
Kroon et al. (1991) and Brueggemann and Jahnke (2000). If set-up times are ignored, it is
rather easy to test whether a feasible solution exists by comparing cumulative demands
(expressed in required capacity) with cumulative capacity. However, if set-up times are
considered, even the feasibility problem is NP-complete (Trigeiro, Thomas et al. 1989;
Salomon, Kroon et al. 1991). The latter indicates the need to develop effective and efficient
approximation techniques to generate good feasible solutions for larger problem
dimensions. Nevertheless, the availability of a correct model formulation offers the
possibility to measure the quality of heuristically generated solutions for small to medium
sized examples.

3.4.3 Heuristic approach

In a small bucket model like SB2, the number of time periods (t = 1…T) have a significant
impact on the problem dimensions and the complexity of the problem. Therefore, it is
obvious to focus on time-based decomposition approaches. The concept of the relax-and-
fix (R&F) heuristic (Dillenberger, Escudero et al. 1994; Stadtler 2003) is a typical example of
a time-oriented decomposition approach. The R&F belongs to the class of mathematical

59
Chapter 3

programming-based heuristics that use a mathematical programming (MP) procedure to

generate solutions. Heuristics of this class have the advantage that modules of many
commercial solvers can be used, possibly with some (minor) customisation. Moreover, the
R&F heuristic provides guidance for the assessment of the quality of generated solutions by
generating both a lower and an upper bound on the optimal objective function value. We
briefly describe a variant of the R&F approach. Numerical results are given in Section 3.5.

Outline Relax-and-Fix
The relax-and-fix algorithm solves sequentially P different mixed integer programming
problems, denoted by MIP p with 1 ≤ p ≤ P . The set of integer variables Q is partitioned into
P disjoint subsets Q1... QP. For model SB2, Q1 consists of all production and idle variables Y
and W associated with the time periods t := {1, ... , t1} , Q2 contains all binary variables
associated with the periods t := {t1 + 1, ...,t 2 } up to QP consisting of all binary variables
associated with the periods t := {t p −1 + 1, ...,T }. In all iterations the variables of a single
subset are defined as integers while all other variables in Q are either relaxed (i.e. defined
as continuous variables) or fixed to the (binary) values found in earlier iterations. We apply
a backward procedure, i.e. in the first iteration the subproblem MIP P is solved in which the
integrality restrictions are imposed on the variables in the subset QP. In other words,
production and idle decisions are only made within the window t = t P −1 + 1, ...,T .In the next
iteration, the integer variables in the subset QP are fixed at their optimal values as found in
iteration P. Next, problem MIP P −1 is solved to find integer values for the subset of binary
variables in Q P −1 and so on. In each problem MIP p , the production and idle variables are
fixed at their optimal values in earlier iterations. We do not apply a common forward
procedure in the R&F algorithm. As demand matrices for small bucket models are usually
sparse (i.e. many if not most entries of the matrix are zero), finding a feasible solution at the
end of the R&F algorithm strongly depends on the solution procedure (i.e. forward or
backward). In a forward procedure, production is postponed in early iterations. If capacities
are tight, the concept of fixing the production and idle variables at their optimal values from
previous iterations, will easily lead to infeasible solutions in case a forward procedure is
applied. It should be mentioned that (only) problem MIP p in the first iteration of the R&F
algorithm is a relaxation of the original problem. This automatically implies that the objective
value of problem MIP p in the first iteration of the R&F procedure provides a valid lower
bound on the optimal objective function value.

60
 Vertical integration of lot-sizing and scheduling in food processing industry

3.5 Numerical analysis

We investigate the computational behaviour of model SB2 and the potential value of a
backward Relax-and-Fix solution procedure. We compare the objective values reached and
the required computing time to run the solution procedures. All problems were solved using
Xpress-Mosel on a personal computer with an Intel Core i3 CPU, 2.13 GHz, RAM 4GB. The
results are given in Table 3.7.

61
Chapter 3

Test cases were generated from small to moderate size. Both the number of items N
and the number of time periods T are varied according to N = 3, 4, 5, 6 and
T = 30, 40, 50, 60, 90,120 . The perishability rate is set at pr = 1.1 and set-up-costs St = 500
for all t in all cases. The inventory-holding costs differ between items i but remain unaltered
between cases.

Demand is expressed in discrete units of production capacity and may occur at any
timeslot in the planning horizon. Demand on a timeslot may be larger than a single unit of
production capacity. The changeover matrix is asymmetric, sequence-dependent and such
that the triangular set-up conditions (1) do not hold. All changeover times ai , j > 0 for i ≠ j .

Table 3.7 shows the results for a set of instances solved by model SB2. The first
column indicates the problem size, i.e. the number of items N and the planning horizon T.
Columns 2 and 3 refer to the objective function value of the optimal solution (Objv) and the
time needed to find the solution (CPU). The results in column 3 confirm that finding optimal
solutions requires high computation times for medium size instances. The search procedure
for optimal solutions was interrupted after eight hours of CPU time, provided that for each
value of N = 3, 4, 5, 6 at least three test cases were solved to optimality. An interrupted
solution procedure is indicated by entry ‘−’ in Table 3.7. The columns 4 to 6 refer to
solutions found by the Relax-and-Fix (R&F) heuristic, i.e. the objective function value found
in the final iteration (Objv_RF), the relative deviation between the optimal objective value
and the objective value of the R&F solution (GAP), and the computation time needed to find
the R&F solution (CPU_RF). For all instances in Table 3.7, the R&F heuristic performed
remarkably well: for 10 out of 18 cases the optimal solution was found, and time savings
were substantial.

3.6 Concluding remarks

This paper studies how to include realistic features of food processing industry (FPI) in
models for simultaneous lot-sizing and scheduling, in particular (i) sequence-dependent set-
ups (including non-triangular set-ups), and (ii) product decay of inventory due to
perishability.

Although big bucket (BB) models are usually associated with medium-term planning
horizons, various extensions of these models are proposed for short-term planning horizons
too. From a computational point of view it is explainable to prefer models with large time
intervals over small bucket (SB) approaches. However, we state that segmentation of the
planning horizon is a key issue for simultaneous lot-sizing and scheduling, particularly in
food processing industry. Using large time periods implies that some basic principles for lot-

62
 Vertical integration of lot-sizing and scheduling in food processing industry

sizing and scheduling (unfoundedly) disappear from sight. We give two reasons to underpin
the latter statement.

Firstly, if the objective for simultaneous lot-sizing and scheduling should include the
best compromise between total set-up costs and total inventory holding costs, a time-
oriented aggregation (e.g. in BB models and its variants) may easily disrupt the general
principle of optimality for lot-sizing (Pochet and Wolsey 2006). If time intervals represent
long periods (e.g. a week or more) multiple batches can be produced in each period.
Consequently, inventory costs for batches manufactured at the start of lengthy periods are
assumed to be equal to inventory costs of lot sizes produced at the end of the same period.
As a result, total inventory costs in (2) or (2b) are underestimated, the main principle of
optimality for lot-sizing may become disrupted, and production schedules will change
accordingly. Secondly, product decay in food processing industry is primarily associated
with the “age” of products. Incorporating issues of perishability like product decay requires
capturing the precise moments of production for manufactured products.

We developed an SB model that (i) can handle sequence-dependent set-ups

(including non-triangular set-ups), (ii) addresses product decay by incorporating age-
dependent holding costs. Small-scale examples are used to demonstrate the impact of non-
triangular set-ups and product decay on the generated solutions. The models show how a
small change in the balance between inventory holding- and changeover costs may
generate significantly different solutions, especially when the triangular set-up conditions do
not hold.

As expected, the computational effort for the SB model is substantial. We performed

exploratory research with a straightforward implementation of a Relax-and-Fix heuristic.
Numerical tests show that the quality of R&F solutions is promising at manageable
computational effort. Additional research is needed to find more enhanced variants of the
R&F and an effective (algorithmic) segmentation of the time horizon, e.g. by exploiting the
sparse demand matrix in a SB approach. Smoothing the heuristic solution by creating some
overlap between successive iterations may be an option for further research (Pochet and
Wolsey 2006). We refer to Escudero and Salmeron (2005) and (Federgruen, Meissner et al.
2007) for an overview of various strategies in an R&F framework.

Another interesting question for further research is to improve computational

performance by adding valid inequalities (VI’s) a priori to the initial formulation. Preliminary
research revealed that adding the inequalities Y j ,t + W j ,t ≤ Y j ,t −1 + W j ,t −1 + V j ,t −1 ∀j , t
tightened the linear relaxation and provided substantially better heuristic solutions in an
R&F framework. The inequalities are based on feasibility conditions within any production
scheme and are added before calling the MP-solver. The proposed model offers a unique
possibility to measure the quality and performance of any (heuristic) approach.

63
Chapter 3

64
 Chapter 4

Integrated planning between procurement

and production

Simplicity always beats complex optimization (Little 2004)

This chapter is based on:

Claassen, G.D.H., and Hendriks, Th.H.B., (2007)

An application of Special Ordered Sets to a periodic milk collection problem
European Journal of Operational Research
Vol. 180 (2), pp. 754 – 769.

Claassen, G.D.H., and Hendriks, Th.H.B., (2007)

Modelling techniques for (non)linear and integer programming
In: Decision Science, Theory and applications, Ch. 9
Eds: G.D.H. Claassen et al.
Wageningen Academic Publishers, pp. 185 – 226
Chapter 4

Abstract
This chapter focusses on the need for integrated decision support between procurement
and production in the Supply Chain Planning Matrix (SCPM), particularly important across
organizational borders. We take different cooperative organizations in dairy industry as a
starting point and focus on the impact of (dominant) batch type production processes on
push-oriented sourcing of perishable raw materials. Various raw milks are processed in
different batches on shared resources. Capacities of special storage facilities are limited on
supply, distribution, and processing level. The study intends to develop and test a pilot DSS
in practice, particularly helpful for non-dominant partners in a food supply chain. We present
an OR-based approach to support milk collection in a special branch of dairy industry. The
annual growth of the sector and the continuous imbalance between milk supply and
demand, has urged the sector to look for a different approach to their daily milk collection
problem. Specific details of the problem environment (i.e. the continuous production on
supply level and the delivery conditions on demand level) gave rise to choose for a short- to
medium-term planning approach. The proposed decision support system has to be
considered as an efficient tool for generating stable milk collection plans which in turn
serves as an effective starting point for the vehicle routing problem. From a computational
point of view it turned out that the application of Special Ordered Sets type 1 (SOS1) was
very useful. Although it appears from literature that the computational advantage of SOS1 is
restricted to supplementary model conditions, this study shows that these conditions are not
necessarily needed.

66
 Integrated planning between procurement and production

4.1 Introduction
The introduction of the so-called milk quotation system for cow’s milk in 1984, implied a
strong stimulus for the annual growth of milk goats for professional use in the Netherlands.
The continuous growth since 1984 was intensified by the favourable profit for the production
of goat’s milk on a biological and professional scale. Nowadays, Dutch goatherds for
professional use produce about 40 million litres of milk yearly. The main part is used for
domestic cheese production but export of fresh milk to Belgium, Germany and the UK is
also quite common. The remainder of the supply is sold at an unattractive price level for
food (milk powder) of young animals and dairy concentrates.

The annual growth of the sector and the increasing imbalance between the
continuous production of milk on supply level and the delivery conditions of dairy factories
has urged the sector to look for a different approach to their milk collection problem. Milk
collection in this sector is usually not set up by processing industry or by (local)
transportation companies. Individual farmers are mostly united in cooperative associations.
All costs and profits are shared with the members of the association. The (yearly)
negotiations with processing industry about the expected amounts of milk to deliver, the
delivery days, the selling prices, the contracts with local transportation companies and the
(daily) construction of milk collection schedules are all covered by the cooperative
association. This study has been done for one of the largest cooperative associations in
goatherd industry in the Netherlands. The main questions were:

1. How to support the vehicle schedulers of the cooperative association in their daily
job to build milk collection schedules such that the (financial) interests of the farmers
are served as well as possible. It is significant to realize that collecting milk
effectively from the supplier farms and deliver the milk to the different parties on
demand level is not restricted to solving vehicle routing problems. Common dairy
factories call for large amounts of raw material and their demand for goat’s milk is
scheduled to arrive at a very limited number of fixed days. As a consequence, the
milk collection schedules should balance milk supply and demand such that the raw
material can be sold at the best possible price level.

2. How to support other decision-makers of the cooperative association in their (yearly)

negotiations with processing industry, (third party) transportation companies and
suppliers.

Although a lot of literature, e.g. (Toth and Vigo 2002; Ghiani, Guerriero et al. 2003;
Gayialis and Tatsiopoulos 2004), has been dedicated to (the application of) vehicle routing
problems and even on milk collection problems in common dairy industry (Basnet, Foulds et
al. 1996; Gerdessen 1996; Butler, Williams et al. 1997), the collection problem for goat’s
milk is characterised by rather specific details. The routing aspect could be classified as a

67
Chapter 4

Periodic Vehicle Routing Problem (PVRP). The PVRP is an important generalization of the
classical Vehicle Routing Problem (VRP). The VRP consists of constructing delivery routes
for a fleet of vehicles at minimum costs. The capacity of each vehicle is fixed and may not
be exceeded. Moreover, each vehicle must return to its departure site. Customers have a
known demand that must be fully satisfied. Each customer is visited exactly once by a
single vehicle (Cordeau, Gendreau et al. 1997). There may be constraints that limit the
distance travelled by each vehicle (Chao, Golden et al. 1995; Cordeau, Gendreau et al.
1997). Typically, the planning period is a single day. The VRP is a hard combinatorial
problem that received a great deal of attention in literature. Usually the problem is tackled
by means of heuristics (Cordeau, Gendreau et al. 1997).

The PVRP generalizes the classical VRP by extending the planning period from a
single day to T days (Chao, Golden et al. 1995). Over this planning horizon the clients are
not to be served on a daily bases, but are characterized in terms of some sort of periodicity
of the demand. Each customer i on demand level specifies a service frequency by a set of
allowable combinations of visit days (Cordeau, Gendreau et al. 1997). Each customer must
be visited at least once but some of them require several visits during the T-days period.
Now the problem consists of simultaneously selecting a visit combination for each customer
and establishing vehicle routes for each day of the planning horizon according to the VRP
rules as outlined above. An integer programming formulation of the PVRP can be found in
(Gaudioso and Paletta 1992). The periodicity of demand implies that it is not possible to
solve the problem on a daily bases and subsequently replicate the solution over time. Chao,
Golden et al. (1995) classifies the PVRP as a multi-level combinatorial optimization
problem. At the first level it is necessary to assign an allowable visit combination to each
customer. At the second level a classical VRP (i.e. assigning vehicles to routes) for each
day of the planning period should be solved. At the third level, a classical Travelling
Salesman Problem (TSP) should be solved. As the TSP has been shown to be NP-hard,
the PVRP is at least as difficult (Chao, Golden et al. 1995). Within this context it is hardly
surprising that most papers on the PVRP reported in literature present heuristic methods.
See for example (Gaudioso and Paletta 1992; Chao, Golden et al. 1995; Cordeau,
Gendreau et al. 1997).

A review of solution approaches for the PVRP can be found in Chao, Golden et al.
(1995). Practical applications of the PVRP are for example in grocery distribution (Golden
and Wasil 1987) but in Chao, Golden et al. (1995) more areas of application can be found.
Many efforts in the literature have been established to extend the basic PVRP model to
incorporate additional constraints or different objectives. However, at our knowledge the
PVRP assumes either pickup or delivery operations, not both. In other words: it either
concerns the construction of pickup routes for raw material(s) from several suppliers to a
single manufacturer or the construction of delivery routes from a single supplier (for
example a warehouse) to several customers. Typically in a PVRP, suppliers or customers
are characterized by some kind of periodicity of visiting days over a T-day planning horizon

68
 Integrated planning between procurement and production

and their geographically dispersed locations. The collection problem also concerns the
construction of routes over a T-day period but in this case both the suppliers and the
customers specify a set of allowable combinations of visit days. Although customers specify
a service frequency, it is not necessary to satisfy the periodic demand completely for every
customer. Moreover, customers may be visited by more than one vehicle from different
routes. Finally, the problem is even more complicated by keeping qualities of the raw
material. In contrast to the common PVRP, emphasis is not at first towards routing costs or
fleet size but towards fitting and balancing milk supply and demand by assigning allowable
visit combination simultaneously to farmers and customers. The goal of this research is
twofold.

First we discuss how an OR-based approach, the related optimization techniques,

structured data queries and additional analysis tools can support a specific milk collection
problem such that several, mostly conflicting, goals of the relevant players (i.e. farmers,
processing industry and transporters) are taken into consideration. Analogous to the PVRP,
the most critical decision is to assign allowable visit combination to farmers and customers,
once this is done the daily routing of vehicles is relatively straightforward. For assigning
allowable and stable visit combinations we propose a mixed integer linear programming
model that is solved by applying the concept of Special Ordered Sets type 1 (SOS1),
introduced by Beale and Tomlin (1970). Although Williams (1990) stated that there is a
great computational advantage to be gained in the SOS-formulation, the questions “why,
when and how to apply Special Ordered Sets of type 1”, have not got much attention in
literature yet.

The second goal of this study is to contribute to the insights of an effective use of
special ordered sets of type 1. From a theoretical point of view we prove that there is no
advantage in branching on sets of variables by using the SOS1 concept or branching on
individual (integer) variables in a commonly applied branch-and-bound procedure. We show
that the efficiency of the SOS1 formulation strongly depends on the ordering of the
variables within each set.

The remainder of this chapter is organized as follows. In the next section, we

describe the problem environment in more detail and focus on the main differences
between milk collection problems for cow’s - and goat’s milk. In Section 4.3 we present a
model formulation that turns out to be quite hard to solve for real sized problems. In Section
4.4 we focus on an efficient use of the SOS1-formulation and the related computational
performance. In the last section we finalize by a general discussion and some concluding
remarks.

69
Chapter 4

4.2 Problem description

Although the yearly supply of goat’s milk is of minor importance for the Dutch dairy industry,
the market for the related end products is growing annually. The exclusive end products,
mainly cheese, are processed by a limited number of dairy factories. Actually, just a few
factories are processing goat’s milk and their production capacity is mainly based on
processing large quantities of common cow’s milk. As a consequence, set-ups for
processing goat’s milk on demand level are usually restricted to one or at most two days
weekly for every factory. However, in view of meeting all predefined quality standards, the
freshness or “age” of the milk at arrival time, is of major importance. This implies that the
raw material has to be collected before the “age” of the (oldest) milk exceeds three days.
This time restriction is fixed and independent of the final destination of the raw material,
inside or outside the Netherlands. So on the one hand, dairy factories call for large amounts
of raw material and their demand is only scheduled to arrive at some fixed days. On the
other hand, looking at supply level, the number of goat’s for professional use are small
compared to common dairy farms, the average milk production yields on individual farms
are substantially less and goat’s farms are geographically spread over the country. So, from
a transportation point of view the complexity of the collection problem for goat’s milk is quite
different from collecting cow’s milk. Especially if we take into account that (cooled) storage
of milk at supply level is restricted to at most three days and the dairy factories only take
delivery of goat’s milk at a small number of fixed days. This in turn enhances the problem
that the transported amount of milk between the supply- and demand level is often out of
balance with the capacity of modern transportation vehicles.

These conflicting interests, together with the annual growth of the sector, urged for a
different approach of the daily milk collection problem in the goatherd sector. It raised the
question to develop an interactive planning tool in order to support the milk collection
problem and attune the imbalance between milk supplies on the one hand and the
individual demand levels of dairy factories on the other hand. The system should have a
major focus on constructing stable, short- to medium-term milk collection and delivery plans
rather than solving the daily VRP (i.e. assigning vehicles to routes) and subsequently, the
Travelling Salesman Problem TSP for each vehicle.

4.3 Model formulation

Part of the system is based on a mixed integer linear programming model. This model takes
a (rolling) planning horizon of two weeks into account. Milk supply and demand is exactly
known for two weeks in advance. Individual farms are clustered to larger entities. This
grouping is primary based on the geographical location of the farms and the available
quantity of milk within a cluster. The main idea is that within each cluster the entire milk

70
 Integrated planning between procurement and production

production will be collected at days, i.e. collection rhythms, still to be determined within the
planning horizon (see Table 4.1).

Table 4.1: some feasible milk collection rhythms; to be repeated every two weeks

Week 1 Week 2

Rhythm Mo1 Tu1 We1 Th1 Fr1 Sa1 Su1 Mo2 Tu2 We2 Th2 Fr2 Sa2 Su2
1
2
3
:
r
:
R

The available amount of milk after one, two or three days should match with the different
carrying capacities of (several) transportation vehicles. A surplus of milk at supply level can
be sold at an unattractive price level to a selected number of surplus companies. Now the
question is not only which cluster should be visited but also when to visit the farms in a
cluster such that the allowed visit days and delivered quantities on demand level are
satisfied as well as possible. In order to meet the most important quality standards of the
collected milk, the period of time between two consecutive visits within a cluster should not
exceed three days. In fact this quality constraint means that the potential number of milk
collection schemes or rhythms, for a two weeks planning horizon, is finite and limited (see
Table 4.1). Collecting milk at Sundays is not allowed. After the milk collection rhythms are
chosen they will be repeated every two weeks. The introduction of these so-called milk
collection rhythms reduced the complexity of the problem considerably. After all, the
problem is now which milk collection rhythm should be assigned to each cluster such that
the individual visit days demand levels are served as well as possible. This problem can be
formulated as a mixed integer linear programming model. Suppose we define:

Indices
c = 1 .. C ~ the different clusters at supply level
b = 1 .. B ~ the different buyers or factories at demand level
r = 1 .. R ~ the available milk collection rhythms
t = 1 .. T ~ the relevant days of the planning horizon

Data
Sc,r,t ~ the milk supply in cluster c on day t according to milk collection rhythm r

71
Chapter 4

Db,t ~ the demand of milk for buyer b on day t

Pb+ ~ penalty for every unit of milk delivered more than the actual demand of
buyer b
Pb− ~ penalty for every unit of milk delivered less than the actual demand of buyer
b

Variables
xc,r,b,t ~ delivered amount of milk from cluster c, at rhythm r for buyer b on day t
+ -
xd b,t , xd b,t ~ surplus or shortage of milk at demand level (buyer b) on day t
Yc,r ~ binary variable in order to assign milk collection rhythms to clusters

Model formulation

Min  ∑ ∑(Pb+ ⋅ xd b+,t + Pb− ⋅ xd b−,t ) (1)

b t 
∑ y c,r = 1 ∀c (2)
r

∑ x c ,r ,b,t ≤ S c,r ,t ⋅ y c ,r ∀ c, r , t (3)

b

∑ ∑ x c,r ,b,t − xd b+,t + xd b−,t = Db,t ∀ b, t (4)

c r

y c,r ∈ {0,1} ∀ c, r (5)

x c ,r ,b,t , xd b+,t , xd b- ,t ≥ 0 ∀ c, r , b, t (6)

The objective function (1) minimizes the total weighted sum of deviations on demand level.
Especially the penalty coefficients Pb+ ∀ b (surplus) are important in order to weight any
amount of milk delivered at an unattractive price level to (i) the subset of surplus companies
or to (ii) buyers that accept deliveries above their contractual maximum amounts. The
constraints in (2) ensure that exactly one milk collection rhythm will be assigned to every
cluster of farmers. The equations in (3) are classical logical conditions between the
continuous variables at the left-hand side and the binary variables at the right-hand side.
They imply that no milk can be transported from a cluster on a day to any buyer if it is not in
accordance with the chosen milk collection rhythm. Moreover, the equations in (3) ensure
that the total amount of milk to be transported from a cluster to the buyers may not exceed
the available quantity on supply level. The equations in (4), together with the objective
function, ensure that demand levels of all buyers are (more or less) satisfied. The deviation
between the delivered amount of milk and the actual demand level is expressed by the
auxiliary variables xd b+,t (surplus) and xd b−,t (shortage).

72
 Integrated planning between procurement and production

4.4 Solving the model

Despite of the limited number of both the predefined milk supply clusters C and the milk
collection rhythms R, the problem turned out to be disappointingly hard to solve. In most
cases practice defined ten to twelve different milk supply clusters. An (arbitrary) upper
bound of ten cpu-minutes for solving a problem is already reached at six potential milk
collection rhythms.

Instead of defining C * R different binary variables and subsequently branch on

individual variables in a branch-and-bound (B&B) tree, the integrality constraints (5) can be
relaxed and it is possible to apply the concept of special ordered sets type1 (SOS1),
introduced by Beale and Tomlin (1970). An SOS1 is defined as a set of variables within
which at most one variable may be non-zero. In this case we defined for each milk supply
cluster c, an SOS1 set S1c of continuous variables such as:

S1c : = {Yc,1, Yc,2, …, Yc,R } together with the conditions at most one
of the variables within this set can be non-zero ∀c (5.1)

Note that it is not necessary to treat the variables Yc,r in (5) as binary variables since the
S1c-conditions in (5.1) together with the constraints in (2) ensure that within each S1c-set
exactly one continuous variable will get a final value of one. As an alternative to define the
variables Yc,r as 0−1 integers ∀ c, r in (5), it is convenient to consider each S1c -set as a
discrete entity or generalisation of a 0−1 variable.

Conditions (5.1) can be dealt with algorithmically through the method of integer
programming (Williams 1993). Treating each set as an entity makes it possible to branch in
a branch-and-bound (B&B) algorithm on entities rather than on individual variables. The
non-zero variable in each feasible S1c -set of (5.1) will lie either to the left, or to the right, of
any marker placed between two consecutive variables within a set. So:

either {Yc,1, Yc,2, …, Yc,j } are all zero

or {Yc,j+1, Yc,j+2, …, Yc,R } are all zero

These two possibilities correspond to a branch in a solution tree as demonstrated in Figure

4.1 in which Pc,k is defined as a subproblem P for a S1c -set in node k of the search-tree
(Williams 1993).

73
Chapter 4

Pc,k

{yc,1, yc,2, …, yc,j } all zero {yc,(j+1), …, yc,R } all zero

Pc,(k+1) Pc,(k+2)

Figure 4.1: The branching procedure in a SOS1 search-tree (Williams 1993)

For any node in the search tree, for example, problem Pc,(k+1), one of the following situations
holds:
- problem Pc,(k+1) is infeasible which implies that the search-tree stops below node Pc,(k+1).
- problem Pc,(k+1) is feasible. Now, two possibilities are left:
{ }
• the subset y c,( j +1) ,…, y c,R is feasible, i.e. at most one of the variables in the
set is non-zero. If the objective function value w for problem Pc,(k+1) is better
than the best bound wb so far, the value for wb is updated (wb := w). The
search tree stops below node Pc,(k+1).
• { }
the subset y c,( j +1) ,…, y c,R is infeasible i.e. at least two variables in this
subset are non-zero. If the objective function value w for problem Pc,(k+1) is
worse or equal to the bound wb found so far, the search tree stops below
node Pc,(k+1). If the value of the objective function w for problem Pc,(k+1) is
better than the best bound wb, the branching procedure is to be continued on
{ }
the subset y c,( j +1) ,…, y c,R . Note that in any node below problem Pc,(k+1), at
{ }
least the variables y c,1 ,…, y c, j are all zero.

In Appendix 4.1 of this chapter we prove by means of complete induction that the upper
bound B for the number of branches B, in case of C different S1-sets (milk supply clusters)
and R different milk collection rhythms, is defined by:

C
B= ∑R
c =1
c -1
(2R - 2) (7)

Increasing the number of clusters C will have a larger impact (exponentially) on the
potential size of the search-tree than the number of milk collection rhythms R. Using
complete induction (see Appendix 4.1) we can also proof that the potential number of
branches B for a common B&B approach, i.e. branching on individual binary variables Yc,r

74
 Integrated planning between procurement and production

for problem (1) to (6), is also equal to (7). However, from literature it appears that there is a
great advantage to be gained in the SOS-formulation, provided that the variables within the
sets have a so-called natural ordering (Williams 1990).

As the upper bound B of an SOS-based search tree is equal to a ‘conventional’

branch-and-bound tree (B&B-tree), any computational advantage of the SOS-formulation
must be based on finding strong and/or early bounds in the search-tree. For that purpose
(Williams 1990) adds that the variables should have a natural ordering within the sets.
Unfortunately, in our case the variables (i.e. the milk collection rhythms) within the sets can
hardly be ordered in a natural way. In the following, we primary focus on the concept of
finding early bounds in a SOS1-based search-tree. Next we will present an alternative
procedure for ordering the variables within the sets in case there exists no natural ordering.

If more than one variable in (5.1) takes a non-zero value, the S1-set is infeasible. In
order to measure this infeasibility analogous to the fractionality of an integer variable, the
variables in each set of (5.1) have to be associated with a monotonic, increasing or
decreasing, set of numbers (a1, a2, …, aR) known as the reference row (Beale and Tomlin
1970; Williams 1990). In the formulation of some applications this set of numbers arises
from a constraint. In case these constraints are not present, the index numbers can be used
in order to associate each variable with its place in the ordering, so a1 =1, a2 =2, …, aR =R.
Now, the fractionality of an infeasible S1c-set in any node of the B&B-tree, can be
calculated as follows (Williams 1993):
R

∑a r
~
y c,r
r =1
R ∀c (8)
∑y
r =1
~
c,r

~
In which y c ,r denotes the solution value of the variables in the current node of the B&B-tree.
Since the numbers ar are monotonic, there will be some ar such that

∑a y r
~
c ,r
ar ≤ r =1
R < ar +1 ∀c
∑
r =1
~
y c ,r

(9)

indicating that the “centre of gravity” of the set has come out between the index r and r+1
(Williams 1993). If the set is infeasible the branching marker will be placed between the
variables Yc,r and Yc,r+1.

75
Chapter 4

Now the problem is how to order the (continuous) variables Yc,r within every S1c-set
such that the branch and bound (B&B) procedure can be executed more efficiently than in
case of branching on the individual (binary) variables Yc,r in problem (1) to (6).

Obviously, finding strong bounds in an early stage of the B&B procedure will have a
significant effect on the efficiency of a B&B algorithm. However, a general strategy for
strong bounds may be hard to find. Nevertheless, we could try to set up the branching-tree
in such a way that the chances for fathoming large(r) parts of the search-tree in an early
stage of the B&B algorithm are increasing. Within this context we will focus on a sorting
procedure for the individual variables within the S1c-sets. According to (9) the position of the
branching marker in an infeasible S1c -set depends both on the values for a1, a2 ,…, aR in
the reference row and on the position of the non-zero variables within the set. Within this
study the reference row itself remains unaltered, so a1 =1, a2 =2 ,…, aR =R. If the actual
position of the decision variables Yc,r within a set is such that the corresponding non-zero
~
variables y c,r of an infeasible S1c -set will be located on the left-hand (or right-hand) side
within a set, the position of the branching marker will be placed in the same area. As a
result, the subsets corresponding to each of the branches in Figure 4.1 will be unequal in
size. This in turn means that the potential depth of the branch related to the largest subset
will be less than the depth of the opposite branch. So, it is likely to expect that the chances
for finding an early solution (i.e. bound) will be larger in a node beneath the branch on the
largest subset. After all, according to the constraints in (2), every S1c -set has to be feasible
in the end. Note that for all potential milk collection rhythms a feasible solution for problem
(1) to (6) can be found.

Next, we focus on an ordering procedure for the decision variables Yc,r within the S1c
~
-sets such that the value of the corresponding non-zero variables y c,r of an infeasible set
will be located on the left-hand (or right-hand) side within the set. Within this ordering
context it is convenient to define some measure of performance for each milk collection
rhythm r on supply level. Suppose we define a parameter DSc,r for every decision variable
Yc,r within an infeasible S1c-set. The value of these parameters should be regarded as an
heuristic fit for applying milk collection rhythm r in cluster c (supply level) to all needs on
demand level. The value for DSc,r is defined as:

T
 B 
DSc,r = ∑ ∑  Db,t − Sc ,r ,t ⋅ y~ c,r


 ∀r in all infeasible S1c - sets . (10)
t =1  b =1 

Now, the actual position from r=1 to R of the variables Yc,r within an infeasible S1c-
set is based on an increasing (or decreasing) value for DSc,r . So, in case of an increasing
ordering for DSc,r , the corresponding (non-zero) variables Yc,r in the linear programming
relaxation (LP-relaxation) will be placed on the left-hand side in the set and vice versa (right

76
 Integrated planning between procurement and production

side) for a decreasing ordering. These ordering strategies will be called S1_LEFT and
S1_RIGHT respectively. The strategy in which the corresponding (non-zero) variables Yc,r of
the lowest values for DSc,r are placed in the middle of the S1-sets, is called S1_MID.

~
Suppose for a cluster c=i the solution values y i ,r of the LP-relaxation for r =1,..,6 are:
~ 
 y i ,1 = 0 , y i ,2 = 0.4 , y i ,3 = 0 , y i ,4 = 0 , y i ,5 = 0.6 , y i ,6 = 0  then the ordering of the variables
~ ~ ~ ~ ~

 
Yi,r for each strategy within the S1i-set is given in Table 4.2. The arrows below the sets
denote the position of the branching marker which values are calculated by (9). The
branching marker defines the two subsets in the SOS1 branching procedure (see Figure
4.1).

Table 4.2 Fictitious example of the ordering strategies within the S1c -set

DS i ,r for r = 1,..6 DSi ,1 = 10 , DSi ,2 = 6 , DSi ,3 = 9 , DSi ,4 = 8 , DSi ,5 = 4 , DSi ,6 = 8

Ref. row {1, 2, 3, 4, 5, 6}

{y i ,5 } {y i ,2 , y i ,4 , y i ,6 , y i ,3 , y i ,1}
S1_LEFT ↑
1.4

{y i ,1 , y i ,6 , y i ,2 } {y i ,5 , y i ,4 , y i ,3 }
S1_MID ↑
3 .6

{y i ,1 , y i ,3 , y i ,4 , y i ,6 , y i ,2 } {y i ,5 }
S1_RIGHT ↑
5 .6

The experiments related to the impact of the number of milk collection rhythms on
the computational effort, are summarised in Figure 4.2. Every marked point in Figure 4.2
represents the average result for three different cases in which individual farms were
grouped into ten different clusters. This grouping remained unaltered between the cases.
The curves represent four different strategies. One of the curves (BIN) is based on a
common B&B approach, i.e. branching on individual binary variables Yc,r for problem (1) to
(6). All other curves are related to the application of different SOS1 branching strategies as
discussed before and demonstrated in Table 4.2.

77
Chapter 4

As expected, Figure 4.2 shows that the computational effectiveness of the S1_LEFT
and S1_RIGHT strategy are comparable. From a computational point of view these
strategies are much better than branching on individual binary variables in a common B&B
approach (the BIN-curve) or the S1_MID strategy. The difference in effectiveness of the
S1_LEFT or S1_RIGHT strategy on the one hand and the BIN or S1_MID strategy on the other
hand becomes more evident in case the number of milk collection rhythms increases.

BIN S1_MID S1_LEFT S1_RIGHT

80

70

60

50
CPU-time (min.)

40

30

20

10

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Number of rhythms
(10 clusters)
Figure 4.2 The impact of the number of collection rhythms on the calculation time

The impact of the number of clusters on the computational effort is demonstrated in

Figure 4.3. The observations in Figure 4.3 are all based on a single case in which we
defined seventeen different milk collection rhythms for an increasing number of clusters on
the horizontal axis. Although there is still a difference in effectiveness between the BIN (or
S1_MID) and the S1_LEFT (or S1_RIGHT) strategy, the computational advantage of the

78
 Integrated planning between procurement and production

latter strategies is less beneficially for an increasing number of clusters. This observation
becomes obvious if we take into account that only the number of milk supply clusters C will
affect the number of global entities (i.e. the S1c-sets) in an SOS1-based B&B algorithm. The
number of milk collection rhythms R, mainly affect the size of each global entity or S1c-set.
Moreover, in equation (7) we already showed that increasing the number of clusters C will
have a larger impact (exponentially) on the potential size of the search-tree than a
comparable increase of R (i.e. the number of milk collection rhythms).

BIN S1_MID S1_LEFT S1_RIGHT

80

70

60

50
CPU-time (min.)

40

30

20

10

0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Number of clusters
(17 rhythms)

Figure 4.3 The impact of the number of clusters on the calculation time

As mentioned before, the S1_LEFT (or S1_RIGHT) strategy aims to affect the position
of the branching marker in an S1c-set such that the two branching subsets in Figure 4.1 will
be split efficiently into two subsets of unequal size. The main idea is to find early solutions
(i.e. bounds) in nodes beneath the largest subset. We already found a theoretical upper
_
bound B for the number of branches B in case there are C different S1c-sets and R

79
 Chapter 4

_
different milk collection rhythms. This upper bound B has been calculated for a case in
which we defined 10 milk supply clusters C and an increasing number of (different) milk
collection rhythms from R= 3 to R=17 in the second column of Table 4.3. The columns 3 to
6 in Table 4.3 represent the explored number of nodes at the end of the B&B algorithm for
each strategy. In the columns 7 to 10 the explored number of nodes are expressed as a
_
percentage of the upper bound B in the second column. The results of column 9 (S1_LEFT)
and 10 (S1_RIGHT) show that the latter strategies are more effective than the S1_MID
strategy in case the S1-setsize increases.

Table 4.3 efficiency B&B search procedure per strategy for C = 10 and increasing R

Total nodes explored Total nodes explored as a percentage of B

R S1_ S1_ S1_ S1_ S1_ S1_

B
BIN MID LEFT RIGHT BIN MID LEFT RIGHT
3 1.18E+05 1040 1291 1382 1408 0.8806395 1.0931784 1.1702344 1.1922504

4 2.10E+06 4062 10611 2616 3234 0.1936914 0.5059724 0.1247407 0.1542093

5 1.95E+07 13615 18668 4726 4761 0.0697088 0.0955802 0.0241971 0.0243763

6 1.21E+08 17285 41628 7142 4345 0.0142931 0.0344226 0.0059058 0.0035929

7 5.65E+08 18316 43478 8457 8245 0.0032421 0.0076959 0.0014969 0.0014594

8 2.15E+09 20559 45287 8084 6077 0.0009574 0.0021088 0.0003764 0.0002830

9 6.97E+09 20272 53201 10415 8124 0.0002907 0.0007629 0.0001493 0.0001165

10 2.00E+10 25404 47496 8924 1182 0.0001270 0.0002375 0.0000446 0.0000059

11 5.19E+10 26962 36206 7418 7086 0.0000520 0.0000698 0.0000143 0.0000137

12 1.24E+11 22602 52491 9520 7938 0.0000183 0.0000424 0.0000077 0.0000064

13 2.76E+11 18924 41598 8845 7288 0.0000069 0.0000151 0.0000032 0.0000026

14 5.79E+11 23071 48194 7456 7452 0.0000040 0.0000083 0.0000013 0.0000013

15 1.15E+12 27477 40730 9077 9753 0.0000024 0.0000035 0.0000008 0.0000008

16 2.20E+12 30301 39470 10030 7734 0.0000014 0.0000018 0.0000005 0.0000004

17 4.03E+12 24299 57393 8852 6682 0.0000006 0.0000014 0.0000002 0.0000002

80
 Integrated planning between procurement and production

4.5 Discussion and concluding remarks

The main purpose of this study was to find effective ways in supporting several decision-
makers on different decision levels in a cooperative association. Management of the
cooperative association realized that, due to the annual growth, the increasing imbalance
between the continuous production on supply level and the fixed delivery conditions on
demand level, could not be solved solely by detailed scheduling of routes and tours on a
daily basis. The current, low-level oriented, approach of the milk collection problem was
hardly viable in the future. It was expected that a first draft of a PC-based system was
necessary in order to show the potential benefits of a different approach and the related
improvement of individual and organizational performance. The system should have a
major focus on short- to medium-term planning rather than building detailed routes and
tours on a daily basis. One of the goals was to build a pilot system for generating stable
milk collection plans. From a computational point of view the optimisation module should be
able to generate plans within a reasonable amount of time.

4.5.1 The pilot decision support system

The pilot system was divided into four major components: a user interface, database,
simulation-, and an optimisation module. The menu-driven user interface and the database
and simulation module were all developed in Microsoft Access. The optimisation part, i.e.
the model in Section 4.3, has been specified and solved by the advanced modelling and
solving language Xpress-Mosel. The modelling component also provides for a set of
procedures and functions that enable a connection to the database by an Open Data Base
Connectivity (ODBC). All necessary data for the input of the model can be retrieved directly
from the data source. Reversely, the output of the optimisation- or simulation routine can be
written directly into the database by the ODBC interface.

Achieving the mission of a DSS, i.e. to help end-users in making better decisions,
implies that such a system does not replace the decision-maker. Only the end-users have
the skills and specialized knowledge to review the quality of the generated plans. The DSS
aims to assist decision-makers and should not be considered as an optimiser but rather as
a tool for generating and storing high-quality plans to be used for further analyses. Within
this context the facilities of a user-friendly and interactive man-machine interface are
essential.

The user interface has been divided into an input and planning i.e. analysing part. A
basic start-up screen offers access to each part. Any other screen of the user interface
includes a link to return to the start-up form. The start-up screen also offers an option to
start a run of the optimisation routine and assign (required) vehicle capacities to routes
(VRP) directly, or to use stored solutions from the past to simulate and compare options for
the VRP.

81
Chapter 4

The input forms enable the modification of fields in existing records in the underlying
database or the addition of new records. The system distinguishes different forms for
suppliers, transportation companies, buyers on demand level, the defined milk collection
clusters and the potential milk collection rhythms. All input forms are provided with
navigation buttons and record selectors, enabling the movement between records in the
database. The supplier form contains text boxes for a unique identification number, address
data, the (default) assigned milk collection cluster, the daily milk production and available
(cooled) storage capacity at supply level. Apart from standard fields for an identification
number and address data, the input form on demand level contains fields for specifying a
daily minimum and maximum amount to deliver, maximum age of the delivered milk (default
three days) and a check box indicating whether the buyer accepts milk on days without
demand (the so-called surplus companies). A push button gives access to a linked subform
in which the daily demand levels are specified. As demand levels are usually based on
contracts of either a weekly or a two weekly repetitive pattern, the planning horizon is set on
a default period of two weeks. A third input form contains adjustable fields for the available
transportation companies, i.e. a list box of (daily) available transportation vehicles for each
company and the related loading capacities. The last input forms contain a list box for all
defined clusters, an overview of all suppliers in a cluster and a subform, comparable with
Table 4.1, for the set of defined milk collection rhythms.

Any generated solution, either obtained by recovering a stored plan of a former run
or running the optimisation routine with (changed) data from the input part, enables access
to the planning and analysing part. This part of the system is roughly divided into output
forms on supply and demand level. The basic screen on supply level contains a drop-down
box in which the defined milk collection clusters are listed including the option for an
overview that takes all clusters together. Selecting one of the listed options will expand the
contents of all relevant fields and subforms in the output screen. The screen is subdivided
into two identical parts for each week of the planning horizon containing orderly information
with regard to the (daily) offered quantities of the cluster(s) and the planned amounts of milk
to deliver from each cluster to the intended buyers in case the proposed milk collection
rhythm would be followed. In a subscreen the user can ask for a (daily) overview of
available transportation vehicles and the (remaining) loading capacity of each vehicle. The
allocation of vehicle capacities to milk collection clusters has to be done by the end-user.
We suppose that only the end-users have the skills and specialized knowledge to recognize
patterns in the location of suppliers, (third party) transportation companies and buyers on
demand level. Nevertheless, the system can be very helpful. It constantly updates the
values of several indicators like the remaining loading capacities of the vehicles and the
average distances between the (different) departure points, the centre of a milk collection
cluster, and the different points of destination. The lay-out of the output screen on demand
level is comparable with the form(s) on supply level. A drop-down box enables the selection
of individual buyers including the option for a general overview that takes all buyers
together. The output form contains orderly information with respect to the (daily) demand

82
 Integrated planning between procurement and production

levels and the planned amounts to deliver to the buyer(s) in case the proposed milk
collection rhythm would be followed.

In the output screens on supply and demand level several (conflicting)

measurements of performance are calculated and presented in the reports. Changing the
data, for example moving farms from one cluster to another, or changing the milk collection
rhythm for a cluster in the generated plan, is possible. However, the consequences of any
modification in the data or the proposed plan will affect the measures of performance too. In
this way the pilot system combines the power of the human judgement and experience on
the one hand with the calculation speed, accuracy and storage facilities of computer
systems on the other hand.

4.5.2 Conclusions
From the start it was clear that the emphasis of the system should be to support decision-
making on different levels within the cooperative association. Vehicle schedulers as well as
managers of the cooperative association were looking for ways that helped them to make
better decisions. The visualization of (modified) plans and the possibility to store plans over
the year enables the decision-maker to ‘optimize’ his / her performance with respect to his
or her mission.

In a way the problem can be viewed as an instance of the periodic vehicle routing
problem (PVRP) with the following characteristics: it concerns pickup and delivery
operations simultaneously. Consequently there is a stronger focus on balancing supply and
demand as opposed to the routing of vehicles. Referring to the multi-level classification of
the PVRP by Chao, Golden et al. (1995), the short- to medium-term planning model turns
out to be a successful approach for the first level of the PVRP in which it is necessary to
assign allowable visit combinations to suppliers as well as customers, such that the
continuous production on supply level will be balanced with periodic demand. The
generated plans also serve as a starting point for the next level of the PVRP. This level
consists of solving several vehicle routing problems (VRP; i.e. the construction of routes
and the assignment of vehicles to routes) for each day of the planning period. Although the
system does not generate detailed solutions for the vehicle schedulers, the plans offer a
solid and stable starting point for the daily VRP. The idea of assigning feasible milk
collection rhythms to clusters of suppliers was adopted in an early stage by the vehicle
schedulers. Once supplier farms are (geographically) grouped into clusters and the
complete milk production within a cluster is assigned to a single rhythm with fixed collection
days (see Table 4.1), the daily routing problem has been simplified substantially. The
overview of available transportation vehicles ordered by the (remaining) loading capacities
or the average distances between the departure point of vehicles, the centre of a milk
collection cluster, and the intended points of destination, turns out to be very useful.
Nevertheless, if the cooperative association ultimately decides to set up a final software

83
Chapter 4

development project, management should reconsider the functional characteristics of the

system again. Extending the current pilot system by an additional module that enables the
construction of starting solutions for the daily VRP, might be helpful for vehicle schedulers
too. However, the construction of this type of system in a real world environment tends to
be a time-consuming and expensive software development project. Third party software
developers have to be contracted for the development of custom-made software.

Especially the stored plans and their information regarding the delivered amounts of
milk (sold at unattractive price levels) to so-called surplus companies or the shortages of
deliveries to regular buyers (at attractive price levels), can be very helpful in order to attune
the future imbalance between milk supply and the individual demand levels of dairy
factories. A profound analysis of these data will be very beneficial for the outcome of the
yearly negotiations on demand level with respect to the expected amounts of milk to deliver
and the desired delivery days weekly. Moreover, the analysis of stored plans can be quite
helpful for the negotiations with third party transportation companies regarding the (daily)
required vehicle capacities in different seasons of the year.

From a computational point of view it turned out that the application of Special
Ordered Sets was useful. The numerical experiments confirm that the efficiency of the
SOS-formulation strongly depends on the ordering of the variables within the sets.
However, we also showed that the computational advantage of the SOS-formulation is not
restricted to cases in which the variables within the sets have a natural ordering. A
reordering procedure of the variables, based on their solution values of the LP-relaxation of
problem (1) to (6), turned out to be very effective. However, it is too premature to conclude
that a natural ordering of the variables within S1-sets is superfluous for an efficient use of
SOS1-formulations. In this study it turns out that the values of the numbers in the reference
row are of minor importance for the computational efficiency of the SOS-formulation. As a
result, the relevance of a reference row defined by the model developer personally might be
omitted in the future for mathematical programming software. Further research in this area
(i.e. numerical results of other cases) has to be done.

84
 Integrated planning between procurement and production

Appendix 4.1
We will proof that the upper bound for the number of branches B in an SOS-based branch-
and-bound search tree is defined by:
C
B= ∑R
c =1
c -1
(2R - 2) (7)

in which C denotes the number of milk supply clusters (S1-sets) and R denotes the number
of different milk collection rhythms. We also proof that the potential number of branches B in
a common branch-and-bound approach, i.e. branching on individual binary variables Yc,r in
problem (1) to (6), equals (7) too.
A) First we focus on the impact of R (the number of milk collection rhythms) on the number of
branches B and prove that B = (2R − 2) in case we define only one milk supply cluster
(C=1)

85
Chapter 4

86
 Integrated planning between procurement and production

By means of complete induction we proved that the relation between the potential number of
branches B and the available number of milk collection rhythms R is equal to B=2R-2 for both
branching principles in case we define only one cluster C=1.
C
B) Next we prove that the potential number of branches is defined by B = ∑R
c =1
c -1
(2R - 2) for an

arbitrary number of clusters c =1 ... C. Note: exactly one rhythm must be chosen in each cluster.

87
Chapter 4

88
 Chapter 5

Mixed Integer (0-1) fractional programming

in Paper Production Industry

Fractional programming gave itself a somewhat questionable reputation in the Operations Research
community by divorcing itself too much from the applications (Schaible and Ibaraki 1983)

This chapter is based on:

Claassen, G. D. H., (2014)

Mixed Integer (0-1) Fractional Programming for Decision Support
in Paper Production Industry
Omega, the International Journal of Management Science
Vol. 43, pp. 21 – 29.

Claassen, G. D. H., and Hendriks, Th. H. B., (2007)

Modelling techniques for (non)linear and integer programming
In: Decision Science, Theory and applications, Ch. 9
Eds: Claassen et al.,
Wageningen Academic Publishers, pp. 185 – 226
Chapter 5

Abstract
This chapter presents an efficient and effective method for solving a special class of mixed
integer fractional programming (FP) problems. We take a classical reformulation approach
for continuous FP as a starting point and extend it for solving a more general class of mixed
integer (0-1) fractional programming problems. To stress the practical relevance of the
research we focus on a real-life application in paper production industry. The constantly
advancing physical knowledge of large scale pulp- and paper production did have a
substantial impact on an existing DSS in which mixed integer (0-1) fractional programming
is introduced. We show that the motivation to solve a real-life fractional programming
problem can provide the basis for a new approach in a new context that has an added value
of its own, even outside the given application area. We describe the main characteristics of
the DSS, the necessity to develop a non-iterative solution procedure and demonstrate the
efficiency of the proposed approach from practical data sets.

90
 Mixed integer (0-1) fractional programming in paper production industry

5.1 Introduction
Amply before the era of decision support systems, Little (1970) launched “The concept of
decision calculus” starting with a provoking phrase: “The big problem with management
science models is that managers practically never use them”. The author presents a set of
guidelines to bridge the increasing gap between (mathematical) theory and the scientific
challenge of its applicability in real-life enterprises. At the time of appearance the impact of
developed theory in practice and the wave of real-world applications (e.g. in decision
support systems) was still in its infancy. After more than three decades, Little (2004)
reflected back on his original paper: ”The good news is that more managers than ever are
using models. The bad news is that many managers do not even realize they are using
models! (But we should ask whether this is really bad)”. In other words: today, (decision
support) models are used! However not by managers themselves, but by management
scientists and management assistants acting as intermediates to frontline managers.

We illustrate the significance of Little’s concept for a novel application of fractional

programming in practice and describe the latest developments of an OR-based DSS for a
continuously changing decision environment in pulp- and paper production industry.
Pressed by changed circumstances, management teams of global enterprises continuously
aim to exploit innovation due to advanced physical knowledge of large scale pulp- and
paper production. Investments in flexible tools for global decision support on different
decision levels within the enterprise (e.g. marketing-, production-, R&D etc.) are of utmost
importance to remain competitive. Once the system demonstrated its added value and
validity to management scientists, it gained the trust of end-users and was preferred to be
adapted to new decision environments. We show that the motivation to solve a practical
problem in a real-world environment forces researchers to find new and efficient
approaches. In this paper we focus in particular on the impact of progressive physical
insight on an existing optimization module in the DSS. An efficient, non-iterative solution
procedure was needed to solve mixed integer (0-1) fractional programming problems in a
real-world situation.

Fractional programming (FP) may be an illustrative field in mathematical

programming demonstrating the separation between theoretical developments and its
applicability in practice. According to Schaible and Ibaraki (1983), research in FP divorced
itself too much from real applications and the majority of FP models analysed in literature
were still waiting for their actual implementation in real-world situations. Meanwhile, several
surveys have been published on FP (Schaible 1995; Stancu-Minasian 1999; Stancu-
Minasian 2006). The extensive survey of Schaible (1995) was published in 1995 and
contains almost twelve-hundred entries. The latest bibliography (Stancu-Minasian 2006)
covers the period 1997 – 2005 with almost 500 entries. Although the interest for
applications of FP has increased since the nineties, the bibliography shows that between
1997 and 2005 less than 10% of the papers are application oriented (including potential

91
Chapter 5

applications and case studies). Hardly 7% of all covered papers (including theoretical
studies) in Stancu-Minasian (2006) are devoted to integer FP. Schaible and Shi (2004) also
stated that integer FP is a somewhat neglected field that deserves more attention. The
problem we describe belongs both to the class of real-life applications and the class of
mixed integer fractional programming problems. To the best of our knowledge, there exists
no simple and non-iterative solution technique for solving mixed integer (0-1) linear
fractional programming problems.

The paper is organized as follows. In the next section we characterize the pulp and
paper industry and give an overview of the progress in OR-based decision support for this
branch of industry. Section 5.3 describes the actual mixed integer (0-1) fractional
programming problem and gives a basic outline of the model formulation which is
necessary to understand the impact of progressive (physical) insight on both the existing
model formulation and the necessity to find an efficient solution procedure for solving this
new problem (Section 5.4). The added value of the FP model, including the efficiency of the
proposed solution technique, will be tested from several practical data sets (Section 5.5).
Discussion and conclusions follow in the Sections 5.6 and 5.7.

5.2 Background
Pulp and paper industry is an extremely large business characterised by low margins and
high capital costs. Large mills can cost hundreds of millions of US$ to construct. As a
consequence, only a few companies are active on this huge and global market. The
company and sponsor of the research is a leading producer of coated fine paper in North
America, Europe and South Africa. The fine paper division (e.g. copy papers, writing papers
etc.) is a business with manufacturing assets in eight countries on three continents and
customers in over 100 countries. Their production capacity is about 6 million tons of fine
paper per annum, produced in 14 different mills in North America (3), Europe (8) and Africa
(3). In 2011 the total sales of the company reached 6.01 billion US$; the operating profit
was 404 million US$.

As margins are low in this capital-intensive branch of industry, a continuous search

for efficient production and decision support is of utmost importance. Faced with rising raw
material and energy costs, pulp and paper producers have to optimise the performance of
their processes to remain competitive. Reducing costs, e.g. for raw materials, is an
important way to increase or at least maintain the annual operating profit.

Several OR-based applications have been described in literature, tailored for pulp
and paper industry. We refer to (Johnston 1980; D'Amours, Rönnqvist et al. 2008) for an
overview of OR applications in pulp and paper industry. The papers provide a good insight
in the progress that has been made in recent decades regarding the application of OR

92
 Mixed integer (0-1) fractional programming in paper production industry

models and -techniques in this branch of industry. Nowadays, decision support for planning
and management in pulp and paper industry includes the complete supply chain from
strategic-, tactical- to operational control level.

Typical studies on a strategic level concern investment studies or models for

optimization of the supply chain i.e. determining the facility location, optimally allocate
suppliers to mills, products to paper machines i.e. mills and machines to markets (Berends
and Romme 2001; Philpott and Everett 2001(a); Philpott, Everett et al. 2001(b); Everett,
Philpott et al. 2002) strategic design of distribution networks (Gunnarsson, Rönnqvist et al.
2006; Pati, Vrat et al. 2008) and/or studies on the conflict between economic optimization
and environmental efficiency (Hua, Bian et al. 2007; Pati, Vrat et al. 2008).

On a tactical level decision-making refers to planning problems in different links of

the supply chain. Carlsson and Rönnqvist (2007) focused on the wood procurement stage
of the supply chain. Medium-term (i.e three months) production planning schedules for pulp
mills are developed in (Bredström, Lundgren et al. 2004). The case study (Bouchriha,
Ouhimmou et al. 2007) discusses a specific production planning (i.e. lot sizing) problem for
a single paper machine of a fine paper mill. Different (synchronized) models for production,
transportation and inventory planning problems in the fine paper industry are studied in
(Martel, Rizk et al. 2005). Chauhan, Martel et al. (2008) deals with tactical demand
fulfilment of sheeted paper in the fine paper industry. The authors propose an integer
programming model to find the stock keeping units of parent rolls in order to minimize
expected inventory holding and trim loss costs.

Typical decision problems on an operational control level refer to short-term

production- and distribution planning problems in single (Alemayehu and Arora 2002) or
multiple successive stages of paper manufacturing for multiple paper machines and
distribution centres (Murthy, Akkiraju et al. 1999; Rizk, Martel et al. 2006).

Although literature provides a good insight in the progress that has been made in
OR-based decision support for pulp and paper industry, none of the studies focuses on the
impact of raw material composition and its technical treatments on the final properties of
paper grades. To the best of our knowledge there exists no other DSS to support this
practical problem adequately.

In the next section we describe the impact of raw material composition and its
technical treatments on the final properties of paper grades and present an outline of the
model formulation. Section 5.3 aims to set the boundaries of the decision environment
which is needed to understand the core of the research in Section 5.4.

93
Chapter 5

5.3 Problem description and model formulation

Trees provide the primary raw material for paper production. Wood is comprised of
cellulose fibres which are bound together by the natural “glue”, called lignin. In the first step
of paper production the cellulose fibres are separated from one another into a mass of
individual fibres called pulp. This is done in a pulp mill by using either chemical or
mechanical processes. Using chemical pulp for paper production is more expensive than
mechanical pulp or recovered paper. However, as the cellulose fibres of mechanical pulps
are more or less “damaged”, the resulting paper has lower strength characteristics. The
chosen pulping process will also affect other properties of the final paper.

Wood fibres can be divided into hardwood (i.e. deciduous trees) and softwood (i.e.
pine trees) fibres. Softwood fibres are longer and coarser than fine hardwood fibres. Usually
softwood pulp is used to provide the required strength when producing light-weight
publication papers. Fine papers (e.g. copy papers, writing papers etc.) are mainly produced
from hardwood pulp, which is reinforced by a minor amount of stronger and more expensive
softwood pulp. Pulp may be fed directly to a paper machine in an “integrated mill” or dried
and pressed into bales to be used as a raw material by paper mills elsewhere.

It is convenient to decompose the process of paper production into three major

steps: pulp production, stock preparation and paper production. We focus in particular on
the decision problems in the stock preparation process. Stock preparation is a critical part in
paper production. It consists of combining and preparing the raw material into a mixture of
several bleached softwood , hardwood and mechanical pulps. Part of stock preparation is
the refining process of bleached pulps. This operation is a crucial step to provide the
surface properties for printing grades according to customer specifications. At the refiners,
the pulp composition passes a system of rotating and stationary blades. Depending on the
beating intensity, fibres are more or less shortened and damaged to give the final mix,
more correctly termed the “furnish”, the exact properties required for a particular type of
paper. After the stock preparation, the furnish is suitable to create a uniform web of paper
on a paper machine.
Given the technical characteristics of a mill (machine size, type of production
equipment and automation level) there are many variables that determine the final quality of
the paper grade. Throughout the past decades the R&D department quantified the paper
quality by introducing several, partly physically, measurable characteristics. These so-called
properties can be divided into several groups, i.e. subsets of properties that are indicative
for the dehydration of the pulp, strength of the paper, optical properties of a paper grade,
fibre dimensions, total costs etc. All property values depend on the mixture of raw materials:
i.e. chemical softwood, hardwood and mechanical pulps. Moreover, a substantial number of
property values depend on the beating intensity of the refiners. Other characteristics are so-
called beating independent properties.

94
 Mixed integer (0-1) fractional programming in paper production industry

Several interrelated decisions must be made to produce paper that meets all
requirements. An outline of the stock preparation problem is given in Figure 5.1. Key
decisions to be made are:

1) The raw material composition, i.e. which mixture of chemical softwood and
hardwood pulps and mechanical pulps to choose.

2) The choice of the available refiners, their beating intensity including the assignment
of combinations of celluloses to individual refiners. The sequence of treatment of
different celluloses in the refiners is not important.

3) The contribution (i.e. mass fractions) of the individual pulp flows from the refiners to
the final furnish for paper production.

Figure 5.1 Stock preparation process

A major issue from practice is that a common principle of an overriding objective is

too restrictive for modern decision support in globally operating enterprises. Management
assistants in paper production industry usually focus on a variety of properties
prop := {1, 2, ..,P} of the paper grades (e.g. total costs, tear index, dewatering rate, bursting
strength, opacity etc.). Depending on the case to study, tools are needed to calculate the
best upper and/or lower bounds for individual property values. Decision support systems

95
Chapter 5

should foster “out of the box thinking” and make decision-makers aware of the impact of
personal trade-offs with respect to requirements of individual properties on other properties.
For example total costs of a paper grade is an important property in the set
prop := {1, 2, ..,P} to minimize. However, optimizing any other property after an acceptable
upper bound has been set on the total costs, may deliver insight as well, and consequently
contribute more to the decision-making process.

Based on progressive physical knowledge, laboratory tests show that the predicted
values for a subset of properties (i.e. all properties related to the compactness of final paper
grades), can be improved if these property values are calculated as a function of the total
number of fibres in the furnish. It appears that the combination of this progressive physical
knowledge with abandoning the common principle of an overriding objective (e.g. total
costs), has a significant impact on the optimization routine in the DSS. To illustrate the
latter, it is convenient to decompose the complete set of properties prop := {1, 2, ..,P} into
two disjoint subsets pmass and pvol. So, prop = pmass U pvol . We define:

Indices
c = 1,…,C ~ the different types of celluloses
p = 1,…,P ~ the different properties

b = 0,…,B ~ the beating intensity of a refining line

r = 1,…,R ~ the available refiners

Variables
x c, r, b ~ mass fraction of cellulose c in the total mixture, refined in line r, at beating
level b

y r, b ~ 1 if refiner r operates at beating intensity level b


 0 else

Property values for the subset pmass (e.g. total costs, opacity, dewatering rate,
brightness etc.) are calculated by using mass fractions, i.e. xc,r ,b . Properties in the subset
pvol are related to the compactness of the final paper grade (e.g. bursting strength, tear
index, coarseness, breaking length etc.). In the past, all property values for pmass and pvol
were calculated by using mass fractions. However, progressive physical knowledge learned
that property values in the subset pvol should be calculated by a non-linear function g (x c ,r ,b ) .
The function g (x c ,r ,b ) is defined as the volume fraction of cellulose c in the total mixture,
refined in line r, at beating level b. Using volume fractions instead of mass fractions may

96
 Mixed integer (0-1) fractional programming in paper production industry

change the contribution of individual celluloses to the total number of fibres in the furnish.
The final values for all properties in a paper grade are given by (1) and (2):

fp = ∑∑∑ PV
c r b
c ,p,b ⋅ x c,r ,b ∀p ∈ pmass (1)

fp = ∑∑∑ PV c , p,b ⋅ g (x c,r ,b ) ∀p ∈ pvol (2)

c r b

in which PVc,p,b is defined as the property value of property p for cellulose c, at beating level
b. The relation between volume fractions and mass fractions is given in (3).

Vc ⋅ x c,r ,b
g (x c,r ,b ) = C
(3)
∑V
c =1
c ⋅ x c,r ,b

in which Vc denotes the number of fibres per gram of cellulose c. The value for the
parameter Vc strongly depends on the type of cellulose. After a few additional data
definitions, an outline of the core problem can be formulated.

Data
Lp ~ Lower bound for property value p
Up ~ Upper bound for property value p
Capr ~ Capacity of refiner r expressed in a percentage of the total flow

Max
x cbr ;y rb
{f }~
p
~ ∈ prop
for some p (4)

s.t.
∑∑∑ x
c r b
c ,r ,b =1 (5)

∑y b
r ,b =1 ∀r (6)

f p ≥ Lp ~
∀p ≠ p (7a)

fp ≤ U p ~
∀p ≠ p (7b)

∑x c
c,r ,b ≤ Capr ⋅ y r ,b ∀ r, b (8)

x c,r ,b ≥ 0 ∀ c, r, b (9)

y r ,b ∈ {0,1} ∀ r, b (10)

97
Chapter 5

The objective function (4) optimizes the desired property value f p~ of the final paper
grade. In case of a minimization problem, f p~ is multiplied by minus one. Equation (5)
ensures that the sum of the fractions of all celluloses in a mixture equals one. As all
variables yr,b are defined to be binary (10), constraints (6) require that each refiner will run
at only one beating level. Constraints (7a) and (7b) put lower- and upper bounds on the
property values in the final pulp. The capacity constraints of the refiners are formulated in
(8). Moreover, constraints (8) state that the total throughput of raw materials in each refiner
at a certain beating level b can be positive only if the value of the corresponding binary
variable yr,b equals one. The conditions (9) and (10) complete the set of restrictions.
~ ∈ p will hardly affect the structure
Replacing Eq (1) by (2) in (7a) and (7b) ∀p ≠ p vol

and complexity of the problem. The original model (4) – (10) remains linear after multiplying
both the left- and right-hand sides of (7a) and (7b) by the denominator in (3). However, for
objective function (4) two cases need to be distinguished. The optimizing property function
f ~ may refer to a mass-fraction-dependent property p ~∈ p (e.g. p~ may express total
p mass
~∈ p
costs, see (4a) below) or to a volume-fraction-dependent property p ~
vol (e.g. p may

express the Burst index, see (4b) below).

  
Max − 
 
∑∑∑ RCost
c r b
c ⋅ x c ,r ,b + ∑∑ ECost
r b
r ,b ⋅ y r ,b 

(4a)

in which RCostc and ECostr,b denote the raw material costs of cellulose c and the total
energy costs of refiner line r, at beating level b, respectively.

 
Max  ∑ ∑ ∑ Burst c ,b ⋅ g (x c ,r ,b ) (4b)
 c r b 

in which Burstc,b denotes the contribution of cellulose c to the Burst-index at beating level b.
~∈ p
If the objective function (4) refers to p mass , the problem is a straightforward
~∈ p ,
mixed integer linear programming problem, e.g. (4a)–(10). However, if (4) refers to p vol

then Eq. (3) implies that the objective function becomes a ratio of two linear functions. Such
models, e.g. (4b)–(10), belong to the class of Linear Fractional Programming or Hyperbolic
Programming problems (Bajalinov 2003; Bazaraa, Sherali et al. 2006). Models for which the
objective function (4) refers to mass-fraction-dependent properties p ~∈ p , or volume- mass
~ ∈ p , will be referred to as (MFM) or (VFM) models,
fraction-dependent properties p vol

respectively. Generally, a continuous linear fractional programming model (LFP) can be

formulated as:

98
 Mixed integer (0-1) fractional programming in paper production industry

 c' x + c0 
max  
x
d' x + d0 
s.t. (LFP)
x ∈S

The solution space or set S := {x | Ax ≤ b; x ≥ 0 } is assumed to be compact, i.e. convex,

non-empty and bounded. Note the difference between a basic (LFP) problem and model
(VFM), i.e. a subset of decision variables in (VFM) are defined as binary variables in (10).

In the next section we follow the reformulation approach introduced in Charnes and
Cooper (1962) for the transformation of a continuous fractional programming problem (LFP)
into an equivalent linear programming model (LP). Next, we extend this approach for
solving mixed integer (0-1) fractional programming problems arising from any (VFM).

5.4 Solving mixed integer (0-1) fractional problems

Fractional programming (FP) can be considered as a separate entity within the field of non-
linear programming. Apart from some isolated papers, a systematic study in this area
started in the early seventies. Meanwhile a rich collection of papers has been devoted to
this field of non-linear programming. For an overview we refer to some extensive review
papers (Schaible and Ibaraki 1983; Schaible 1995; Stancu-Minasian 1999; Stancu-Minasian
2006).

The study of fractional programs with a single ratio dominated literature for a long
time. Dinkelbach (1967) introduced a very popular and general parametric approach that
can be applied to all types of (non)-linear, (integer) fractional problems. Generally, for
integer fractional programming problems parametric approaches are used (Barros 1998).
However, parametric approaches like Dinkelbach (1967), and its variants, require an
iterative evaluation of a parametric function. From a practical point of view, these
approaches were not preferred for a single ratio linear fractional problem. The focus was
directed towards taking advantage of important properties of the (VFM), i.e. a single ratio
problem and to exploit special structures for solving the mixed (0-1) integer problem by a
fast, non-iterative, solution procedure. Such an approach is proposed in Robillard (1971) for
a special class of (0-1) fractional programs with a single term in the objective function. The
algorithm takes advantage of an assumed special structure of the feasible set. However, the
required structure of the constraints is too restrictive for problem (VFM). Moreover, the
approach needs a special purpose branch-and-bound algorithm.

The relationship between problem (LFP) and linear programming is also exploited in
Granot and Granot (1977). The authors develop cutting planes which can be systematically

99
Chapter 5

generated if some variable of the original problem (LFP) is not integer. However, the
approach needs a special purpose cutting plane algorithm.

Alemayehu and Arora (2002) described an alternative approach in which a mixed

(0,1) linear fractional problem is reformulated into a bi-level FP problem. The concept is
demonstrated on a small scale example. Bi-level programming involves two optimization
problems where the constraint region of the first level problem is implicitly determined by
another optimization problem (Calvete and Galé 2004). However, the reformulation in
Alemayehu and Arora (2002) needs an iterative/nested solution approach too. To avoid
these obstacles we start from a classical reformulation approach for solving continuous
fractional programming models like problem (LFP) and extend it for mixed integer (0-1)
fractional problems.

A single ratio of linear functions is neither convex nor concave. However, any local
maximum of problem (LFP) is global (Bazaraa, Sherali et al. 2006). Likewise, a local
minimum is also a global minimum over the set S. Moreover, if the solution space is
compact, then the objective function has both a minimum and a maximum at an extreme
point of the feasible area (Bazaraa, Sherali et al. 2006). As the optimal solution for a
(mixed) integer linear FP problem is a vertex of the convex hull for the set of feasible
discrete solutions, it gives rise to apply a solution procedure that moves from one extreme
point to an adjacent and use a branch-and-bound technique to eliminate non-discrete
solutions.

In 1962 Charnes and Cooper introduced their classical paper in which a continuous
model (LFP) is transformed into an equivalent linear programming model (LP). The model
(LP) needs exactly one additional variable and only one additional constraint. With
reference to the general problem (LFP), the reformulation approach is based on the
definition of a vector w and a scalar t of decision variables:

x
w= (11)
d' x + d0

1
t= (12)
d' x + d0

From (11) and (12) it follows that:

w = t⋅x (13)

The basic idea is to reformulate model (LFP) by means of (11) and (12) such that a
linear programming model arises in terms of the variables w and t. If this new model can
be solved for all relevant values of t, the solution of any fractional problem derived from

100
 Mixed integer (0-1) fractional programming in paper production industry

(LFP) can be found by (13). Next, the reformulation method will be illustrated for cases in
which the denominator in the objective function of (LFP) is positive over the entire set S.
So, d ' x + d0 > 0 for all x ∈ S.

The objective function of the model (LFP) can be rewritten as follows:

 c' x + c0   x 1 
max   = max c ' ⋅ + c0 ⋅ 
x
d ' x + d0  x
 d' x + d0 d' x + d0 

Using definition (11) and (12) this can be reformulated as:

max { c ' w + c 0 t } (14)

w ,t

w
Using (13), i.e. x = , the constraints Ax ≤ b of the set S can be written as:
t

Ax ≤ b ⇒ A ⋅ w / t ≤ b ⇒ A w ≤ b t . So,

Aw − b t ≤ 0 (15)

Definition (12) of the (new) variable t in (14) and (15) needs to be added:
1
t= ⇒ t (d ' x + d 0 ) = 1 ⇒ d ' t x + t d 0 = 1. Using w = t ⋅ x :
d' x + d0

d 'w + t d 0 = 1 (16)

Note, we assume d ' x + d 0 > 0 . So, t > 0.

Reformulating the non-negative constraints x ≥ 0 of the set S in terms of the variables w

and t using w = t ⋅ x and t > 0 gives:

x ≥ 0 ⇒ w / t ≥ 0 , t > 0. So,

w ≥ 0, t > 0 (17a)

If we suppose t ≥ 0, then the following linear programming model is equivalent to the

original fractional model (LFP), provided that d ' x + d0 > 0.

max { c ' w + c 0 t } (14)

w ,t

s.t.
Aw − b t ≤ 0 (15)
d 'w + t d 0 = 1 (16)
w ≥0,t ≥0 (17b)

101
Chapter 5

Note that t ≥ 0 in (17b) is just for the form’s sake of linear programming. If d ' x + d 0 > 0 ,
x ∈ S , S is compact and t > 0 (see 12), then the optimal value t = t * > 0 for model (14) –
(16) including (17a) will also satisfy model (14) – (16) including (17b).

In summary, any finite maximum of the fractional programming problem (LFP) can
be found by solving the LP model in Table 5.1.

Table 5.1 General structure of the transformed model (LP) for continuous model (LFP)

(LFP) (LP)
 c' x + c0 
max   max { c ' w + c 0 t }
x
d' x + d0  w ,t

Ax ≤ b Aw − b t ≤ 0

d'w + t d0 = 1
x ≥0
w ≥0,t >0

After solving model (LP), the solution x of any fractional problem derived from (LP)
can be found by (13). The denominator d ' x + d 0 of problem (LFP), should be either strictly
positive (or strictly negative) for all possible values of d and d 0 . If not, then there exists a
solution for the non-negative variables x for which the denominator d ' x + d 0 = 0 . In such
cases the (transformed) problem (LP) in Table 5.1 is obviously not defined. According to
physicists of the R&D department, both the data PVc,p,b (i.e. the property values) and Vc (the
number of fibres per gram of cellulose) comply to the assumption that these values are
always positive.

Applying the reformulation approach of Table 5.1 to a mixed integer FP model, e.g.
model (VFM), implies that the integrality constraints (10) must be relaxed. As a
consequence, solutions of model (LP) are mostly infeasible for the original mixed integer
fractional problem (VFM). A method must be found to fulfil the binary conditions (10) of the
original problem.

w x 
The vector w of continuous variables in Table 5.1 can be partitioned into w =  
w y 
 
x y
where the vector w refers to the continuous variables xc,r,b in (9) and w to the binary
y
variables yr,b in (10) of problem (VFM). According to (13), the vector w in model (LP) is

102
 Mixed integer (0-1) fractional programming in paper production industry

= t ⋅ y . If t > 0 (Table 5.1) and y must be binary according to (10), then

y
defined by w
y
branching on individual variables in w and t, is not applicable for problem (LP) because the
solution ratios w ( y
t ) must be binary. It should be mentioned that the variable t will always
be positive (i.e. basic) in an optimal vertex of problem (LP). If the denominator d ' x + d 0 ≠ 0,
x ∈ S and S is a compact set, then according to (12) the optimal value for t = t * > 0 .

y
Now, suppose we decompose the vector w of continuous variables in model (LP)
y
into r=1…R disjoint subsets. For each refiner r, we distinguish a subset W r ,b : = { w ry,0 , … ,
w ry,B }. Indices b = 0, … B refer to the beating level. Note that all variables w ry,b ∀r, b are
continuous in problem (LP). If the optimal value for t = t * > 0 , ∑y
b
r ,b = 1 ∀r in (6) and

y r ,b ∈ {0,1} ∀r , b in (10), then at most one of the variables in each subset W y

r ,b for r =1…R
can be non-zero. Now it is convenient to treat these disjoint subsets W y
r ,b ∀r as discrete
entities or generalisations of a 0−1 variable and apply the SOS1 branching concept,
originally introduced in (Beale and Tomlin 1970) and extensively discussed in the previous
y
chapter. In problem (LP) we define the sets W r ,b of continuous variables:

W y
r ,b := { w ry,0 , … , w ry,B } ∀r together with the conditions that at most one
of the variables { w ry,0 , … , w ry,B } can be non-zero ∀r . (14)

If we can prove that the optimal value t = t * > 0 equals the non-zero value in each subset
W y
r ,b ∀r in (14), then we may solve problem (VFM) by adding (14) to problem (LP) of
Table 5.1. Problem (LP) including (14) will be referred to as problem (VFMLP).

Lemma
Assume that for problem (VFM), (x, y)' ∈ S , S is a compact set and the assumption with
respect to the denominator d ' x + d 0 ≠ 0 holds, then the optimal value t = t * > 0 for problem
(VFMLP) equals the non-zero values in the subsets W y
r ,b ∀r .

Proof
Applying the reformulation approach as summarized in Table 5.1 to problem (VFM), implies
that the constraints in (6) are transformed into ∑w
b
y
r ,b −t = 0 ∀r . If at most one of the

103
Chapter 5

variables { w ry,0 , … , w ry,B } ∀r can be non-zero according to (14), then the optimal value
for t = t * equals the non-zero values, i.e. ∑w
b
y
r ,b , in each subset W y
r ,b ∀r .

The conditions (14) can be dealt with algorithmically through the method of integer
programming (see Chapter 4). Treating each set as an entity instead of a collection of
variables makes it possible to apply a different branching scheme. We refer to Chapter 4 for
a complete description of the concept. The single, non-zero variable in each feasible
y
W r ,b -set of (14) will lie either to the left, or to the right, of any marker placed between two
consecutive variables within a set:

either { w ry,0 , … , w ry, j } or { w ry, j +1 , … , w ry,B } are all zero

In summary: instead of defining R · B different binary variables yr,b in the original

model (VFM), the integrality constraints (10) can be relaxed. The reformulation approach of
Table 5.1 can be applied and (14) must be added to model (LP). The conditions in (14) can
be handled algorithmically. The optimal solution for w* and t* will automatically satisfy the
*
w
conditions (10) of the original problem after the substitution x = x =
*
.
t*

5.5 Numerical results

We used several data sets from practice and compared the solutions of the former mass
fraction based model (MFM), with the solutions derived from the volume fraction based

model (VFMLP). In all cases some property p~ ∈ pvol was optimized. Characteristics of the
cases are summarized in Table 5.2.

The first column specifies the type of case (i.e. case number, minimization /
maximization problem, optimized property and beating-dependent BD or beating-
independent BI property). The second, third and fourth columns refer to the available
number of celluloses, the relevant number of properties and refiner numbers, respectively.
The revolution interval for each refiner is indicated in column five. The columns six and
seven indicate the lower- and upper bounds on the flow constraints (i.e. the fractions of the
contributions to the final furnish). All cases are based on step sizes of 50 units in the
revolution interval of column five. The problem size, i.e. the numbers of constraints (m),
variables (n) and binary variables (Nr_bin) of the MFM are given in the last three columns of
Table 5.2, respectively.

104
 Mixed integer (0-1) fractional programming in paper production industry

Table 5.2. Case characteristics

min (max) /

Revolution
Case_nr /

Prop_nr /

Nr_prop

UB-flow
LB-flow
BD (BI)

interval
Ref_nr

Nr_bin
Nr_cel

n
0 / min / 7 / BD 7733 8194 482
16 21 29 0 – 12000 0.7
30 0 – 12000 0.7
1 / min / 7 / BD 9050 9571 563
16 23 27 2000 – 6000 1.0
29 0 – 12000 0.7
30 0 – 12000 0.7
2 / min / 11 / BD 1239 1453 242
5 9 25 0 – 6000 0.1 1.0
26 0 – 6000 0.2 1.0
3 / max / 12 / BD 1651 2010 402
4 29 25 0 – 10000 1.0
26 0 – 10000 1.0
4 / max / 53 / BI 9074 9571 563
18 23 27 2000 – 6000 1.0
29 0 – 12000 0.7
30 0 – 12000 0.7

The solutions for all cases of Table 5.2 are given in Table 5.3. The abbreviations
used in the second column (Model) of Table 5.3 refer to the type of model used (i.e. mass-
or volume fraction based model). The columns 3 to 8 refer to the optimal objective value
(Obj-value), the time needed to solve the problem (cpu), refiner number (r), chosen beating
level (b) in the revolution interval, the selected cellulose number (c) and the fraction of
cellulose c in the total mixture (xc,r,b). All values in the column (xc,r,b) are expressed in, or
converted to, mass fractions. The last column (Furnish) shows the share (i.e. percentage) of
each refiner flow to the total furnish. Table 5.3 shows that cpu-times for the volume fraction
based model (VFMLP) are less for all cases. Except for the third case (3/max/12/BD), the
chosen beating levels are substantially different for the MFM and VFM. The latter statement
holds in particular if we take the refiner flows (see Furnish) into consideration too.

105
Chapter 5

Table 5.3. Optimal solutions using two models

Case Model Obj-value cpu r b c xc,r,b Furnish

0 / min / 7 / BD MFM 352.50 58.1 29 6100 36 0.65

76 0.05 0.70
30 3500 33 0.27
36 0.03 0.30
VFM 56.74 9.2 29 7200 31 0.15
36 0.47 0.62
30 2600 52 0.09
53 0.06
68 0.23 0.38

1 / min / 7 / BD MFM 355.80 1008.9 27 3000 33 0.20

44 0.05 0.25
29 6650 31 0.15
36 0.44
39 0.04
44 0.05 0.68
30 850 36 0.03
39 0.04 0.07
VFM 56.24 13.6 27 3000 33 0.04
44 0.05 0.09
29 1000 31 0.02
36 0.09
39 0.05
44 0.05 0.21
30 6200 31 0.60
36 0.06
39 0.04 0.70

2 / min / 11 / BD MFM 91.46 6.8 25 6000 29 0.38 0.38

26 6000 54 0.62 0.62
VFM 11.54 3.9 25 3000 29 0.70 0.70
26 6000 52 0.30 0.30

3 / max / 12 / BD MFM 1.61 7.8 25 0 61 0.43 0.43

26 3000 31 0.45
53 0.12 0.57
VFM 0.17 1.9 25 0 31 0.12
61 0.11 0.23
26 3000 31 0.72
53 0.05 0.77
Continued on the next page

106
 Mixed integer (0-1) fractional programming in paper production industry

Case Model Obj-value cpu r b c xc,r,b Furnish

4 / max / 53 / BI MFM 7.29 38.2 27 5150 31 0.04

36 0.44
43 0.24
44 0.05 0.77
29 1950 31 0.06
39 0.04
44 0.05 0.15
30 7250 39 0.04
43 0.04 0.08
VFM 0.93 9.0 27 3050 31 0.24
33 0.04
36 0.07
44 0.05 0.40
29 2200 36 0.13
39 0.04
44 0.05 0.22
30 7300 31 0.34
39 0.04 0.38

Further analysis and observations of the results in Table 5.3 are summarized in
Table 5.4. The second column in Table 5.4 refers to the difference in the optimal beating
level of the refiners between the two models, MFM and VFM. The degree of difference is
expressed in four qualitative expressions, i.e. “strong” (represented by ++ ; more than 1000
units), “substantial” (represented by + ; between 500 and 1000 units), “moderate”
(represented by ± ; between 0 and 500 units), and “none” (represented by −).

The fractions in the third column of Table 5.4 (Furnish) show the differences
between the compositions of the mixtures using the two models. The latter is expressed by
two subcolumns: the “common” fractions in MFM and VFM, i.e. the total fraction of
celluloses selected by both models independent of the refiner choice and beating level:
 VFM 
∑ min∑∑ x ∑∑ x
MFM
c ,r ,b 
c ,r ,b
, , and the total fraction of completely different celluloses
c r b r b

(“other”) in the furnish of the VFM.

For example, in the case “3/max/12/BD” the fraction in the column “common” equals
0.11 + 0.45 + 0.05 = 0.61 for c=61, c=31 and c=53 respectively (see Table 5.3). The
fraction of the furnish denoted by “other” in Table 5.4 equals zero because both models
select identical celluloses. For the case “0/min/7/BD” in Table 5.4 the column “common”
fraction equals 0.47 for c=36. The total fraction of “other” celluloses for the VFM in Table 5.4

107
Chapter 5

equals 0.15 + 0.09 + 0.06 + 0.23 = 0.53 for c=31, c=52, c=53 and c=68 respectively (see
Table 5.3).

Table 5.4 shows that in all instances at least (100-61=) 39% of the furnish (sub
column “common” for 3/max/12/BD) is different for both models.

The last two columns in Table 5.4 show the results of the reverse solutions (i.e. the
optimal furnish for the VFM fixed in the MFM and vice versa). Except for one case, all
reverse solutions are infeasible in the alternative model (i.e. objective value “-- “ ). The
optimal objective values of the MFM and VFM are given between brackets in column four
and five respectively.

Table 5.4. Analysis of the solutions using two different models

Difference in beating levels Furnish Reverse solution

per refiner_nr
VFM in MFM MFM in VFM
25 26 27 29 30 common other status objv. status objv.

0 /min / 7 / BD ++ + 0.47 0.53 Inf. -- Inf. --

(352.5) (56.74)

1 /min / 7 / BD − ++ ++ 0.52 0 Inf. -- Inf. --

(355.80) (56.24)

2 /min / 11 / BD ++ − 0.38 0.30 Feas. 156.56 Inf. --

(91.46) (11.54)

3 /max / 12 / BD − − 0.61 0 Inf. -- Inf. --

(1.61) (0.17)

4 /max / 53 / BI ++ ± ± 0.48 0.04 Inf. -- Inf. --

(7.29) (0.93)

5.6 Discussion
Although we focussed in particular on the reformulation and solution approach of an OR-
model inside a DSS, it should be mentioned that the core of the system consists of three
main building blocks: a user interface, a simulation- and an optimization routine.

The first release of the DSS was handed over in 1990. At that time the development
and use of Windows applications on personal computers was just evolving. From the

108
 Mixed integer (0-1) fractional programming in paper production industry

beginning it was clear that the user interface is the lubricant between decision-makers on
the one the hand and the underlying database for data storage, the simulation and
optimization routines on the other hand. Between 1990 and 2005 the user interface hardly
changed and desired major changes were postponed. Finally a clear picture and blueprint
emerged for new features and improved ease of control. Unlike the previous release, the
final development of the user interface was outsourced to a software company. The latter
secured the inevitable maintenance for continuous use of the DSS in daily practice.

The simulation module provides a fast and systematic tool to support understanding
and insight regarding the impact of (technical) settings on all properties of a paper grade.
Based on recipes that were stored in the past, the simulation module enables end-users to
study the impact of changes in a recipe (i.e. the contribution of different combinations of raw
materials and additives in the pulp flows) and/or to study (altered) settings of the technical
equipment (i.e. number of refiners, beating intensity and pulp flows) on the final property
values. Various indicators are immediately calculated and visualised. In this way decision-
makers become aware of their trade-offs between various targets. The simulation module
makes clear how difficult it is to find a (feasible) solution that meets all requirements. On the
other hand, upper and lower limits on property values or settings of the technical equipment
are rarely treated as hard constraints in practice. Depending on the case to study, upper
and lower bounds are mainly seen as aspiration levels rather than hard limits.

The optimization module fosters “out of the box thinking”. It provides a powerful tool
to find feasible solutions and the best (surprising) recipes for any available set of raw
materials. Moreover, it provides an innovative way of decision support for purchasing (new)
pulps on the market, for assigning available pulps to different paper grades and for attuning
available stock levels of raw materials to changing production targets for different paper
grades. The results of the optimization routine are mainly used to obtain alternative recipes
for different paper grades. Usually, these recipes are adapted to daily practice in the
simulation module. Tests by practical experience showed that the tendencies predicted by
the system fit very well with the final properties of the paper grades on a paper machine.

In the past twenty years the DSS has become a valuable, regularly used resource
which played a significant role in all kind of projects. Nowadays, the DSS supports
significantly more mills and, depending on the plant, revolution intervals [LL, … , UL] can
vary by choosing different values for the lower limit LL, the step size s and upper limit UL of
the refiners, i.e. [LL, LL+s, LL+2s, …, UL-s, UL]. Although the decision to apply an iterative
approach of stepwise refinement for revolution intervals (i.e. increasing LL, decreasing s
and decreasing UL) in successive runs was initiated from a computational point of view (i.e.
to reduce the number of binary variables in model (4)-(10)), it turns out that especially this
approach is of unexpected and remarkable importance for practice. The approach fosters
understanding and enables end-users to study the impact of different combinations of raw
materials at different technical settings in successive runs on property values.

109
Chapter 5

As another iterative procedure for purely technical reasons was not preferred, the
focus to solve the mixed integer (0-1) fractional programming problem was directed to a
non-iterative approach that takes advantage of the problem characteristics and exploits the
special structure between the original non-linear mixed integer model and a linear
reformulation. The proposed solution procedure in Section 4 neither needs an iterative
procedure nor a special purpose algorithm. From a computational point of view the
approach turns out to be very effective. The calculation time for optimizing volume-fraction-
dependent properties is on average even faster than optimizing mass-fraction-dependent
properties. The experimental results show that the distinction between p ~∈ p and
mass
~∈ p
p vol improved the added value of the optimization routine in the DSS because the
~ ∈ p without any loss of
generated solutions meet the true physical requirements in case p vol

computational efficiency.

Determination of the decision-maker’s preferences between different property

values of a final paper grade is currently achieved by interaction between the decision-
~ are translated
maker and the system. Objectives (i.e. targets values) for properties p ≠ p
into lower and upper tolerance limits Lp and Up in (7a) and (7b). In each run a single, not
necessarily the same, property function f ~p is optimized in (4) while (re)setting lower and
upper tolerance limits on other properties. However, tolerance limits in (7a) and (7b) are
vague or imprecise in practice. These values are merely considered as aspiration levels
and rarely treated as hard values. Moreover, frequently several valid combinations of
aspirations levels exist. Additional research is needed to find approaches that reach a
higher overall aspiration level in the initial stage of the iterative solution process. If
imprecise aspiration levels are introduced to different valid combinations of property values,
the problem turns into a fuzzy, multi-choice, multi-objective, mixed integer (0-1) FP problem.
This problem cannot be solved by applying conventional linear goal programming
techniques (Chang 2007). The author proposed approaches that provide a way to solve
multi-choice aspiration levels in a linear programming context. Chakraborty and Chandra
(2005) approached a blending problem with imprecise specifications as a multi-criteria
decision-making problem and applied fuzzy set theory while Ahlatcioglu and Tiryaki (2007)
introduced interactive fuzzy programming approaches to obtain an overall satisfactory
balance for linear FP problems. The studies (Chakraborty and Chandra 2005; Ahlatcioglu
and Tiryaki 2007; Chang 2007) may be a starting point for new approaches that reach a
higher overall aspiration level in a mixed integer (0-1) FP context.

5.7 Conclusions
We focused on the impact and relation between progressive physical insights and desired
new functionalities from management on an OR-module in paper production industry. The

110
 Mixed integer (0-1) fractional programming in paper production industry

choice of management to upgrade the system for future decision support in new decision
environments may be indicative for the added value and validity of the system in practice.

The experimental results show that the generated solutions are effective and more
accurate than formerly used mass fraction based solutions as they meet the true physical
requirements. The described extension of the classical reformulation approach by Charnes
and Cooper (1962) for a more general class of mixed integer (0-1) FP problems is, to the
best of our knowledge, a novel contribution in (0-1) fractional programming. Moreover, it
neither requires an iterative evaluation of a function in commonly applied parametric
approaches for fractional programming problems, nor a special purpose algorithm. The
branching concept of Beale and Tomlin (1970) may even be available in modern, state-of-
the-art, mathematical programming packages. This broader availability contributes to the
adaptability of the system in practice. Inevitable application-oriented maintenance in the
future will hardly be disturbed by locally developed (special purpose) solution techniques.

Without any value judgement on the (theoretical) value and progress in the field of
(mixed integer) FP, we showed that the motivation to solve a real-life mixed integer FP
problem can provide the basis for a new approach in a new context that has an added value
of its own, even outside the given application area. Any mixed integer 0/1 linear FP problem
that contains common constraints like (6) (i.e. out of a set of decisions, at most one decision
variable may be positive) can be solved by the proposed combination of methods.

Future research and improvements will focus on two main issues, i.e. to contribute
to an on-going trend in paper production industry to use alternatives and additives for
predominant expensive wood fibres as raw materials and secondly to optimize various
conflicting MFM and/or VFM properties simultaneously.

Acknowledgement
The author gratefully acknowledges the valuable comments and suggestions on earlier
versions of the paper by all anonymous reviewers, including the associate editor who
processed the article and lifted the review process to a stimulating and constructive
experience. Thanks to all involved employees of the R&D department in pulp and paper
industry for the fruitful discussions we had in all these years of collaboration.

111
Chapter 5

112
 Chapter 6

General discussion

Operations Research (OR) is preoccupied with efficiency, not effectiveness. Effectiveness is

evaluated efficiency. In other words efficiency is concerned with doing things right; effectiveness
with doing the right thing (Ackoff 2001).
Chapter 6

6.1 Introduction
Literature shows that research in the field of Decision Support Systems (DSS) enjoyed
its strongest growth in the first two decades after its inception in 1971. Since its peak in
1994 there is a consistent decline of annual DSS publications (Eom, Lee et al. 1998;
Arnott and Pervan 2005; Eom and Kim 2006; Arnott and Pervan 2008). Barely four
decades after its birth, Carlsson and Turban (2002) and Arnott and Pervan (2008)
evidenced a trend in which the term DSS matured to a point of losing its identity and
might disappear as a stand-alone field. Arnott and Pervan (2008) stated that the
practical contribution of the broad field of DSS research, which includes model-driven
DSS, faces a crisis of relevance due to a long-term issue, i.e. the tension between
academic rigor and professional relevance. The reviews of Framinan and Ruiz (2010)
and Mula, Peidro et al. (2010) confirmed the identified gap between theory and the use
of (mathematical programming) models in practice. Arnott and Pervan (2008) defined,
among others, professional relevance (i.e. the disconnection of DSS research from
practice), case study research (currently under represented), low industry support, and
the presence of DSS in ‘A’ journals other than the journal ‘Decision Support Systems’ as
key issues for the field to focus on in the future.

As the field of decision support systems aims to be an application-oriented

discipline, the strategy of what is referred to as “application-driven theory” (Cooper and
McAlister 1999; Cooper 2005) is taken as the preferred approach for this thesis.
“Application-driven” refers to a bottom-up approach which means that the relevance of
the research should both be initiated and obtained from concrete problems in real-life
environments. The intended successful use of the proposed approaches should, where
possible, be represented by tests of adequacy. Simultaneously, the contribution to
“theory” aims to be a recognizable part of the research effort. This implies that obtained
understanding and insights from problems in practice can be generalized to and provide
the basis for new approaches. The following two-sided research objective was defined
to reflect this goal.

To support medium- to short-term planning problems by optimization-based models and

solution techniques such that:

i) The applicability and added value of (prototype) systems is recognized and

carried by decision-makers in practice
ii) The proposed approaches contribute to knowledge, understanding and insights
from a model building and solving point of view.

In Chapter 1 we described the focus of the research which, resulted in five

recurring research premises. This sharable set of premises constitutes the connecting
link between the general objective and its translation into four research questions
(RQ’s), which are all addressed in the preceding chapters. Table 6.1 shows the relation
between the research premises and their contribution to the different chapters. We
merely show the key elements of the research premises in Table 6.1 and refer to
Chapter 1 for their complete description.

114
 General discussion

Each chapter refers to at least two premises for different planning issues in the
supply chain planning matrix (SCPM) of Figure 1.2. Besides the first overarching
premise, i.e, professional relevance and applicability, all other premises refer to model
building and/or solving in optimization-based DSS. To make the research premises
visible, their key elements will be printed in italics in the next sections.

Table 6.1 Relation between research premises (Pm) and chapters (Chn)

Premise Ch2 Ch3 Ch4 Ch5

Professional relevance
P1 ( )
and applicability

P2 Aggregation
Decomposition and/or
P3
reformulation
P4 Vertical integration

P5 Horizontal integration

The first goal of this final chapter is to reflect on the RQ’s and to summarize the bare
findings of each chapter in Section 6.2. In Section 6.3 we aim to take some distance
from the RQ’s. The main goals of Section 6.3 are i) to position the main findings in the
current context of research and literature which is for instance particularly relevant for
the studies in Chapter 2 and 4, and ii) to take the research premises as a guideline for
an integrated discussion of the findings. Both Sections 6.2 and 6.3 provide the basis for
the last goal of this chapter, i.e. to summarize the main conclusions and to define some
directions for future research (Section 6.4).

6.2 Reflection on research questions and main findings

In Chapter 1 we formulated four research questions (RQ’s). This section describes how
the RQ’s are addressed in the preceding chapters and what we learned from the
studies. The related main findings in each chapter will be summarized in bullet points in
each subsection. Findings will be grouped into three main categories: i) model-building,
ii) model-solving, and iii) professional relevance and applicability.

6.2.1 Research question 1

Chapter 2 addresses RQ1: “How to apply aggregation, decomposition and reformulation
in model-based DSS at planning and scheduling level such that the aspect of decision
support is recognized and appreciated by decision-makers in practice, and which level

115
Chapter 6

of aggregation is needed to integrate production planning (i.e. lot-sizing) and scheduling

problems in a single model?

The chapter consists of two parts. Part I refers to a case study for production
planning and detailed scheduling problems at a bottleneck production facility in food
processing industry. Part II refers to a literature research for integrated production
planning (i.e. lot-sizing) and scheduling.

The aim of the case study was to develop, implement and test a pilot DSS, able
to deliver solutions recognized and carried by decision-makers in practice. The latter aim
implies that a straight-forward aggregation on time, product type, resources or product
stage (Wijngaard 1982) was not preferred. The key to develop a solvable approach for
regular use was to identify and take advantage of specific problem characteristics.
Experience from practice offered a way to cluster the numerous jobs of the order book
into a restricted number of families of jobs. Each cluster of jobs shared a set of
comparable operations with comparable machine setups. Whenever a production line is
prepared to produce an item in a family, all other items in the same cluster can be
produced with negligible changes in setups. In order to reduce the complexity of the
problem, planning tasks were separated i.e. decomposed into two hierarchical levels: (i)
production planning over a short- to medium-term rolling horizon, and (ii) sequencing of
jobs at a daily level second. However, decomposition was still insufficient to solve the
daily problems within an acceptable time frame. Clustering while retaining information at
order level could be exploited in a reformulation approach by the inclusion of (combined)
generalized- and variable upper bound constraints which gave very tight lower bounds
and sparse search trees.

Decision-making at this production level implies the consideration of several,

mostly incommensurable, objectives. The suggested solution, i.e. to assign penalties
and/or weights to different criteria in a single objective function, was effective to combine
different goals in this case study. For the acceptance of the generated production plans,
it was of particular importance to generate recognizable and valued production
sequences for the shop floor. At this lowest (i.e. short-term) level, several logical
sequencing rules from practice were considered and programmed, all ranked in a given
order of importance.

The main benefit of the approach is the constant and initial quality of the
generated plans including the time needed to generate these schedules. Hence,
decision-makers could i) postpone their planning tasks, ii) conveniently cope with rush
orders or planned maintenance and iii) easily generate alternatives or revised plans
when unforeseen disturbances occur. Moreover, the graphical presentation and
overview of the planned working schedule enabled order acceptance to make use of
remaining capacity.

Basic understanding on how the production schedules are calculated turned out
to be an important issue for acceptance and usability of the generated plans. Decision-
makers usually have more information at their disposal than is modelled in a DSS. Not

116
 General discussion

all observations in practice can be captured in a (simplified) model particularly in a

barely stable decision environment. Therefore decision-makers need the opportunity to
modify automatically generated plans manually and use human judgement and
experience such that the solution is tuned to the actual situation. Hence, the DSS should
not be considered as an optimizer but rather as a tool for generating high-quality plans
to be used for further analysis. Within this context the various utilities of a user-friendly,
graphical, and fully interactive user interface, was of major importance.

To summarize the findings of the case study in Part I of Chapter 2:

Model-building (RQ1, Part I)

• The separation of planning and scheduling implies that some (unknown) capacity
must be reserved at planning level in order to compensate capacity losses due to
changeovers at sequencing level.
• Decomposition without aggregation was insufficient to solve the generated
problems within an acceptable time frame. A combined approach based on i) taking
advantage of specific problem characteristics (i.e. a case-based clustering
procedure instead of aggregation) and ii) the identification of special model
structures (i.e. a reformulation which offered the inclusion of combined generalized-
and variable upper bound constraints) resulted in very tight lower bounds and
sparse search trees.

Professional relevance and applicability (RQ1, Part I)

• The research clearly demonstrates the anticipating value of earlier case-based DSS
research, funded and tested by industrial practice. Meanwhile, the applied principle
of decomposition at production level became a commonly accepted starting point in
the framework of APS in Figure 1.2 (Stadtler and Kilger 2008; Stadtler, Fleischmann
et al. 2012).
• A user-friendly, graphical, and fully interactive user interface is of major importance
for both the development and adoption of automated systems in practice.
• Personalized and customized modules are particularly important on lower decision
levels

The proposed approach in Chapter 2 may reduce the computational burden and
provide adequate decision support in specific cases, but it is hardly a generic solution for
the intended vertical integration between lot-sizing and scheduling. Stadtler and Kilger
(2008) stated for instance, if products (lot-sizes) have to compete for scarce resources
(e.g. flow lines with sequence dependent setup costs and times), a separation into two
planning levels is inadequate. Due to specific characteristics in FPI, e.g. non-triangular
setups and product decay, the need for simultaneous lot-sizing and scheduling may be
even more relevant for this branch of industry.

117
Chapter 6

As the case study in Chapter 2 was based on an earlier study, a literature research
on modelling developments for simultaneous lot-sizing and scheduling was carried out
too. The research was restricted to contributions that are directive for the identified
problem characteristics, i.e. setup carry-overs, sequence dependent setup costs and
times, relaxation of the triangular setup conditions and product decay.

According to literature, two main classes of models can be distinguished, i.e. Small
Bucket (SB) and Big Bucket (BB) models. In SB models, the planning horizon is divided
into a finite number of small time periods. Conversely, in BB approaches the planning
horizon is divided into longer periods, usually of equal length, and in each period
multiple products may be produced. As a consequence, SB models are usually
associated with short-term planning horizons and BB models with medium term planning
horizons. We noticed a tendency in literature in which special Big Bucket (BB) models
are proposed for short-term time horizons too; particularly within the (broad) context of
sequence dependent changeovers and triangular setup conditions. Despite of the
aggregation in time, these BB models, including intermediate variants like the General
Lot-sizing and Scheduling Problem (GLSP) or block planning approaches, consider both
the size and the production sequence of lots within these larger time-intervals.

To summarize the findings of the literature research in Part II of Chapter 2:

Model-building (RQ1, Part II)

• There exists a noticeable trend in simultaneous lot-sizing and scheduling in which
Big Bucket (BB) approaches or hybrid variants are preferred to Small Bucket (SB)
models.
• Non-triangular setups are hardly considered, particularly not in SB models.
• A substantial number of proposed SB models introduce an artificial product to
represent idleness of resources. For these models the changeover matrices must
comply with very strict, usually unrealistic, conditions to cope with sequence-
dependent changeover times and non-triangular setups. If these conditions are not
met, the setup state of the production facility is not correctly carried over across the
boundaries of idleness.
• Block planning approaches can be regarded as a practical variant of the General
Lot-sizing and Scheduling Problem (GLSP). However, in the concept of block
planning the production sequence of (variable) batch sizes is pre-defined (Lütke-
Entrup, Günther et al. 2005; Günther, Grunow et al. 2006; Bilgen and Günther 2010;
Baumann and Trautmann 2012). These approaches may be difficult to apply if the
triangular setup conditions do not hold.
• Surprisingly little research has been devoted to include issues of product decay in
traditional lot-sizing and scheduling models.
• In contrast to BB models, SB approaches offer the timeframe to attune short-term
physical distribution planning to production planning and scheduling, e.g. by
assigning demand to specific time slots in a 24-hours production environment.

118
 General discussion

• Although the separation between production planning (i.e. lot-sizing) and scheduling
in successive hierarchical phases is commonly accepted, e.g. in APS software,
these planning steps are closely linked areas which should (ideally) be considered
simultaneously.

The findings of the literature review laid the foundation for additional research on a
complete vertical integration of planning and scheduling tasks. The goal was to develop
a single model for both planning tasks simultaneously, to study its behaviour, the
complexity, and to investigate the impact of non-triangular setups and product decay on
optimal production schedules.

6.2.2 Research question 2

Chapter 3 addresses RQ2: “How to integrate production planning (i.e. lot-sizing) and
scheduling problems in a single model, such that common assumptions regarding the
triangular setup conditions are relaxed and issues of product decay and limited shelf
lives are taken into account?”

Although BB approaches may have a computational advantage, Chapter 3 argues

that segmentation of the planning horizon (i.e. aggregation over time) is a key issue for
simultaneous lot-sizing and scheduling in food processing industry. Defining large time
intervals in BB approaches implies that the general principle of optimality for lot-sizing
may unfoundedly disappear from sight. Moreover, product decay is primarily associated
with the “age” of manufactured products and consequently relates to the segmentation
of the time-horizon. Therefore, two consistent SB models are developed to demonstrate
the impact of non-triangular setups and product decay on the generated solutions.
Small-scale examples are used to demonstrate the impact of minor changes in the
balance between inventory-holding and changeover costs.
The developed models are potentially very large formulations. Basic complexity
analysis for the developed models shows that particularly the segmentation of the time
horizon in SB approaches has a substantial impact on problem sizes. Solving the
developed models for large (i.e. real-size) problem instances, requires effective and
efficient approximations techniques. Exploratory research was conducted based on a
Relax-and-Fix (R&F) heuristic in which the principle of decomposition was applied to the
solution procedure. Numerical results of small- to medium-sized problem instances are
presented.

To summarize the main findings of the study in Chapter 3:

Model-building (RQ2)
• If the objective for simultaneous lot-sizing and scheduling should include the best
compromise between total setup costs and total inventory-holding costs, a time-
oriented aggregation (like in BB models and its variants) easily disrupts the general

119
Chapter 6

principle of optimality for lot-sizing. As a result, total inventory-holding costs are

underestimated and lot-sizes become too large.
• Product decay in food processing industry is primarily associated with the “age” of
manufactured products, which argues the need to capture the precise moments of
production. If multiple batches are planned in larger time intervals, it implies that all
lot-sizes in each period share the same moment of production. As a consequence,
the age of manufactured products on stock is underestimated.
• Product decay has an impact on the remaining shelf life of products. This aspect
was included by an age-dependent component in the inventory-holding costs.
Numerical results show how a small change in the balance between inventory - and
changeover costs may generate significantly different solutions, especially when the
triangular setup conditions do not hold.

Model-solving (RQ2)
• The developed models are potentially very large formulations. Computation times
grow very fast, both with the number of products N and (particularly) with the number
of periods T in the planning horizon.
• Although R&F algorithms in (mixed) integer programming literature are commonly
presented as forward procedures, a backward R&F procedure is favourable for
simultaneous lot-sizing and scheduling. Demand matrices for SB models are usually
sparse (i.e. many, if not most entries of the matrix are zero). Numerical tests confirm
that in a forward procedure, production will be postponed in early iterations. If
capacities are tight, the concept of fixing production and idle time at their optimal
values from previous iterations will easily lead to infeasible solutions in a forward
solution procedure.
• The quality of the R&F solutions is promising at manageable computational effort.
However, solving real-size problem instances may not be possible yet. Nevertheless,
the availability of a correct MP model for the given problem description offers at least
the possibility to measure the quality of small- to medium-sized problems solved by
any (other) heuristic.

Professional relevance and applicability (RQ2)

• Although the boundaries between planning (i.e. lot-sizing) and scheduling are fading
in literature, there is a need for more practical cases of simultaneous lot-sizing and
scheduling, particularly in food processing industries.

6.2.3 Research question 3

Chapter 4 addresses RQ3: “How to model and solve an integrated planning problem
between procurement and production, both on a midterm and short-term planning level,
in an inter-organizational supply chain?

120
 General discussion

Both production and distribution planning of (end) products are part of the APS
framework and issues of integration between both phases have been the concern of
research. However, surprisingly little research has been devoted to issues of horizontal
integration between procurement and production. The lack of both a midterm distribution
and a short-term transportation module between procurement and production in the APS
framework of Figure 1.2 may be an illustrative observation within this context.
Comparable problems of coordination may manifest between procurement and
production particularly in push-oriented supply chains. The case study in Chapter 4
focused on horizontal coordination and integration between the phases procurement
and production for a milk collection problem in practice, which is of particular importance
in inter-organizational supply chains. The aim was to develop a pilot DSS that lifted
decision support for a “weaker” partner in a food supply chain (i.e. a stakeholder who is
not in charge of planning process) to a higher level, and to illustrate the importance of
horizontal coordination and integration between the phases procurement and production
in an APS framework.

Initially, the case was presented by the stakeholder as a complex, daily vehicle
routing problem. Problem analysis revealed that the problem can be classified as an
extension of the Periodic Vehicle Routing Problem (PVRP). However, the basic PVRP in
literature assumes either pickup or delivery operations, not both simultaneously like the
case study in Chapter 4.

In order to solve the PVRP in a practical setting, the complete problem was
decomposed into more tractable subproblems on different levels, i.e. to separate the
daily routing problem from a new medium-term planning problem. On the higher
planning level, numerous supplier farms were aggregated such that total supply within a
cluster met (multiple) vehicle loading capacities. Based on limited storage capacities at
supplier level and additional requirements for the freshness of raw milk, feasible
collection frequencies (rhythms) for aggregated supply were introduced (see Table 4.1).
The geographical location of supplier farms was the starting point for aggregation on
supply level. A model was developed to generate stable collection schedules. The
continuous supply of relatively small amounts from many suppliers had to be balanced
with strict delivery conditions at processing level (i.e. large amounts of raw milk
scheduled to arrive at processing facilities on a limited number of fixed days in the
planning horizon). The aim of the model was to assign a single collection rhythm to each
cluster such that the total, weighted deviation (i.e. surplus and shortage) of desired
processing levels at fixed days in the planning horizon was minimized.

The computational complexity of the problem could be reduced by taking

advantage of specific, application-based properties and to exploit them in a specific
branch-and-bound scheme. The improved computational efficiency of the branching
concept made it possible to solve the generated problems exactly for real-size problem
instances.

The applied aggregation on the higher planning level turned out to be very
beneficial for the required disaggregation at the lower planning level, i.e. the daily

121
Chapter 6

vehicle routing problem. Once supplier farms were geographically grouped into clusters
and the aggregated supply within a cluster was assigned to a single collection rhythm
with fixed collection days, the (initial) daily routing problem was considerably easier to
solve for vehicle schedulers.

Besides the added value on the mid- and short-term level, the planning model
turned out to be a very helpful strategic tool for the cooperative association at supply
level (i.e. the “weaker” partner). When periodic delivery conditions are set by stronger
partners in an inter-organizational network (in this case stakeholders at processing level)
the generated plans can be used effectively by a weaker partner (e.g. for their regular
negotiations with both processing and transportation companies).

The visualization of (modified) plans including the possibility to store plans over
the year enabled decision-makers to ‘optimize’ their performance with respect to his or
her planning tasks. Within this context, the various facilities of a user-friendly and
interactive man-machine interface were essential. The user interface was divided into an
input, planning, simulation and analysing part. Changing the data, e.g. moving supplier
farms to other clusters or changing the milk collection rhythm for a cluster was possible.
However, the impact of any modification of the data in the simulation module was
immediately visualized by several (conflicting) indicators in the output screens, both on
supply and demand level.

To summarize the findings of the study in Chapter 4:

Model-building (RQ3)
• The case study demonstrated that an additional planning phase (i.e. distribution)
between procurement and production contributes considerably to horizontal
integration in the SCPM, particularly in a push-oriented, inter-organizational food
supply chain.
• The main problem to solve was a special variant of the Periodic Vehicle Routing
Problem (PVRP) which concerns pickup and delivery operations, simultaneously.
The focus for this variant of the PVRP in practice should be on decomposition of the
problem into more tractable sub problems on different hierarchical levels.
• Although aggregation on higher planning levels is often associated with an
(undesired) loss of information, the applied aggregation at medium-term planning
level was very beneficial for the (inevitable) disaggregation at the lower planning
levels.

Model-solving (RQ3)
• Although literature on the tactical PVRP focuses primarily on heuristic methods
(Mourgaya and Vanderbeck 2007; Francis, Smilowitz et al. 2008; Baldacci,
Mingozzi et al. 2011), we showed that real-sized problems can be solved using

122
 General discussion

exact methods at the highest level in the PVRP. Francis, Smilowitz et al. (2008)
confirmed the latter finding in their book on vehicle routing problems and refer to the
study in Chapter 4 as an example for the applicability of exact methods.
• We focused on the algorithmic side for the given case. Compared to a conventional
branch-and-bound scheme for integer programming, we showed that there is no
advantage in the SOS1-branching scheme itself. The potential sizes of search trees
for branching on single variables or sets of variables are equal. Consequently, any
potential benefit of the SOS1-concept is due to the efficiency of the search
procedure.
• We showed that the efficiency of the SOS1-concept primarily depends on a
(re-)ordering procedure of the variables within the sets rather than on the weights
associated with each variable in the set.
• Numerical tests confirm that substantial computational advantages can be gained
by applying an SOS1-based solution procedure, provided that a (re-)ordering of the
variables within the sets is considered.
• The case study showed that an efficient use of the SOS1-based solution procedure
is not necessarily restricted to problems with supplementary model conditions. We
showed that a natural ordering of the variables within the sets (Williams 1990; ILOG
2009), is not necessary to make their use worthwhile and/or applicable in a broader
context. In addition to the latter statement, we also refer to the reflection on RQ4 in
the next Section 6.2.4.

Professional relevance and applicability (RQ3)

• The case study demonstrated that the adoption and added value of DSS in practice
can be increased by including both a simulation module and scenario management
tools.
• The necessity of a (separate) reference row or weights associated to the variables
might be omitted in future implementations of the SOS1-based branching scheme.

6.2.4 Research question 4

Chapter 5 addresses RQ4: “How to support decision-makers in practice if crucial
properties of end products simultaneously depend on (endogenous) types of raw
materials with different chemical or physical properties and (endogenous) technical
settings of processing units?”

The study refers to a specific characteristic of process industries where technical

settings of processing units have a variable (physical) impact on material flows which in
turn determine the final properties of end products. If the related planning problems are
treated in the context of mathematical optimization, they may lead to (mixed integer)
non-linear problems, which are often hard to solve.

123
Chapter 6

Pressed by changed circumstances and constantly advancing physical

knowledge of large scale pulp and paper production, the goal was to revise and upgrade
an existing, locally used DSS, to a tailored and flexible tool within the enterprise. The
cornerstones simplicity, ease of control, adaptability and completeness from Little’s
seminal paper were taken as a guideline for developing models and techniques that are
actually used in practice (Little 1970; Little 2004).

One of the main concerns from practice was that the principle of a single
overriding objective is too restrictive for future decision support and had to be
abandoned. Management scientists of corporations usually focus on a variety of
objectives. The study revealed that the aimed extension towards multi-objective decision
support, together with new physical insight for calculating properties of end products due
to process operations, had a significant impact on the optimization module. From a
practical point of view, the method to solve the non-linear programming problem should
neither be based on an iterative solution procedure nor a locally developed special
purpose algorithm.

The proposed solution procedure takes advantage of the problem characteristics

and gives rise i) to apply and extend a classical reformulation approach for continuous
linear fractional programming (FP) problems to a more general class of mixed integer
(0-1) FP problems and ii) to exploit the special structure between the original non-linear
mixed integer model and the continuous, linear reformulation by applying the concept of
Special Ordered Sets type 1 (SOS1).

Although Chapter 5 focuses in particular on the reformulation and solution

approach, the DSS consists of four main building blocks, i.e. the user interface, a
scenario manager, a simulation- and optimization routine. The user interface is the
lubricant between decision-makers on the one hand and the underlying database for
data storage, scenario manager tools, simulation and optimization routines on the other
hand. Between 1991 and 2005 the user interface had hardly changed and desired
(major) changes were postponed. Finally a clear picture and blueprint emerged for new
features and improved ease of control and ease of communication. Unlike earlier
releases, the final development of the user interface was outsourced to a software
company. The latter secures the inevitable maintenance (i.e. adaptability) for continuous
use of the DSS in daily practice.

Scenario manager tools were developed to store, structure and analyse multiple
solution scenarios such that it benefits the understanding of underlying patterns. The
simulation module provides a fast and systematic tool to support understanding and
insight regarding the impact of (technical) settings on all properties of a paper grade.
Based on stored recipes in the past (e.g. by the scenario manager), the simulation
module enables end-users to study the impact of changes in a recipe (i.e. the
contribution of different combinations of raw materials and additives in adjustable
fractions in the pulp flows) and/or to study the (altered) settings of the technical
equipment (i.e. number of refiners, beating intensity and pulp flows) on the final property

124
 General discussion

values. From the perception of the end-user, the simulation module also makes clear
how difficult it is to find a (feasible) solution that meets all requirements.

The optimization module provides a powerful tool to find feasible solutions and
the best (unexpected) recipes for any available set of raw materials. Moreover, it
provides an innovative way of decision support for purchasing (new) pulps on the
market, for assigning available pulps to different paper grades, and for attuning available
stock levels of raw materials to (changing) production targets for different paper grades.
The results of the optimization routine are mainly used to obtain alternative recipes for
different paper grades. Usually, these recipes are stored as base scenarios and adapted
to daily practice in the simulation module.

To summarize the findings of the study in Chapter 5:

Model-solving (RQ4)
• The applied combination of methods neither requires a commonly applied iterative
evaluation of a parametric function for (non-linear) fractional programming (FP)
problems, nor a special purpose algorithm.
• The applied concept was easy to implement in a DSS and may even be available in
modern, state-of-the-art, mathematical programming packages.
• Numerical results show that the proposed approach can be applied to problems
with real-life dimensions. In this particular case even without a loss of computational
efficiency.
• We proved that the concept of Special Ordered Sets type 1 (SOS1) can extend a
classical reformulation approach for continuous FP problems to a specific class of
mixed integer (0-1) FP problems.

Professional relevance and applicability (RQ4)

• From a decision-maker point of view, the simulation module exceeds the added
value of the optimization module. It enables end-users to study and explain the
impact of minor (technical) changes. Practical use of the system in real-life shows
that simulation contributes significantly to basic understanding and insights of the
underlying problem.
• The optimization module mainly fosters “out-of-the-box thinking”. Minor changes in
(technical) settings may result in completely different solutions which are less easy
to grasp and accept in practice.
• The study showed once again the importance of solid scenario management tools
for the adoption and regular use of the DSS in real-life.
• The study showed that a broader availability of the SOS1 branching concept in
modern mathematical programming packages contributes to the adaptability of
systems in practice. Inevitable application-oriented maintenance in the future will be
hardly disturbed by locally developed special purpose solution techniques.

125
Chapter 6

The next Section 6.3 provides a wider discussion of findings along the main objective of
the research including the defined research premises in Chapter 1. The main goal of
Section 6.3 is to position the bare findings of Section 6.2 in the current context of
insights, to take some distance and place them in a slightly broader perspective. Based
on our experiences and gained insights in developing model-based DSS we finally
revisit current developments in decision support for industrial practice, i.e. Advanced
Planning Systems (APS) as described in Section 1.2 of Chapter 1.

6.3 Discussion of findings

Given the need for computerized decision support, the objective of this study was i) to
contribute to the applicability of DSS in process industries using model-based
approaches and ii) to acquire additional optimization-based insights and contribute to
new approaches, both from a modelling and solving point of view. Carlsson and Turban
(2002) stated that the term decision support systems may be seen less and less
frequently but its basic concepts, aims, and added value for practice are still valid.

In Section 1.4 of Chapter 1 we motivated why a case-based approach was taken

as a starting point for this thesis. The general goal of this section is to contribute to the
question “How to shape model-based DSS studies such that they actually persuade
decision-makers for using normative approaches for decision support in practice and
simultaneously contribute to existing knowledge, understanding and insights from a
theoretical point of view?”. An answer to this question may be to combine the strong
elements of normative models with (descriptive) observations in practice, i.e. to tune
normative models both to specific situations and to the needs of decision-makers in
practice. The studies in Chapter 2, 4 and (particularly) 5 are illustrative examples within
this context. Starting point in each case was not to focus solely on the most obvious
component in a model-based DSS but to develop blueprints of usable systems in
practice.

A set of five research premises was introduced in Chapter 1, providing the basis
for a recurring link between the general objective and its translation into research
questions. In Section 6.1 we separated the overarching research premise P1
“professional relevance and applicability” from the other premises P2 – P5, all referring
to “model building and/or solving”. The separation of premises into these two headings
will be the starting point for the following discussion. Analogous to the preceding section,
references to research premises are printed in italics.

6.3.1 Model building and solving

Although automated decision support can be very valuable for (programmable) decision
problems, it can easily demonstrate its weakness too in complex decision-making,
particularly with respect to model-driven DSS. The core of that weakness may originate
from OR scientists who are primarily concerned with developing and solving (normative)

126
 General discussion

models, i.e. to identify the best decision to take and to describe how decision-makers
ought to make decisions. According to Ackoff (2001), management scientists are
preoccupied with “doing things right”, but simultaneously may neglect to design models
of what the decision-making process is really about or what decision-makers actually do,
i.e. “doing the right thing”. The study in Chapter 5 on a class of fractional programming
problems may be an illustrative example within this context. Problem classes are often
preferred to be narrowed down by scientists to one specific problem that suits personal
research interests best rather than trying to model and solve the real problem in the
environment in which it is embedded. According to Williams (2013) it is surprising that
comparatively little attention has been paid in literature to the problem of formulating and
building mathematical programming models and deciding when (normative) models are
applicable. According to our view, an application-oriented field like model-based DSS
needs to apply OR knowledge and experience such that systems are developed in
which decision-making based on preferred and/or historical lines of thoughts in practice
are predisposed by adopting combined normative and descriptive approaches with
recognized, favoured added value and outcome. Our aim is to contribute to this major
issue. We take the premises P2 – P5 as guidelines and project them to modelling and
solving the described problems in the previous chapters.

Aggregation (P2), decomposition and reformulation (P3) are commonly applied

to reduce primarily the computational complexity and secondly the need for detailed
data. Aggregation usually includes a loss of information and may be done at the
expense of accuracy. The principles of decomposition and (a priori) reformulation are
much broader and a clear distinction between them is less delineated. Generally,
decomposition schemes are based on breaking up the original problem into smaller,
more tractable subproblems and may refer to the scope of the initial problem, the
proposed model(s), and/or the applied solution technique(s). Liberti (2009) defined a
reformulation of a mathematical program as a formulation which shares some properties
with, but is in some sense better than, the original program. Reformulations are
widespread in mathematical programming and important with respect to the choice and
efficiency of solution algorithms (Liberti 2009). The aim of this section is not to contribute
to a fundamental discussion regarding clear definitions and/or general statements on
when to use aggregation, decomposition and/or reformulation but to demonstrate where
and how (a combined use of) these principles contributed to the studies in the preceding
chapters.

Towards complete vertical integration at production level

A major drawback of aggregation is that it is done at the expense of exactness (Stadtler,
Fleischmann et al. 2012). Aggregated solutions may be difficult or even impossible to
disaggregate into feasible, detailed plans on lower decision levels. Consequently, the
generated solutions may be less recognized and appreciated by decision-makers, even
after a cumbersome (ex post) disaggregation procedure. However, the unavoidable
computational complexity of non-aggregated planning and scheduling models may
neither be an option, even after applying a decomposition approach by separating

127
Chapter 6

planning tasks from scheduling (see Figure 1.2 in Chapter 1). Clear examples within this
context refer to the studies in Chapter 2 and 3.

The average number of (expected) orders in the order book for the case study in
Chapter 2 was too large to solve the problem within an acceptable time frame. At model
construction level a clustering could be exploited in a reformulation approach which
takes advantage of a favourable model structure. Instead of defining continuous
production variables expressed in absolute amounts, production variables were defined
as a fraction of demand at order level. The latter definition and the related inclusion of
both generalized upper bounds (GUB) and variable upper bound (VUB) constraints was
crucial to make the difference between a “weak” and a “strong” model formulation. The
VUB constraints enriched the model formulation in Chapter 2 and induced tight LP
relaxations which tend to give answers that are integer in the binary variables. Several
studies, mostly devoted to the facility location problem, confirmed that the inclusion of
variable upper bounds can give tight lower bounds and sparse search trees (Schrage
1978; Christofides and Beasley 1983; Vanroy 1986). The applied combination of
decomposition, clustering while retaining detailed information at order level, and
reformulation was crucial to solve the problems in Chapter 2 within a few minutes on a
PC.

Unfortunately, the applied combination of clustering and reformulation is not

generically applicable because i) the number of jobs in the order book should be
substantially larger than the number of defined clusters, ii) processing times of the
majority of jobs in the order book should be considerably smaller than daily production
capacity, and iii) only a minority of jobs in the order book may have processing times
that exceed daily capacity. A quadratic penalty function for the starting time of each job
within its feasible time frame (see Figure 2.2) was (in this specific case) sufficient to
prevent an excessive split of jobs over more than a single production day. However, the
overall conclusion with respect to the case study in Chapter 2 is that the applicability of
the approach highly depends on taking advantage of specific, case-based
characteristics. Moreover, the main drawback of a hierarchical decomposition approach
at production stage in Figure 1.2, remains an issue. This issue may be of minor
importance if changeovers are not sequence dependent (Roosma and Claassen 1996).
However, if capacity losses in planning and scheduling problems are caused by
sequence dependent changeover matrices, whether or not including the triangular setup
conditions, the hierarchical decomposition approach easily results in infeasible or
suboptimal solutions (Kreipl and Pinedo 2004; Shah and Ierapetritou 2012).

Complete vertical integration at production level

Although undesired losses of production capacity due to sequence dependent setup-
times could be reduced in the case study of Chapter 2 by penalising the number of
clusters in the time horizon, an unknown portion of daily capacity had to be reserved for
changeovers. Both Stadtler and Kilger (2008) and Stadtler, Fleischmann et al. (2012) in
their books suggested a comparable approach. However, reserved capacity at the

128
 General discussion

higher decision level is either under- or overestimated and may become a serious
problem, particularly in case the triangular setup conditions do not hold.

An obvious way to address such problems is to release the general principle of

decomposition in Figure 2.1 and to accomplish a complete vertical integration (P4)
between planning and scheduling at production stage. The main issue for model-
building and -solving is that integrated models are much more complicated to develop
and usually much harder to solve (Kallrath 2002).

As has been shown in Chapter 3 and confirmed in literature (Salomon, Solomon

et al. 1997; Méndez, Cerdá et al. 2006; Pochet and Wolsey 2006; Stadtler and Kilger
2008), computational effort increases very fast, particularly with the number of time
periods in the planning horizon. This well-known issue turns out to be an incentive for
the construction of integrated models which are based on aggregation in time, i.e. a
segmentation of the planning horizon into large time intervals. Due to a significantly
smaller number of time periods, the related big bucket (BB) models and their practical
variants have a major computational advantage (Méndez, Cerdá et al. 2006; Baumann
and Trautmann 2012). As a result, there exists a trend in simultaneous lot-sizing and
scheduling that moves away from small bucket (SB) to big bucket (BB) models or
intermediate variants. In Chapter 2 and 3 we argue that applying aggregation over time
in model-building at production level is ineffective for both a complete vertical integration
(P4) at production level and the need for horizontal coordination and integration (P5)
between the phases production and physical distribution in the SCPM of Figure 1.2.
Modelling approaches that take a continuous time axis as a starting point, i.e. consider
no time buckets at all, may end up in the most appropriate models but these approaches
will usually result in the largest computational effort (Méndez, Cerdá et al. 2006; Stadtler
and Kilger 2008). Compared to BB models, an SB approach takes the closest
approximation for a continuous representation of time.

If aggregation and decomposition are not eligible at model construction level,

other approaches must be found to address planning (i.e. lot-sizing) and scheduling
problems simultaneously, e.g. reformulation or a decomposition approach applied to
solution techniques. In their book on production planning and mixed integer
programming, Pochet and Wolsey (2006) stated that only a very few reformulations exist
concerning those models. The authors stated that the most commonly applied
optimization approach in solving such problems is to integrate existing approaches for
single-item problems, using a decomposition technique. We conducted exploratory
research on a decomposition-based solution technique. The applied heuristic considers
a decomposition of variable definitions in the time horizon, but it can be changed easily
into a decomposition approach on product and/or constraints level. The main advantage
of an R&F procedure is its broad applicability in mathematical programming. Dedicated,
special purpose algorithms may perform better, but these algorithms usually have to be
redesigned or even abandoned in case certain (minor) features change (Hax and
Candea 1984; Günther and van Beek 2003).

129
Chapter 6

The need for horizontal coordination and integration

In Chapter 3 we argued that an SB approach at production level has the flexibility to take
issues of horizontal coordination between production planning and physical distribution
into consideration. In Chapter 4 we addressed the problem of horizontal coordination
and integration between procurement and production. Fleischmann and Meyr (2003)
stated that the bill of materials (BOM) in consumer goods industries is generally rather
flat, which makes procurement usually unproblematic. According to the authors, mostly
only a few raw materials with a low value are sourced from a handful of suppliers in
consumer goods industry. As a consequence, supplier lead times are relatively short
and reliable. We argue in Chapter 4 that, despite of a flat BOM with reliable supplier lead
times, issues of coordination and (horizontal) integration should not be restricted to
material flows between the stages production and distribution in push-oriented supply
chains. Even if perishable raw materials are produced at a constant level, they are
usually processed in (semi-) batch type production processes, mostly on shared or multi-
purpose equipment.

As shown in Chapter 4, integrated (midterm) distribution and (short-term)

transportation planning between procurement and production level becomes even more
complicated if procurement and processing of raw materials is carried out by different
companies. Stadtler (2005) stated that issues facing an inter-organizational supply chain
are mainly addressed in research areas associated with the integration of individual
organizations. The author concluded that advanced planning across company borders is
still in its infancy. Akkermans, Bogerd et al. (2003) mentioned that it becomes
increasingly apparent that supply chains, rather than individual organizations, compete.
Consequently, there is an increasing demand for collaborative architectures in decision
support software (Akkermans, Bogerd et al. 2003). Stadtler (2005) confirmed that
hierarchical coordination is possible and prevailing in intra-organizational supply chains,
but the real challenge arises in inter-organizational supply chain where hierarchical
coordination is no longer possible. The author stated that the centralistic view of
hierarchical planning underlying today’s commercial planning systems like APS, may
even be a questionable assumption. It might be suitable in intra-organizational supply
chains but not across organizational borders (Stadtler 2005). The case study in
Chapter 4 confirms the latter view. The interests of partners in the chain were clearly
contradictory. As a consequence, the weaker partner focuses on his own planning
domain, is reluctant to share data, and constantly tries to strengthen his position. In
addition, the case study demonstrated the added value of an additional (distribution)
phase between procurement and production in a push-oriented, inter-organizational food
chain (see Figure 1.2).

Model-building for horizontal coordination and integration

From an OR point of view, the problem in the proposed distribution phase of Chapter 4
can be classified as a periodic vehicle routing problem (PVRP). We refer to Golden,
Raghavan et al. (2008) for an overview of papers including the most important advances
of (new) variants in the vehicle routing domain. The extension of the PVRP in Chapter 4

130
 General discussion

concerns the link between the stages procurement and production in the APS
framework by considering both pickup and delivering conditions in the PVRP.

Recently, Baldacci, Mingozzi et al. (2011) confirmed that over the past 30 years,
PVRP literature focused primarily on heuristic methods and no exact methods have
been proposed so far. The authors stated that the PVRP contains many variants and
special cases in terms of objectives or sets of additional constraints, strongly specific to
the application area. The Tactical Planning Vehicle Routing Problem (TPVRP) is most
closely related to the study in Chapter 4. In its general form, the TPVRP is a strategic
model because, in practice, the routes of a solution for a T-day planning period remain
unchanged for several months. Only a very few publications can be found in
optimization literature on solving the TPVRP (Mourgaya and Vanderbeck 2007;
Baldacci, Mingozzi et al. 2011).

A decomposition approach was applied to reduce the complexity of the PVRP.

Literature confirmed that the PVRP is a multi-level optimization problem (Chao, Golden
et al. 1995; Mourgaya and Vanderbeck 2007). Commonly, three levels are defined.
Firstly, allowable visit combinations should be assigned to each “customer”. In the
second level vehicles are assigned to routes, i.e. a classical vehicle routing problem
(VRP). Thirdly, the classical Travelling Salesman Problem (TSP) remains to be solved at
daily decision level. Mourgaya and Vanderbeck (2007) applied a comparable approach
as described in Chapter 4 by ignoring the third problem level of the PVRP too. The
authors confirmed that problem-specific criteria are usually more important in the eyes of
the practitioners than solving the classical third level for the PVRP in literature. The
authors stated that the emphasis for PVRP problems in practice should be primarily
directed on the first level such that the remaining VRP and TSP problems at the levels
two and three are easier to solve.

In contrast to the studies in Chapter 2 and 3, the case study in Chapter 4 showed
that a loss of information due to aggregation can also be very helpful at lower decision
levels. Once the aggregated supply is assigned to a single collection scheme with fixed
collection days, both the VRP and the TSP are considerably easier to solve. Total
supply within each cluster is known in advance and can be assigned to available vehicle
capacities. Moreover, after disaggregation at the lowest decision level, all suppliers
within each cluster share the same collection days which makes the TSP problems
considerably easier to solve.

Wider applicability of special ordered sets type 1

From a computational point of view, the first level planning model in Chapter 4 was
disappointingly hard to solve by standard solution techniques. Mourgaya and
Vanderbeck (2007) confirmed that comparable problems with wide time windows are
much harder to solve than problems with very tight time windows. The authors applied a
heuristic to solve specific balancing problems at the highest decision level for the PVRP.
We focused on the algorithmic side.

131
Chapter 6

Problem specific properties were exploited and incorporated in a non-standard

way i.e. an adapted application of Special Ordered Sets type 1 (SOS1) to solve the
generated problems exactly. Literature stated that there is a great advantage to be
gained in the SOS1 formulation, provided that the variables within the sets have a
natural ordering (Williams 1990; Ashford and Daniel 1992). Documentation of modern
mathematical programming software states that if there is no ordered relationship
among the variables (such that weights cannot be specified or would not be meaningful),
other formulations should be used instead of a special ordered sets (ILOG 2009). The
background, understanding, and added value of “a natural ordering” remains
undiscussed and seems to be based purely on computational experiments (Beale and
Tomlin 1970; Williams 1990; Ashford and Daniel 1992; Williams 1993). The aim in
Chapter 4 was to understand the potential benefits of the SOS1-concept and, if possible,
to find a way that makes their use worthwhile in a broader context. In Chapter 4 we
showed that an effective use of the SOS1 solution procedure is not necessarily
restricted to problems with supplementary model conditions, i.e. a natural ordering of the
variables within the sets.

The applicability of SOS 1 is very common in mathematical programming. As a

matter of fact, treating SOS1 restrictions algorithmically, can be applied to all proposed
models in the chapters 2, 3, 4 and 5. A notable extension for their use refers to a subset
of models in Chapter 5. The gained insight in Chapter 4 was the basis for an efficient
use of SOS1 in a reformulation approach for a specific class of mixed integer non-linear
problems in Chapter 5.

Abandoning the principle of an overriding monetary objective (e.g. total costs),

implied that the some problems in Chapter 5 changed into a special class of mixed
integer, (0-1) fractional programming problems. Several extensive reviews showed that
fractional programming (FP) is an illustrative field in mathematical programming
demonstrating the separation between theoretical developments and its applicability in
practice (Schaible and Ibaraki 1983; Stancu-Minasian 1999; Schaible and Shi 2004;
Stancu-Minasian 2006). From a practical point of view, the focus was directed towards
taking advantage of specific problem characteristics and to exploit the structure between
two continuous linear models for solving the original mixed integer (0-1) FP problem. As
indicated in Chapter 5 we firstly applied a reformulation approach to the continuous
relaxation of the original, mixed integer non-linear problem. Next, we proved that the
SOS1 concept can extend Charnes’ and Cooper’s reformulation approach for
continuous FP problems to a specific class of mixed integer (0-1) FP problems. We
showed that the combination of concepts can solve problems in practice without a loss
of computational efficiency.

Although the case in Chapter 5 refers to an industry-specific problem, the study

demonstrates that the motivation to solve real-life problems can provide the basis for a
new approach that has an added value of its own, even outside the given application
area. Other mixed integer (0-1) fractional programming problems containing common
(convexity) constraints like (6) in Chapter 5 can be solved by the proposed combination

132
 General discussion

of methods. The solution procedure neither requires an iterative evaluation of a function

in commonly applied parametric approaches, nor a special purpose algorithm. The
SOS1 branching concept may even be available in modern, state-of-the-art,
mathematical programming packages, which contributes to the adaptability of the
approach in daily practice. Inevitable application-oriented maintenance in the future will
hardly be disturbed by locally developed (special purpose) solution techniques.

6.3.2 Professional relevance and applicability of DSS in practice

From an applicability point of view it is paradoxical that Arnott and Pervan (2008)
evidenced a trend in which the professional relevance and contribution of DSS research
is facing a crisis of relevance while on the other hand Entrup (2005) found an opposing
interest from practice for commercial, OR-based Advanced Planning Systems (APS).
Industrial practice started to demand for APS particularly due to the added functionality
of optimization-based decision support (Rudberg and Thulin 2009; Ivert and Jonsson
2010).

The dominant aspect of building models and solving them in optimization-based

DSS has been addressed in the previous Section 6.3.1. In this section we pay attention
to other aspects, i.e. effective user interfaces and additional modules and tools to
analyse the generated solutions beyond the classical sensitivity analysis. Based on the
findings in the case studies of Chapter 2, 4 and 5 we believe these aspects are still
underestimated or simply viewed as obvious. The importance of these modules for
improving the applicability and adoption of model-based approaches makes it necessary
to pay some attention to these design issues. Following the findings in Chapter 2, 4 and
5, we describe our experiences regarding i) user interfaces, and ii) simulation and
scenario management tools.

User interface
Models may be the dominant component in a model-driven DSS, but the final stage of a
modelling process is the analysis, which includes the delivery of solutions in a usable
form and to enhance the ability to analyse and understand the problem (Kallrath 2004).
In the end, decision-making must be executed by end-users who have the final insight in
the problem, know the real constraints and have the ultimate feeling regarding the
feasibility of generated plans.

The case studies in Chapter 2, 4 and (particularly) 5 showed that tailored user
interfaces are crucial for the applicability and adoption of DSS in practice. Within this
context it should be mentioned that the case study in Chapter 2 was developed at the
time that DSS applications on a PC were just evolving. The developed graphical user
interface (GUI) was state of the art at that time and showed the importance of a tailored
GUI. The user interfaces for the case studies in Chapter 4 and 5 were built using widely
available and more general development tools. The developed user interfaces
demonstrated how significant a tailored user interface including a period of testing in

133
Chapter 6

practice is for i) the communication between system and non-technical specialists, ii)
integrating new technology into decision-maker’s (daily) tasks, iii) the elicitation of
specific domain knowledge needed to identify and exploit special structures during
model construction and/or to take advantage of specific characteristics to solve
problems faster and iv) the adoption of an automated system in practice. Our experience
in the case studies was that using widely available general-purpose development
software simplifies the development time of effective user interfaces. However,
personalization remains an area that must be addressed.

Nowadays, tools for developing user interfaces are widely available and the
continuous growth of visualization tools will benefit the process of solution delivery in
model-driven DSS applications. Power and Sharda (2007) confirmed that the goal of
making these systems accessible to non-technical specialists implies that the design
and capabilities of the user interface are important to the success of the system.
According to Kallrath (2004), standard interface-design factors mean that users can
quickly adopt new DSS with less training and more confidence. However, while
standards are advantageous from a developing point of view, both Kallrath (2004) and
Power and Sharda (2007) confirmed that personalization of user interfaces is important
and should be addressed by developers and researchers. A high-quality user interface
will not guarantee the success of a DSS. However, a poor user interface may be a
missed opportunity to test the added value including the applicability of the proposed
system in real life. It may even be a serious threat for the survival of a DSS. Framinan
and Ruiz (2010) even highlighted in their literature review on manufacturing scheduling
systems the need to shift the research pattern and increase the investigation on areas
such as user interfaces, data management, and other tools and methods for a better
design and implementation of manufacturing systems.

In all of the preceding case studies, end-users were actively involved in the
development of the (G)UI. Test phases in practice were crucial to spot functional
convenience of the complete system for end-users. The effort devoted to personalize
the (G)UI and test the DSS for a certain period was of striking importance to make ill-
structured domain knowledge more tractable and to exploit semi-structured knowledge
and understanding from practice in the process of developing models and solution
techniques. Based on our experiences, building blueprints of DSS (i.e. the user interface
including all underlying components) in a laboratory environment and restrict feedback
from practice to a confrontation of the generated results, is barely sufficient to convince
daily practice for the added value of a DSS and to mirror its intended professional
relevance and applicability. From all case studies we learned that the ultimate way to
reduce the tension between descriptive decision-making and the adoption of normative
approaches in practice, is to associate blueprints of model-based DSS with a serious
period of testing, i.e. shadowing, in a real-life environment. The choice of management
to upgrade and revise an existing, customized DSS in Chapter 5 for future decision
support in new decision environments may be indicative for the validity, added value and
adoption of the developed DSS with all its components in practice.

134
 General discussion

Simulation and scenario management

Although several types and a variety of overlapping terms are used to identify
simulation, we define simulation as an approach for imitating the behaviour of an actual
or anticipated physical system (Power and Sharda 2007). The authors give an overview
of simulation modules in model-driven DSS either as a dominant or an additional
component in DSS and typify simulation as a descriptive tool that can be used for both
prediction and exploration of the behaviour of a specific system. Jacobs and Weston
(2007) confirmed that simulation will be an increasingly important element of integrated
and extended enterprise planning systems. Kallrath (2004) stated the importance of
scenario management and typify scenario management tools as a trend to store and
analyse multiple solution scenarios such that it benefits the understanding of underlying
patterns. Particularly the studies in Chapter 4 and 5 showed that the classical
architecture of model-based DSS as described in Chapter 1 (Sprague Jr 1980), should
be extended by embedding additional descriptive models, e.g. a simulation module, and
profound scenario manager tools.

From a decision support perspective, the main difference between the system in
Chapter 2 and those described in the chapters 4 and 5 are the added functionality of
simulation and scenario management. Both in Chapter 4 and Chapter 5 simulation is
used as an additional descriptive component embedded in the DSS. Simulation modules
turned out to be powerful tools i) to assist decision-makers in calculating alternative
plans and to anticipate on the impact of changes in practice, ii) to study the
consequences and support the awareness of those changes on threshold values for
different indicators, and finally iii) to support the assessment of specific actions. Tools for
scenario management, i.e. to store, visualize, systematically keeping track of generated
solutions, and particularly to combine and analyse different scenarios (e.g. on input
settings and generated solutions) were mainly experienced in practice as indispensable
tools to support the insight and understanding of the underlying problem. The results of
the optimization routine and stored plans of the past by the scenario manager were
mostly used as a point of reference to study the impact of any change and/or to obtain
alternative (favoured) plans.

The case studies in Chapter 2, 4 and 5 showed that a model-based DSS should
not be considered as an optimizer but rather as a tool to ‘optimize’ the insights and
performance of decision-makers. The main added value of an optimization module is its
ability to foster out-of-the-box thinking and lift decision-making in practice to a higher
level. Particularly its integration with simulation and scenario manager is important for
real use in practice. Changing “optimal” solutions is necessary in practice. However,
simultaneously quantifying the impact of those changes on different (conflicting)
indicators is even more important to lift decision-making to a higher level and, in the end,
convince decision-makers to adopt systems in practice.

135
Chapter 6

6.3.3 Advanced Planning Systems

Meanwhile the basic concepts of model-based DSS did find their application in
commercial software suites, called Advanced Planning Systems (APS). The historical
development and market penetration of these systems in Figure 1.1 (see Chapter 1)
shows a clear interest from industrial practice for automated decision support. Although
APS are viewed in general as an extension of transactional-oriented ERP systems, APS
modules originate from many in-house developed DSS that aid planners at various
levels in the decision hierarchy (Rudberg and Thulin 2009; Ivert and Jonsson 2010).
Starting point for APS modules is to introduce a standardized way of performing
planning tasks, i.e. to find an effective architecture for modules that are on the one hand
easy to use and maintain, and on the other hand deliver realistic planning results for
contemporary practice. Nowadays, modern APS provide generic modules to convert
planning tasks into abstract mathematical models (Stadtler, Fleischmann et al. 2012).
For the translation of these models into software modules, Stadtler and Kilger (2008)
advocate and expect in their book on APS that vendors will provide a similar
mathematical modelling language (MPL) for all modules.

The introduction of MPL’s for the PC market, as for instance used in the case
study of Chapter 2, was indeed a major step forward to simplify the construction of
optimization models and allowed for a crucial separation between data and models.
However, as indicated in the previous sections, solving real-life problems often requires
that problem-specific characteristics are exploited either with respect to modelling and/or
solving the related models. Compared to tailored systems like the blueprints in Chapter
2 and 4 and the custom-made DSS in Chapter 5, a potential drawback of standardized
and generic (APS) modules may be that problems in practice differ between companies
and it might be hard to create generic modules that fit many companies or specific
markets, even within the unified framework of today’s APS.

A striking observation we experienced in developing model-based DSS is a trend

that moves away from generic and easy-to-use modelling languages (MPL) to new
generations of mathematical programming suites in which generic MPL’s are combined
with classical programming languages. This move towards open modular architectures
for optimization software was, particularly in the study of Chapter 5, experienced as one
of the most convenient developments in MPL-technology. It combines the strengths of
two concepts, i.e. a mathematical modelling and a classical programming language, into
a single environment. As opposed to a (traditional) modelling language and a (classical)
programming language there exists no difference between a modelling statement (e.g.
expression of a constraint) and a procedure that actually solves a problem (Ciriani,
Colombani et al. 2002; Ciriani, Colombani et al. 2003; Colombani, Daniel et al. 2004).
Yunes, Aron et al. (2010) defined the need for more flexibility and efficient model-solver
integration as a central trend in the optimization community.

The reached synergy between a generic MPL and a classical programming

language enables developers to interlace, for instance, specific data handling, modelling
statements, built-in or external solving procedures, predefined or new user-written

136
 General discussion

subroutines in a single environment, all according to the required needs for embedding
model-based decision support in real-life environments. For instance, the required
interpolation and extrapolation routines for calculating property values in Chapter 5 were
conveniently realized by i) defining simple and fast executable SQL-commands to
retrieve the basic data from any database and ii) programming the necessary routines in
the same environment. From a developer’s point of view, the move (backwards) to
programming languages was experienced as extremely convenient and great
advantages would have been possible if these architectures had been available at the
time that the systems as described in Chapter 2 and 4, were developed.

Based on our experiences, these new architectures provide the required

extension of functionality for developers but simultaneously may decrease their general
use by a wider audience of less-specialized experts. Extensive support from highly
trained modellers remains necessary, which may result in great (consultancy)
dependence and spending much time in the development and maintenance of those
systems.

6.4 Main conclusions and future research

The first study in this thesis on optimization-based decision support appeared at the time
that research in DSS prospered and since its peak in 1994 the exposure of the research
field has been in consistent decline. The (name of the field) field may face a crisis of
relevance but the question remains: “Is the former added value of an application-
oriented research field like optimization-based DSS less important nowadays?”

We believe the opposite holds. Developments in optimization-based DSS will

continue, regardless of what name(s) the field is going to appear. Complex and
integrated decision-making is for instance still done semi- or completely manually in
contemporary (food) processing industry. Many sources in the preceding chapters
emphasized the need for developing (integrated) decision models for each planning step
in the supply chain planning matrix of Figure 1.2. The current generation of enterprise
systems responded to this need and extended information-oriented enterprise systems
by offering optimization-based tools as a top layer on transaction-oriented (ERP)
systems in so-called Advanced Planning Systems (see Figure 1.1)

This final section aims to look back briefly on the essentials, i.e. to draw some
main conclusions, and to look forward on what remains, i.e. to define some directions for
future research. We continue to make a distinction between model-building and -solving
on the one hand and professional relevance and applicability on the other hand.

Model-building and -solving

As stated earlier, an important issue or tension in an applied research field is the
extent to which the academic field leads or follows industrial practice. Based on the
studies in the chapters 2 and 3 and the current state of affairs in enterprise systems, we

137
Chapter 6

conclude on the one hand that today’s APS apply the same (hierarchical) planning
approach as in the first case study of this thesis, i.e. by separating planning tasks from
scheduling. APS systems do not provide modules for simultaneous planning and
scheduling (Stadtler, Fleischmann et al. 2012). On the other hand, many sources from
literature argue that both decisions should be made simultaneously, especially in
processing industries.

Despite the progress that has been reached to eliminate the main drawback of
hierarchical production planning approaches, many of the developed models and
methods in literature for simultaneous planning and scheduling are not intended to be
generically applicable and/or solvable for problems of realistic sizes. We believe that
more industry-specific solutions are needed which try to incorporate specificities of
different production environments into models. Starting from our experiences in
modelling and solving problems, the key to develop solvable approaches in practice may
be i) to use knowledge and experience from practice and take advantage of specific
characteristics in different problem domains during model-construction, i.e. to find tighter
models and stronger (problem-specific) valid inequalities, and/or ii) to identify and exploit
special problem structures for solving the related models using existing -, novel - , and
combined solution techniques.

We conclude that most of the lot-sizing literature addresses (extensions of)

problems for discrete manufacturing environments. Although recent reviews on lot-sizing
confirm the need for more practical variants of models addressing typical characteristics
in processing industries (Jans and Degraeve 2008; Quadt and Kuhn 2008; Clark,
Almada-Lobo et al. 2011), we believe it deserves serious consideration to relax the
existing focus on optimality for lot-sizing, particularly on lower decision levels. In many
production planning models the quality of a plan is historically evaluated by the value of
a single objective function in which possibly several terms or criteria (if recognized) are
expressed in a monetary unit. In practice, decision-makers on lower decision levels are
foremost interested in the generated plans. The quality of these plans is usually not
measured in value of the objective function. An extreme orientation towards optimal
solutions may carry the risk of ignoring the human nature of multi-objective decision-
making. We believe that in many practical situations decision-makers are not infinitely
sensitive to changes in the objective value, particularly when the differences become
small. Instead, (s)he may intend to include other criteria. To give an example,
maximizing the freshness of delivered products may be on the one hand hard to express
(efficiently) in a monetary unit, but on the other hand it has to be done at the expense of
total production costs, e.g. by splitting lot-sizes.

The chapters 1 and 4 indicate that surprisingly little research has been devoted
to issues of coordination and integration between the building blocks “procurement” and
“production” in the supply chain planning matrix (SCPM) of Figure 1.2. We do not share
the view that a very limited number of required raw materials in processing industries,
with relatively low value and reliable lead times, would make procurement unproblematic
for this branch of industries. The study in Chapter 4 clearly shows that sourcing needs
more attention in processing industries, particularly in push-oriented, inter-organizational

138
 General discussion

supply chains. Based on the findings in Chapter 4, we conclude that the valorisation of
raw materials needs additional planning in order to fit the quantities of raw materials at
supply level to strict delivery conditions at processing level. Within this context, the study
in Chapter 4 should not been considered as an isolated example. Comparable
considerations hold for the application area in Chapter 5.

Although the planning model in Chapter 4 enables different partners in a supply

chain to sell and/or to source raw materials at more attractive price levels, the proposed
approach limits to balancing quantities of material flows between successive links in a
supply chain. We believe that the valorisation of raw materials could be improved even
more if the composition of raw materials is considered in planning too. From a collection
point of view, milk is for instance still viewed as being “just white” but the suitability of
different types of milk for various processing operations and the quality of the related
final products is strongly affected by changes in supply and the composition of raw milk
(Banaszewska, Cruijssen et al. 2013). Taking the composition of raw materials for
collection problems in food industry into consideration may imply that further research on
the class of Periodic Vehicle Routing Problems (PVRP) should focus on a close
integration between the Tactical Planning Vehicle Routing Problem (TPVRP) and the
related Vehicle Routing problem (VRP) problems.

It should be mentioned that the interrelated impact of processing operations

including the composition of materials flows on the properties and quality of final
products is even more obvious in the application area of Chapter 5. The latter enhances
the need for additional research on a closer integration and coordination between
procurement and production.

In Chapter 4 we focused on an alternative algorithmic implementation of Special

Ordered Sets type 1 (SOS1). In Chapter 5 we proved that the SOS1 concept can extend
a classical reformulation approach for continuous fractional programming (FP) problems,
to a more general class of mixed integer (0-1) FP problems. Based on our findings we
conclude, in contrast to literature, that a natural ordering of the variables within the sets
is not necessary to make their use worthwhile. We believe that the necessity of a
separate (user defined) reference row or weights associated to the variables in the sets
may be omitted for an effective use of the SOS1 branching scheme. In Chapter 4 we
applied a problem specific procedure for (re-)ordering the variables within the sets which
resulted in substantial computational advantages. Re-ordering the variables within the
sets, solely on their continuous value in each node of the search tree might be a
promising option to generalize the concept to an effective and broader use in
mathematical programming software. This requires, however, further research and
extensive computational tests.

Professional relevance and applicability of optimization-based DSS

Attention in OR literature was, and to a certain extent still is, focused on developing
models and algorithms and its use is supposed to be a distinctive step in the decision-

139
Chapter 6

making process. However, the final stage in decision processes includes the delivery
and analysis of the generated solutions in a usable form. We believe that an important
contribution in bridging the gap between theory and practice is to recognize and
combine the strong elements of normative models in optimization-based decision
support (i.e. what practice ought to do) with descriptive decision-making (i.e. what
practice actually does), such that systems arise that provide in what practice should and
can do.

Within the context of complex decision-making we foremost conclude that

optimization mainly fosters out-of-the-box thinking. The integration of (normative)
optimization with (descriptive) simulation and scenario management in a single
environment provides a valuable combination of tools for successive questioning, to
clarify available options and to obtain a greater understanding of what is possible in
practice. In addition, customization and personalization should be addressed for the
adoption of DSS in practice.

Finding the right balance between generic models and solution techniques in
APS on the one hand and tailored DSS for specific (problems in) industries on the other
hand, may be a major challenge for the future, particularly for decision support on lower
hierarchical levels. Within the latter context it is remarkable that despite of the industrial
interest for APS, almost no systematic research has been conducted regarding
adoption, implementation, usage and/or failures of APS in practice (Lin, Hwang et al.
2007; Wiers 2009; Ivert and Jonsson 2010; Ivert 2012). We believe it is crucial to
examine the requirements of industries and develop either industry-specific solutions
(Entrup 2005), and/or open architectures that allow for industry-specific approaches.

More real-life case studies will contribute to get insights and understanding of the
strong and weak aspects of APS systems including the needs from industrial practice. In
their book on APS, Stadtler, Fleischmann et al. (2012) confirmed this view by their
statement “New findings in research and good business practices should find their way
into future developments of APS”.

140
 Summary

Samenvatting

References
Summary
Nowadays, efficient planning of material flows within and between supply chains is of
vital importance and has become one of the most challenging problems for decision
support in practice. The tremendous progress in hard- and software of the past decades
was an important gateway for developing computerized systems that are able to support
decision making on different levels within enterprises. The history of such systems
started in 1971 when the concept of Decision Support Systems (DSS) emerged. Over
the years, the field of DSS has evolved into a broad variety of directions. The described
research in this thesis limits to the category of model-driven or optimization-based DSS.

Simultaneously with the emergence of DSS, software vendors recognized the

high potentials of available data and developed Enterprise Systems to standardize
planning problems. Meanwhile, information oriented systems like MRP and its
successors are extended by the basic concepts of optimization based decision support.
These systems are called Advanced Planning Systems (APS). The main focus of APS is
to support decision making at different stages or phases in the material flow, i.e. from
procurement, production, distribution to sales (horizontal-axis), on different hierarchical
aggregation levels (vertical-axis) ranging from strategic (long-term) to operational (short-
term) planning. This framework of building blocks decomposes planning tasks
hierarchically into partial planning problems. This basic architecture of the planning
processes in APS is known as the Supply Chain Planning Matrix (SCPM).

Compared to, for instance, discrete parts manufacturing, planning tasks are
much more complicated in processing industries due to a natural variation in the
composition of raw materials, the impact of processing operations on properties of
material flows, sequence dependent change-over times, the inevitable decline in quality
of product flows and relatively low margins. These specific characteristics gave rise to
focus on optimization-based decision support in the domain of processing industries.
The problems to be addressed in this field call for (inter-related) decisions with respect
to the required raw materials, the production quantities to be manufactured, the efficient
use of available resources, and the times at which raw materials must be available.

Although different APS modules can interact directly, coordination and

integration is often restricted to the exchange of data flows between different modules.
Given the need for specific integrated decision support, the research presented in this
thesis focusses particularly on medium to short term decision support at production
stage in processing industry, including the vertical and horizontal integration and
coordination with adjacent building blocks in the SCPM.

Extensive reviews from literature show that the gap between research and
practice of DSS is widening. As the field of DSS was initiated as an application oriented
discipline, the strategy of what is referred to as “application-driven theory” was taken as
the preferred approach for this thesis. “Application-driven” refers to a bottom-up
approach which means that the relevance of the research should both be initiated and

142
 Summary

obtained from practice. The intended successful use of the proposed approaches
should, where possible, be represented by tests of adequacy. Simultaneously, the
contribution to “theory” aims to be a recognizable part of the research effort, i.e.
obtained understanding and insights from problems in practice should provide the basis
for new approaches. Based on the preceding considerations we defined the following
general research objective:

General research objective

To support medium- to short term planning problems by optimization-based models and
solution techniques such that:

i) The applicability and added value of (prototype) systems is recognized and

carried by decision makers in practice
ii) The proposed approaches contribute to knowledge, understanding and insights
from a model building and – solving point of view.

In order to link the general objective with the different studies in the thesis, we defined
five, recurring research premises, i.e. Professional relevance and applicability (P1),
Aggregation (P2), Decomposition and reformulation (P3), Vertical integration at
production level (P4), and Horizontal coordination and integration (P5).

The overarching premise P1 refers to the first part of the research objective. All
other premises refer to the second part of the research objective, i.e. model building
and/or – solving. Several planning issues are studied to give substance to the research
objective and each study is connected to at least two research premises.

Study 1: Planning and scheduling in food processing industry

The main question in Chapter 2 was:” How to apply aggregation, decomposition and
reformulation in model-based DSS at planning and scheduling level such that the aspect
of decision support is recognized and appreciated by decision makers in practice, and
which level of aggregation is needed to integrate production planning (i.e. lot-sizing) and
scheduling problems in a single model?

The study consists of two parts. The first part of the study refers to a case study
for the bottleneck packaging facilities of a large dairy company. The goal was to
develop, implement and test a pilot DSS which was able to deliver solutions recognized
and carried by decision makers at lower decision levels. The latter aim implied that a
straight-forward aggregation on time, product type, resources or product stage, was not
preferred. The key to develop an approach for regular use was to identify and take
advantage of specific problem characteristics. Clustering of numerous jobs, while
retaining information at order level, could be exploited in a reformulation approach. The
inclusion of (combined) generalized- and variable upper bound constraints gave very
tight lower bounds and sparse search trees.

143
 An extensive test phase in daily practice showed that the main benefit of the
DSS was the initial quality of the generated plans including the time needed to generate
these schedules. Hence, decision makers could i) postpone their planning tasks, ii)
conveniently cope with rush orders or planned maintenance and iii) easily generate
alternatives or revised plans when unforeseen disturbances occur. Moreover, the
graphical presentation and overview of the (future) working schedule enabled order
acceptance to make use of remaining capacity.

The study also showed that planning problems in practice cannot be captured
exhaustively by a (simplified) model. Decision makers need the opportunity to modify
automatically generated plans manually and use human judgement and experience
such that the solution is tuned to the actual situation. Hence, the DSS should not be
considered as an optimizer but rather as a tool for generating high quality plans to be
used for further analysis. Within this context the various options of a user-friendly,
graphical, and fully interactive user interface, were of major importance.

Although the case study clearly demonstrates the validity of earlier case based
DSS research for current days APS, the proposed approach is hardly a generic solution
for a complete vertical integration between lot-sizing and scheduling. If lot-size decisions
are strongly affected by the sequence of jobs, production planning and scheduling
should be performed simultaneously.

As the described case refers to an earlier study and today’s APS do not provide
modules for integrated lot-sizing and scheduling, the second part of the study gives an
overview of developments in literature regarding lot-sizing and scheduling models and
assess their suitability for addressing sequence-dependent setups, non-triangular
setups and product decay. The review shows a tendency in which so-called Big Bucket
(BB) models are currently proposed for short term time horizons too. However, we argue
that segmentation of the planning horizon is a key issue for simultaneous lot-sizing and
scheduling. The advantage of BB models may become a major obstacle for i) the
effectiveness of simultaneous lot-sizing and scheduling, and ii) addressing specific
characteristics in food processing industry.

Study 2: Vertical integration of lot-sizing and scheduling in food processing

industry
Chapter 3 focused on a complete integration of lot-sizing and scheduling decisions in a
single model. The main question was:” How to integrate production planning (i.e. lot-
sizing) and scheduling problems in a single model, such that common assumptions
regarding the triangular setup conditions are relaxed and issues of product decay and
limited shelf lives are taken into account?”
The literature research in Chapter 2 revealed that the computational advantage
of time oriented aggregation in BB models may become a major obstacle in addressing
the identified characteristics in FPI. In addition, product decay is primarily associated
with the “age” of products and consequently relates to the segmentation of the time-

144
 Summary

horizon. Therefore, two SB models are developed to demonstrate the impact of non-
triangular setups and product decay on the generated solutions. Small scale examples
were used to demonstrate how a small change in the balance between inventory - and
changeover costs may generate significantly different solutions, especially when the
triangular setup conditions do not hold.

The developed models are potentially very large formulations and, as expected,
hard to solve. Exploratory research was conducted with a Relax-and-Fix (R&F) heuristic.
The heuristic is based on a decomposition of the time horizon. Numerical results of
small to medium sized problem instances are promising. However, solving real-size
problem instances is not possible yet.

Study 3: Integrated planning between procurement and production

The case study in Chapter 4 focussed on the need for horizontal coordination and
integration between the phases procurement and production, which is of particular
importance in inter-organizational supply chains. The main question was:” How to model
and solve an integrated planning problem between procurement and production, both on
a mid-term and short-term planning level, in an inter-organizational supply chain? The
research question was projected on an illustrative milk collection problem in practice.

The aim was to develop a pilot DSS that lifted decision support for a “weaker” partner in
a food supply chain to a higher level, and to illustrate the importance of horizontal
integration between the phases procurement and production in an APS framework.
Problem analysis revealed that the problem can be classified as an extension of
the Periodic Vehicle Routing Problem (PVRP). The problem was decomposed into more
tractable sub problems on different hierarchical levels, i.e. the daily (vehicle) routing
problem was separated from a medium-term planning problem. On the higher planning
level, numerous suppliers were aggregated such that total supply within a cluster met
(multiple) vehicle loading capacities. The continuous supply of relatively small amounts
from many suppliers had to be balanced with strict delivery conditions at processing
level. A model was developed to assign a single (stable) collection rhythm to each
cluster such that the total, weighted deviation of desired processing levels on various
days in the planning horizon was minimized.

The applied aggregation on the higher planning level turned out to be very
beneficial for the required disaggregation at the lower planning level. Once supplier
farms were geographically grouped into clusters and the aggregated supply within a
cluster was assigned to a single collection rhythm with fixed collection days, the (initial)
daily routing problem was considerably easier to solve for vehicle schedulers.

The computational complexity of the problem was reduced by exploiting

application-based properties algorithmically in a specific branch-and-bound scheme, i.e.

145
a customized approach of Special Ordered Sets type 1 (SOS1) This approach made it
possible to solve the generated problems exactly for real-size problem instances.

The various facilities of a user-friendly and interactive man-machine interface

(i.e. an input, planning, simulation and analysing module) turned out to be essential.
Decision makers could easily change the data, and the generated plans, in a separate
simulation module. However, the impact of any modification was immediately visualised
by several (conflicting) indicators in the output screens, both on supply and demand
level.

Study 4: Mixed Integer (0-1) Fractional Programming in Paper Production Industry

The study in Chapter 5 focussed on the impact of technical settings of production units
on material flows. The main question was:” How to support decision-makers in practice if
crucial properties of end products simultaneously depend on (endogenous) types of raw
materials with different chemical or physical properties and (endogenous) technical
settings of processing units?

The goal of the study was to revise and upgrade an existing, locally used DSS,
to a tailored and flexible tool for decision support within the enterprise. The study
revealed that the aimed extension towards multi-objective decision support, together
with new physical insight for calculating properties of end products due to process
operations, had a substantial impact on the optimization module.

The proposed solution procedure takes advantage of the problem characteristics

and gives rise i) to apply and extend a classical reformulation approach for continuous
linear fractional programming (FP) problems to a more general class of mixed integer
(binary) FP problems and ii) to exploit the special structure between the original non-
linear mixed integer model and the continuous, linear reformulation by applying the
concept of Special Ordered Sets type 1 (SOS1).

Although Chapter 5 focusses in particular on the reformulation and solution

approach, the DSS consists of four main building blocks, i.e. the user interface, a
scenario manager, a simulation- and optimization routine. The optimization module
provides a powerful tool to find feasible solutions and the best (unexpected) recipes for
any available set of raw materials. Moreover, it provides an innovative way of decision
support for purchasing (new) pulps on the market, for assigning available pulps to
different paper grades, and for attuning available stock levels of raw materials to
(changing) production targets for different paper grades. The results of the optimization
routine are mainly used to obtain alternative recipes for different paper grades. Usually,
these recipes are stored as base scenarios and adapted to daily practice in the
simulation module.

146
 Summary

Main conclusions and future research

Based on the studies in the Chapters 2 and 3 we conclude that no generically applicable
models and/or solution approaches exist for simultaneous planning and scheduling in
processing industries. More industry-specific solutions are needed incorporating
specificities of different production environments into those models. The key to develop
solvable approaches for contemporary practice may be i) to use knowledge and
experience from practice and take advantage of specific characteristics in different
problem domains during model construction, and/or ii) to identify and exploit special
problem structures for solving the related models.

We conclude that surprisingly little research has been devoted to issues of

coordination and integration between “procurement” and “production”. The studies in the
chapters 4 and 5 confirm that sourcing of (raw) materials flows needs more attention in
processing industries, particularly in push-oriented, inter-organizational networks. The
valorisation of raw materials can be improved even more if the composition of raw
materials is considered too in future planning problems at production level.

In the second part of this thesis we focused on extensions for the applicability of
Special Ordered Sets type 1 (SOS1), both from an algorithmic (Chapter 4) and
modelling (Chapter 5) point of view. We conclude that the concept of SOS1 can extend
a classical reformulation approach for continuous fractional programming (FP) problems,
to a specific class of mixed integer (0-1) FP problems. Moreover, we conclude that a
natural ordering of the variables within the sets is not necessary to make their use
worthwhile. A separate (user defined) reference row or weights associated to the
variables in the sets might be omitted for an efficient use of SOS1 in commercially
available mathematical programming packages. However, this requires further research
and extensive computational tests.

147
148
Samenvatting
Het efficiënt plannen van goederenstromen binnen en tussen logistieke netwerken is
uitgegroeid tot een van de grootste uitdagingen voor beslissingsondersteuning in de
dagelijkse praktijk. De enorme vooruitgang in hard- en software van de laatste decennia
heeft de ontwikkeling van software modules voor beslissingsondersteuning op
verschillende niveaus binnen ondernemingen in een stroomversnelling gebracht. De
oorsprong van dergelijke systemen dateert uit 1971 toen het concept van “Decision
Support Systems” (DSS), ofwel beslissingsondersteunende systemen, werd
geïntroduceerd. Sindsdien heeft het onderzoeksveld van DSS zich in tal van richtingen
ontwikkeld. Het onderzoek in deze dissertatie beperkt zich tot de categorie van
modelgestuurde, of op de optimalisering gebaseerde, systemen.

Gelijktijdig met het ontstaan van DSS onderkenden softwareverkopers de

mogelijkheden die de beschikbare gegevens boden en ontwikkelden bedrijfssystemen
om planningproblemen te standaardiseren. Intussen werden informatie-georiënteerde
systemen als MRP en de daaraan gerelateerde opvolgers uitgebreid met de
basisbeginselen van beslissingsondersteuning gebaseerd op optimaliseringsmodellen.
Deze systemen worden “Advanced Planning Systems” (APS) genoemd. APS richten
zich met name op de toepassing van wiskundige technieken uit de besliskunde voor het
modelleren en kwantitatief ondersteunen van besluitvormingsprocessen in bedrijven.
Hiertoe werd een raamwerk van bouwstenen gedefinieerd waarin twee dimensies
worden onderscheiden. Op de horizontale as worden verschillende fasen in de
goederenstroom onderscheiden, d.w.z. van inkoop, productie, distributie tot verkoop. Op
de verticale as worden verschillende hiërarchische beslissingsniveaus onderscheiden,
variërend van strategische (lange termijn) tot operationele (korte termijn) besluitvorming.
Dit raamwerk van bouwstenen verdeelt planningstaken in deelproblemen en staat ook
wel bekend als de “Supply Chain Planning Matrix” (SCPM), ofwel de goederenstroom-
planningsmatrix.

Planningstaken zijn in de procesindustrie veelal gecompliceerder dan in,

bijvoorbeeld, de stukgoedindustrie. Het verschil in complexiteit wordt onder andere
veroorzaakt door de natuurlijke variatie in de samenstelling van grondstoffen, het effect
van (chemische) bewerkingen op de eigenschappen van (half)producten, volgorde-
afhankelijke omsteltijden, de onvermijdelijke kwaliteitsachteruitgang van materiaal-
stromen en de relatief lage marges op eindproducten. Planningsproblemen in dit veld
vragen om (onderling samenhangende) beslissingen met betrekking tot de benodigde
(hoeveelheden) van verschillende grondstoffen, de te produceren hoeveelheden, het
efficiënt gebruik van beschikbare hulpbronnen, alsmede de tijdstippen waarop
grondstoffen beschikbaar moeten zijn.

Ofschoon interactie tussen verschillende APS modules mogelijk is, blijven

coördinatie en integratie veelal beperkt tot het uitwisselen van data. Gezien de behoefte
aan geïntegreerde beslissingsondersteuning richt het onderzoek in dit proefschrift zich

149
vooral op de ondersteuning van middellange- tot kortetermijnbeslissingen in de
productiefase voor de procesindustrie, inclusief de verticale en horizontale integratie met
aangrenzende bouwstenen in de SCPM.

Uitgebreide literatuuroverzichten laten zien dat de kloof tussen het onderzoek en

de praktijk van DSS steeds breder wordt. Juist omdat onderzoek naar DSS een
toepassingsgerichte discipline beoogt te zijn, werd Cooper’s concept van “Applications
driven theory”, ofwel toepassingsgedreven theorievorming, als uitgangspunt gehanteerd
voor dit proefschrift. “Toepassingsgedreven” verwijst naar een bottom-up benadering,
hetgeen betekent dat de relevantie van het onderzoek wordt ontleend aan en getoetst in
concrete praktijksituaties. Tegelijkertijd beoogt de bijdrage aan de “theorie” van de
besliskunde een herkenbaar deel van de onderzoeksinspanning te zijn. Begrip van en
inzicht in praktijkproblemen moeten de basis vormen voor nieuwe benaderingen.
Gebaseerd op de voorgaande overweging is de volgende onderzoeksdoelstelling
gedefinieerd:

Algemeen onderzoeksdoel

Het ondersteunen van middellange- tot kortetermijnplanningsproblemen met modellen

en oplossingstechnieken die zijn gebaseerd op de OR (Operations Research), zodanig
dat:

i) De toepasbaarheid en toegevoegde waarde van (prototype) systemen wordt

herkend en gedragen door beslissers in de praktijk,
ii) De voorgestelde benaderingen bijdragen aan kennis, begrip en inzicht in het
bouwen van modellen en het oplossen van de gegenereerde problemen.

Om het breed gedefinieerde onderzoeksdoel met de beschreven studies in dit

proefschrift te verbinden, zijn een vijftal uitgangspunten gedefinieerd, te weten:
professionele relevantie en toepasbaarheid (P1), aggregatie (P2), decompositie en
herformulering (P3), verticale integratie op productieniveau (P4) en horizontale
coördinatie en integratie (P5).

Het overkoepelend uitgangspunt P1 refereert aan het eerste deel van het
onderzoeksdoel. Alle andere uitgangspunten hebben betrekking op het tweede deel van
de onderzoeksdoelstelling. Verschillende planningsproblemen uit de SCPM zijn
bestudeerd, waarbij elk onderzoek met tenminste twee van de gedefinieerde
uitgangspunten is geassocieerd.

Studie 1: Planning en roostering in de voedingsmiddelenindustrie

De hoofdvraag in hoofdstuk 2 was: “Hoe kunnen aggregatie, decompositie en

herformulering worden toegepast op plannings- en roosteringsproblemen, zodanig dat
de toegevoegde waarde van modelgebaseerde beslissingsondersteuning in de praktijk
wordt herkend en gedragen, en welk aggregatieniveau is nodig om productieplanning
(d.w.z. seriegrootteplanning) en roostering in één enkel model te integreren?”

150
 Samenvatting

De studie omvat twee delen. Het eerste deel verwijst naar een casus uit de
praktijk voor de planning en roostering van verpakkingsfaciliteiten op de
productieafdeling van een grote zuivelfabrikant. Het doel was een blauwdruk DSS te
ontwikkelen, te implementeren en te testen waarmee de praktijk in staat zou zijn om
oplossingen te genereren die door beslissers op lagere beslissingsniveaus werden
herkend en gedragen. Deze laatste doelstelling impliceerde dat rechttoe rechtaan
aggregatie naar tijd, producttype, beschikbare productiefaciliteiten en/of
productiestadium niet de voorkeur had. Het antwoord voor het ontwikkelen van een
werkbare aanpak voor regelmatig gebruik was om casus gebonden
probleemkarakteristieken te identificeren en tijdens de modelvorming uit te buiten. Het
gericht clusteren van grote aantallen orders, zonder dat daarbij informatie op
orderniveau verloren ging, kon worden benut in een herformuleringsaanpak waardoor
een combinatie van zogenaamde “Generalized- and Variable Upper Bound constraints”
(GUB, VUB) in de modelvormingsfase bereikbaar werd. De aanpak leidde tot zeer
strakke ondergrenzen in een gangbare, impliciete aftelmethode en ijle zoekbomen.

Een uitgebreide testfase in de praktijk liet zien dat de constante kwaliteit van de
gegenereerde oplossingen gecombineerd met de vereiste rekentijd die nodig was om de
startoplossingen te genereren, het grootste voordeel van het DSS waren. Beslissers
konden daardoor i) hun planningstaken uitstellen, ii) gemakkelijker omgaan met
spoedorders of gepland onderhoud, en iii) eenvoudig alternatieven genereren dan wel
de opgestelde plannen herzien wanneer zich onvoorziene verstoringen voordeden.
Bovendien stelde de grafische weergave en het overzicht van het (toekomstige)
werkplan de afdeling orderacceptatie in staat om restcapaciteit beter te benutten.

Het onderzoek toonde ook aan dat planningsproblemen uit de praktijk niet
volledig in een (vereenvoudigd) model beschreven kunnen worden. Beslissers hebben
behoefte aan mogelijkheden om automatisch gegenereerde planningen aan te passen
ten einde menselijk inzicht en ervaring aan te wenden voor afstemming op de feitelijke
situatie. Het DSS moet derhalve niet worden beschouwd als een optimalisatie
gereedschap in de meest letterlijke zin van het woord, maar veeleer als hulpmiddel om
kwalitatief hoogwaardige (start)oplossingen te genereren voor aanvullende analyse. De
verschillende hulpmiddelen in een gebruikersvriendelijk, grafisch en volledig interactief
gebruikersinterface waren daarbij van groot belang.

Ofschoon het belang van eerder casus-gebaseerd DSS-onderzoek voor huidige

APS systemen overduidelijk wordt aantoond in de eerste twee hoofdstukken, is de
voorgestelde aanpak in hoofdstuk 2 nauwelijks een algemene oplossing voor de
beoogde verticale integratie tussen seriegrootte- en roosteringsproblemen. Als
beslissingen over seriegroottes sterk worden beïnvloed door de productievolgorde van
de geplande series is integratie van productieplanning en roostering noodzakelijk.

Aangezien de beschreven casus in hoofdstuk 2 betrekking heeft op een eerdere

studie en de huidige APS niet in modules voorzien voor geïntegreerde seriegrootte-
bepaling en roostering, richt het tweede deel van de studie in hoofdstuk 2 zich op een

151
literatuuronderzoek naar modelontwikkelingen voor simultane seriegroottebepaling en
roostering. Voorts beoogt de studie de geschiktheid van modellen te evalueren ten
aanzien van volgorde afhankelijke omstellingen, omstelmatrices waarvoor de driehoeks-
voorwaarden niet gelden, en de bederfelijkheid van geproduceerde producten. Het
onderzoek toont een tendens aan waarin middellangetermijn modellen, zogenaamde
“Big Bucket models” (BB), ook worden voorgesteld voor problemen met een kortetermijn
planningshorizon. Wij betogen echter dat de segmentatie van de planningshorizon (i.e.
het aggregatieniveau in de tijdshorizon) cruciaal is voor simultane seriegroottebepaling
en roostering. Het voordeel van BB modellen kan een groot obstakel vormen voor i) de
effectiviteit van simultane seriegrootteplanning en roostering, en ii) het aanpakken van
specifieke probleem karakteristieken uit de voedingsmiddelenindustrie.

Studie 2: Verticale integratie van seriegrootteproblemen en roostering in de

voedingsmiddelenindustrie.

Hoofdstuk 3 is gericht op volledige integratie van seriegrootteproblemen en roostering in

één enkel model. De hoofdvraag was: “Hoe kunnen productieplanning (d.w.z. planning
van seriegroottes) en roostering in één enkel model worden opgenomen, zodanig dat de
gebruikelijke aannames over de driehoeksongelijkheden worden losgelaten en de
bederfelijkheid en beperkte houdbaarheid van producten modelmatig worden mee-
genomen?”

Het literatuuroverzicht in hoofdstuk 2 bracht aan het licht dat het rekentechnische
voordeel van tijdsgeoriënteerde aggregatie in BB modellen een groot struikelblok kan
vormen voor het aanpakken van de vermelde probleemkarakteristieken in de
voedingsmiddelenindustrie. Bederfelijkheid is primair geassocieerd met de “leeftijd” van
producten, en is bijgevolg gerelateerd aan de segmentatie van de tijdshorizon. Derhalve
zijn twee SB modellen ontwikkeld. Deze modellen laten zien welk effect zowel
bederfelijkheid als het loslaten van de driehoeksvoorwaarden1 hebben op de
gegenereerde oplossingen. Aan de hand van illustratieve voorbeelden wordt duidelijk
hoe kleine veranderingen in de balans tussen voorraad- en omstelkosten tot wezenlijk
verschillende oplossingen kunnen leiden. Dit laatste geldt vooral als de driehoeks-
ongelijkheden niet gelden.

De ontwikkelde modellen kunnen in een praktische context tot zeer grote

problemen leiden, en zijn – zoals verwacht – uiterst moeilijk oplosbaar. Een verkennend
onderzoek werd verricht met een Relax-and-Fix (R&F) heuristiek. De heuristiek is
gebaseerd op een decompositie van de tijdshorizon. De numerieke resultaten van kleine
tot middelgrote problemen zijn bemoedigend. Het is echter nog niet mogelijk om
problemen van realistische probleemomvang op te lossen.

1
De driehoeksvoorwaarden beschrijven de aanname dat de totale omstelkosten en -tijden
tussen twee achtereenvolgende productieseries voor verschillende producten niet toe-
nemen door een derde product tussen betreffende productiehoeveelheden te produceren.

152
 Samenvatting

Studie 3: Geïntegreerde planning tussen verwerving en verwerking.

De casus in hoofdstuk 4 richt zich op de noodzaak tot horizontale coördinatie en

integratie tussen de verwerving en verwerking van de grondstoffen tot eindproducten.
Deze integratie is van bijzonder belang voor situaties waarin de interactie tussen
verschillende productieketens centraal staat. De hoofdvraag was: “Hoe kan de
integratie van een planningsprobleem voor de verwerving en verwerking van
grondstoffen in een logistieke keten die uit meerdere partijen (i.e. coöperaties) bestaat,
zowel op het niveau van korte- als middellangetermijn worden gemodelleerd en
opgelost? De casus betrof een melkophaalprobleem.

Het doel was wederom een blauwdruk DSS te ontwikkelen dat beslissings-
ondersteuning voor een “zwakkere” ketenpartner naar een hoger niveau kon tillen.
Voorts beoogt de studie te illustreren hoe belangrijk horizontale integratie tussen de
fasen verwerving en verwerking in een APS raamwerk is.

De probleemanalyse toonde aan dat het probleem kan worden geclassificeerd

als een uitbreiding van het zogenaamde “Periodic Vehicle Routing Problem” (PVRP).
Via decompositie naar hiërarchische niveaus ontstonden beter oplosbare
deelproblemen. Het dagelijks routeringsvraagstuk werd daarmee gescheiden van een
middellangetermijn planningsprobleem. Op het hoogste planningsniveau werden de
talrijke leveranciers geaggregeerd tot clusters zodanig dat het totale aanbod van melk
binnen een cluster de maximale transportcapaciteit van (meerdere) vrachtwagens
benaderde. De continue productie van relatief kleine hoeveelheden melk door vele
leveranciers moest in balans worden gebracht met strikte aflevercondities op
verwerkingsniveau. Een model werd ontwikkeld voor het toewijzen van één enkel
(stabiel) ophaalschema aan elk cluster, zodanig dat de totale gewogen afwijking van de
gewenste hoeveelheden op diverse verwerkingslocaties op verschillende dagen in de
planningshorizon zo laag mogelijk was.

De toegepaste aggregatie op het hogere planningsniveau bleek van grote

waarde te zijn voor de uiteindelijke disaggregatie op het laagste planningsniveau. Nadat
alle toeleverende bedrijven geografisch waren geclusterd en het geaggregeerde aanbod
van melk aan één schema (per cluster) met vaste ophaaldagen was gekoppeld werd het
dagelijkse routeringsvraagstuk aanzienlijk eenvoudiger op te lossen voor de planners.

De rekentechnische complexiteit van het probleem kon worden gereduceerd

door specifieke eigenschappen van de casus algoritmisch te gebruiken voor de
ontwikkeling van een (specifiek) branch-and-bound schema. Hierdoor werd het mogelijk
om problemen van realistische omvang exact op te lossen.

De verschillende functionaliteiten die via een gebruikersvriendelijk en interactief

gebruikersinterface werden aangereikt (i.e. een invoer-, plannings-, simulatie- en
analysemodule) bleken essentieel te zijn. In de simulatiemodule konden eindgebruikers
bijvoorbeeld eenvoudig zowel de data als de gegenereerde planningen veranderen. Via

153
de uitvoerschermen werd het effect van zulke veranderingen direct zichtbaar gemaakt
via diverse (veelal conflicterende) indicatoren.

Studie 4: Gemengd geheeltallige (binaire) fractionele programmering in de

papierindustrie.

Het onderzoek in hoofdstuk 5 concentreert zich op het effect van machine-instellingen

op goederenstromen. De hoofdvraag was: “Hoe kunnen beslissers in de praktijk worden
ondersteund als cruciale eigenschappen van eindproducten gelijktijdig afhangen van (te
bepalen) typen grondstoffen met verschillende chemische en fysische eigenschappen
en van (te bepalen) instellingen van procesbewerkingseenheden in een productie-
proces?”

Het doel van het onderzoek was om een bestaand, lokaal gebruikt DSS door te
ontwikkelen tot een flexibel instrument voor beslissingsondersteuning binnen de
onderneming. Het onderzoek bracht aan het licht dat de beoogde flexibiliteit in
doelstellingsfuncties voor de mathematische programmering, gecombineerd met nieuwe
fysische inzichten voor het berekenen van eindproducteigenschappen na
procesbewerkingen, een aanzienlijk effect had op de optimaliseringsmodule.

De voorgestelde methode om deze categorie van niet-lineaire probleem op te

lossen maakt gebruik van probleemkarakteristieken en leidt ertoe dat een klassieke
herformuleringsaanpak voor de continue lineaire fractionele programmering (FP) – na
uitbreiding – ook kan worden toegepast op een meer algemene klasse van gemengd-
geheeltallige (binaire) FP-problemen. De speciale samenhang tussen het
oorspronkelijke, niet-lineaire, gemengd-geheeltallige model en de continue, lineaire
herformulering werd benut door gebruik te maken van Special Ordered Sets type 1
(SOS1).

Ofschoon hoofdstuk 5 zich met name richt op de herformulerings- en

oplossingsaanpak bestaat het DSS uit vier bouwstenen, namelijk het gebruikers-
interface, een scenario manager, een simulatie- en een optimaliseringsmodule. De
optimalisatiemodule voorziet in een krachtig hulpmiddel voor het zoeken naar toegelaten
oplossingen en de beste (onverwachte) recepturen voor elke beschikbare verzameling
van grondstoffen. Bovendien biedt de module op innovatieve wijze ondersteuning bij de
besluitvorming over de aankoop van (nieuwe) pulpsoorten, voor het toewijzen van
beschikbare pulpsoorten aan diverse papiersoorten, en voor het afstemmen van
beschikbare grondstofvoorraden op (veranderende) productiedoelen. De resultaten van
de optimalisatieroutine worden vooral gebruikt om alternatieve recepturen te genereren.
Gewoonlijk worden deze recepturen als basis-scenario’s bewaard en in de
simulatiemodule afgestemd op de dagelijkse praktijk.

Voornaamste conclusies en toekomstig onderzoek

Op basis van het onderzoek in de hoofdstukken 2 en 3 concluderen we dat geen

algemeen toepasbare modellen en/of oplossingstechnieken bestaan voor simultaan

154
 Samenvatting

plannen en roosteren in de procesindustrie. Meer industrie-specifieke oplossingen zijn

nodig die de specifieke kenmerken van verschillende productieomgevingen in die
modellen opnemen. De sleutel tot het ontwerp van effectieve oplossingen voor de
hedendaagse praktijk zou kunnen zijn i) het gebruik van praktijkkennis en -ervaring en
het benutten van domeinspecifieke probleemeigenschappen tijdens het bouwen van het
model, en/of ii) het identificeren en benutten van speciale probleemstructuren.

We concluderen dat opvallend weinig onderzoek is gewijd aan de coördinatie en

integratie tussen de verwerving en verwerking van grondstoffen. De studies in hoofdstuk
4 en 5 bevestigen dat planningsproblemen met betrekking tot het verwerven van
grondstoffen in de procesindustrie meer aandacht verdienen, met name in “push”-
georiënteerde logistieke netwerken waarbij de interactie tussen verschillende
productieketens van belang is. De verwaarding van grondstoffen kan nog verder worden
verbeterd indien bij productieplanningsproblemen ook de samenstelling van de
grondstoffen in beschouwing wordt genomen.

In het tweede deel van dit proefschrift richtten we ons op begrip en inzicht in,
alsmede de uitbreiding van, de toepasbaarheid van Special Ordered Sets type 1
(SOS1), zowel vanuit algoritmisch (hoofdstuk 4) als vanuit modelmatig oogpunt
(hoofdstuk 5). We concluderen dat met SOS1 een klassieke herformuleringsaanpak
voor continue fractionele programmeringsproblemen kan worden uitgebreid tot een
klasse van gemengd-geheeltallige (binaire) FP-problemen. Bovendien concluderen we
dat het gebruik van SOS1 ook zonder een natuurlijke ordening van de variabelen binnen
de sets rekentechnisch interessant is. Voor een efficiënt gebruik van SOS1 in
commerciële software voor de mathematische programmering zou de door de gebuiker
te definiëren referentierij kunnen vervallen. Dit laatste vereist echter nader onderzoek en
uitgebreide rekentests.

155
156
 References

References
Ackoff, R. L. (1977). Optimization + objectivity = optout. European Journal of Operational
Research 1, 1-7.

Ackoff, R. L. (2001). OR: After the post mortem. System Dynamics Review 17, 341-346.

Agarwal, R., and Lucas Jr, H. C. (2005). The information systems identity crisis: Focusing
on high-visibility and high-impact research. MIS Quarterly: Management Information
Systems 29, 381-398.

Ahlatcioglu, M., and Tiryaki, F. (2007). Interactive fuzzy programming for decentralized two-
level linear fractional programming (DTLLFP) problems. Omega 35, 432-450.

Akkermans, H. A., Bogerd, P., Yücesan, E., and Van Wassenhove, L. N. (2003). The
impact of ERP on supply chain management: Exploratory findings from a European
Delphi study. European Journal of Operational Research 146, 284-301.

Alemayehu, G., and Arora, S. R. (2002). A short note on the mixed (0, 1) linear fractional
programming problem. Indian Journal of Pure and Applied Mathematics 33, 895-
901.

Allahverdi, A., Gupta, J. N. D., and Aldowaisan, T. (1999). A review of scheduling research
involving setup considerations. Omega 27, 219-239.

Almada-Lobo, B., Klabjan, D., Carravilla, M. A., and Oliveira, J. F. (2007). Single machine
multi-product capacitated lot sizing with sequence-dependent setups. International
Journal of Production Research 45, 4873-4894.

Almada-Lobo, B., Oliveira, J. F., and Carravilla, M. A. (2008). A note on "the capacitated
lot-sizing and scheduling problem with sequence-dependent setup costs and setup
times". Computers & Operations Research 35, 1374-1376.

Amorim, P., Antunes, C. H., and Almada-Lobo, B. (2011). Multi-Objective Lot-Sizing and
Scheduling Dealing with Perishability Issues. Industrial & Engineering Chemistry
Research 50, 3371-3381.

Amos, F., Ronnqvist, M., and Gill, G. (1997). Modelling the pooling problem at the New
Zealand Refining Company. Journal of the Operational Research Society 48, 767-
778.

Angus, J. Simultaneous determination of production lot size and process parameters under
process deterioration and process breakdown. Omega.

Anthony, R. N. (1965). "Planning and control systems; a framework for analysis," Division of
Research, Graduate School of Business Administration, Harvard University,
Boston,.

157
Arnott, D., and Pervan, G. (2005). A critical analysis of decision support systems research.
Journal of Information Technology 20, 67-87.

Arnott, D., and Pervan, G. (2008). Eight key issues for the decision support systems
discipline. Decision Support Systems 44, 657-672.

Arora, S. (2002). A short note on the mixed (0, 1) linear fractional programming problem.
Indian J. pure appi. Math 33, 895-901.

Ashford, R. W., and Daniel, R. C. (1992). Some Lessons in Solving Practical Integer
Programs. The Journal of the Operational Research Society 43, 425-433.

Bajalinov, E. B. (2003). "Linear-fractional programming: theory, methods, applications and

software," Springer.

Baldacci, R., Mingozzi, A., and Roberti, R. (2011). New route relaxation and pricing
strategies for the vehicle routing problem. Operations Research 59, 1269-1283.

Banaszewska, A., Cruijssen, F. C. A. M., van der Vorst, J. G. A. J., Claassen, G. D. H., and
Kampman, J. L. (2013). A comprehensive dairy valorization model. Journal of dairy
science 96, 761-779

Barros, A. I. (1998). "Discrete and fractional programming techniques for location models,"
Springer.

Basnet, C., Foulds, L., and Igbaria, M. (1996). FleetManager: a microcomputer-based

decision support system for vehicle routing. Decision Support Systems 16, 195-207.

Baumann, P., and Trautmann, N. (2012). A continuous-time MILP model for short-term
scheduling of make-and-pack production processes. International Journal of
Production Research 51, 1707-1727.

Bazaraa, M., Sherali, H., and Shetty, C. (2006). Nonlinear Programming Theory and
Application. . Jon Wiley and Sons, New York.

Beale, E. M. L., and Tomlin, J. A. (1970). Special facilities in a general mathematical

programming system for non-convex problems using ordered sets of variables. OR
69, 447-454.

Belvaux, G., and Wolsey, L. A. (2001). Modelling practical lot-sizing problems as mixed-
integer programs. Management Science 47, 993-1007.

Berends, P. A. J., and Romme, A. G. L. (2001). Cyclicality of capital-intensive industries: A

system dynamics simulation study of the paper industry. Omega 29, 543-552.

Bilgen, B., and Günther, H. O. (2010). Integrated production and distribution planning in the
fast moving consumer goods industry: a block planning application. OR Spectrum
32, 927-955.

158
 References

Bouchriha, H., Ouhimmou, M., and D’Amours, S. (2007). Lot sizing problem on a paper
machine under a cyclic production approach. International Journal of Production
Economics 105, 318-328.

Bredström, D., Lundgren, J. T., Rönnqvist, M., Carlsson, D., and Mason, A. (2004). Supply
chain optimization in the pulp mill industry––IP models, column generation and
novel constraint branches. European Journal of Operational Research 156, 2-22.

Brueggemann, W., and Jahnke, H. (2000). The discrete lot-sizing and scheduling problem:
Complexity and modification for batch availability (vol 124, pg 511, 2000). European
Journal of Operational Research 126, 688-688.

Butler, M., Williams, H. P., and Yarrow, L. A. (1997). The Two-Period Travelling Salesman
Problem Applied to Milk Collection in Ireland. Computational Optimization and
Applications 7, 291-306.

Calvete, H. I., and Galé, C. (2004). Solving linear fractional bilevel programs. Operations
Research Letters 32, 143-151.

Carlsson, C., and Turban, E. (2002). DSS: directions for the next decade. Decision Support
Systems 33, 105-110.

Carlsson, D., and Rönnqvist, M. (2007). Backhauling in forest transportation: Models,

methods, and practical usage. Canadian Journal of Forest Research 37, 2612-2623.

Chakraborty, M., and Chandra, M. K. (2005). Multicriteria decision making for optimal
blending for beneficiation of coal: A fuzzy programming approach. Omega 33, 413-
418.

Chakraborty, T., Giri, B. C., and Chaudhuri, K. S. (2009). Production lot sizing with process
deterioration and machine breakdown under inspection schedule. Omega 37, 257-
271.

Chang, C. T. (2007). Multi-choice goal programming. Omega 35, 389-396.

Chao, I. M., Golden, B. L., and Wasil, E. (1995a). An improved heuristic for the period
vehicle routing problem. Networks 26, 25-44.

Chao, I. M., Golden, B. L., and Wasil, E. A. (1995b). A new heuristic for the period traveling
salesman problem. Computers and Operations Research 22, 553-565.

Charnes, A., and Cooper, W. W. (1962). Programming with linear fractional functionals.
Naval Research Logistics Quarterly 9, 181-186.

Chauhan, S. S., Martel, A., and D'Amour, S. (2008). Roll assortment optimization in a paper
mill: An integer programming approach. Computers and Operations Research 35,
614-627.

Chen, H.-K., Hsueh, C.-F., and Chang, M.-S. (2009). Production scheduling and vehicle
routing with time windows for perishable food products. Computers & Operations
Research 36, 2311-2319.

159
Christofides, N., and Beasley, J. E. (1983). Extensions to a Lagrangean relaxation
approach for the capacitated warehouse location problem. European Journal of
Operational Research 12, 19-28.

Christofides, N., and Beasley, J. E. (1984). The period routing problem. Networks 14, 237-
256.

Christopher, M. (2005). "Logistics and supply chain management : creating value-adding

networks," Financial Times/Prentice Hall, Harlow [etc.].

Ciriani, T. A., Colombani, Y., and Heipcke, S. (2002). Optimisation Modelling Innovations. In
"Proceedings of" VI Congresso Nazionale SIMAI", Chia Laguna (Sardinia, Italy)",
pp. 248. Citeseer.

Ciriani, T., Colombani, Y., and Heipcke, S. (2003). Embedding optimisation algorithms with
Mosel. Quarterly Journal of the Belgian, French and Italian Operations Research
Societies 1, 155-167.

Claassen, G. D. H., and Beek, P. v. (1993). Planning and scheduling packaging lines in
food-industry. European Journal of Operational Research 70, 150-158.

Claassen, G. D. H., and Hendriks, Th. H. B. (2007a). An application of special ordered sets
to a periodic milk collection problem. European Journal of Operational Research
180, 754-769.

Claassen, G. D. H., Hendriks, Th. H. B, and Hendrix, E.M.T. (2007b). "Decision science :
theory and applications," Wageningen Academic Publishers, Wageningen.

Claassen, G. D. H., & Hendriks, Th. H. B. (2010, August). OR-based Decision Support in
Paper Production Industry. In: Bridging the Gap in Decision Support Systems -
Challenges for the Next Decade. Eds: Respicio, A., et al. Amsterdam: IOS Press,
(Frontiers in Artificial Intelligence and Applications 212), pp. 573-582.

Claassen, G. D. H. (2014). Mixed integer (0–1) fractional programming for decision support
in paper production industry. Omega 43, 21-29.

Claassen, G. D. H., Gerdessen, J. C., Hendrix, E. M. T., and Vorst van der, J.G A.J. (2014).
On production planning and scheduling in food processing industry: modelling non-
triangular setups and product decay. Under review.

Claassen, G.D.H. and Hendrix, E.M.T., (2014) On Modelling Approaches for Planning and
Scheduling in Food Processing Industry In: ICCSA 2014, Part II, Lecture Notes in
Computer Science 8580 Eds: B. Murgante et al. Springer, pp. 47 – 59

Clark, A., Almada-Lobo, B., and Almeder, C. (2011). Lot sizing and scheduling: Industrial
extensions and research opportunities. International Journal of Production Research
49, 2457-2461.

160
 References

Colombani, Y., Daniel, B., and Heipcke, S. (2004). MOSEL: A modular environment for
modeling and solving optimization problems. Modeling Languages in Mathematical
Optimization 88, 211-238.

Constantino, M. (1996). A cutting plane approach to capacitated lot-sizing with start-up

costs. Mathematical Programming 75, 353-376.

Cooper, W. W., and McAlister, L. (1999). Can research be basic and applied? You bet. It
better be for B-schools! Socio-Economic Planning Sciences 33, 257-276.

Cooper, W. W. (2005). Origins, uses of, and relations between goal programming and data
envelopment analysis. Journal of Multi-Criteria Decision Analysis 13, 3-11.

Cordeau, J.-F., Gendreau, M., and Laporte, G. (1997). A tabu search heuristic for periodic
and multi-depot vehicle routing problems. Networks 30, 105-119.

Cordeau, J.-F., Laporte, G., and Mercier, A. (2001). A unified tabu search heuristic for
vehicle routing problems with time windows. Journal of the Operational Research
Society, 928-936.

Cornuejols, G., Sridharan, R., and Thizy, J.-M. (1991). A comparison of heuristics and
relaxations for the capacitated plant location problem. European Journal of
Operational Research 50, 280-297.

D'Amours, S., Rönnqvist, M., and Weintraub, A. (2008). Using Operational Research for
supply chain planning in the forest products industry. INFOR 46, 265-281.

Dillenberger, C., Escudero, L. F., Wollensak, A., and Wu, Z. (1994). On practical resource-
allocation for production planning and scheduling with period overlapping setups.
European Journal of Operational Research 75, 275-286.

Dinkelbach, W. (1967). On nonlinear fractional programming. Management Science 13,

492-498.

Drexl, A., and Haase, K. (1995). Proportional lotsizing and scheduling. International Journal
of Production Economics 40, 73-87.

Drexl, A., and Kimms, A. (1997). Lot sizing and scheduling - Survey and extensions.
European Journal of Operational Research 99, 221-235.

Entrup, M. L. (2005). "Advanced Planning in Fresh Food Industries : Integrating Shelf Life
into Production Planning," Physica-Verlag Heidelberg, Heidelberg.

Entrup, M. L., Gunther, H. O., Van Beek, P., Grunow, M., and Seiler, T. (2005). Mixed-
Integer Linear Programming approaches to shelf-life-integrated planning and
scheduling in yoghurt production. International Journal of Production Research 43,
5071-5100.

Eom, H. B., and Lee, S. M. (1990a). Decision Support Systems Applications Research - a
Bibliography (1971-1988). European Journal of Operational Research 46, 333-342.

161
Eom, H. B., and Lee, S. M. (1990b). DSS Applications developments research: Leading
institutions and most frequent contributors (1971 - April 1988). Decision Support
Systems 6, 269-275.

Eom, H. B., and Lee, S. M. (1990c). A Survey of Decision Support System Applications
(1971-April 1988). Interfaces 20, 65-79.

Eom, S. B., Lee, S. M., and Kim, J. K. (1993). The intellectual structure of decision support
systems (1971-1989). Decision Support Systems 10, 19-35.

Eom, S. B., Lee, S. M., Kim, E. B., and Somarajan, C. (1998). A survey of decision support
system applications (1988-1994). Journal of the Operational Research Society 49,
109-120.

Eom, S., and Kim, E. (2006). A survey of decision support system applications (1995-2001).
Journal of the Operational Research Society 57, 1264-1278.

Eppen, G. D., and Martin, R. K. (1987). Solving multi-item capacitated lot-sizing problems
using variable redefinition. Operations Research 35, 832-848.

Erlenkotter, D. (1978). A dual-based procedure for uncapacitated facility location.

Operations Research 26, 992-1009.

Escudero, L. F., and Salmeron, J. (2005). On a fix-and-relax framework for a class of

project scheduling problems. Annals of Operations Research 140, 163-188.

Everett, G., Philpott, A., and Cook, G. (2002). Capital planning under uncertainty at Fletcher
Challenge Canada. In "56th Appita Annual Conference, Rotorua, New Zealand 18-
20 March 2002: Proceedings", pp. 429. Appita Inc.

Federgruen, A., Meissner, J., and Tzur, M. (2007). Progressive interval heuristics for multi-
item capacitated lot-sizing problems. Operations Research 55, 490-502.

Ferreira, D., Morabito, R., and Rangel, S. (2009). Solution approaches for the soft drink
integrated production lot sizing and scheduling problem. European Journal of
Operational Research 196, 697-706.

Fleischmann, B. (1990). The discrete lot-sizing and scheduling problem. European Journal
of Operational Research 44, 337-348.

Fleischmann, B. (1994). The discrete lot-sizing and scheduling problem with sequence-
dependent setup costs. European Journal of Operational Research 75, 395-404.

Fleischmann, B., and Meyr, H. (1997). The general lotsizing and scheduling problem. Or
Spektrum 19, 11-21.

Fleischmann, B., and Meyr, H. (2003). Planning hierarchy, modeling and advanced
planning systems. Handbooks in Operations Research and Management Science
11, 455-523.

162
 References

Forslund, H., and Jonsson, P. (2010). Integrating the performance management process of
on-time delivery with suppliers. International Journal of Logistics Research and
Applications 13, 225-241.

Framinan, J. M., and Ruiz, R. (2010). Architecture of manufacturing scheduling systems:

Literature review and an integrated proposal. European Journal of Operational
Research 205, 237-246.

Francis, P. M., Smilowitz, K. R., and Tzur, M. (2008). The period vehicle routing problem
and its extensions. In "The vehicle routing problem: latest advances and new
challenges", pp. 73-102. Springer.

Fransoo, J. C., and Rutten, W. (1994). A typology of production control situations in

processs industries. International Journal of Operations & Production
Management 14, 47-57.

Garey, M. R., Johnson, D. S., and Sethi, R. (1976). The complexity of flowshop and jobshop
scheduling. Mathematics of operations research 1, 117-129.
Gaudioso, M., and Paletta, G. (1992). A heuristic for the periodic vehicle routing problem.
Transportation Science 26, 86-92.

Gayialis, S. P., and Tatsiopoulos, I. P. (2004). Design of an IT-driven decision support

system for vehicle routing and scheduling. European Journal of Operational
Research 152, 382-398.

Gerdessen, J. C. (1996). Vehicle routing problem with trailers. European Journal of

Operational Research 93, 135-147.

Ghiani, G., Guerriero, F., Laporte, G., and Musmanno, R. (2003). Real-time vehicle routing:
Solution concepts, algorithms and parallel computing strategies. European Journal
of Operational Research 151, 1-11.

Golden, B. L., and Wasil, E. A. (1987). Computerized vehicle routing in the soft drink
industry. Operations Research 35, 6-17.

Golden, B. L., Raghavan, S., and Wasil, E. A. (2008). "The vehicle routing problem
[electronic resource]: latest advances and new challenges," Springer.

Gopalakrishnan, M., Miller, D. M., and Schmidt, C. P. (1995). A framework for modeling
setup carryover in the capacitated lot-sizing problem. International Journal of
Production Research 33, 1973-1988.

Gopalakrishnan, M. (2000). A modified framework for modelling set-up carryover in the

capacitated lotsizing problem. International Journal of Production Research 38,
3421-3424.

Gorry, G. A., and Morton, M. S. S. (1971). Framework for Management Information

Systems. Sloan Management Review 13, 55-70.

Granot, D., and Granot, F. (1977). On integer and mixed integer fractional programming
problems. Annals of Discrete Mathematics 1, 221-231.

163
Gunnarsson, H., Rönnqvist, M., and Carlsson, D. (2006). A combined terminal location and
ship routing problem. Journal of the Operational Research Society 57, 928-938.

Gunnarsson, H., Rönnqvist, M., and Carlsson, D. (2007). Integrated production and
distribution planning for Södra Cell AB. Journal of Mathematical Modelling and
Algorithms 6, 25-45.

Gunnarsson, H., and Rönnqvist, M. (2008). Solving a multi-period supply chain problem for
a pulp company using heuristics—An application to Södra Cell AB. International
Journal of Production Economics 116, 75-94.

Gunther, H. O., Grunow, M., and Neuhaus, U. (2006). Realizing block planning concepts in
make-and-pack production using MILP modelling and SAP APO. International
Journal of Production Research 44, 3711-3726.

Günther, H. O., and van Beek, P. (2003). "Advanced planning and scheduling solutions in
process industry," Springer, Berlin [etc.].

Gupta, D., and Magnusson, T. (2005). The capacitated lot-sizing and scheduling problem
with sequence-dependent setup costs and setup times. Computers & Operations
Research 32, 727-747.

Haase, K., and Kimms, A. (2000). Lot sizing and scheduling with sequence-dependent
setup costs and times and efficient rescheduling opportunities. International Journal
of Production Economics 66, 159-169.

Haase, K. (1996). Capacitated lot-sizing with sequence dependent setup costs. Or

Spektrum 18, 51-59.

Harris, F. W. (1913). How much stock to keep on hand. Factory, The Magazine of
Management 10.

Hax, A. C., and Candea, D. (1984). "Production and inventory management," Prentice-Hall,
Englewood Cliffs, N.J.

Hopp, W. J. (2004). 50th Anniversary Article Fifty Years of Management Science.

Management Science 50, 1-7.

Hua, Z., Bian, Y., and Liang, L. (2007). Eco-efficiency analysis of paper mills along the Huai
River: An extended DEA approach. Omega 35, 578-587.

Huiskonen, J., Niemi, P., and Pirttila, T. (2003). An approach to link customer
characteristics to inventory decision making. International Journal of Production
Economics 81-2, 255-264.

ILOG, C. I. (2009). V12. 1 User’s Manual for CPLEX. International Business Machines
Corporation.

Ivert, L. K., and Jonsson, P. (2010). The potential benefitsofadvanced planning and
scheduling systems in sales and operations planning. Industrial Management and
Data Systems 110, 659-681.

164
 References

Ivert, L. K., and Jonsson, P. (2011). Problems in the onward and upward phase of APS
system implementation: Why do they occur? International Journal of Physical
Distribution and Logistics Management 41, 343-363.

Ivert, L. K. (2012). Shop floor characteristics influencing the use of advanced planning and
scheduling systems. Production Planning and Control 23, 452-467.

Jacobs, F. R., and Weston, F. C. (2007). Enterprise resource planning (ERP) - A brief
history. Journal of Operations Management 25, 357-363.

Jans, R., and Degraeve, Z. (2008). Modeling industrial lot sizing problems: A review.
International Journal of Production Research 46, 1619-1643.

Johnston, R. E. (1980). OR in the Paper Industry. The International Journal of Management

Science 9, 43-50.

Jonsson, P., Kjellsdotter, L., and Rudberg, M. (2007). Applying advanced planning systems
for supply chain planning: Three case studies. International Journal of Physical
Distribution and Logistics Management 37, 816-834.

Jordan, C., and Drexl, A. (1998). Discrete lotsizing and scheduling by batch sequencing.
Management Science 44, 698-713.

Kallrath, J. (2000). Mixed integer optimization in the chemical process industry: Experience,
potential and future perspectives. Chemical Engineering Research and Design 78,
809-822.

Kallrath, J. (2002). Planning and scheduling in the process industry. OR Spectrum 24, 219-
250.

Kallrath, J. (2004). "Modeling languages in mathematical optimization," Springer.

Kanyalkar, A. P., and Adil, G. K. (2005). An integrated aggregate and detailed planning in a
multi-site production environment using linear programming. International Journal of
Production Research 43, 4431-4454.

Kanyalkar, A. P., and Adil, G. K. (2007). Aggregate and detailed production planning
integrating procurement and distribution plans in a multi-site environment.
International Journal of Production Research 45, 5329-5353.

Kanyalkar, A. P., and Adil, G. K. (2010). A robust optimisation model for aggregate and
detailed planning of a multi-site procurement-production-distribution system.
International Journal of Production Research 48, 635-656.

Kanyalkar, A. P., and Adil, G. K. (2011). Integrated planning in a multi-site procurement,

production and distribution system under rolling schedule. International Journal of
Services and Operations Management 9, 162-182.

Karimi, B., Ghomi, S., and Wilson, J. M. (2003). The capacitated lot sizing problem: a
review of models and algorithms. Omega-International Journal of Management
Science 31, 365-378.

165
Karmarkar, U. S., and Schrage, L. (1985). The deterministic dynamic product cycling
problem. Operations Research 33, 326-345.

Kopanos, G. M., Puigjaner, L., and Georgiadis, M. C. (2012). Simultaneous production and
logistics operations planning in semicontinuous food industries. Omega 40, 634-
650.

Kreipl, S., and Dickersbach, J. T. (2008). Scheduling coordination problems in supply chain
planning. Annals of Operations Research 161, 103-122.

Kreipl, S., and Pinedo, M. (2004). Planning and scheduling in supply chains: an overview of
issues in practice. Production and Operations management 13, 77-92.

Kuik, R., Salomon, M., and Vanwassenhove, L. N. (1994). Batching decisions - structure
and models. European Journal of Operational Research 75, 243-263.

Lee, I. S., and Yoon, S. H. (2010). Coordinated scheduling of production and delivery
stages with stage-dependent inventory holding costs. Omega-International Journal
of Management Science 38, 509-521.

Liberti, L. (2009). Reformulations in mathematical programming: Definitions and

systematics. RAIRO - Operations Research 43, 55-85.

Lin, C., Hwang, S., and Wang, E. M. (2007). A reappraisal on advanced planning and
scheduling systems. Industrial Management and Data Systems 107, 1212-1226.

Little, J. D. C. (1970). Models and Managers - Concept of a Decision Calculus.

Management Science Series B-Application 16, B466-B485.

Little, J. D. C. (2004). Comments on "Models and managers: The concept of a decision

calculus". Management Science 50, 1854-1860.

Lummus, R. R., and Vokurka, R. J. (1999). Defining supply chain management: a historical
perspective and practical guidelines. Industrial Management & Data Systems 99,
11-17.

Lütke-Entrup, M., Günther, H. O., van Beek, P., Grunow, M., and Seiler, T. (2005). Mixed
integer linear programming approaches to shlef life integrated planning and
scheduling in yogurt production. International Journal of Production Research
Volume 43, 5071-5100.

Maravelias, C. T., and Grossmann, I. E. (2003). A new continuous-time state task network
formulation for short term scheduling of multipurpose batch plants. Vol. 14, pp. 215-
220.

Maravelias, C. T., and Sung, C. (2009). Integration of production planning and scheduling:
Overview, challenges and opportunities. Computers & Chemical Engineering 33,
1919-1930.

Martel, A., Rizk, N., D'Amours, S., and Bouchriha, H. (2005). Synchronized production-
distribution planning in the pulp and paper industry. pp. 323-350.

166
 References

Méndez, C. A., Cerdá, J., Grossmann, I. E., Harjunkoski, I., and Fahl, M. (2006). State-of-
the-art review of optimization methods for short-term scheduling of batch processes.
Computers & Chemical Engineering 30, 913-946.

Menezes, A. A., Clark, A., and Almada-Lobo, B. (2011). Capacitated lot-sizing and
scheduling with sequence-dependent, period-overlapping and non-triangular setups.
Journal of Scheduling 14, 209-219.

Meyr, H. (2000). Simultaneous lotsizing and scheduling by combining local search with dual
reoptimization. European Journal of Operational Research 120, 311-326.

Mitroff, I. I. (1974). "The subjective side of science : a philosophical inquiry into the
psychology of the Apollo moon scientists," Elsevier Scientific Pub. Co.; American
Elsevier Pub. Co., Amsterdam New York.

Møller, C. (2005). ERP II: A conceptual framework for next-generation enterprise systems?
Journal of Enterprise Information Management 18, 483-497.

Mourgaya, M., and Vanderbeck, F. (2007). Column generation based heuristic for tactical
planning in multi-period vehicle routing. European Journal of Operational Research
183, 1028-1041.

Mula, J., Peidro, D., Díaz-Madroñero, M., and Vicens, E. (2010). Mathematical
programming models for supply chain production and transport planning. European
Journal of Operational Research 204, 377-390.

Murthy, S., Akkiraju, R., Goodwin, R., Keskinocak, P., Rachlin, J., Frederick, W., Yeh, J.,
Fuhrer, R., Kumaran, S., Aggarwal, A., Sturzenbecker, M., Jayaraman, R., and
Daigle, R. (1999). Cooperative multiobjective decision support for the paper
industry. Interfaces 29, 5-30.

Myers, G. C. (2002). Thermomechanical pulping of loblolly pine juvenile wood. Wood and
Fiber Science 34, 108-115.

Olhager, J., and Prajogo, D. I. (2012). The impact of manufacturing and supply chain
improvement initiatives: A survey comparing make-to-order and make-to-stock
firms. Omega 40, 159-165.

Pati, R. K., Vrat, P., and Kumar, P. (2008). A goal programming model for paper recycling
system. Omega 36, 405-417.

Paulapuro, H. (2000a). "Book 8: Papermaking part 1: stock preparation and wet end,"
Helsinki [etc.]: Fapet Oy [etc.].

Paulapuro, H. (2000b). "Book 18: Paper and board grades," Helsinki [etc.]: Fapet Oy [etc.].

Philpott, A., and Everett, G. (2001(a)). Supply Chain Optimisation in the Paper Industry.
Annals of Operations Research 108, 225-237.

167
Philpott, A., Everett, G., and Aoude, S. (2001(b)). Capital planning in the paper industry
using COMPASS. In "Proceedings of the 36th Annual Conference of the ORSNZ,
Christchurch", Vol. 78, pp. 87.

Pochet, Y., and Wolsey, L. A. (2006). "Production planning by mixed integer programming,"
Springer, New York ; Berlin.

Porkka, P. P., Vepsalainen, A. P. J., and Kuula, M. (2003). Multiperiod production planning
carrying over set-up time. International Journal of Production Research 41, 1133-
1148.

Power, D. J., and Sharda, R. (2007). Model-driven decision support systems: Concepts and
research directions. Decision Support Systems 43, 1044-1061.

Quadt, D., and Kuhn, H. (2008). Capacitated lot-sizing with extensions: A review. 4OR 6,
61-83.

Rizk, N., Martel, A., and D'Amours, S. (2006). Multi-item dynamic production-distribution
planning in process industries with divergent finishing stages. Computers and
Operations Research 33, 3600-3623.

Robert Jacobs, F., and Ted' Weston Jr, F. C. (2007). Enterprise resource planning (ERP)-A
brief history. Journal of Operations Management 25, 357-363.

Robillard, P. (1971). (0, 1) hyperbolic programming problems. Naval Research Logistics

Quarterly 18, 47-57.

Rohde, J. (2004). Hierarchical supply chain planning using artificial neural networks to
anticipate base-level outcomes. OR Spectrum 26, 471-492.

Rohde, J., Meyr, H., and Wagner, M. (2000). "Die supply chain planning matrix." Darmstadt
Technical University, Department of Business Administration, Economics and Law,
Institute for Business Studies (BWL).

Roosma, J., and Claassen, G.D.H. (1996). Production planning in the fodder industry. OR
at work. Taylor & Francis, London, 133-146.

Rudberg, M., and Thulin, J. (2009). Centralised supply chain master planning employing
advanced planning systems. Production Planning & Control 20, 158-167.

Salomon, M., Kroon, L. G., Kuik, R., and Vanwassenhove, L. N. (1991). Some extensions
of the discrete lotsizing and scheduling problem. Management Science 37, 801-812.

Salomon, M., Solomon, M. M., VanWassenhove, L. N., Dumas, Y., and DauzerePeres, S.
(1997). Solving the discrete lotsizing and scheduling problem with sequence
dependent set-up costs and set-up times using the Travelling Salesman Problem
with time windows. European Journal of Operational Research 100, 494-513.

Schaible, S. (1995). Fractional Programming. In "Handbook of Global Optimization" (R.

Horst and P. Pardalos, eds.), Vol. 2, pp. 495-608. Springer US.

168
 References

Schaible, S., and Ibaraki, T. (1983). Fractional programming. European Journal of

Operational Research 12, 325-338.

Schaible, S., and Shi, J. (2004). Recent developments in fractional programming: single-
ratio and max-min case. Nonlinear analysis and convex analysis, 493-506.

Schrage, L. (1975). Implicit representation of variable upper bounds in linear programming.

In "Computational practice in mathematical programming", pp. 118-132. Springer.

Schrage, L. (1978). Implicit representation of generalized variable upper bounds in linear

programming. Mathematical Programming 14, 11-20.

Shah, N. (2005). Process industry supply chains: Advances and challenges. Computers
and Chemical Engineering 29, 1225-1235.

Shah, N. K., and Ierapetritou, M. G. (2012). Integrated production planning and scheduling
optimization of multisite, multiproduct process industry. Computers and Chemical
Engineering 37, 214-226.

Shim, J. P., Warkentin, M., Courtney, J. F., Power, D. J., Sharda, R., and Carlsson, C.
(2002). Past, present, and future of decision support technology. Decision Support
Systems 33, 111-126.

Simon, H. A. (1960). "The new science of management decision," 1st/Ed. Harper, New
York,.

Soman, C. A., Van Donk, D. P., and Gaalman, G. (2004a). Combined make-to-order and
make-to-stock in a food production system. International Journal of Production
Economics 90, 223-235.

Soman, C. A., van Donk, D. P., and Gaalman, G. J. C. (2004b). A basic period approach to
the economic lot scheduling problem with shelf life considerations. International
Journal of Production Research 42, 1677-1689.

Soman, C. A., van Donk, D. P., and Gaalman, G. J. C. (2007). Capacitated planning and
scheduling for combined make-to-order and make-to-stock production in the food
industry: An illustrative case study. International Journal of Production Economics
108, 191-199.

Sox, C. R., and Gao, Y. B. (1999). The capacitated lot sizing problem with setup carry-over.
Iie Transactions 31, 173-181.

Sprague Jr, R. H. (1980). A framework for the development of decision support systems.
MIS Quarterly: Management Information Systems 4, 1-26.

Sprague, R. H., and Watson, H. J. (1979). Bit by Bit - toward Decision Support Systems.
California Management Review 22, 60-68.

Stadtler, H. (2003). Multilevel lot sizing with setup times and multiple constrained resources:
Internally rolling schedules with lot-sizing windows. Operations Research 51, 487-
502.

169
Stadtler, H. (2005). Supply chain management and advanced planning––basics, overview
and challenges. European Journal of Operational Research 163, 575-588.

Stadtler, H., Fleischmann, B., Grunow, M., Meyr, H., and Sürie, C. (2012). "Advanced
Planning in Supply Chains : Illustrating the Concepts Using an SAP® APO Case
Study," Springer Berlin Heidelberg, Berlin, Heidelberg.

Stadtler, H., and Kilger, C. (2008). "Supply Chain Management and Advanced Planning :
Concepts, Models, Software, and Case Studies," Springer-Verlag Berlin Heidelberg,
Berlin, Heidelberg.

Stadtler, L., Schmitt, A., Klarner, P., and Straub, T. (2010). "More than Bricks in the Wall:
Organizational Perspectives for Sustainable Success," Gabler Verlag / GWV
Fachverlage GmbH, Wiesbaden, Wiesbaden.

Stancu-Minasian, I. (1999). A fifth bibliography of fractional programming*. Optimization 45,

343-367.

Stancu-Minasian, I. (2006). A sixth bibliography of fractional programming. Optimization 55,

405-428.

Suerie, C. (2006). Modeling of period overlapping setup times. European Journal of

Operational Research 174, 874-886.

Suerie, C., and Stadtler, H. (2003). The capacitated lot-sizing problem with linked lot sizes.
Management Science 49, 1039-1054.

Tan, C., and Beasley, J. E. (1984). A heuristic algorithm for the period vehicle routing
problem. Omega 12, 497-504.

Toth, P., and Vigo, D. (2002). Models, relaxations and exact approaches for the capacitated
vehicle routing problem. Discrete Applied Mathematics 123, 487-512.

Transchel, S., Minner, S., Kallrath, J., Loehndorf, N., and Eberhard, U. (2011). A hybrid
general lot-sizing and scheduling formulation for a production process with a two-
stage product structure. International Journal of Production Research 49, 2463-
2480.

Trigeiro, W. W., Thomas, L. J., and McClain, J. O. (1989). Capacitated lot sizing with setup
times. Management Science 35, 353-366.

Tsai, J.-F. (2005). Global optimization of nonlinear fractional programming problems in

engineering design. Engineering Optimization 37, 399-409.

Umble, E. J., Haft, R. R., and Umble, M. M. (2003). Enterprise resource planning:
Implementation procedures and critical success factors. European Journal of
Operational Research 146, 241-257.

van der Vorst, J. G. A. J., Beulens, A. J. M., and van Beek, P. (2000). Modelling and
simulating multi-echelon food systems. European Journal of Operational Research
122, 354-366.

170
 References

Van Wassenhove, L. N., and Vanderhenst, P. (1983). Planning production in a bottleneck

department. European Journal of Operational Research 12, 127-137.

Vanroy, T. J. (1986). A cross decomposition algorithm for capacitated facility location.

Operations Research 34, 145-163.

Wiers, V. C. S. (2009). The relationship between shop floor autonomy and APS
implementation success: Evidence from two cases. Production Planning and Control
20, 576-585.

Wijngaard, J. (1982). On aggregation in production planning. Engineering Costs and

Production Economics 6, 259-265.

Williams, H. P. (1990). "Model building in mathematical programming," Wiley, Chichester,

United Kingdom.

Williams, H. P. (1993). "Model solving in mathematical programming," Wiley.

Williams, H. P. (2013). "Model Building in Mathematical Programming," Wiley.

Woerlee, A. P. (1991). "Decision Support Systems for Production Scheduling: (Beslissings-

ondersteunende Systemen Voor Korte Termijn Produktieplanning)," Erasmus
Universiteit.

Wolsey, L. A. (1997). MIP modelling of changeovers in production planning and scheduling

problems. European Journal of Operational Research 99, 154-165.

Wolsey, L. A. (2002). Solving multi-item lot-sizing problems with an MIP solver using
classification and reformulation. Management Science 48, 1587-1602.

Yunes, T., Aron, I. D., and Hooker, J. N. (2010). An Integrated Solver for Optimization
Problems. Operations Research 58, 342-356.

171
Print
GVO printers & designers B.V.

Cover
Kim and Li-An Claassen

172