Data-Driven Optimization Of Hot Rolling

Processes

Dissertation

zur Erlangung des Grades eines

D o k t o r s d e r I n g e n i e u r w i s s e n s ch a f t e n

der Technischen Universität Dortmund

an der Fakultät für Informatik

von

Christian Jung

Dortmund

2019
Tag der mündlichen Prüfung:
12. Dezember 2019

Dekan:
Prof. Dr.-Ing. Gernot A. Fink

Gutachter:
Prof. Dr. Günter Rudolph, Technische Universität Dortmund
Prof. Dr. Thomas Bartz-Beielstein, Technische Hochschule Köln
Acknowledgments
This work would not have been possible without the help and support from many
friends, colleagues and family. Being a part-time researcher is always a balancing
act which is only possible with continuous support from many people.
I would like to express my gratitude to my supervisor Prof. Dr. Thomas Bartz-
Beielstein for his strong support over all these years. The great atmosphere at his
institute and the fruitful discussions were an important factor for the successful
completion of this thesis.
My supervisor at TU Dortmund, Prof. Dr. Günter Rudolph, made it possible to
carry out my doctoral research with a part-time position. He provided valuable
guidance and gave generous support. The ideas and suggestions improved the
quality of my research tremendously.
I would like to thank Prof. Dr. Thomas Schwentick for being my mentor and
guiding me through the last phase of my PhD.
A very special thanks to my colleagues at the Institut for Data Science, Engineering,
and Analytics (IDE+A), especially Martin Zaefferer, Martina Friese, Andreas
Fischbach and Jörg Stork for great discussions about connected research topics,
further research ideas and valuable feedback.
A very special and warm thanks goes to the following persons for their advice
and discussions during the doctoral seminars: Beate Breiderhoff, Oliver Flasch,
Sebastian Krey, Olaf Mersmann, Steffen Moritz, Boris Naujoks, Quoc Cuong Pham,
Stefanie Raupach, Margarita Rebolledo, Frederik Rehbach, Viktoria Schaale, Katya
Vladislavleva.
I am grateful to Gregor Schneider for his continuous support and for making this
work possible.
I cannot express how thankful I am to my family for their endless support and love
they gave me through all of my live and especially during my academic education.

1
Contents
1. Introduction 7
1.1. Innovation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2. Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

I. Data-Driven Surrogate Optimization of Hot Rolling Pro-

cesses 11
2. Hot Rolling 13
2.1. Hot Strip Mill Process . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2. Main Components of a Hot Strip Mill . . . . . . . . . . . . . . . . 15
2.3. Process Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3. Objectives and Performance Figures 25

4. Data Driven Meta-Model Based Optimization 27

4.1. Surrogate Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.2. Simulation Environment . . . . . . . . . . . . . . . . . . . . . . . 29
4.2.1. From Real-World Process to Simulation . . . . . . . . . . . 31
4.2.2. Implementation . . . . . . . . . . . . . . . . . . . . . . . . 33
4.3. Analytical Roll Force Model MF . . . . . . . . . . . . . . . . . . . 35
4.3.1. Flow Curve . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3.2. Models Used in Industry . . . . . . . . . . . . . . . . . . . 39
4.3.3. Problem Description . . . . . . . . . . . . . . . . . . . . . 40
4.4. Flow Curve Parameter Optimization . . . . . . . . . . . . . . . . 41
4.4.1. The Sequential Parameter Framework . . . . . . . . . . . . 42
4.4.2. Experimental Setup . . . . . . . . . . . . . . . . . . . . . . 43
4.4.3. Optimization Process . . . . . . . . . . . . . . . . . . . . . 45
4.4.4. The Influence of Expected Improvement . . . . . . . . . . 47
4.4.5. Influence of Local Optimization and Initial Designs . . . . 47
4.4.6. Impact of the Fitness Landscape . . . . . . . . . . . . . . 50
4.4.7. Comparison of the Resulting Roll Force Deviations . . . . 53
4.5. Parameter Optimization for Known Materials . . . . . . . . . . . 56

3
Contents

4.6. Parameter Optimization for Multiple Mills . . . . . . . . . . . . . 56

4.7. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

II. Online Optimization 63

5. Introduction to Online Optimization 65
5.1. Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.2. Initialization Problems . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3. Demands and Properties . . . . . . . . . . . . . . . . . . . . . . . 69
5.3.1. Amount of Data . . . . . . . . . . . . . . . . . . . . . . . . 69
5.3.2. Arrival of Data . . . . . . . . . . . . . . . . . . . . . . . . 71
5.3.3. Batch Versus Sequential . . . . . . . . . . . . . . . . . . . 71
5.3.4. Source of Data . . . . . . . . . . . . . . . . . . . . . . . . 72
5.3.5. Pre-Processing . . . . . . . . . . . . . . . . . . . . . . . . 72
5.3.6. Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.4. Performance Indicator . . . . . . . . . . . . . . . . . . . . . . . . 79
5.5. State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

6. Datasets for Online Optimization 83

6.1. Default Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.2. Rolling Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.3. Requirements and Assumptions . . . . . . . . . . . . . . . . . . . 90

7. Algorithms 91
7.1. Passive Aggressive . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.1.1. Comparison of the Three Different Variants . . . . . . . . 93
7.1.2. Influence of Aggressiveness Parameter . . . . . . . . . . . . 93
7.1.3. Handling Categorical Variables . . . . . . . . . . . . . . . 96
7.1.4. Performance . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.2. Recursive Least Squares . . . . . . . . . . . . . . . . . . . . . . . 100
7.2.1. Influence of Exponential Decay . . . . . . . . . . . . . . . 102
7.2.2. Handling Categorical Variables . . . . . . . . . . . . . . . 104
7.2.3. Performance . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.3. Online Support Vector Regression . . . . . . . . . . . . . . . . . . 108
7.3.1. Introduction to Support Vector Regression . . . . . . . . . 108
7.3.2. Incremental Support Vector Regression . . . . . . . . . . . 112
7.3.3. Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.3.4. Parameter Influence . . . . . . . . . . . . . . . . . . . . . 122
7.3.5. Influence of the Storage Size and Storage Management . . 122
7.3.6. Selecting the Kernel Function . . . . . . . . . . . . . . . . 127

4
 Contents

7.3.7. Handling Categorical Variables . . . . . . . . . . . . . . . 129

7.3.8. Training Effort . . . . . . . . . . . . . . . . . . . . . . . . 130
7.3.9. Memory Requirement . . . . . . . . . . . . . . . . . . . . . 132
7.3.10. Implementation . . . . . . . . . . . . . . . . . . . . . . . . 135

8. Online Parameter Optimization 137

8.1. Necessity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
8.2. Time Constraint . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
8.3. Drift Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.4. Online Parameter Optimization for Support Vector Regression . . 141

9. Performance 147
9.1. Extrapolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
9.2. Performance on Default Datasets . . . . . . . . . . . . . . . . . . 149
9.3. Performance on Rolling Datasets . . . . . . . . . . . . . . . . . . 149

10.Conclusion 157
10.1. Recommendations for Other Real-World Problems . . . . . . . . . 158
10.2. Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

III. Summary 161

11.Summary 163

5
1. Introduction
Complex real-world problems require adequate handling and analysis. There
is no best method for all kind of problem available and experts are needed to
understand and adapt algorithms on a specific problem. This is also known as
the No-Free Lunch theorem described by Wolpert and Macready [WM97].
In contrast to artificial problems, many real-world problems have to deal with
uncertainties and inaccuracies. Hot rolling is an important example of a real-world
problem, where multiple of these aspects have to be considered.
Finite element methods (FEM) are available to calculate exact solutions for certain
conditions. Due to the long calculation time required for FEM methods and the
necessity to dynamically adapt to certain conditions, FEM methods cannot be
used for prediction. Instead of FEM models, simple analytical models have been
developed. They are validated and optimized for all occurring situations. One
example of such an analytical model is a material model which describes the
mechanical properties [Zhi10] and the temperature behavior [KW99, YS16] of the
material during the rolling process. For regular material grades the description
and behavior during rolling is well known and the accuracy is very high.
guessed based on similar materials. One way to improve the description is to
take expensive laboratory measurements. Another way is to use data from
rolling trials and to build a data-driven meta model for the description of the
material. The meta-model will be embedded in a simulation environment and
will be validated against the analytical model. This can save costs and time
and can easily be extended also to optimize existing material descriptions. The
optimization is done offline, when enough data for a certain material is available.
Offline hereby refers to the fact that there will be no restriction to time or
any other limitation to the optimization. Since the description of a material is
highly interconnected with other analytical process models, the question is, if
this data-driven optimization procedure can significantly improve the accuracy of
the material description for the whole range of products. The roll force required
during rolling is one important parameter which is strongly correlated to the
material and the geometries. Therefore, the prediction errors of the roll force will
be used to measure the quality of the description.
Residual prediction inaccuracies will still be observable for all kind of analytical
models, even after optimization. To achieve the highest possible prediction

7
1. Introduction

accuracy, online algorithms are used. Online refers to a property of the algorithm to
update its coefficients on arrival of new samples. Basically, the update mechanism
is executed incrementally instead of calculating the coefficients from the whole
dataset.
Online algorithms are especially of interest, if the distribution of a dependent
variable or the target value changes. This behavior is called drift [SG86, WK96]
and will cause the prediction accuracy to decrease. It cannot be avoided and has
many reasons, e.g. sensor drift or mechanical clearance.
Online Support Vector Regression (SVR) is a promising algorithm which was
developed at the early years of the century by Martin [Mar02] and Ma and
Theiler [MTP03] as an extension to Support Vector Machines (SVM). Despite the
high popularity of SVM in general, which have been applied in various fields like
on text recognition [Kat17] or protein domain classification [HNK14], only a few
publications (see e.g. [OMBH07, OJB11]) devoted for online SVR on real-world
problems are available. The published algorithm allows the SVR algorithm to be
incrementally updated instead of a complete recalculation of all parameters from
the whole dataset. Also, a procedure for the removal of a sample is available but
still some questions are open which have to be solved for real-world applications:

• How many samples should be stored?

• How to select samples which should be removed?
• Which kernel should be used?
• How to select constraint and insensitive loss for the online SVR?
• How to handle categorical parameters?

To answer these questions, the available online SVR algorithm based on Ma

and Theiler was extended. To validate the results, the performance on publicly
available datasets is shown. The impact on the kernel choice, storage size and
SVR parameters are shown on data collected from various hot rolling mills.
A general problem for many online algorithms is the handling of categorical
variables. In classical learning methods, specific coding schemes are used and
additional variables are added to the dataset. This concept is only feasible if
the number of different occurrences of the categorical variable is known prior
to modeling. This will in general not be valid for online algorithms, where no
information about the categorical is known in advance. Four concepts for online
learning with categorical variables are addressed in this theses. Their influence on
the performance for three different online learning algorithms is discussed on data
generated during hot rolling.

8
 1.1. Innovation

1.1. Innovation
Parts of this thesis are based on the following publications:
• Metamodel-based optimization of hot rolling processes in the metal industry
[JZBBR17].
• Extending Support Vector Regression for On-line Learning Methods on
Real-world Data [JBBR18].
The main innovation was achieved on the following topics:
• Kriging application on real-world industrial data.
• First data-driven optimization procedure for determination of flow curve
parameter.
• Data-driven optimization across multiple rolling mills.
• Comparison of online algorithms on real-world data from rolling mills.
• Extension of Online SVR with strategies for storage management.
• Analysis of Online SVR for different storage sizes.
• Creation of strategies to handle categorical variables in online algorithms.
• Online parameter optimization, especially for SVR.

1.2. Outline

Figure 1.1.: Hot rolling optimization overview. Analytical models are used to predict
and describe the rolling process. These models can be used for various
types of rolling mills. The analytical models are interconnected and will
be corrected with online models which are using feedback from the rolling
process to minimize the prediction error.

9
1. Introduction

Figure 1.1 shows an overview of the optimization procedure for hot rolling. Several
analytical models are used to describe the rolling process. All those models are
closely interconnected with each other. The analytical models are mostly well
established physically-based models but also empirical models are used. Online
models are optimizing the prediction accuracy with feedback received from the
rolling mills. The main reason for this correction is the high dynamic behavior of
the process and an inaccurate description for special situations.
This thesis starts in Part I with an introduction to the rolling process and a
description of possible sources of errors in Section 2. The focus will be on
hot rolling of steel and aluminum but most of the proposed methods, models
and algorithms can directly be transferred to cold rolling. Further, the main
analytical models and their influence on the prediction are introduced. After
introducing the objectives in Section 3, a data driven approach for meta-model
based optimization is presented in Section 4. This will be used to optimize
the analytical models for various scenarios. Following the offline optimization,
this thesis continues with the main contribution on online algorithms and their
optimization in Part II. After an introduction to online algorithms in Section 5, the
dataset used for performance evaluations are described in Section 6. A discussion
and description of the online algorithms is given in Section 7. A simple first
order algorithm, i.e., Passive Aggressive, will be compared to the more advanced
Recursive Least Squares algorithm. Afterwards the online SVR algorithm is
presented and discussed. Extension made for the online SVR are presented and
their impact on the performance will be shown. The performance for all online
algorithm will be compared on several datasets generated during hot rolling but
also the performance to well established state-of-the-art algorithms is shown. The
thesis ends with a summary on the achievements and gives some recommendations
for other real-world processes.

10
 Part I.

Data-Driven Surrogate
Optimization of Hot Rolling
Processes

11
2. Hot Rolling

Hot rolling refers to a process where heated material is plastically deformed

between two or more steel rolls in order to produce a thinner product. A typical
layout for a steel and aluminum mill is shown in Figure 2.1. Today’s rolling
mills require a huge variety of different materials and geometries to be processed
within close tolerances. To achieve this, the prediction of the process behavior
and controlling of process parameter should be as accurate as possible. Various
interconnected physical-based software models are used for the process prediction.
The most relevant models and their tasks are introduced in Section 2.3.
Various mill types have been established in the market for different requirements.
The major difference is the target product type which can be a plate or a strip.
The mills for the production of a plate are consequently called plate mills. The
most common mill types for steel or aluminum production are hot strip mills
where the material is coiled at then end of the production. Plate mills usually
produce much thicker products than hot strip mills.
Another difference is the maximum width which can be rolled. Plate mills are
able to produce much wider products than hot strip mills. This is achieved with
turning tables which allow the product to be rotated by 90 degrees. Table 2.1
shows common geometries for the different aluminum and steel mills. Although
aluminum and steel as well as plate mills and hot strip mills differ in detail,
the main process flow and utilized components are common to all mill types.
Therefore, these components are introduced in the following section.

Table 2.1.: Typically used product geometries for plate mills and hot strips mills for
aluminum and steel. The ranges are defining the common values for the
input and target geometries in mm.
Steel Mills Aluminum Mills
Plate Strip Plate Strip
Input Thickness [mm] 200-300 400-600
Width [mm] 600-2000 600-2000
Thickness [mm] 5 - 50 1 - 25 4 - 100 2 - 15
Target
Width [mm] 600 - 5500 600 - 2500 600 - 5500 600 - 2500

13
2. Hot Rolling

(a) Hot Strip Mill for Steel

(b) Hot Strip Mill for Aluminum

Figure 2.1.: Hot Rolling mills for steel (top) with coil box and aluminum (bottom).
The concept is similar. For steel mills, additional a cooling line is installed
after the last finishing mill stand. Due to higher initial thickness the total
length of aluminum mills is usually much longer than for steel mills. The
coilbox can be used for steel mills to further decrease the distance between
reversing mill and finishing mill and to improve the temperature behavior.

14
 2.1. Hot Strip Mill Process

2.1. Hot Strip Mill Process

The layout of a common hot strip mill for steel and aluminium is depicted in
Figure 2.1. The figure is reduced to represent only the main components. The
rolling process for steel and aluminum is very similar. The work flow of the process
is from the left to the right side.
The process starts with the charging of the furnace. Here, slabs which are usually
at room temperature are charged and reheated to temperatures around 1200°C
for steel mills and 500°C for aluminum mills. The exact temperature depends on
the specific material and target properties. If multiple material grades are charged
within the same furnace a balancing for all different properties has to be made. In
order to reduce energy costs, the heating time is tried to be kept at a minimum.
The initial products are ingots, which are casted into a crystallizer shape, or
traditional slabs which are produced through a continuous casting process. For
aluminum hot strip mills, ingots are usually used. They are additionally milled
and sawed to compensate material flow effects during rolling. Due to the wide
range of aluminum materials, different shapes have been established for hard and
soft alloys.
After the target temperature in the furnace has been reached, the product will be
discharged from furnace and transported to the first mill stand where the first
deformations take place. The mill stand is usually a so-called quarto reversing
mill, because the stand consists of four rolls. Reversing refers to the change of
rolling direction of two consecutive reductions. Here, the thickness of the product
is reduced to an intermediate thickness for hot strip mills or to the final thickness
for plate mills. The reduction is done in a multiple steps, so-called passes. For
plate mills the next process step is the hot plate leveler, which is not depicted
in Figure 2.1. Here, the plate is straightened before it will be transported to a
cooling bed.
For hot strip mills, the rolling will continue in the finishing mill, where the product
is rolled in several stands to the final thickness. The additional cooling line after
the last finishing mill stand is only used for steel mills because of the higher
temperatures for steel. Furthermore, the cooling strategy will have an huge impact
on the strength of the product[TLWM13]. Finally, the product is coiled in the
downcoiler, respectively belt wrapper.

2.2. Main Components of a Hot Strip Mill

After the hot strip mill process has been described, this sections will briefly
introduce the different tasks of the main components:

15
2. Hot Rolling

• A reheating furnace
• Reversing mill
• Finishing mill
Afterwards, a short summary with the major tasks and the problems is given in
Table 2.2. For a more detailed description, the specific literature is recommended.
A good overview for hot rolling is given, e.g., in [GB00, Hin03, HS78, Web73].

Gas Furnace
The furnace will usually be charged with cold material and has the task to re-
heat the product until a certain temperature is reached. Moreover, the furnace
is divided into several heating zones with the task to homogeneously heat the
material. The temperature distribution within the material is of major importance
especially for steel rolling. The reasons for this is the different heat conductivity.
Aluminum has a much higher conductivity than usual steel grades and therefore
the temperature distribution for aluminum grades is much more homogeneous
than for steel grades. If the surface of the slab is much hotter than the core, the
deformation will not be equally among the thickness cross-section and undesirable
effects might occur. Sometimes the reason for a high temperature gradient among
the thickness is the desire to save energy. To minimize gas consumption and to
optimize the temperature distribution, a predictive control is used to continuously
predict the temperature of the slab at distinct locations over the thickness, width
and length. The reheating procedure will cause an inhomogeneous temperature
distribution within the product. Temperature deviations over the thickness are
critical for the rolling process and have to be limited. The length and width
directions can usually be ignored for the prediction. The furnace model will
predict 5-10 discrete points over the thickness and according to this prediction
and some other constraints, the gas consumption is controlled.
The reheating process lacks of precise monitoring since only the surface tempera-
ture of the product and the gas temperatures within the furnace can be measured
online. A common way to improve the temperature predictions is to use prepared
slabs with thermocouples positioned at multiple locations within the slab [SK96].
They monitor the heating process and allow further tuning of the predictive models.

Hot Reversing Mill

A reversing mill is usually the first mill after the product has been discharged.
Reversing is referring to a situation, where the rolling direction will change after

16
 2.2. Main Components of a Hot Strip Mill

Figure 2.2.: Illustration of a width reduction followed by a thickness reduction in a

roughing mill according to [WSDQS98].

each deformation. For plate mills there are usually one or two reversing mills.
A conventional hot strip mill usually also consists of one or two reversing mills,
called roughing mills. In contrast to reversing mills used for hot strip mills, plate
mills can handle much higher forces and torques, because they are build and
used for much wider products. Modern plate mills are able to produce plates of
approximately 5m width while the usual production width of hot strip mills is up
to 2m.
In a roughing mill, the reheated product is rolled down to a thickness of 30mm
- 50mm, dependent on the final thickness of the product and capability of the
proceeding mill. The reduction from the initial thickness down to this final
thickness is conducted in multiple steps, so-called passes. A pass is a specific
deformation of the product and may consist of a width reduction followed by a
thickness reduction. The width reduction is done in a vertical mill, called edger,
which is an optional part of the roughing mill. The edger is used to allow a
certain variation of slab width and to guarantee a specific target width of the
product. This will be achieved by controlling the position of the edger rolls
dynamically during each pass. After each width reduction in the edger a thickness
reduction will be made. Therefore, the width can only be influenced indirectly.
The combination of a width and a thickness reduction is depicted in Figure 2.2.
In the following figures, index 0 will always denote the initial geometry, index 1,
the intermediate geometry, i.e. the geometrical description between edger and
rougher. Index 2 will denote the final geometry after deformation. The initial
width w0 is reduced in the edger to w1 and the thickness will increase to h1 . The
width reduction creates a dog-bone shaped product which is depicted in Figure 2.3.
The dog-bone will cause a higher material spreading as a result of the thickness
reduction from h1 to h2 . The efficiency of the edger is usually between 40% and
60%. A width reduction of 100mm (w0 − w1 ) in the edger will cause a higher
spreading during the thickness reduction. The remaining width reduction, i.e.

17
2. Hot Rolling

hmax h0 h1

w0
2
∆w w1
2 2

Figure 2.3.: Example of a dog-bone shape which is created after the width reduction
and before the slab enters the horizontal mil. w0 is the entry width into
the edger, w1 is the exit width out of the edger, ∆w is the width reduction.
h0 defines the entry thickness, h1 the exit thickness out of the edger. This
might be higher than h0 . hmax is the maximum height of the dog bone
area.

w0 − w2 will then only be 40mm - 60mm. This efficiency is highly correlated

with geometry, materials and temperature distributions within the material. The
spreading behavior of the material is not equally throughout the length of the
slab. To compete this effect, the width reduction over the length is dynamically
adjusted by the edger. This is done especially at the beginning and the end of the
pass. During the beginning and the end of a pass the material has different flow
properties causing an unequally spread among the length. A typical behavior is
the creation of a so-called fish tail which is depicted in Figure 2.2. A colder core
part of the product will increase the material flow to the edges of the product.
This will result in a higher spreading of the material. Unfortunately, the core
temperature is only a predicted value and cannot be measured directly on-line.
The thickness reductions in reversing mills are optimized in terms of productivity
and may consists of five to seven passes for steel roughing mills or up to 30-40
passes for plate mills or aluminum roughing mills. Here the prediction of the
roll force is essential to achieve a proper quality of the product. The following
measurements may be retrieved from a reversing mill:

• Rolling force.
• Drive torque.
• Roll gap - distance between the rolls.
• Rotational speed of work rolls.
• Bending force (for profile and flatness control).

18
 2.2. Main Components of a Hot Strip Mill

• Axial shifting (for profile and flatness control).

Usually, additional measurements are installed in front or behind the mill. This
may include width, profile, thickness and speed measurements. For plate mills
usually also the profile will be measured.

Finishing Mill
Standard finishing mills consists of multiple stands, i.e., six or seven for steel and
three to five for aluminum. In contrast to reversing mills, the rolling direction is
continuous and therefore this part of the hot strip mill is sometimes also called
hot continuous mill. The finishing mill will constantly reduce the thickness of the
product in each stand and will finally wind the hot strip. The final product is
then called a coil.
For steel mills, an additional cooling line is installed between the last mill stand
and the coiling process (see Figure 2.1). The cooling line will apply a decent
amount of water on the product to reach the target coiling temperatures. The
additional cooling is of major importance for the production of advanced steel
grades.
In aluminum rolling, no cooling line is used. During rolling, all stands are active
at the same time. Therefore the speed of all finishing stands should be exactly
controlled during rolling. Otherwise the mass flow is not constant and more
material leaves one stand than enters the subsequent stand. As a consequence, the
material would accumulate in front of one stand resulting into a stop of the rolling
process. To stabilize the rolling process, a certain amount of tension between
the stands is used and controlled. For the start of rolling in the finishing mill,
i.e. before any thickness controller is used, roll force prediction errors will cause
a thickness deviation at the beginning of the coil. Later on, the thickness is
controlled by measurement feedback. In contrast to the hot reversing mills, the
prediction for a certain material is only updated on a product-to-product basis
instead of a pass-to-pass basis. Fewer corrections during rolling are required, if the
predictions are improved. Therefore, good predictions will cause a stable rolling
process. The measurements which are taken for finishing mills are basically the
same measurements which are also taken for reversing mills. Additionally, the
interstand tension will be used. Table 2.2 summarizes this section.

19
2. Hot Rolling

Table 2.2.: Tasks of the main components and possible problems.

Component Task Problem
Homogeneous heating of Lack of measurements
Furnace slabs / ingots to a speci- and validation. High im-
fied temperature pact on process.
Rolling in multiple steps Temperature distribu-
Hot Reversing Mill to target thickness and tion not measurable. No
target width direct width control.
Rolling from intermedi- Very accurate speed and
Finishing Mill ate thickness to final force prediction neces-
thickness. sary.

2.3. Process Models

Various models are used in hot and cold rolling which have the task of predict-
ing important process parameters. The term model hereby refers to computer
programs which are describing the physical process as exactly as possible. These
programs are normally developed by implementing the physics behind the process.
Some assumptions made in the software are simplifications, the result of some
finite element (FE) studies or some data driven approaches. All of those simplifi-
cations are made to accelerate the computations. Because this will be a general
description the term product will address plate and coil production throughout
this section. Figure 2.4 shows the interaction of process models for the description
of a single pass. Various analytical process models are participating in the state
description Sti for time instance ti . Furthermore, they depend on each other and
share information. At time instance t0 the product starts accelerating towards
the mill. During this time, i.e., from t0 to t1 , the product will cool down. This
will be taken into account by the temperature model MT , which will calculate
the expected entry temperature distribution ϑ(t1 ) at t1 . This temperature will be
used by the force model MF to calculate an expected roll force which will then
again be used by other models to update the predictions. The principle will be
the same for all state descriptions until the pass has finished and the product
reached its end position at t2 .
The accuracy of the physical models which are describing the rolling process are of
major importance. According to their predictions the mill will be operated. Some
guidelines may be set from process experts but the calculations and predictions
are executed with those process models. One of the major aspects of the rolling
process for a single product are e.g. the number of passes in a hot reversing mill.
The value can be set from outside but it is not guaranteed that it will be possible

20
 2.3. Process Models

Figure 2.4.: Interactions of various process models for the product description of a single
pass. It is assumed that all relevant information at time instance t0 are
known and summarized in the state description St0 . For simplicity, the
target thickness h1 , the transport and rolling speed v(ti ) are also assumed
to be known. Four relevant times are denoted in the time line. t1 denotes
the time where the product will have the initial contact with the rolls
and the deformation starts. From t0 to t1 the most relevant model is the
temperature and material model. They will calculate the temperature
loss of the product and will calculate a new temperature distribution ϑti
within the product. Because the temperature will change also temperature
dependent data Px from the material model MM are changed. The force
model and spread model will also update their calculations Fx and w1 since
both are dependent on the temperature model and the material model.
From t1 to t2 the deformation takes place and almost all models participate
in the product description. After the pass has finished at time instance t2
the further temperature loss has to be calculated until the product reaches
the reversing point at time instance t3 .

21
2. Hot Rolling

to reach a certain target thickness with the specified number of passes. There are
limitations like maximum reduction within each pass, i.e. geometric constraints,
and also mill capacities which have to be considered. These are e.g. maximum roll
force and maximum drive torque. Those limits are considered within the physical
models to calculate the number of passes which are required to reach the target
thickness. As a result an instruction is given how many passes and how much
reduction each pass should have. Additionally, temperatures, width, speed and all
other necessary parameter are calculated. The total instructions for each pass are
combined in the pass schedule for each product.
For plate rolling the width can be achieved with the help of turning tables. They
are turning the product so the passes can be rolled in cross-direction. After some
so-called spreading passes the product is turned again but is now much broader
than before. The calculation of the lengthening and broadening effect has to be
very precise in order to achieve also a specific target width. Limits, which are
considered here are e.g. the maximum length of the product where turning is still
possible. If the real length is larger than the calculation it will be impossible to
turn the product again and thus the final geometries cannot be reached anymore.
The following section describes the main physical models and their tasks. For mod-
els which are considered in some of the experiments, a more detailed description
is available in the corresponding section.

Material Model MM
The description of the material properties is one of the most critical aspects for
hot rolling. The material will determine the required roll force which is used for
deformation, the temperature behavior and many more aspects relevant in terms of
quality. According to the Worldsteel Association1 more than 3.500 different steel
grades are currently known. Each material has an unique physical and chemical
property which depends on the chemical composition. Some of them are very
similar and might not show significant differences during rolling. These materials
can then be treated together as one material group. Nevertheless, different groups
behave completely different and it is essential to predict their behavior as accurate
as possible.
The main material characteristic which are used for the hot rolling process are:
• Deformation resistance,
• heat capacity,
• conductivity and
• linear expansion.
The deformation resistance is the most relevant characteristic for the prediction of
1
https://www.worldsteel.org

22
 2.3. Process Models

the roll force. The heat capacity and conductivity are influencing the temperature
behavior of the material. The linear expansion determines the relationship between
hot and cold geometries. This is an important property since the material is rolled
at high temperatures and they are sold when they are cold.
Those properties are calculated online for the actual process state since most of
those properties are described by polynomial functions which are dependent on
the actual temperature.
Unfortunately, there is no general description of those properties in dependence of
the chemical content available. Thus, different approximations for each material
group will be used. Beside the parameter for each material, also the specific
approximation formula might be different for each material. As a consequence,
the predictive accuracy is also correlated to the material group.

Temperature Model MT
The temperature model uses information from the material model to calculate
the temperature behavior for each product at defined process states and positions
within the product. Each cooling device and its influence on the material can be
parameterized. Beside the active cooling devices like descaler, cooling lines or
transferbar cooler, the temperature losses of the product through radiation and
temperature induction caused by the rolling are also taken into account. Although
the temperature distribution over the whole material thickness, i.e. surface to
core, will be calculated, the validation can only be made by comparison with
measured surface temperatures. The measuring will be done by pyrometers which
are measuring the intensity of certain wavelength occurrences.
For hot rolling of steel the occurrence of iron oxide impurities might cause abnormal
temperature measurements. Also the environment might be full of steam and water
which causes measurement inaccuracies. Therefore, temperature measurements
should only be used with caution. Pyrometers are also used for the hot rolling
of aluminum but here the pyrometers are regularly synchronized by a contact
temperature measurement.

Spread Model MW
Spread is defined as material flow in width direction. This is usually a challenging
task for hot strip mills with an edger installed. The material flow is dependent
on temperature, material and geometries and these are also the dependencies for
additional spread caused by the dog bone shape. There are several descriptions
for the dog bone shape available but the validation of those are difficult, since
no shape measurements can be installed. The dog-bone shape will be highly
correlated with the additional spreading which occur.

23
2. Hot Rolling

Additionally, the usual hot strip mill will have only an exit side measurement
gauge installed but material spread will also occur in each backwards pass. The
optimization is then called a hidden state optimization, since several states in
between cannot be measured and validated.
Aluminum hot strips will usually be trimmed between the last stand of the
finishing mill and the belt wrapper. Therefore the prediction does not have to
be as accurate as in the case of the steel mill. However, the milled ingots usually
have a defined shape to prevent an excessive side alligator effect. Side alligator is
referring to an effect, where additional material flow to the edges is created at
the top and the bottom of a product. This will cause the edge shape to look like
an alligator mouth. The main target for aluminum is a uniform width over the
total length. This includes a rectangular shape of the beginning and the end of
the product. The edgers are used to maintain such a shape but it can only work
if the predicted width is within some close range.

Force Model MF
Accurate prediction of the force is important to have a stable process. The force
is a major contributor to the strategic decision how many passes will be used.
Good predictions also prevent damages to the mill because they ensure that no
reductions are made which would require roll forces higher than the mill capabil-
ity. The force is calculated based on the deformation resistance of the material
which is unique for each material. The so-called flow curve are dependent on the
deformation itself, the deformation rate and the temperature. A more detailed
description of the roll force is given in Section 4.3.1.

24
3. Objectives and Performance
Figures
The data-driven optimization can be used to optimize all parameters which are
used for predictions in the rolling process. Typical objectives are parameters for
the prediction of the roll force and roll torque but also geometrical predictions
for width, thickness or profile can be optimized. The following measurements are
usually taken during the rolling process and can be optimized:
1. Roll force
2. Roll torque
3. Thickness
4. Width
5. Profile
6. Surface temperature
7. Flatness
The algorithm and concepts described in Section 4 are used to optimize parameters
which are influencing the prediction of those values. The prediction of the roll
force, e.g., will be made with an analytical model MF , which was introduced
in Section 2.3 and will be explained in more detail in Section 4.3.1. It depends
on material parameters which can be optimized. Therefore, the optimization
may be conducted by comparing the prediction of the analytical model and
the corresponding measurement. If the parameters can be guessed or if initial
parameters are available, the optimization will be easier in most cases. Therefore,
for data-driven optimization, the following tasks for the optimization can be
summarized:
• General data-driven optimization for unknown initial parameter (Section
4.4).
• Optimization for known initial parameter (Section 4.5).
• Optimization across multiple mill types (Section 4.6).
The online optimization described in Section 5 will usually optimize the residual
prediction error, i.e., the deviation between an analytical model and the measure-
ment of the corresponding variable. It can therefore be seen as the logical next

25
3. Objectives and Performance Figures

step in optimization of this variable.

To measure the quality of an optimization it is important to define some metric
which describes how good an approximation is. Therefore, a metric is defined
which penalizes prediction errors. A common objective used is the root mean
squared error (RMSE). For a given data set s = {{x1 , y1 }, {x2 , y2 } · · · {xN , yN }}
and prediction values f (x) = {yˆ1 · · · yˆN } with N being the total number of
observations, the RMSE is defined as root of the average squared error.
⌜ ⌜
1 ∑︂
N ⃓ 1 ∑︂
N
⃓ ⃓
RMSE = (ŷ − y)2 = ⎷ (3.1)
⃓
⎷ e2
N i=1 N i=1 i

Here, ei are the residuals defined as the difference between the measurement and
the prediction. Another quite common measure is the mean absolute error (MAE).
There is no common rule which one is the better so usually the selection is done in
dependence of the dataset and what should be expressed. In [TC14] the authors
are discussing about the pros and cons of the RMSE against MAE and in which
scenarios it might be beneficial to use MAE instead of RMSE. One good argument
to use the MAE is its interpretability. Process experts can directly see from the
analysis how well an algorithm is performing.
The MAE is defined as

1 ∑︂
N
MAE = |ei | (3.2)
N i=1

and the MSE is defined as

1 ∑︂
N
MSE = e2 (3.3)
N i=1 i

where N again is the total number of samples, y the true target value and ŷ
defines the predicted value. The RMSE will be the main performance indicator in
Section 4 and the MAE will be usually used for the datasets specific to rolling in
Section 5.

26
4. Data Driven Meta-Model Based
Optimization
Several physically based software models are used to describe the complex rolling
process. The models are interconnected with each other and the optimization
of a single aspect would not be sufficient to optimize the whole rolling process.
Some of these software models are describing the different aggregates of the rolling
mill. Besides the introduced process models in Section 2.3 also other models
will influence the process prediction. An example is the furnace model which is
modeling the temperature heat up over time and tries to predict the temperature
distribution in the products. It will influence the initial state description St0 (see
Figure 2.4). Other models are more focusing on the material behavior during
rolling and are trying to describe phenomena which occur during rolling. This
might be the spread behavior, roll force or any other kind of properties related
to the finished product. All those different models are combined to achieve the
best process description. The reason why they have to be combined can be seen
in the prediction of the roll force. The roll force will induce temperature during
rolling and therefore temperature and roll force are not independent from each
other. Since many material related parameter are also temperature dependent,
the changed temperature will also cause a change in this behavior.
Therefore the best approach is to model the whole rolling process for every single
product. Additional information about previous rolling also affects the current
behavior and also have to be considered. Those information are, e.g., thermal
state of the rolls and wear of the rolls.
Consequently, the only possible solution is to model the whole scenario but focus
only on the outcome on some parts. An example would be to focus on the roll
force model but to have defined performance criteria for the temperature. If the
temperature criteria would be violated, a penalty would be added to the roll force
performance.
There are many FEM studies describing single deformation behavior of certain
materials, e.g. [RPS17], or FEM studies which are analyzing the general flow
behavior but unfortunately this would require too much calculation time. Fur-
thermore, most of these studies are only valid in very special cases and clearly
defined working points.

27
4. Data Driven Meta-Model Based Optimization

Simplifications have to be made and most often analytical models are used to
describe the whole process online.
This chapter is focusing on concepts to optimize these analytical models with the
use of so-called surrogate models. It starts with the definition of surrogate models,
the description of the created simulation environment and the used optimization
algorithms. Examples from real-world are shown in Sections 4.4, 4.5 and 4.6.
Some parts of this chapter were taken from [JZBBR17].

4.1. Surrogate Modeling

A surrogate model (or meta-model) is a substitute of a usually more complex
model [Søn03]. Complexity refers to the computational budget, e.g., the number
of function evaluations. Simulation runs on the surrogate-model are expected
to be less time consuming than simulation runs on the original complex model.
Therefore, surrogate-model based optimization is used when parameter optimiza-
tions have to be performed but the run time or the total number of function
evaluations on the original model are too high. The original model might be any
kind of model, e.g. a finite element model, and is assumed to approximate the real
behavior of the system. For some applications, the real behavior of the system
might also be obtained by conduction a physical experiment. The surrogate model
is designed to approximate the complex model as good as possible. Well known
surrogate models are artificial neural networks, linear models, Kriging, random
forest. A detailed overview of surrogate model based numerical optimization is
presented by Jin [Jin03] and Jones [Jon01].
During parameter optimization with surrogate-based models usually also some
fitness evaluations with the original complex model have to be made. The inte-
gration of both kinds of models in the optimization process is known as model
management. A simple proposal how to combine the usage of both model types
can be found in the paper by And [DT97] where the authors are defining pattern
search algorithm incorporating two different kinds of models. When combining
two different kinds of models a trade-of between computation time and accuracy
usually has to be made.
For the problem of hot rolling, a FEM simulation of the whole process is not
feasible since the simulation of a single product would already take too long, i.e.
some hours. Furthermore, a FEM study of a single product would not meet our
requirement since uncertainties originating from chemistry or the heating process
cannot be considered and conclusions from this product to the general material
behavior are not legible. The optimization would only be valid for some discrete
working areas for temperature, deformation and deformation rate. FEM simula-

28
 4.2. Simulation Environment

tions can be used to optimize special parts with limited ability of generalization of
the process. An example of such analysis is shown in [RPS17] where the authors
are using the FEM study to describe different behavior of stress in dependence of
the rolling direction for a AISI 303 stainless steel. In [RZD12] and [LMZZ15] the
authors are analyzing the shape of the dog-bone during vertical rolling for specific
cases.
Since FEM simulations are not applicable the process will be described using
analytical models. For the process description several analytical models are com-
bined as described in Section 2.3. The idea is now to use a data-driven surrogate
model for the process description. Data-driven models benefit from measurements
which were taken from the real process. This data is used to build a data-driven
model from various analytical models and retrieve the fitness according to the
real measurements. For this reason throughout this chapter the complex or
expensive model is denoted to be the combination of these analytical models.
The principle is illustrated in Figure 4.1. A product description is used by the
analytical models (1) to update the description and the settings for the process (2).
All measurements collected by the process are then used to update the product
description for further process steps. All those information are also used to build
and optimize data-driven surrogate model. The requirement to have a variation
of chemistry and pass schedules leads to the necessity to build a simulation of the
rolling process incorporating multiple products. In order to optimize the analytical
models, the products are ideally rolled with as much variation as possible. This
way, the analytical models are optimized for a wide range of different process
situations. Before the concrete analytical model, i.e. the roll force model MF ,
which is object of the optimization scenario, is presented, an introduction to the
simulation environment is given.

4.2. Simulation Environment

The simulation environment is driven by the fact to optimize any kind of prediction
which is used for the rolling process. In fact, every prediction with corresponding
measurement can be an objective of the optimization. This includes force, torque,
temperature and also geometric predictions like width, thickness or profile. Also
parameters like friction, which is influencing the force prediction, can be optimized.
The main objective throughout this chapter will be the roll force, i.e., reducing
the roll force prediction error. For some easier models, which are not strongly
interconnected, simplified optimization procedures may be chosen. Pacing infor-
mation like reversing times, which specify the regular delay between passes caused
by the basic automation system and some simple geometry information is one

29
4. Data Driven Meta-Model Based Optimization

Figure 4.1.: Data-driven surrogate modeling for the description of the rolling process.
The process is described by an initial state and product description (1)
which is used by various analytical models (2). They will update their
predictions and settings for the real process. These settings and predictions
will also be available to the surrogate model. Cyclically, or after some
deformation has finished, the measurements from the real process (3) will
be used to update the product description and to calculate new settings and
predictions for the following process. The data-driven surrogate model is
used to simplify this description. The solid lines describe the online process
and the dashed lines are used to describe the data-driven approach used for
the offline environment.

30
 4.2. Simulation Environment

example of these much easier tasks.

4.2.1. From Real-World Process to Simulation

The data which are used for optimization are collected from a real-world process.
The analytical model which describes this process was specifically adapted for this
rolling mill and is parameterized as accurate as possible. The same parameters
which describe the specific rolling mill are used within the simulation scenario.
This means, drive configuration, roller table layout, the position and number of all
furnaces are configured and adapted within the simulation environment to fit the
real mill. This is the reason why there is a different version of the analytical model
for each mill available. It may also contain some plant specific configurations and
include a list for the materials which are typically rolled on this mill. The data
from the real-world process are collected and stored into a database to be able to
use them for the simulation. Those data contain:

• Product description: For every product which is rolled, some initial geometry
information from product planning are received. Usually those data are
coming from the casting process but they may also be updated, e.g. when a
manual measurement was taken. The product planning also defines some
target geometries (width, thickness), target temperatures and material prop-
erties. Along with the geometry information is also a chemical description
of the material. Usually the most important 20 or 30 elements are given to
characterize the product material.
• Operator interactions: Before or during rolling, the operator may influence
the rolling process based on his observations or based on some change within
the planning. Some of these changes have to occur prior to rolling and some
of them may occur at any time. Possible changes may restrict reduction,
speed and cooling and can have huge impact to the process.
• Measurements: There are many measurements installed in rolling mills.
The most important ones are geometry measurements (width, thickness),
temperature measurements and measurements of roll force, roll torque and
speed. The reliability and quality of these measurements are not always
suitable to actively use them and therefore a separate handling for those
measurements is used for the process. Only the pre-processed measurements
are stored.
• Time information: Time information are important because they define the

31
4. Data Driven Meta-Model Based Optimization

temperature loss for each product. Especially if there is some delay during
production these time information have to be used to update the further
process steps.
• State descriptions: The roll state is an important parameter especially for
calculation of the flatness and the profile of a product. The roll state has to
be considered during calculation. Because the roll state is highly dynamic it
will be stored at specific process steps, e.g. before the first pass is rolled.
Especially the state descriptions are important, because the accurate prediction of
the process requires a calculation of all previous rolled products since roll change.
At roll change a stable description of the rolls is possible because they are mounted
at room temperature and no temperature gradient is existing. Because the roll
state is stored for every product before rolling it is possible to simulate specific
products out of a much larger rolling campaign. An optimization of specific
material groups is therefore possible and legible.
The exact same version of the rolling model is used which is also used for the
real-world process. The only difference is, that the data is not coming from
different sources like product planning department, operator, gauges. Instead,
they are all retrieved from the database. Algorithm 1 shows the pseudo-code of
the pass schedule calculations for the real-world process which was illustrated in
Figure 2.1 and described in Section 2.1. Algorithm 2 shows the adaptions made
for the simulation model.
Additionally to the data above also the so-called pass schedule for each individual
calculation is stored. The pass schedule contains all relevant information for the
production for every single deformation step, called pass. The description of one
pass contains all predictions and settings for the basic automation system. The
main values are roll gap, speeds, force, torque and geometric information. They
are used for the process and may or may not be changed by the operator. Usually
the process will follow these changes. The simulation model has to use the same
settings like the real-world process. This means the simulation model will retrieve
the measured reductions and calculate the passes analogues to the settings of the
real process.
This ensures that the calculations from the real-world process can be compared
to the calculations of the simulation model.
If continuous mills such as hot strip mills are used, feedback after every single
pass and update step in line 7 of Algorithm 1 is not necessary for the real-world
process. Similar, for the simulation model step in line 7 up to line 9 will not
be required. Instead, measurements after the whole rolling process will be send
to the model and the update will occur for the next rolled product. Although
the handling is easier, it complicates the online process, because only a certain
deviation can be compensated by the basic automation system.

32
 4.2. Simulation Environment

Algorithm 1: Rolling mill process algorithm.

1 collect initial information from product planning and calculate pass schedule;
2 show setup information to operator and update pass schedule with each change
or measurement;
3 let N be the number of total passes in a reversing mill;
4 for n = 1 to N do
5 send setup for pass n to basic automation system and start rolling;
6 collect measurements from process and send to online model;
7 use measurements to update settings and predictions for pass n;
8 if n + 1 < N then
9 update predictions and setup for pass n + 1

When the simulation scenario is used, it is important to put certain constraints on

the parameters. These constraints might be just boxed constraints but can also
be highly nonlinear constraints, dependent on the application. If those constraints
are violated the simulation will fail or terminate unexpectedly because the usual
working areas are violated.
Before any optimization can start the proper selection of products which should
be analyzed have to be made. This is basically nothing else than an extraction of
given data out of the database which was filled during rolling. After each run of
the simulation the prediction of the interesting value is compared with the real
measurement and a performance figure is calculated. Here, the RMSE defined in
Section 3 is used. The number of measurements N used for the calculation can
be equal to the number of products which have been rolled, but usually is much
higher. This depends on the parameter which should be optimized. The roll force
is calculated on multiple points along the length of a slab. Usually there are at
least three discrete points over the length with corresponding measurements and
predictions available. In an aluminum roughing mill with 21 passes there would
be 63 prediction errors available which can be used for optimization. When the
force optimization is conducted in a finishing mill with four stands, only 12 values
are then available and used for optimization. Optimization for other parameters,
where no difference between the three length coordinates is expected, will result
in a lower number of measurements for each product.

4.2.2. Implementation
The implementation of the simulation environment is done in C++ and R [R C17].
The database used to store calculations and to retrieve the input data is SQLite.

33
4. Data Driven Meta-Model Based Optimization

Algorithm 2: Modified algorithm used for the simulation model.

1 initialize model and read optimization parameters from file or database;
2 foreach product to be simulated do
3 use all information available to calculate complete pass schedule;
4 use time information to take into acount delay between discharge and start
of first pass;
5 let N be the number of total passes in a reversing mill;
6 for n = 1 to N do
7 retrieve all information for pass n from database and recalculate pass n
’as rolled’;
8 store calculations in seperate database;
9 use time information to take into acount delay between pass n and
n + 1;

10 compute performance parameter on specified parameter, e.g. RMSE of roll

force;

SQLite was chosen, because it will not require a server architecture and can easily
be transferred to any other machine. The extracted data can then also be imported
to other databases like Oracle. In fact, in the online process both databases are
used. SQLite is used more for daily analysis and the Oracle database is used for
long-term analysis.
For the purpose of flow curve parameter optimization on a specific material, the
Oracle database is used and recent rolling results for this material are extracted
and imported to SQLite. The principle is illustrated in Figure 4.2.
Within each optimization run a batch script is called from R which triggers the
simulation. The simulation will cause the process to calculate new prediction
results. These results with their corresponding measurements from the real-world
process are stored in a separate database. The evaluation of the performance
figure is done after the simulation finishes. Therefore, the RMSE is passed to the
optimization algorithm which is then modifying one or more parameters. These
parameters are used within in the next simulation run. The optimization of
parameters can be done in various optimization frameworks. Here, two different
optimization strategies are used which are motivated by different use-cases.
If the initial parameters are available and the parameter range is known but the
dependencies of the internal models is unclear first a parameter screening with
the sequential parameter framework SPO [BBLP05] is made. The framework will
be introduced in Section 4.4.1.
If initial parameters are available and parameter ranges are known then the initial
screening is skipped and instead a local optimization is used. The local optimiza-
tion will search for the best parameter in a region of the initial parameters with

34
 4.3. Analytical Roll Force Model MF

Figure 4.2.: Extraction from data in order to use them in the simulation framework.
All data necessary for the calculation are stored into a long term archive
database. The data is filled by the real process. Only the material which
should be optimized is extracted from this database and transferred to a
local SQLite database which is used by the simulation.

given boxed constraints. However, if the best parameter still are not achieving a
target performance then the SPO still can be used to search for better solutions.
Both optimization scenarios are described in Section 4.4 and Section 4.5 . The
first one will contain flow curve parameter optimization for a material which was
not known at all before and therefore the initial guess of the material behavior
achieved a poor performance.
The second optimization scenario uses some flow curve parameter already adapted
to a certain amount. Still, the performance is not sufficient and has to be im-
proved. For the second optimization scenario only the local optimizer will be
used. The first algorithm contains the full parameter optimization scenario with
initial screening of the parameter range. Both scenarios have the analytical model
in common. Therefore the used analytical model with known relationships and
used parameters are introduced in the next section. Afterwards both optimization
scenarios are discussed in detail.

4.3. Analytical Roll Force Model MF

In order to understand the optimization problem it is necessary to have some
knowledge about the assumptions and dependencies of the analytical models.
After a first short introduction about the currently used analytical model, the
optimization problem will be specified and different optimization strategies are
discussed and compared.

35
4. Data Driven Meta-Model Based Optimization

4.3.1. Flow Curve

The prediction of roll force is a central aspect in the pass schedule calculation. As
mentioned in Section 2.3 and 4.2.1, the pass schedule contains specifications and
settings for each process step which occurs during rolling. It therefore contains
settings of thickness reduction and width reductions for each individual pass and
the corresponding predictions based on these reductions, i.e. roll force and torque.
Besides, also the speed is contained in the settings and is usually varied from start
to end of each pass.
To predict the required roll force for a given reduction the so-called flow curves are
used. They describe the resistance of a the material during plastic deformation.
Intuitively, this resistance depends on

• reduction,
• temperature,
• speed and
• material properties.

The flow curve is exactly using these parameters for the description of the resistance.
The dependency of the reduction is usually expressed as effective, logarithmic
deformation φ and is expressed as:
h0
φ = ln ,
h2
where h0 is the input thickness of a pass and h2 is the target or output thickness of
the pass as illustrated in Figure 2.2. The logarithmic deformation is used because
its dependency to the roll force can more easily expressed in this way.
The temperature is usually denoted as ϑ and the speed is included as deformation
rate φ̇ = dφ
dt
which is the first order derivation of the deformation against time. The
material dependency is unfortunately not expressed as a continuous formulation
among the chemical compositions. Each material has its own flow curve and
corresponding parameters. A material is here defined as a chemical composition
close to a somehow specified target composition for an alloy. Usually, the ranges
for these chemicals are very small so that no huge difference within one material
is expected. This is at least valid for the usual steel grades. Unfortunately, there
are different standards for steel grades available and every company may has an
own definition. For aluminum grades almost every company uses the ANSI or AA
standard1 for the definition of the grades. The problem within these grades is
that unique grades sometimes allow high variation of important chemical elements.
1
https://www.aluminum.org/standards

36
 4.3. Analytical Roll Force Model MF

For alloy AA5182 wide ranges of magnesium are allowed having a huge impact on
the roll force [SS16].
The flow curves are not directly expressing the roll force but they describe the
deformation resistance kf during plastic deformation:

kf = AKφ Kϑ Kφ̇ . (4.1)

Here, A ∈ R+ is a constant factor, and K(·) are functions of the corresponding

variables φ, φ̇ , and ϑ, respectively.
The first flow curve formulas were developed in the early 50s of the last century
by Geleji and Ekelund and the dependency Kϑ was only linear [HS78]. These
formulas were mainly developed for low carbon steel. Afterwards also formula
with higher order terms were developed for medium and high alloyed carbon
steels [H+ 72, HZEK72]. A good overview and description of the different flow
curves can be found in [HS78, Hin03, TNR81].
Today most common flow curve models were originally developed by Hensel and
Spittel [HS78]. They were extended at the University of Freiberg to have a better
accuracy of the flow stress for high deformation grades. Thus, these extensions
are called Freiberger Approach. Some of the available flow curve models, which
are typically used in process models for hot rolling, are presented in Table 4.1.
Their corresponding equations read as follows.
m4
kf = kf,0 A0 A1 em1 ϑ A2 φm2 e φ A3 φ̇ m3 (4.2)
m4
kf = A0 em1 ϑ φm2 e φ φ̇ m3 (4.3)
m4
kf = A0 em1 ϑ φm2 e φ (1 + φ) m5 ϑ m7 φ
e φ̇ m8 ϑ (4.4)
kf = kf,0 A1 em1 ϑ
A2 φ m2
A3 φ̇
m3
(4.5)

The multipliers Ai (i = 0, 1, 2, 3) can be reduced to one parameter, A. The param-

eters mj (j = 1, 2, . . . , 8) are defining the exponential behavior of the materials in
dependence of the temperature ϑ, the deformation φ, and the deformation rate φ̇ .
The parameters ϑ, φ, and φ̇ are defining the working point in each deformation.
The value kf,0 used in the equations (4.2) and (4.5) is the basic deformation and
is calculated by empirical formulations based on the chemistry.
Each material is classified according to its chemistry and gets its own parameter
set M with parameter values m1 to m8 and A0 to A3 , respectively. For the
classification the chemistry is compared to a default target chemistry. This means
for each flow curve a default chemistry is stored. Then, the distance of the actual
chemistry to this target chemistry is calculated. In order to prevent a wrong
classification, a weighting of important elements for each material is also made.
This ensures that the important, characterizing elements of the material, are more
significant than other elements. The procedure is depicted in Figure 4.3.

37
4. Data Driven Meta-Model Based Optimization

Table 4.1.: Overview of typical equations for the regression of the material flow stress.
The multipliers Ai are reduced to one parameter, A. Parameters mj are
defining the exponential behavior of the materials in dependence of the
temperature ϑ, the deformation φ or the deformation rate φ̇ which defines
the working point. Parameter kf,0 is defined by a simple equation based
on the chemistry. Entries in the column ”Equation” refer to the equations
defined on p. 37.
Eq. Name #Params Parameter List
(4.2) Freiberg 1 5 A, m1 , m2 , m3 , m4
(4.3) Freiberg 4 5 A, m1 , m2 , m3 , m4
(4.4) Freiberg 8 7 A, m1 , m2 , m4 , m5 , m7 , m8
(4.5) Hensel Spittel 4 A, m1 , m2 , m3

Figure 4.3.: Process for selecting a suitable flow curve. The chemistry of the product
which is going to be rolled is send to the material database. Then, a
ranking according to the distance to a target chemistry of available flow
curves is made. The most suitable flow curve parameter together with the
information about the used equation is then send back to the process model.

38
 4.3. Analytical Roll Force Model MF

For each flow curve also the valid ranges of operation for the parameter-set M is
stored. This is done to have a more suitable description for phase transformations
within the material and to distinguish between hot and cold rolling. For aluminum
hot rolling usually the temperatures are not lower than 250 degree Celsius. For
the region below other parameter sets are used. Summarizing, the parameters
kf , φ, φ̇ , and ϑ, the multiplicative variables Ai , mj , (i = 0, 1, 2, 3; j = 1, . . . , 8),
and related functions kf,0 , Kφ , Kφ̇ , Kϑ , are used.
Nowadays, hundreds of different materials are known. The parameter kf is almost
linearly correlated with the roll force and roll torque. Thus, the model prediction
quality and herewith the process stability and product quality are correlated to
the parameter-set M. It is therefore important to optimize those parameters in
order to increase the model quality and to ensure a stable process with maximum
throughput and product quality. A standard procedure for obtaining these pa-
rameters is the measurement of the deformation resistance in a laboratory. Those
measurements can be used for a regression onto one of the formulas shown in
Equations (4.1) to (4.5). Of course, other formulas exist, and might be used for
regression. Especially when trying to model the deformation of micro alloyed or
high alloyed steel or when complex materials with phase transformations should
be described these other models might be more suitable.

4.3.2. Models Used in Industry

There are many analytical models for the description of the roll force. One of the
most common analytical models, which is used for the calculation of the roll force,
is based on the elementary rolling theory [Web73, Hin03]. Some of the limitations
of that theory are compensated with correction functions. For example, one of
the requirements for the elementary theory is the complete plastic deformation
during each pass.
As described in Table 2.1, Aluminum ingots have usually a thickness of above
500mm. For thick products like this, the assumption of complete homogeneous
deformation is not valid anymore. The analytical model holds compensation and
correction function which were empirically determined to correct these inaccuracies.
For the calculation of the roll force in each deformation step, the contact zone
between product and rolls are divided into strip-like elements and the force balance
for each stripe is calculated. The solution of this equation system yields to the basic
differential equation of the plastically deformation theory which was developed by
Karman in 1925 [Kar25]. The flow resistance for each deformation is a major part
of this equation system and is highly dependent on the material, the deformation
φ, deformation rate φ̇ and temperature ϑ. It is calculated with the flow curve

39
4. Data Driven Meta-Model Based Optimization

discussed in the previous section. Furthermore, the forces induce temperature into
the material so the calculation of the next deformation depends on the previous
deformations. This is valid not only for reversing mills. The principle is the same
for finishing mills, where the next deformation is made in a separate stand.
The roll force is linearly correlated to the parameter kf . Intuitively also an
optimization without simulating the whole process would be suitable. But since
the roll force is inducing and changing the temperature of the product the next
deformation would take place at different temperatures and originally calculated
values are not valid anymore. Therefore, if the parameter of the flow curve
changes, the whole process has to be simulated again. After the new simulation,
the calculated roll forces based on the new flow curve can be compared with
the original feedback, i.e., measurements of roll force, torque, temperature, and
speed. If the parameter are only optimized with data of a single product, the
parameters φ, φ̇ and ϑ may not vary enough. The resulting optimization might
not be generalizing good enough for other areas of deformation. Therefore it is
preferable to consider a campaign with a wide variation of product geometries,
temperature ranges, and deformations.
Currently, the available concepts for the optimization of flow curve parameters
are mostly dealing with determination of those parameters in laboratory rather
than optimizing those parameters with real process data. Traditionally, the
parameters are measured with small samples of one piece in laboratory and are
then generalized for every material which is close to the sample in terms of material
composition. Some companies are modifying the flow curve parameters with linear
models. That is, they are determining the prediction accuracy of their model and
are varying some of the influence parameters. Most of the research in this area is
on the development of suitable flow curve equations especially for high and micro
alloyed steel rather than using a data driven approach for the optimization of those
parameters. Lin [YL08],e.g., analyzes the deformation resistance of 42CrMo steel
for high strain rates, Hernandez [HMR96] describes the deformation resistance
within the austenite phase of micro alloyed steels and in [MRS+ 09], the authors
describe the deformation behavior for Ti-alloyed austenitic steel.

4.3.3. Problem Description

Flow curve parameter determination in laboratory as described in Sec. 4.3.1
requires several weeks and costs several thousand Euro for each material. Some-
times, this is not affordable and therefore not a suitable way to determine those
parameters. Hence, cheaper estimation of flow curve parameters have to be found.
Due to their highly nonlinear behavior, the flow curve equations cannot be solved
directly. Furthermore, a different roll force would result in a different temperature

40
 4.4. Flow Curve Parameter Optimization

balance of the product and thus the temperature in the next pass differs from
the original calculation. Because these aspects cannot be neglected, the whole
process has to be simulated when testing new parameter sets for the flow curve of
a specified material. The calculations of the roll forces and roll torques within
this simulation are afterwards compared with the measurements to get a quality
criterion for the new parameter set. The detailed description of the simulation
scenario is presented in Section 4.2.
It is important that the simulation scenario behaves in the same way as the
online process. Therefore, the simulation uses feedback of the measured speed,
the reduction, and temperature to calculate the new predictions for roll force and
roll torque. This enables the simulation to achieve the same working point as in
the online process. Another problem can be the amount of data. The simulation
of a whole batch where only one material group was rolled consists of thousands
of different deformation steps and will therefore be highly expensive in terms of
simulation time. Several optimization algorithms require boxed constraints of the
optimized parameters. In our case, parameters A and mi are dependent on each
other. The only limitation which can be set is a plausible region for the resulting
value kf of the basic deformation. In hot mills, the maximum basic deformation
value for kf is usually below 300 mm N
2 , but always positive. Then, for a given

maximum working range of the deformation, deformation rate, and temperature,

the feasibility of the parameter set can be tested. Due to the fact that every
company has usually its own classification system it might be that materials
which are grouped together in one company are separated in other companies. In
this case, the optimization, which has been done in the first company cannot be
directly used for other companies and has to be renewed every time. Summarizing,
it is desirable to optimize the process in order to

• reliably estimate valid flow curves,

• reduce lab costs,
• save time,
• determine parameters in their working environments, and
• make the process more flexible and adapt to new (material) changes quickly.

4.4. Flow Curve Parameter Optimization

For the development of new materials or if the material behavior cannot be
deduced from similar grades, the flow curve parameters are unknown and no
proper parameter guess is available. Therefore, an initial screening with the SPO
framework can be used prior to a local optimization.

41
4. Data Driven Meta-Model Based Optimization

Algorithm 3: SPO-based hot mill simulation tuning.

// phase 1, collect initial knowledge about the optimized
process:
1 let A be the hot mill simulation model we want to tune;
2 generate an initial design DES = {x(1) , . . . , x(n) } of n parameter vectors;
3 let k = k0 be the initial number of replications for determining estimated
responses;
4 foreach x ∈ DES do
5 run A with x to determine the estimated response y of x;
// phase 2, building, using and improving a surrogate model:
6 while stop criteria not reached do
7 build surrogate model Y (x) based on DES and {y (1) , . . . , y (|DES|) };
8 optimize the model w.r.t some cost function and constraints, thus produce
a set DES’ of d new parameter vectors ;
9 run A with each x ∈ DES’ to determine the response;
10 extend the design by DES = DES ∪ DES’;
// phase 3, final exploitation and fine tuning:
11 use local optimizer for the best p parameter sets x1...p ∈ DES, without
constraints

4.4.1. The Sequential Parameter Framework

One framework for surrogate-model based optimization is sequential parameter
optimization (SPO) [BBLP05]. SPO combines methods from classical Design of
Experiments (DoE) and modern Design and Analysis of Computer Experiments
(DACE) [BB03, BBPV04] based on Kriging models. The algorithm was adapted
for hot mill parameter optimization(︂ task. Algorithm
)︂ 3 presents the pseudo code
of SPO. Note, that the notation x(i) , y (i) for the data from the i-th pass is
used. This data is passed to the surrogate model. The algorithm was divided
into three phases. Within the initial stage, SPO explores the search space of the
optimization problem A. The optimization is treated as a black box. For the
initial phase a certain pre-defined number of design points x is passed to A. The
strategy for the creation if the initial design is usually a space filling design, e.g.
Latin hypercube sampling. With every design point the black box optimization
problem is called and the objective function returns some output y representing
the performance value. In the second phase SPO tries to determine a functional
relationship between x and y. SPO thus uses a model Y (x) as surrogate for
the hot mill simulation model A. As mentioned above, the chosen model type
is Kriging. Kriging is frequently used for surrogate-model based optimization,
because it provides a powerful and flexible predictor. It also provides an estimate

42
 4.4. Flow Curve Parameter Optimization

of the variance or error of each prediction. The observations are interpreted

as realizations of a stochastic process. A Gaussian kernel is used to model the
correlation between observations [SWMW89].
In the sequential improvement loop SPO optimizes the surrogate model Y (x)
over the considered space of input variables by means of a cost function. Once
the new set of design points DES’ has been selected, the required evaluations
of DES’ are performed. Based on DES’, the surrogate model Y (x) is updated.
In the second phase at line 8 of the of Algorithm 3, a search on the surrogate
model is performed. During that search the inequality constraints of the mill
parametrization problem have to be considered. As the constraints are not
expensive to evaluate, they are evaluated together with the surrogate model itself.
For the inequality constrained optimization, we use the popular method developed
by Powell [Pow94, Pow88], which does not require any derivatives of the objective
function to be available. During this optimization step 8, the next point x to
evaluate in the sequential loop of SPO is determined. For expensive, global, black-
box optimization Jones [JSW98] introduced efficient global optimization (EGO).
EGO exploits the information given from a Kriging model, i.e., the predicted
mean and variance, to compute the expected improvement (EI) of a given solution.
EI can hence be used as a cost function during step 8, as an alternative to the
predicted value of the Kriging model.
In the last phase of the optimization (line 11), the well known downhill simplex
algorithm introduced by Nelder and Mead [NM65] is used to improve the best
found results by a local optimization procedure. The algorithms uses the downhill
simplex implementation in the nloptr R package. During local refinement,
constraints are disregarded because they no longer play a role in the region of
good solutions.

4.4.2. Experimental Setup

For the first mill problem, the parameter optimization was based on Equation
(4.2). The feasible range was 0 ≤ kf ≤ 300. That is, solutions that result into
negative kf values or kf values larger than 300 are considered to be infeasible.
The usual working point for the test data was in the following range:

0 ≤ φ ≤ 0.5, 0 ≤ φ̇ ≤ 600, 500 ≤ ϑ ≤ 600.

With that said, the optimization problem to be solved in this study is defined as
follows:
• Parameters to be changed are the flow curve parameter vector m and the
consolidated parameter A of the flow curve.

43
4. Data Driven Meta-Model Based Optimization

Table 4.2.: Upper and lower bounds for the parameter set M introduced in section 4.3.1
which was used during the optimization. All parameters are of type FLOAT.

Factor Low High

A 0 2
m1 -0.01 0
m2 -0.3 0.4
m3 0 0.2
m4 -0.1 0.1

• The target is to minimize the deviation of simulated roll force from the
measured roll force.
• Computational constraints: The evaluation of the objective function is
expensive.
The parameters which represent the search space and were subject to optimization
in this study are summarized in Table 4.2.
The success and improvement of the optimization which is the central aspect of
the roll force optimization problem was validated against a parameter set which
was used in practice so far. This parameter set was selected by experts according
to the closest distance to a known material. With closest distance the relationship
according to chemistry is meant. This is a valid approach since beside the roll
force deviation of the model calculation an additional online adaption is used
to compensate errors. These online models are described and analyzed in the
next chapter. However, a good and solid model calculation is always the best
precondition for good rolling quality. The residuals which cannot be compensated
with the data-driven approach have to be compensated directly during rolling.
The selected flow curve will therefore be used as a baseline reference for the
optimizations. The SPO toolbox (SPOT) was selected to conduct the experi-
ments [BBZ12]. Parameter used for SPOT are set as follows:
seq.predictionModel.func: The chosen surrogate model is Kriging, based on code
by Forrester et al. [FSK08].
init.design.size: The initial design consists of 40 candidate solutions, which are
created by Latin Hypercube Sampling (LHS).
auto.loop.nevals: Overall, the number of evaluations of the objective function
are varied (hot mill simulation) with different sequential design. The se-
quential design either consist of 10 evaluations of the objective function or

44
 4.4. Flow Curve Parameter Optimization

60 evaluations of the objective function. Therefore, the total number of

evaluations were 40, 50 or 100 which includes the initial design of 40.
init.design.repeats: Since the objective function is deterministic each solution is
evaluated exactly once.
seq.design.new.size: In each sequential step, one new solution is evaluated on
the target function.
seq.predictionOpt.method: The sequentially created models are optimized by
Latin Hypercube Sampling (LHS) and Constrained Optimization by Lin-
ear Approximation (COBYLA). The COBYLA implementation from the
NLOPT library, included with the nloptr R-package is used.
seq.design.size: The sequential step size. LHS evaluates 200 points with the
model.
seq.predictionOpt.budget, seq.predictionOpt.restarts COBYLA has a budget
of 1 000 evaluations of the surrogate model, and will restart if it converges
prematurely.
Table 4.3 summarizes the these chosen values for each parameter.

Table 4.3.: Spot settings. A detailed description of each parameter can be found in
Section 4.4.2
Parameter Value
seq.predictionModel.func spotPredictForrester
seq.predictionOpt.func spotModelOptim
seq.predictionOpt.method "NLOPT_LN_COBYLA"
seq.predictionOpt.budget 1 000
seq.predictionOpt.restarts TRUE
seq.design.size 200
seq.design.new.size 1
init.design.size 40
init.design.repeats 1
auto.loop.nevals 40/50/100
seq.infill spotInfillExpImp/NA

4.4.3. Optimization Process

The description of the optimization process is shown in Figure 4.4 and is separated
into three stages. The first stage is the initial design (I) and is followed by an

45
4. Data Driven Meta-Model Based Optimization

Figure 4.4.: Flow chart of the simulation presented in this section. First of all an initial
design (I) is evaluated followed by an sequential design (S). Afterwards a
local optimization (L) is performed.

46
 4.4. Flow Curve Parameter Optimization

sequential design (S) based on the results of the first stage. The last stage consists
on local optimization (L) based on the second stage. In order to distinguish
between the different settings the notation I/S + n · L is used to characterize
the settings for the different stages. I/S describes the size of the initial design
respectively the sequential design and n describes the number of points used for
the local optimization. the number of evaluations in the local optimization stage
is characterized by L. An experiment conducted with 40/10 + 5 × 100 has a design
size of 40, 10 sequential designs and based on these results a local optimization
on the 5 best points is made.

4.4.4. The Influence of Expected Improvement

With the used Kriging model the predicted mean and variance are used to calculate
the expected improvement (EI). The idea of EI is to generate new designs in areas
where only little data is available and the uncertainty of the model is very high.
Furthermore, also designs in areas with an expected high performance is generated.
For both proposals the constraint is always considered. The question is, if it will
be beneficial to use the EI infill criterion in the sequential design. To answer
this question, a scenario with an initial design size I=40, sequential design size
S=10 or S=60 and 10 repeated evaluations was chosen. The experiments with a
sequential design size of 60 outperformed the experiments with a sequential design
size of only 10. No statistically significant difference between the experiments
with EI and without EI can be seen so EI seems not to improve the results even
with larger sequential design. In fact it is quite the reverse: The performance of
the model-based optimization decreases if EI is used. This is owed to the fact
that the total number of evaluations is higher.
The experiments with a sequential design size of 10 had a median RMSE of 1235
(no EI) and 1352 (EI) and the experiments with the greater design size of 60 had
a median RMSE of 644 (no EI) and 1016 (EI). The results are summarized in
Figure 4.5(a) . To answer the question of the influence of EI it was shown that EI
seems to drop or at least not improve the optimization. A possible reason for this
might be that the problem is too simple and that exploring new regions is not
beneficial. The major improvement seems to come from the initial design and the
local optimization afterwards which is subject of the study in the next section.

4.4.5. Influence of Local Optimization and Initial Designs

The same comparison as in the previous section was made for the additional
refinement of the best-found results with local optimization. Here, the best five
parameter sets found by the Kriging-based optimization were selected and opti-

47
4. Data Driven Meta-Model Based Optimization

(a) Analysis of the Kriging simulation with (b) Analysis of the results after local opti-
an Initial Design Size of 40 and a sequential mization of the previously results found with
design Size of 10 and 60. Shown are the Kriging. Shown is the best point in each seed
results for the SPOT runs with and without after 100 local optimization runs of the 5 best
EI based on 10 and 60 evaluations in the points.
sequential design phase. The number in the
labels are indicating the total number of
evaluations on the original problem.

(c) Zoom of the analysis after local optimiza-

tion shown in Figure 4.5(b).

Figure 4.5.: Comparison of the results for Kriging with and without Expected Improve-
ment after sequential design (a) and after the local optimization (b). The
blue dotted line shows the reference value for the original parameter set. In
order to compare the local optimizations, Figure (c) shows the boxes with
different scaling.

48
 4.4. Flow Curve Parameter Optimization

Figure 4.6.: Comparison of simulation results for an initial design of 40 points with
different local optimization afterwards against a pure local optimization with
a random starting point. As a reference box also a result for a simulation
with a sequential design and local optimization is shown (50 + 5 × 100). The
boxplots are representing the results for 10 different seeds. The number
of parameter evaluations in case of a previous initial design is 40 + 5 × L,
where L denotes the number of local optimization steps. The blue dotted
line shows the reference value for the original parameter set.

mized with a Downhill Simplex approach. Each of the five parameter sets received
a budget of 100 evaluations for further improvement. The local optimization
improved the median of the RMSE to below 500 in all experiments whereby
only one experiment without the local optimization had a resulting RMSE below
1000. This was the case for an initial design size of 40, a sequential design size
of 60 without EI. The local optimization yields the major improvement indepen-
dent of the budget of the model-based optimization. The results are shown in
Figure 4.5(b). Thus, the question is whether model-based optimization can be
skipped, and whether a local optimization directly after the initial design is more
profitable. In addition, it was also checked whether it can be beneficial to choose
completely randomized starting points for the local optimization. The results of
this investigation is shown in Figure 4.6. To have a decent comparison the number
of evaluations should also be taken into account. The boxplot is showing the
statistic for the best point produced by each of the 10 random number generator

49
4. Data Driven Meta-Model Based Optimization

Table 4.4.: Numerical values for the box plots shown in Figure 4.6. Shown are the
number of evaluations in the initial and sequential design and during the
local optimization. Box plot parameters listed are minimum, lower quartile
(Q1), median, upper quartile (Q3) and maximum value.
Init. Seq. Local Min Q1 Median Q3 Max
40 5x100 342 352 363 374 401
40 1x100 352 378 437 498 527
40 1x500 331 331 334 340 350
40 10 5x100 342 352 361 374 401
40 10 1x490 331 331 338 351 376
100 390 479 585 652 787
140 348 442 506 580 731
300 331 366 390 470 523
540 331 331 345 371 400

seeds. To analyze the importance of the initial design and the local optimization
the experiments from Table 4.5 were performed.

The completely local optimization (with random starting point) performs similarly
good as the two approaches where the five best points are each optimized locally.
The best approach seems to be (40 + 1 × 500).
As a result of the investigation in this section it can be concluded that it is more
profitable to optimize just the single best point found in an initial design.

4.4.6. Impact of the Fitness Landscape

However, the completely randomized, local approach is not much worse than the
approaches with an initial design. This indicates that the problem landscape is
rather simple. If the landscape would be univariate the optimizations would always
result in the same best parameter set independent of the seed. There can be seen
huge differences in the parameter set which are shown in Table 4.6. This gets more
clear if we keep in mind that the first parameter is a multiplicative component and
is therefore linearly correlated to the flow resistance. The corresponding contour
plot is shown in Figure 4.7. While the contour plot depicts several basins in the
fitness landscape, they are all of some considerable size. Note that the contour
plots of other parameter combinations result into even simpler landscapes, most
giving the impression of being unimodal.

50
 4.4. Flow Curve Parameter Optimization

Table 4.5.: Experiments, which were performed to analyze the impact of the initial
design and local optimization.
Name Description Evaluations
40+5x100: Initial design of 40 points with 100 local opti- 540
mization steps of the best 5 points.
40+1x100: Initial design of 40 points with 100 local opti- 140
mization steps of the best point.
40+1x500: Initial design of 40 points with 500 local opti- 540
mization steps of the best point.
40/10+5x100: Initial design of 40 points, model-based optimiza- 550
tion with 10 points, and 100 local optimization
steps of the best 5 points.
40/10+1x490: Initial design of 40 points, model-based optimiza- 540
tion with 10 points, and 490 local optimization
steps of the best point.
100 Local: Random start point with 100 local optimization 100
steps.
140 Local: Random start point with 140 local optimization 140
steps.
300 Local: Random start point with 300 local optimization 300
steps.
540 Local: Random start point with 540 local optimization 540
steps.

51
4. Data Driven Meta-Model Based Optimization

Table 4.6.: Parameter comparison for local optimizations with an arbitrary starting
point and a budget of 540 evaluations. Only seed 2 stopped earlier, i.e.,
after 431 evaluations. Shown is the best parameter of each seed with the
corresponding RMSE value. The parameters are used to calculate the flow
stress according to 4.2.
Seed A m1 m2 m3 m4 RMSE
Unit e-03 e-02 e-02 e-02 kN
1 1.2 -4.8 -0.2 8.2 2.6 331
2 1.2 -4.8 -0.3 8.1 2.6 331
3 0.9 -4.6 -2.7 11.0 2.8 339
4 1.2 -5.0 -5.7 9.8 2.3 335
5 1.0 -5.3 -18.0 14.0 1.8 371
6 0.7 -4.2 -3.2 11.0 2.9 359
7 0.8 -3.5 16.0 5.9 3.9 400
8 1.2 -4.8 -0.5 8.2 2.6 331
9 1.4 -5.5 -6.0 11.0 2.7 352
10 2.9 -5.7 0.8 -0.58 0.95 461

Predicted RMSE [kN] [kN]

−0.0040

8000

−0.0045

−0.0050 6000
m1

−0.0055

4000

−0.0060

2000

−0.0065

−0.0070 0
0 1 2 3 4

Figure 4.7.: Contour plot for parameters A and m1. All other variables are kept constant
at the same value as for the best case. A random Forest model was created
with all available simulation results and is used here for the prediction of
the RMSE, denoted by the background coloring. Each black point indicates
the result of a simulation. The white triangle marks the simulation with
the best RMSE.

52
 4.4. Flow Curve Parameter Optimization

4.4.7. Comparison of the Resulting Roll Force Deviations

Figure 4.8 shows the deviations of the pure model roll force (red), the adapted roll
force (blue) and the measured roll force (green) of the original parameter set, before
optimization. Figure 4.9 shows the same for the optimized parameter set. The
model roll force is highly correlated with the calculation of the parameter kf which
was described above. The calculation of the roll force was done in a production
scenario. According to previous knowledge of the rolling process the system
learned the roll force deviation for this material and made an online regression for
the correction of the calculation. Furthermore a pass to pass adaption was used
to correct the roll force deviation. That is the reason why the adapted roll force is
often close to the measured roll force. Due to limitations to the adaption system
some large deviations remain, especially when the model roll force estimate is way
off.
Typically, model corrections of up to 30% are tolerated in the online scenario.
This limitation is used because something else has to be wrong if calculation
errors are exceeding this value. Furthermore, the adaption is reset from time to
time. In case of the original parameter set, the deviation were more than 30%
and therefore we still have a residual deviation for the rolling force.
In both figures, Figure 4.8 and Figure 4.9, the values for the pass number 15 are
missing. This is due to a special strategy during rolling of this material. In this
early stage of commissioning there were still some difficulties during measurement
preparation. Usually, pure model deviations without any optimization of about
10% are tolerated and deviations of less than 5% are considered to be good. In
Figure 4.10, the roll force deviation of the model with the original, standard
parameter set is shown. Next to it, the roll force deviation of the model with
the new parameter set is plotted. It can be seen, that the deviation is very good
in the first passes but still shows some deterioration in the last passes. This
phenomenon accounts for wrong temperature calculation and the process was
often delayed which was not correctly modeled during this phase of the process. If
these parameters are optimized in a later stage of the project, the deviation in the
last passes will also be reduced to a suitable amount. When the calculated roll
force deviates from the actual measurement the automation system is correcting
the roll gap position based on the current roll force. This is done in order to
compensate for the mill spring. The mill spring is the spring which occurs in
the stand when we suddenly have to apply a roll force. This means that the gap
between both rolls will be higher and we have to adjust the gap setting to reach
a specific thickness. If we consider this opening as a linear function of the roll
force and assume a value of 1mm at a roll force of 6 000kN then we would have
an error of 0.5mm when the prediction is 3 000kN away from the measured roll
force. When optimizing this prediction, the length of the final product, which is

53
4. Data Driven Meta-Model Based Optimization

Figure 4.8.: Roll force of pure model (red), model with online adaption (blue) and
measured roll force (green) with original parameter set. The online adaption
is typically a simple regression based optimization which learns the difference
between model and measured of the last 2 month of production. Additionally
the difference of the last pass between prediction and measurement is taken
into account. The adaption is used as an additive offset to the model
prediction. Because it can be reset and should only dynamically react on
slight deviations from unknown source the correction amount is limited.
This limitation can be seen especially in passes 12-14 where we still have
a residual error. The missing measurement in pass 15 is due to a special
strategy and problems during the measurement preparation in this pass.

Figure 4.9.: Roll force of pure model (red), model with online adaption (blue) and
measured roll force (green) after optimization. Here we have used the
parameters of our offline simulation, i.e., the result of the local optimization.
This figure clearly demonstrates that the performance of the model has
vastly improved and that the online regression could have been switched
off.

54
 4.4. Flow Curve Parameter Optimization

Figure 4.10.: Relative roll force deviation plotted for each pass. The error is plotted as
the relative roll force error Measured Force - Calculated Force. The
top plot shows the prediction error with the original parameter-set and the
bottom one uses the optimized parameter-set. The drift which is present
in the original parameter set is reduced with the optimized parameters.
Furthermore, also the IQR is reduced with the usage of the optimized
parameter.

55
4. Data Driven Meta-Model Based Optimization

in an acceptable range, can be increased.

4.5. Parameter Optimization for Known Materials

There are several situations where materials are well known but still achieve
a performance which is not sufficient. Those situations can be caused by the
following problems.
• Measurement equipment or test scenario in laboratory.
• Different chemical analysis than in laboratory, but within the specifications.
• Process description not fully available.
Sometimes, also parameters of a similar material are used to have an initial guess
of the behavior. If the deviations are not caused by some failure in other parts of
the description and can surely be addressed to the flow curve, then an optimization
of the parameters have to be made. It will now be assumed that the performance
is already within some reasonable limits and that the optimization is only used
for fine tuning of the parameters.
Then, instead of initial screening of the whole region of interest, a simple local
optimization with the downhill simplex algorithm can be performed. The algo-
rithm and concept of the previous section can still be used. The result of the
local optimization of such a scenario is shown in Figure 4.11 where the minimum
achieved RMSE over the number of local optimizations is plotted. The data were
coming from a aluminum finishing mill and the material which was optimized was
a magnesium based alloy.
It can be seen that the major improvements were achieved during the first 100
evaluations. The initial RMSE off 11.5% was already in a reasonable region.
During the optimization it was decreased to a value of 8.6%. This value might
be further decreased by the use of online optimization algorithm which will be
introduced in Part II.

4.6. Parameter Optimization for Multiple Mills

Up to now the optimization procedure was conducted on real-world data for a
single mill type only. All mill specific errors are compensated now by the flow
curve parameters. Common mill specific errors are

• measurement errors (device specific),

56
 4.6. Parameter Optimization for Multiple Mills

Figure 4.11.: Minimum RMSE over a local optimization with 200 evaluations. The
performance is measured as total RMSE of the roll force for all simulated
products.

• construction errors (stand specific),

• different constitution of same material grade and
• input errors (production planning or calibration).

Especially input errors should usually be avoided and are sometimes very hard
to detect. An example of a trivial input error is the roll diameter which is
coming from the mill operator or the roll shop where the rolls are grinded. Also
the ingot description may cause deviations which are not always easy to detect.
For aluminum rolling, where the ingots are usually faced, at least the geometry
description should be accurate. Nevertheless, deviations of the ingot description
are very common. For steel rolling the chance of input errors is much higher since
the slabs are not faced.
Other problems for aluminum rolling may origin in the used emulsion. The used
emulsion is often kept secret and is very specific to the mill owner and mill type.
Therefore, no complete description of it is available but the impact on friction
and rolling is quite high [SLSJ02].
As a consequence, optimized parameters for one rolling mill might not be ideal for
another mill of the same type. The problem gets worse if different mill types are
considered. E.g. the usual temperatures in aluminum roughing mills are 400-500
degree C and usual target temperatures for the finishing mill are between 250 and
350 degree C. So the optimization of parameters with data from the roughing mill
are optimized in a different region compared to the usual rolling in finishing mill.
Furthermore, when rolling in finishing mill multiple stands are rolling the product
at the same time and a specific tension between those mills is used and controlled.

57
4. Data Driven Meta-Model Based Optimization

This tension can also have a high impact on the roll force.
So, two questions arises for the optimized parameters.

• How is the performance of the optimized parameters on other mills of the

same type?

• How is the performance of the optimized parameters on other mill types?

For answering the first question a new experiment has to be conducted since no
data on other mills were available for the material optimized above.
First, a flow curve is optimized with real-world data in the same way as described
in the previous section. The data is coming from an aluminum hot strip finishing
mill. To answer the first question, the optimized flow curve parameter are used
on another finishing mill. Here, real-world data of the rolling process for the same
material is used. The chosen performance parameter is the RMSE of the roll force
deviation for the simulated products.
The box plot in Figure 4.12 shows the achieved performance for both mills. The
boxes are representing the force prediction errors of all products and all stands for
two different mills, A and B. Further, the Figure shows the result for flow curve
parameter optimized on the specific mill, A(Opt) and B(Opt) and the performance
when the parameters are optimized on the other mill, A(BestB) and B(BestA).
The simulated material was a manganese aluminum alloy. The RMSE of the
first mill (FM A) cannot be compared to the RMSE of the second mill (FM B)
since the product variety, the measurement equipment and the working areas in
terms of φ, φ̇ and ϑ are different. This is also the reason why the initial RMSE
shows huge differences, although the same original flow curve parameter are used.
These parameters were chosen because the material behavior was expected to
be close to another known material group. The optimization of the flow curve
with real-world data from Mill A lead to an enormous improvement of roll force
prediction accuracy. This can be seen in the second box (A Opt.). The opti-
mization procedure presented in the sections before could not only improve the
mean and median prediction error but also the standard deviation of the accuracy.
While the prediction error with the original parameter had a mean of -4 413kN
and a standard deviation of 3 359kN, the optimized parameters achieved a mean
prediction error of 161kN with a standard deviation of 1 766kN.
Unfortunately, the same parameters did not manage to improve the prediction
accuracy on another mill of the same type (B Best A). Although the standard
deviation was improved from 4 991kN to 2 028kN, the mean prediction error
increased from 2 732kN to 4 690kN.
The reason for this behavior can be caused by the material description and classi-
fication. The considered material was a magnesium based alloy. The deformation
resistance and herewith the roll force is highly dependent on the magnesium

58
 4.6. Parameter Optimization for Multiple Mills

Figure 4.12.: Optimization of the flow curve among different mills of the same type. In
this experiment, the flow curve was optimized on two aluminum finishing
mills (A and B) separately. Then, roll force prediction error on the other
mill was measured. Additionally the residual errors of the original flow
curve on both mills is shown.

content. The range for a specific type of aluminum alloy may be e.g. 4.5%±0.5%
which can cause differences in roll force of approx. 10%-20%. This variation of
magnesium can lead to higher variance in the prediction and incorrect description
of the flow curve. A possible solution would be either to create subgroups for this
alloy with closer ranges for elements with high impact on roll force or to create a
flow curve formula which reflects theses changes.
Also mill related phenomena like the composition of the applied emulsion or
measurement inaccuracies can lead to these effects.
Since also all errors which occur in mill A are compensated by the optimization,
a suitable solution would also be to optimize the multiplicative factor for the flow
curve. Reconsider formula 4.1 which is the basic description of all flow curves.
The multiplicative factor A can be used here to optimize the flow curve prediction.
This is reasonable because the basic shape of the flow curve which is reflected in
the variance of the prediction has improved for both mills. The factor can now be
optimized among different mills or just for every mill. An optimization among
different mills would decrease the performance for mill A but improve it for mill
B. The resulting residuals can then be compensated by online algorithms which
are discussed in detail in the next chapter.
Because the flow curve should represent the material behavior and not mill char-
acteristics the answer to the second question is of great importance. To test the

59
4. Data Driven Meta-Model Based Optimization

Table 4.7.: Summary of the boxplots shown in Figure 4.13.Beside the quartile values,
the median the minimum and maximum, also the standard deviation is given.
All values are in kN.
Name Median Std. Dev. Min Max 25% 75%
RM (Orig.) 112 7764 -3173 23506 -2379 12239
RM (HSM Opt.) -6977 7835 -9637 23300 -8881 -1528

Figure 4.13.: Boxplot for the roll force prediction error when using a flow curve from a
similar material (left) and a flow curve optimized from the finishing mill
(right).

optimized flow curve parameter, they were used for roll force prediction of an
aluminum roughing mill (RM). The same material as in the previous test was
used for the simulation, i.e., a magnesium based aluminum alloy. Figure 4.13
shows the residual roll force error for an aluminum roughing mill. The box on
the left side represents the errors when a flow curve of a similar material is used.
Here, the material was selected due to best knowledge of process experts. These
predictions have been compensated by other algorithms during rolling. The box
on the right side shows the residual prediction errors from an optimized flow curve.
The optimization of the flow curve was done on a finishing mill.

The summary of the boxes is given in Table 4.7. It can be seen that the standard
deviation cannot significantly be reduced in comparison to the original curve.
Both flow curve parameters show high residual roll force deviations which have
to be compensated by other algorithms. Otherwise the rolling process will be
unstable. The reason for this behavior is due to the different regions of the flow
curve which are used in both mills. In a finishing mill, the usual temperatures are
below 400 degree Celsius. This is almost the minimum temperature for roughing

60
 4.7. Summary

mills.
A possible solution would be to extend the optimization scenario to consider
multiple mills. Then, the number of samples available for both mills have to be
considered and a global performance value has to be selected. Another solution
would be to optimize the flow curve for the selected mill types. Then different
parameter sets will be generated dependent on the temperature or other criteria.

4.7. Summary
The procedure presented in the previous section showed a model-based opti-
mization for flow curve parameters which are typically used in rolling mills. A
combination of global and local optimization approached were used. Data from
various sources and mill types were used to conduct the experiments. For the first
study, data was collected from an aluminum rolling mill where no information
about the material was known prior to the first rolling trials. The second study is
based on data from various type of mills, i.e. two different aluminum finishing
mills and a roughing mill. Here, some knowledge about the material was available
and an initial setting of flow curve parameter was used. Both studies showed a
significant improvement in model prediction accuracy which makes it possible
to evaluate and optimize flow curves without conducting time consuming and
expensive measurements in a laboratory.
Especially the study in Section 4.6 showed the generalizability of the optimized
parameters. When optimizing parameters on a specific mill they may not be
suitable for the usage on other mills since all plant specific errors are compensated
during the optimization procedure. Instead, conducting experiments for multiple
plants and especially multiple mill types results in optimized parameters for the
material. This also ensures that a broad range in terms of validity of the flow curve
which was used by the algorithm. When only data of a single mill is available
then the optimization procedure should be slightly modified. In roughing mills
the temperature usually does not drop below 400 °C for aluminum and 1 000 °C
for steel. Therefore, the temperature may not be optimized for low temperatures.
A possibility to overcome this problem is to neglect the temperature parameter
during the optimization and to optimize only the remaining parameters. This is
especially helpful when default parameters for a material are available and the
task is only to optimize those parameters. In case no parameters are available a
default value can be retrieved from similar materials where a similar material has
to be chosen by process experts.
After optimizing the flow curve parameters they should be visualized and com-
pared against known flow curves of similar materials.

61
4. Data Driven Meta-Model Based Optimization

Finally, the surrogate-modeling method employed here may profit from compu-
tationally cheaper information that can be gained from a coarse variant of the
objective function. In detail, a parametrization of the mill model may be tested
with less data. This results into faster evaluation times, but yields a less accurate
estimation of the quality. This cheaper information can be used in tandem with
the more accurate, expensive data to train a better surrogate-model. To that end,
Co-Kriging [FSK07] can be employed. Co-Kriging allows to exploit correlation
between coarse and fine target functions. This may result into a much improved
surrogate-model of the hot-mill, without the requirement of additional expensive
evaluations.

62
 Part II.

Online Optimization

63
5. Introduction to Online
Optimization
Industrial processes require a high flexibility without losing robustness and gener-
alizability. Process predictions have to be fast and accurate. Offline models as
described in Chapter 4 are suitable for improving the general process models but
fail to adapt to dynamic changes. For these dynamic adaptions, online models
are perfectly applicable. Throughout this thesis, the term online does not mean
that the process is connected to the internet, but instead refers to the on-the-fly
adaption of the algorithms.
This chapter presents several concepts and algorithms to improve predictions
without reducing the robustness that are provided by offline algorithms. Typi-
cal online optimization mechanisms, which are traditionally used within many
industrial processes, are building clusters and are using average values within each
cluster. The clusters are build either by discrete values directly or by partitioning
of continuous variables. Prominent values which are used within the context
of hot rolling are e.g. the material or the used reheating furnace. The most
important continuous values are thickness, width, profile and temperature. For
each unique combination of theses values an average prediction error is collected.
Problems which arise with this concept lie in the inhomogeneous boundaries. Most
often, building several groups out of a continuous variable does not make any
sense and using a regression is a traditional approach to solve this problem. One
of the promising candidates for industrial online optimization is online support
vector regression which is an enhancement to classical support vector regression.
Standard support vector regression lacks the possibility to incrementally add
new observations without retraining of the complete data. It has already been
successfully applied to hot rolling problems [BYHB10, HELR08]. Theoretically,
regular support vector regression can also be used for online optimization. Appli-
cations that are not time critical can integrate new samples by building a new
model with the complete data set. In this way, dynamic effects can be taken into
account with the classical approach. But if new samples are arriving very fast
the computation time for model building is too long. This is where online SVR
can be used. It offers incremental and decremental learning. This chapter will
give an introduction to online optimization. Afterwards, the datasets used for

65
5. Introduction to Online Optimization

the optimization are presented in Section 6. Section 7 will introduce three online
algorithms and show their individual performance. After a solution for online
parameter optimization is presented in Section 8 the performance of the three
presented algorithms are compared in Section 9. Parts of these chapters have
been taken from [JBBR18].

5.1. Definition
Before the main requirements and the methods are presented, a central question
has to be answered first.
• What exactly is online optimization?
There are several definitions of online optimization available:
• ”Online learning is the process of answering a sequence of questions given
(maybe partial) knowledge of the correct answers to previous questions and
possibly additional available information.”[SS12]
• ”... an algorithm runs online if it makes a decision (computes a partial
solution) whenever a new piece of data requests an action.”[GKR+ 01]
• ”An on-line algorithm is one that receives a sequence of requests and performs
an immediate action in response to each request.”[Kar92]
All of those definitions have in common that there needs to be an immediate
response on arrival of a question / request. The information about the past is very
limited. In the case of hot rolling Online optimization procedures will continuously
receive data and try to optimize one or multiple objectives with limited or without
any information about the past. Let yt = (y1t , . . . , ymt )T define the m objectives
at time t and ŷt defines the predictions of the online algorithm at time t. Then
the optimization procedure tries to minimize residuals:

et = yt − ŷt . (5.1)

Additionally, the procedure incorporates new samples and updates its prediction
for the future. For the prediction of future values it may or may not incorporate
additional input data xt = (x1t , x2t , · · · xnt )T received at the same time t. It will
be further assumed that only a limited amount of data from the past can be stored
and used for prediction.
Since some aspects may be common also to other optimization procedures, another
important question is:
• What exactly is the difference to classical (offline) optimization?

66
 5.2. Initialization Problems

In contrast to traditional (offline) optimization, no or only few samples are already

known and may be used for initialization of the algorithm.
Another important difference to classical optimization is the update procedure.
New data will be incorporated in an incremental way. For classical optimization,
usually the whole dataset has to be used to update the predictions.

5.2. Initialization Problems

The potential problem, which might occur if no data has been retrieved before,
is called cold-start problem. Otherwise it will be called a warm-start problem
[GG01, Gon98, MTP03].
The fact that data may or may not be available does not mean that no information
about the samples is known prior to training. Information about the number of
dependable variables, type of variables and their usual range is assumed to be
known. This assumption can be made because process know-how is mandatory for
the task of optimization in the industrial environment. The previous assumption
allows furthermore to scale the data prior to training. Scaling is used in most
common algorithms like artificial neural networks (ANN) or kernel methods.
Scaling refers to a linear transformation of data. After scaling, the data will have
pre-defined upper and lower boundaries. Additionally to scaling, also normalization
is used in many offline algorithms. Since no information about the variable
distribution is known, normalization cannot easily be used in online algorithms.
When the performance of cold-start and warm-start problems are compared there
will be a significant difference at the beginning. The performance of cold-start
problems are inherently very poor at the beginning and will increase significantly
after some data have been processed.
Example 1. To illustrate this behavior, a one-dimensional test data problem is
analyzed. The selected model for this case is a simple linear regression model.
Let us assume the error is linear correlated as shown in Figure 5.1(a). Here, the
error is shown over the dependable variable x. The first five samples are labeled
separately. For sake of simplicity the regression formula y = 0.15 + 0.1 · x + ϵ has
been used while the noise term is uniform randomly generated within an interval
of [−0.4, 0.4]. Figure 5.1(b) shows the residual error over the sample numbers.
This error is used to train the online algorithm. A regular linear regression model
without any outlier detection or learning rate was used. The warm-start model
was fed in advance with the first ten training samples which were randomly
generated. The error in case of a cold-start for the first sample has exactly the
same value as the first labeled value shown in Figure 5.1(a). The second error is
then the difference between the first value and the second value. Afterwards the

67
5. Introduction to Online Optimization

(a) Test data (b) Prediction error

Figure 5.1.: Comparison of residuals between warm-start and cold-start. The one
dimensional test data was generated according to y = 0.15 + 0.1 · x + ϵ.
The y-axis value in Figure 5.1(a) represents the target value which should
be predicted by the online algorithm, while the x-axis value represents
the value of the dependent variable. The values were randomly generated
and the first five data points are labeled separately. Figure 5.1(b) shows
the prediction errors over the sample numbers. The red curve shows the
warm-start online algorithm, which was trained already using the first 10
samples. The blue curve represents the residual error term when no training
was applied before (cold-start). The online algorithm was a linear regression
model in both cases.

68
 5.3. Demands and Properties

residual prediction errors results from the linear regression model of the processed
samples. This simple example nicely illustrates the effect. After five samples
have been received, the difference between warm-start and cold-start is almost
negligible. The point or sample number where the difference between cold-start
and warm-start gets marginal is highly dependent on the number of features in
the dataset, the number of categorical variables and the complexity of the dataset.
Regularly, it will take much longer than in this case.
Detection of outliers is almost impossible during the initialization of the algorithm
and thus the model is very sensitive to them at the beginning. Inherently, the
performance will also be very sensitive to possible outliers at the start. To prevent
this, different strategies can be implemented. A simple strategy would be to use
only a scaled version of the gradient error term for learning. Learning rates are
also quite common in stochastic gradient descent (SGD) [Zei12] where the step
size is influenced by the learning rate or neural networks [ST17].

5.3. Demands and Properties

There are different types of algorithms which may be applicable for real-world
online optimization. The industrial application that generates the data stream
will define the most relevant criteria for the selection of the specific algorithm.
There are large varieties of industrial applications where online algorithms may
be used. Different kinds of streaming data, like weather data, stock data, energy
costs may be used for forecasting or general prediction. Also in the field of
hot and cold rolling, many different data streams occur. Some criteria typical
to the field of hot rolling and some assumptions are further described in this section.

5.3.1. Amount of Data

Algorithms used for the task of online optimization should assume to have an
endless data stream. One aspect of this is, that algorithms may use already
received data which are stored in memory and on hard disk. The total available
memory is presumed to be limited.
The number of features is also assumed to be limited. Features are specifying here
the number of different dependent variables. If yt is defining our target variable
at time t and xt = (x1 (t), x2 (t), · · · , xn (t)) defines the input data at time t, then
n corresponds to the number of features. This limitation of n has to be presumed
because a higher number of features would also require more memory. How much

69
5. Introduction to Online Optimization

more memory has to be allocated with a higher number of features also depends on
the algorithm type. Some algorithms will store received samples or intermediate
calculation results like kernel matrices. Some algorithms will only store coefficients
for each feature. Nowadays, the algorithm used for hot rolling should not allocate
more than approximately 100 MB. With a restriction to this amount it can be
assured that at least the data can easily be kept in memory. The reason for such
a small amount of memory is due to the fact that additional memory allocations
are needed for the process where these algorithms are embedded.
Storing data on hard disk has to be avoided since it will inherently increase
computation time because accessing data in memory is much faster. To have a fair
comparison every algorithm is allowed to allocate the same amount of memory.
Storage size refers to the internal memory allocated by the stored samples.
The storage size will define also the possibility of the algorithm to react to changes.
If a large amount of samples is used by the algorithm, the performance is usually
very robust and will adapt to current changes very slowly. Vice versa, within a
small storage each sample gets a higher weight. The adaption to current changes
is performed very fast with the drawback of lower robustness.
Once the amount of data which the algorithm is allowed to allocate hast been
determined the question is which samples should be kept in memory. If the algo-
rithm requires access to already received samples, a strategy for the management
of these samples has to be evolved. For some applications this selection is straight
forward.

Example 2. Assume a manufacturing process which is always producing the same

product and all dependable variables are assumed to be similar, i.e. xt ≈ xt+1 . Fur-
ther denote the actual predictions for the process as yt = f (xt , xt−1 , · · · , xt−s+1 )+ϵ,
where ϵ represents some random noise and s defines the number of samples which
are stored. Then the quality of the prediction is mainly influenced by the current
settings at time t and the actual process state. The dependency or correlation to
the past production is not given. Therefore it is favorable to store only the latest
samples.
For many other processes, like the highly flexible rolling process, this is not true.
For special materials, which are only rolled on rare occasions, the observations
of the actual behavior is strongly correlated with the last time this material
was rolled. On conventional hot strip mills, a change of material may occur in
every second or third product. Surely, the performance will also depend on the
current process state, so a mixture between both data would be beneficial. The
management, which defines the samples that should be stored, is therefore highly
correlated to the concrete real-world problem and its complexity.

70
 5.3. Demands and Properties

5.3.2. Arrival of Data

Another important question is how fast the samples are arriving. This is the most
limiting requirement for the selection of the algorithm type [Kru02].
Typical applications in hot rolling are using TCP telegrams for the communication
and have cycle times of some milliseconds up to some minutes. Cycle time is
defined as the average distance between the arrival of two consecutive telegrams.
Cycle times of more than one minute typically occur when the optimization
problem should consider only data on slab-to-slab basis. An example of such a
scenario is the width prediction in finishing mills or the correction of the so-called
zero point.
Although the data may be received very fast, the necessity for an algorithm update
is not always given. During rolling of a single pass the measurements collected
are usually arriving very fast, i.e. in some milliseconds. But usually, the variance
during this pass is handled by real-time control mechanisms. Those controls are far
more capable to react in real-time but are not designed for complex calculations.
Online algorithms are usually applicable if no real-time requirement is given. Pre-
processing of those measurements and the calculation of averages over a certain
time period is usually done for the hot rolling problems. Afterwards, a statistical
summary will be used to update the online algorithm or the data is processed as
a batch. This leads to the next criteria for algorithm selection: The number of
samples which is used to update the algorithm.

5.3.3. Batch Versus Sequential

If the algorithm is receiving streaming data, the data may occur sequentially or it
may always be a block of data which is received. There are special algorithms
which are using the whole chunk of data (batch) to update their prediction and
some are using every single data sample for updating. An example of batch
learning algorithms is presented e.g. in [LZHM12] where the authors are using
a principal component regression on batches for iterative updates. In principle,
all algorithms which are able to be updated with a single sample can also be
updated with the whole batch. Then, all individual samples of the batch are fed
to the model. Algorithms, which are using batches have to collect samples before
updating. They cannot be updated with every received sample.

71
5. Introduction to Online Optimization

5.3.4. Source of Data

On the first glimpse it does not seem to be important where the data is coming
from. But there is a huge difference, if the underlying process is just periodically
sending measurements which will be processed, or if some event is causing the
processing of data, which will be referred to as event driven data processing.
Periodically measurements are for examples temperature measurements or energy
consumption. This is defined as time driven data processing [Mis18]. The rolling
process is an example for the event driven processing of measurements. Only
if products are rolled the algorithms will be fed with data. When monitoring
the arrival time only, it will be impossible to distinguish between event based or
periodically triggering.
If there is a continuous measurement and periodically some data will be collected
and sent to the online optimization algorithm, then frequency, amount of data
and other behavior of the stream is assumed to be always the same during normal
production. For some time, it may look similar to the collection of continuous
measurements as in the case of weather data. At some point, there will be a
disruption in this data stream. For real-world scenarios this is often the case
due to some maintenance or scheduled downtime where parts of the production
equipment is not available. If this time is known in advance the algorithm may
perform some more cost-intensive optimization which would not be easily possible
during the continuous production. It may also be achieved during the production
if the cost-intensive optimization can be outsourced to other machines. Then
again, the time for takeover of the improved optimization has carefully be selected.
For rolling mills, there are always some scheduled break-down events. This can
be e.g. some roll change or some regular scheduled maintenance.

5.3.5. Pre-Processing
The pre-processing of data can be realized in several ways and can get very
demanding. It can and should be installed on the sender side or on the receiver
side of the data. The general process of filtering and processing for the hot rolling
application is almost identical for all scenarios discussed throughout this thesis.
The description of the process is illustrated in Figure 5.2 and comprehends the
following steps.

1. Calculation of Setup
First the automation system calculates settings based on the information about
the initial product and the target product. With these information, the schedule

72
 5.3. Demands and Properties

Figure 5.2.: One cycle for updating the online algorithm. First, the analytical process
models are calculating and predicting the process parameters (1) based
on the information (x) on the product and the corrections received from
the online algorithm. They are used to start the rolling process (2). As
soon as the rolling starts all measurements are collected and sent to the
process system (3). There, the raw values are pre-processed and validated
(4). If they are valid the prediction error is calculated and fed to the online
algorithm which will update the subsequent predictions (5).

73
5. Introduction to Online Optimization

how to roll this product is summarized in the pass schedule. It will contain every
process step. For each of these process steps, the setup is send to the mill and the
basic automation system (1). This may already include a correction term from
the online algorithm.

2. Rolling
With the settings and predictions from the process models, the rolling of the next
pass will start (2). During rolling, the operator or the basic automation system
may adapt settings in case of unexpected situations.

3. Collecting Measurements
During rolling all measurements are send to the measurement pre-processing
system where they are collected (3). The most relevant measurements are roll
force, roll torque, thickness, width and temperatures.

4. Filtering and Validation

From all of these measurements outliers will be removed by simple statistical
filters. One example would be to a apply a filter based on the mean and the
standard deviation. All measurements which are not within the range of µ ± 2σ
are excluded directly (4). Here, µ defines the mean of the process variable and
σ the standard deviation. The predictions for all parameters in Step 1 are only
done on several discrete positions of the product, i.e. three or five discrete length
positions. After removal of outlier, a statistical summary for these positions will
be made. The mean, minimum, maximum and standard deviation are calculated
individually for each position. Therefore, only measurements which are close to
these length positions are taken into consideration. If, e.g., a length position
of 10m was calculated, then only measurements for 10m ± 2m are taken into
consideration.
These values are used by the process models to decide whether the measurements
are valid and can be used for optimization. If they are within a pre-defined range,
the average value will be used for calculation of the prediction error which is sent
to the online algorithm.

5. Update
If necessary, the data will be scaled to defined upper and lower boundaries before
the online algorithm will update the corrections (5) and a new cycle may start.
In general, a filter can also be applied on the receiver side of the online algorithm.
But then the online algorithm should know some details about the process to be

74
 5.3. Demands and Properties

able to detect outliers.

Another difficulty is the selection of variables which are used for regression. This
is a similar and well known problem also for offline algorithms. An introduction
to variable selection is, e.g., given in [GE03]. An additional problem for variable
selection in online algorithms is the time dependency. If a variable shows a time
dependent behavior it is usually not known in advance and is especially not known
for cold-start problems.
Sometimes process experts know, that a certain variable has an influence on the
prediction performance but the level of significance is not known. So process
knowledge is used for variable selection. After some samples have been received
the decision can be re-evaluated based on these samples.

5.3.6. Data Types

Data streams may contain continuous and categorical variables. A categorical
variable usually only has a limited number of possible values. These values may
be numeric but can also contain arbitrary characters. The different values are
called levels throughout this thesis to distinguish them from the continuous, nu-
meric values. Categorical variables are also called discrete variables or qualitative
variables [Agr07]. If the number of levels is two than it is a dichotomous variable.
If it has more than two levels the categorical variable is called polychotomous
[Enc08].
For hot rolling, all categorical variables are assumed to by polychotomous. Exam-
ples of categorical variables for hot rolling are alloy code, coiler number, furnace
number or time information like the working day or the season.
If there is an order within the different levels of the categorical variable, the
categorical is also called ordinal variable. This is usually the case for variables
which are the result of a grouping based on numerical values. Consider e.g. the
categorical variable size of a person with values small, medium and tall where the
grouping is done according to the absolute size. Surely, the categorical can be
seen as ordinal. Basically the same is often done for hot rolling to build groups for
width, thickness or temperature. Although the categorical itself may be ordinal it
does not mean the effect on the optimization problem has the same relationship.
It may be, that wide and narrow products have the same effect on the problem
but medium products have a different effect. If no direct relationship between
the different levels of a categorical variable can be made, the categorical may be
converted to an ordinal variable. But this is not always possible and desirable.
There are numerous methods available to handle categorical variables [PPP17,
CCWA03]. The simplest one is dummy coding, where each level of a categorical

75
5. Introduction to Online Optimization

Table 5.1.: Contrast coding and dummy coding for categorical variables for the cate-
gorical variable Furnace with three levels (F1,F2,F3). Two columns (C1,C2)
have to be added to the dataset to represent this categorical variable.
Coding scheme Furnace C1 C2
F1 -1 -1
Contrast F2 1 0
F3 0 1
F1 0 0
Dummy F2 1 0
F3 0 1

variable is added as a new feature to the dataset and is coded with 1 if this level
is present or 0 if it’s not. In total for c numbers of different levels c − 1 additional
variables are added to the dataset.
An alternative coding scheme is contrast coding. In contrast coding each level of
the categorical is assigned a specific weight and the sum of the weights for each
category has to be 0. The numerical weight expresses the relationship between
the different levels, i.e. the mean value for this group. Usually the coding scheme
is applied in such a way that a chosen referential level will be assigned a value of
-1. But also other codings are possible. A coding example for a rolling mill with
three different furnaces F 1, F 2, F 3 and two possible coding schemes are given in
Table 5.1. Note that only two variables, i.e. C1 and C2, have to be added and
F1 is the reference furnace. The regression coefficient for C1 in case of dummy
coding can be interpreted as the effect of F2 compared to F1. In case of contrast
coding C1 expresses the difference of F2 and F1 to the overall mean. There a also
other coding schemes which can be used. For the case of ordinal variables special
coding schemes like polynomial coding or Helmert coding can be applied [Wen04].
If there is an interaction between the categorical variables and other variables
the previously described coding schemes alone will not be suitable to improve
the predictions. Interaction terms have to be added to the regression formula.
This has to be considered during the design of the algorithm, so a profound
understanding about the process is needed.
For example in the case of roll force optimization in the hot rolling process the cat-
egorical Furnace is considered to be independent of all other dependent variables.
But the categorical variable, which is used for the alloy specification, has usually
a strong correlation to other parameters. For one alloy there might be a linear
correlation between the error of the roll force calculation and the temperature
where another material has a descriptional error in dependence of the deformation
speed. The reason for such differences can be found in the analytical model which

76
 5.3. Demands and Properties

uses different parameter settings to predict the material behavior. Therefore,

interaction terms of alloy specification and the parameters which are describing
the flow curve, i.e. deformation, deformation speed and temperature, have to be
added to the dataset.
A major drawback of the these coding schemes is their extensive memory require-
ment. For offline optimization, because of the features selection and dimensionality
reduction possibility, it may be possible to use this kind of coding schemes. For
online optimization where the number of different levels of such categorical vari-
ables may be very high, this cannot be applied.
Another problem is the lack of knowledge about the different levels of the cate-
gories. Materials not yet observed by the algorithm might be rolled and would
introduce a new level of this categorical. This cannot be handled by most common
algorithms properly.
Instead of adding a new feature to the data, a new instance of the algorithm
can be used on each different level of such categories. A new instance refers to a
new, independent model which only receives data, if a specific combination of all
categorical variables are occurring in the data stream. The memory requirements
will also be very high but there is no need to have all different versions in memory.
The process will know in advance which furnace will be used for the next product
and which material is going to be rolled. These information are used to load the
specific algorithm version prior to rolling. A problem with this approach is that
current process trends which are independent from the categorical variable will
not be recognized by each model instance. This will make drift detection more
complicated. The scenario is depicted in the upper part of Figure 5.3.
Another possibility is to use a baseline model and only use a specific correction
model for each different occurrence of the categorical variables. The baseline
model will consider each drift and the specific models are only compensating the
difference to this baseline. Somehow the idea is similar as the dummy coding
scheme. But instead all numerical variables are also used in the correction models.
This is depicted in the lower part of Figure 5.3. A more detailed description of
this strategy is given in the corresponding algorithm section.
Summarizing, the following different treatments of categorical variables will be
analyzed.
Remove: Removing the categorical variable. The model will then be
trained only with the numerical values.
Numerical: Each level of the categorical variable will be represented by a
numerical value.
Parallel 1: Creating a model instance for each combination of levels.
Parallel 2: Creating a correction model instance for each combination of
levels and use a baseline instance.

77
5. Introduction to Online Optimization

Figure 5.3.: Two variants for handling categorical variables. In the scenario depicted at
the top, for each unique combination of the different levels of each categorical
variable C1 · · · Cp with number of levels l, an own model instance will be
used. This specific model is then trained and used for prediction with all n
numerical variables. In the lower part, a baseline model is trained and only
the deviation to the baseline is handled by the specific model instances for
prediction and training.

78
 5.4. Performance Indicator

For comparisons and also for offline algorithms however, there are additional
possibilities.

Dummy 1: Regular dummy or contrast coding.

Dummy 2: Dummy coding including all interaction terms.

5.4. Performance Indicator

If the main interest is not the exact value of the total performance but the dynamic
behavior, then MSE or MAE described in Section 3 are not always suitable. If
the dataset contains a large number of samples then the deviations at the end of
the dataset will have no huge impact on the performance. Instead, a windowed
version of the MSE and MAE should be used.
For offline algorithms, the data is usually split into three different sets [RH95]. The
training set is used for training of the model, i.e. the initialization. The validation
set which is used to compare different models and to validate the performance.
And the test set which should confirm the measured generalization performance
and is used because of the possibility of over fitting. The training set is usually
a randomly chosen subset of the whole data set and covers a much smaller size
than the test set. The size of the set also depends on the model type and the
total amount of data available.
For validation and comparison of online algorithm performance this kind of
approach is inappropriate. The main field of application for online algorithms are
dynamically changing optimization problems. The selection of randomly chosen
data is not reflecting the real process behavior and cannot be used for performance
evaluations. Only if the process is stable without any kind of change in data the
same procedure as for offline algorithms can be chosen. But then the usage of
online algorithms lose their advantage of being able to adapt to changes.
For a fair comparison of online and offline algorithms the offline algorithms will be
trained on consecutive data only. Otherwise the time dependency of the features
would be averaged out of the data.
For a real-world application the online algorithms are used for both, warm-start
and cold-start optimization problems. The ability to quickly react on observed
errors is also a quality criteria which should be used for proper model selection.
Ideally, the algorithm can react very fast to dynamic changes in the data whereby
being robust in its predictions.

79
5. Introduction to Online Optimization

5.5. State of the Art

Modern hot rolling mills will not achieve a high performance without any kind of
online algorithms. Different strategies have been evolved to achieve best possible
solutions. Several years ago, where data storage was very expensive, the data
collected during rolling was not stored for more than 4-6 weeks. Databases were
also not widely used and statistical analysis was hardly applicable.
Process experts discovered inaccuracies and were usually able to define critical
products, i.e. products where the prediction shows inaccuracies or high deviations.
The usual method to compensate these inaccuracies was to define groups of prod-
ucts where these errors occur. Continuous variables like width and thickness were
subdivided into groups together with an alloy identifier and other classification
variables [FF09].
The usual approach was now to store a sliding average for the n-dimensional
array and to correct the prediction based on the observed errors. The idea of this
approach was driven by multiple reasons. On the one hand side, the complexity
of this approach is very low and correction terms can easily be calculated by
hand. On the other hand side, there was a risk to influence other regions where
no similar prediction inaccuracies could have been observed.
Since cost for data storage, the processing power of the machines and the automa-
tion systems became more complex and capable, databases were used to store
more and more process related data. This enabled the usage of sophisticated
statistical analysis. Consequently, online algorithms were also getting smarter and
more complex. Analysis of certain behavior over several weeks or even month are
now possible and parameters explaining the prediction variance can be detected
more easily. The usage of more complex algorithm like artificial neural networks
(ANN), Bayesian networks [LRT01] and regression algorithms [ÖÖ00, AY01] were
implemented. ANN was and still is a black box algorithm for most people and
it never was used on a very broad basis. However, ANN is applied to several
optimization problems in the hot rolling area but is most often used offline to
improve the general prediction [DSP+ 19, SK08]. Sometimes they are also used as
so-called long term adaption which is characterized by a slow learning rate [FF09].
Long term adaptions have the task to correct almost constant prediction errors.
Common algorithms for long term adaption are moving averages and recursive
regression but also other may be used. Among the other algorithms, ANN is
the most popular algorithm.The easiest approaches for the moving averages are
given in the following. With the residual ek = yk − ŷ k at time instance k, êk+1 as
n−1
prediction error with corresponding weights w, wi = 1 and with window size n
∑︁
i=0

80
 5.5. State of the Art

we get:

Moving average: êk+1 = ek +ek−1 +···+ek−n+1

n
n−1
= 1 ∑︁
n
ek−i .
i=0
wk ·ek +wk−1 ·ek−1 +···+wk−n+1 ·ek−n+1
Weighted moving average: êk+1 = n
n−1
= 1 ∑︁
n
wk−i ek−i .
⎧ i=0
⎨e ,
k k=1
Exponential moving average: êk+1 = ⎩ .
wk · ek + (1 − wk )êk , k > 1

The window size is defining the number of samples which are taken into account.
The proper choice of weights is usually done empirically by process experts. Beside
the correction of long-term trends and effects, there is a strong need to quickly
react on actual deviations. This cannot be guaranteed with most of the long
term approaches alone. Additional algorithms are implemented and so the usual
approach is to have another class of adaptation algorithms. Consequently it is
called short term adaption because it will react instantaneously to actual deviation.
Usually a simple moving average algorithm will be used. In comparison to the
long term adaption, the number of samples n, which are used for calculation, is
much lower. As a consequence the average value will react much faster on higher
deviations. The combination of both adaptation classes is usually additive.

81
6. Datasets for Online Optimization
The concept and methods discussed and presented in this chapter are applicable
for most real-world optimization problems. Although the special focus is on the
complex rolling process, the methods are also tested for other datasets which are
publically accessible. There are three main resources which are used. The Delve
datasets1 hosted by the University of Toronto, the UCI datasets2 with more than
400 different datasets and the StatLib datasets3 . In the following, first a brief
description of the default datasets is given. Afterwards the datasets used in the
field of hot rolling are described.

6.1. Default Datasets

In order to compare the implemented algorithm and validate the implementations
the algorithms are compared to other real-world datasets which are publicly
accessible. Ikonomovska [Iko12] compared different algorithms, mainly regression
trees, on a large variety of real-world regression datasets which can ideally be
used as comparison to our implementations. The datasets are available on her
homepage4 or from the original source. Most of the examples are taken from the
UCI database, the StatLib database or the Delve database. Table 9.1 summa-
rizes the reference datasets used throughout this thesis. Clearly the size of these
datasets is significantly smaller than in usual online applications. The datasets
are also used to discuss the different parameter effects of the corresponding al-
gorithm. The datasets Abalone and Mv delve are also used to show the impact
on the different treatment of categorical variables. The usage of these default
datasets should also emphasize the wide application range of the online algorithms.

1
http://www.cs.toronto.edu/ delve/data/datasets.html
2
https://archive.ics.uci.edu/ml/datasets.html
3
http://lib.stat.cmu.edu/datasets/
4
http://kt.ijs.si/elena ikonomovska/

83
6. Datasets for Online Optimization

Table 6.1.: Overview of samples and data types in the default datasets
Problem Samples Numerical Nominal
Abalone 4 977 7 1
Cal housing 20 500 8 0
Elevators 16 599 18 0
House 8L 22 784 8 0
House 16H 22 784 16 0
Mv delve 40 967 7 3
Pol 15 600 48 0

Table 6.2.: Number of samples and variable types of the rolling datasets.
Problem Samples Numerical Nominal Mill Type Material
Width 93 935 34 1 Roughing Steel
Mill
PlateForce 21 840 8 1 Plate Mill Steel
StripForce 25 721 9 2 Hot Strip Aluminum
Mill

6.2. Rolling Datasets

There are three rolling datasets used for testing various requirements scenarios. A
short description for each dataset is given in the following section. Table 6.2 is
giving a summary of the used rolling datasets.
For all rolling datasets, the task is to compensate the difference between a
measurement from the real process and a prediction made by an analytical model.
The target value for the online algorithm is then the residuals e calculated as the
difference between measurement y and prediction ŷ. The predictions made by the
analytical model and the measurement will not be included in the dataset since
the focus will not be to replace the analytical model, but to improve the residuals.

Width Dataset
Width control and prediction is very complex and has multiple dependencies. The
Width dataset contains approximately 94 000 samples of a steel roughing mill. The
target variable is the width prediction error during rolling. An analytical model
MW , briefly introduced in Section 2.3 was used for the prediction of the width.
Basically a slightly modified version of the Nippon Kokan [TOY+ 82, THW+ 83,
OANY80] model was used. The modification was made to optimize the predictions
for high width reductions and to distinguish between advanced steel grades. This

84
 6.2. Rolling Datasets

model was already tuned so that only the residuals should be compensated online.
The width measurement is taken at the exit side of the mill with an installed
width gauge. The residuals e are therefore the difference between the prediction ŷ
of the analytical model and the measured width y of the installed width gauge.
The dataset consists of 34 numerical features which are describing the geometrical
deformations, temperatures, roll state and the material. Additionally, the furnace
number is included as a categorical variable. For the material description 21 of
the most important chemical elements are used.
The dataset is well suited for the analysis of dynamic behavior and drift. A
common issue in roughing mills is that the roll diameter of the edger rolls are
measured at the wrong position. Edger rolls usually have a conical shape to apply
a certain force towards the roller table. The diameter at the top is therefore usually
a bit larger than the bottom diameter. When the roll diameter is measured at the
wrong position, the automation system will have a bias in the roll gap calculation
and the expected width reduction will not be achieved. This phenomenon is visible
in this dataset, approximately at samples number 40 000. Here, either the assumed
diameter or the conical shape of the rolls are different to the actual used rolls.
After roll change, the width reduction was higher than expected which caused the
measured width after the mill to be more narrow than expected. Somehow the
behavior changed over the next samples and the initial accuracy was reached after
approximately 20 000 samples. The residuals represent only the deviation of the
analytical model. During the collection of the process data, an online model was
used to compensate this deviation.
The categorical variable ”furnace” cannot be used as ordinal variable since each
furnace has an own predictive model associated to. The number of furnaces is
known a priori and can therefore also be used in the traditional dummy coding
schemes. Nevertheless also the performance when converting this variable to a
numerical value is shown.
To get an idea about the target value distribution Figure 6.1 shows the target
value and the trend of the target value. It can be seen that there is a sudden
drop of the target value and afterwards a trend towards zero. Zero in this case
means, that the width prediction of the analytical model and the measured width
are exactly the same. The main goal for the online optimization is to reduce the
variance in this prediction. The standard deviation of the analytical model is
approximately 4.5mm. A standard deviation of below 3mm would be ideal.
Throughout this thesis this dataset will be referred to as Width.

PlateForce Dataset
The PlateForce dataset contains roll force errors for a heavy plate steel mill. It
will be referenced as PlateForce throughout this thesis.

85
6. Datasets for Online Optimization

Figure 6.1.: Scatter plot of the target value. The blue solid line represents the average
value.

The target value represents the predicted roll force error e. The errors are collected
as the difference of measured roll force y and the predicted roll force ŷ in the body
part of the plate, i.e. e = y − ŷ The body part has the big advantage that the
force is usually very stable since the temperature in the body part is also very
stable. On both ends of the plates the temperature losses are higher and therefore
the forces may show larger differences than in the body part. For each plate an
average value of the body will be calculated for each pass. The predicted values
are calculated by an analytical model MF which was briefly described in Section
4.3.
The measured values were pre-processed and filtered so it is assumed that no
measurement errors are present in the data. The dataset has eight numerical
features and one categorical. The categorical variable represents the material
description which is a mapping of the alloy against some target values. There are
12 different levels of the categorical variable within the whole dataset. Since the
data is only an extract of the production, it might be possible that new materials
are rolled. The only way of handling this categorical variable is therefore to remove
it, convert it to a numerical value or add parallel instances of the algorithms. The
numerical values represent the description of the deformation, i.e., geometrical
attributes, temperatures and resistance values. A total number of 21 840 samples
are collected. More than 90 % of the examples contain measurement for one
dominating material. The flow curves are validated for those rolled materials.
The data can therefore be seen as the logical next step in improving roll force
prediction after a flow curve has been found.
Figure 6.2 shows boxplots for the target value, divided into groups according to
the rolled material. It can be seen that the target value has different deviations for

86
 6.2. Rolling Datasets

Figure 6.2.: Target value boxplots for different occurence of the material group for the
PlateForce dataset. High deviations may indicate inaccurate description of
the material.

87
6. Datasets for Online Optimization

the material groups. For some material groups, e.g. the first one, the prediction
is already very good. The mean deviation is approximately zero and the standard
deviation is very small. For other materials, a high mean deviation and much
higher standard deviation can be observed. This should be corrected by the online
algorithm. From Figure 6.2, it can already be seen that neglecting the material
group will probably not be sufficient. The high deviation and large variance
observed for material 10 indicates an inaccurate prediction. This material was
used for thermo-mechanical rolling practice where material is first rolled to an
intermediate thickness and afterwards cooled to a specific temperature. If this
temperature is reached, the rolling continues. This procedure is well established
and is used to achieve specific mechanical properties, especially for pipeline steels
[WZZ16, TWW+ 18].

StripForce Dataset
The previous rolling datasets originate from a single stand mill. The StripForce
dataset is coming from a continuous aluminum finishing mill with four stands. The
target value is the roll force error e which is calculated as the difference between
the measured value y and the predicted value ŷ of the roll force model MF . The
measured value is taken this time from different length positions over the length
for each stand. Usually five different length positions are taken into consideration.
The measured value will be calculated as average value at the corresponding length
coordinate. The predicted value is calculated by an analytical model which is
similar to the one described in 4.3 but was modified and tuned for this type of
mill. The target value does not correspond to the real error which was present
during rolling, since other online algorithms were used to minimize this error.
The data contains nine numerical features and two categorical features. The
categorical variables are the material group and the stand number. 24 different
materials, from soft alloys like pure aluminum to some hard Mg-based aluminum
alloys, have been rolled and are present in this extract. The numerical features
are geometrical values, temperatures and the inter stand tensions. In total 25 721
samples are in this dataset.
In order to see the influence of the different material groups on the target value,
Figure 6.3 shows a boxplot of the target value over all different material groups.
The plot does not distinguish between the different mill stands, because the main
influence is assumed to be correlated to the material.
Obviously, the distribution of the target value is strongly correlated to the mate-
rial group and the removal of this categorical will probably not be beneficial. Such
high deviations would cause severe problems to the rolling process. For the real
process, simple online corrections have been applied to compensate these errors.

88
 6.2. Rolling Datasets

Figure 6.3.: Boxplot for the target value over each material group in the StripForce
dataset. The target value is the difference between the measured roll force
and the prediction made by the roll force model MF . High target values
indicate a poor performance of the prediction.

89
6. Datasets for Online Optimization

6.3. Requirements and Assumptions

After the general description of the demands and properties and the discussion of
the different datasets some limitations have to be set for the development of new
methods and algorithms. The following restrictions have to be considered for the
application of the online algorithms that are introduced in Section 7.1, 7.2 and
7.3. Parts of this list are taken from [BHKP10].
• The dataset consists of a limited amount of columns / features.
• Datasets are received one at a time. They may be pre-processed and used
for batch learning.
• The data consists of some factors with a limited number of levels. The
datatype for all other data are floating point numbers.
• The data stream is endless.
• The number of features is fixed through the entire stream although there
may be some new levels of known factors.
• Algorithm should use a limited amount of memory.
• New samples should be incorporated in a limited amount of time (≤ 100ms).
• Prediction should be available at any time in a limited amount of time (≤
10ms) as long as no optimization is running. A parameter optimization may
be conducted if there will be a scheduled maintenance. During maintenance,
no process prediction is required.
During this thesis, there will not be distinguished between batch learning and
learning of sequential samples one-by-one. If data are arriving in batches, these
batches can be split into single samples and used for training.

90
7. Algorithms
The following chapter will describe three different online algorithms which were
implemented and extended for the hot rolling process. First, a simple first order
algorithm, Passive Aggressive will be introduced in Section 7.1. Afterwards
the popular Recursive Least Squares algorithm is discussed in Section 7.2.
This chapter ends with the discussion of the properties and the extensions made
for the online Support Vector Regression in Section 7.3.

7.1. Passive Aggressive

The Passive Aggressive (PA) algorithm was developed by Crammer et al. [CDSSS03].
It is a simple first order online algorithm. First order is referring to algorithms that
are taking the error directly into account without considering higher order losses.
The algorithm was first developed for binary classification and was extended
afterwards to the regression case.
Let xt = (x1t , x2t , · · · xnt )T define the input data vector at time t of n features
and yt define the target value at time t. Further, let wt = (w1t , w2t , · · · , wnt )T
define the coefficient vector for those features at time t. Then the prediction of yt
is calculated using the dot product as:

ŷ t = wt · xt (7.1)

After receiving the true target value yt and calculating the absolut prediction error,
the coefficients wt are now updated according to the ϵ-insensitive loss function
⎧
⎨0 |w · xt − yt | ≤ ϵ
lϵ (w, (xt , yt )) =
⎩|w · xt − yt | − ϵ otherwise

where ϵ ≥ 0 is a parameter which controls the sensitivity to any prediction error.

The parameters w are updated using the solution of the optimization problem.
1
wt+1 = argmin ||w − wt ||2 s.t. lϵ (w, (xt , yt )) = 0. (7.2)
w∈RN 2

91
7. Algorithms

The detailed solution of this optimization problem is presented in [CDK+ 06] and
yields:

wt+1 = wt + sign (yt − ŷ t ) τt xt (7.3)

with
lt
τt = . (PA)
||xt ||2

Each update of the coefficient vector w would now ensure, that the constraint held
by the actual sample is satisfied. For real-world optimization problems with the
presence of noise and following measurement inaccuracies, this update strategy
is too aggressive and can lead to dramatic consequences. To avoid this problem,
two variants of this algorithm were developed[CDK+ 06] by reformulation of the
optimization problem:

1
wt+1 =argmin ||w − wt ||2 + Cξ (7.4)
w∈RN 2
1
wt+1 =argmin ||w − wt ||2 + Cξ 2 (7.5)
w∈RN 2

The introduced parameter C is the so-called aggressiveness parameter and defines

how aggressive the updates should be. The variable ξ, ξ ≥ 0 is a slack variable
and controls the objective function, i.e. l (w, (xt , yt )) ≤ ξ. The optimization
problem in Equation 7.4 is called PA-I and 7.5 is called PA-II. The update of the
coefficient vector will still be calculated using Equation 7.3, but parameter τt is
now calculated according to:

{︄ }︄
lt
τt =min C, (PA-I)
||xt ||2
lt
τt = 1 (PA-II)
2C
+ ||xt ||2

From Equation PA-I it can be seen that all three variants will perform similar if
the parameter C is chosen to be very high. For PA-I this will already be the case
if C > ||xltt||2 whereas in PA-II the algorithm will continuously tend towards the
PA algorithm with increasing C.
In the remainder of this section the different settings and variants of the PA
algorithm will be compared and their effect on real-world scenarios are discussed.

92
 7.1. Passive Aggressive

7.1.1. Comparison of the Three Different Variants

To compare the performance of all three PA algorithm variants (PA, PA-I, PA-II)
four of the default datasets were used to train the algorithm. Figure 7.1 shows
the MSE over the whole datasets. For PA-I and PA-II the algorithm was used
with various values for the parameter C, but only the algorithm which achieves
the best performance over the whole dataset in terms of MSE is shown.
It can be seen from Figure 7.1 that in general the original algorithm (PA) has
the worst performance. This is obvious because both variants must have at least
the same performance if parameter C is chosen to be very high. The datasets
elevator(Fig. 7.1(a)), Pol (Fig: 7.1(b)) and House 8L (7.1(d)) show no dynamic
behavior and the prediction is almost constant over time. For all shown datasets
the variant PA-II achieves the best performance, followed by PA-I and the original
algorithm. The dataset Cal housing (Fig. 7.1(c)) shows a high dynamic behavior
over the number of samples but also here, the PA-II is superior to the other variants.

7.1.2. Influence of Aggressiveness Parameter

The parameter C for variants PA-I and PA-II is of great importance and has major
impact on the performance of the algorithms. To illustrate its impact, Figure 7.2
will show the prediction of two different parameter settings for each algorithm
type on some artificially generated data. The data was generated according to:

⎧
⎨2.6 + 0.4x + ϵ, x
≤5
y=⎩ , (7.6)
2.6 + 0.2x + ϵ, x > 5

where ϵ is the realization of a normally distributed random variable with mean

zero and variance 0.1. The function in Equation 7.6 will further be referred to
as gap5. Samples were fed to the algorithm in ascending order of x. Therefore
the algorithm suddenly has to face a target value shift at x = 5. For the PA-I,
parameter C was chosen to be 1 and 0.01. During initial learning and also at the
sudden drop in data, the performance with parameter C = 0.01 was worse than in
the case of a C = 1. This can easily be explained because the parameter is defining
the maximum incremental update rate for learning of new behavior. The same
behavior can be seen when looking at the performance of PA-II in Figure 7.2(b).
The only difference is, that in comparison to PA-I, the initial learning is much
faster even for the same Parameter C = 0.01. For PA-II, a small value of C will
cause a small value of τt which is responsible for the update procedure. To analyze
the effect on our datasets, both variances of the PA algorithm were trained with

93
7. Algorithms

(a) Elevators (b) Pol

(c) Cal housing (d) House 8L

Figure 7.1.: Mean squared error (MSE) performance for four selected default datasets.
Figure 7.1(a) shows the elevator dataset, 7.1(b) the Pol dataset, 7.1(c) the
calhousing dataset and 7.1(d) the house8L dataset.

variation of parameter C between 0.001 and 100, focusing on the range between
0.1 and 1.0. Figure 7.3 show the performance on the default datasets Elevators
and Cal housing. The performance of the best and the worst algorithm setting for
PA-I and PA-II is shown. Following values for the parameter C were analyzed:

94
 7.1. Passive Aggressive

(a) PA-I (b) PA-II

Figure 7.2.: Prediction with PA-I ((a)) and PA-II ((b)) on test data according to
Equation 7.6. Two different settings of hyperparmeter C, i.e. C = 0.01 and
C = 1, are shown to demonstrate its impact.

(a) Elevators (b) Cal housing

Figure 7.3.: MSE Performance of both PA variants with different settings for the aggres-
siveness parameter C. Figure 7.3(a) shows the performance on the Elevators
dataset and Figure 7.3(b) shows the performance on the Cal housing dataset.

95
7. Algorithms

Algorithm 4: PA algorithm for categorical variables with parallel algo-

rithm instances.
1 let P be a vector of k categorical variables;
2 let i be the new sample with Pi = (P1 (i), P2 (i), · · · , Pk (i))T ;
3 let S be a set of models M (P ) for each combination of the k categoricals;
4 foreach new sample i do
5 update coefficients w of base model with all numerics in current sample i
based on the chosen PA-variant (PA, PA-I, PA-II);
6 if M (Pi ) ∈ / S then
7 create new instance of algorithm with same variant (PA, PA-I, PA-II);
8 initialize the coefficients w, the tolerated loss ϵ and the aggressiveness
parameter from the base model.;
9 update w of specific model M (Ci ) ;

Values of C: 0.001, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 10, 100

The reason for choosing these values is because the best performance is expected
between 0.1 and 1.0. It can be seen, that the parameter C has a huge impact on
total performance as well as on the dynamic behavior.

7.1.3. Handling Categorical Variables

The usage of different dummy coding schemes as described in Section 5.3.6 for
the PA algorithm is affecting the performance on the datasets depending on the
influence of the categorical variables and its correlation. For online processes
only the parallel algorithm instances, the conversion to numeric or the removal
of the categorical variable is applicable since the different levels are not known.
If they are known in advance, regular dummy coding may also be used. The
pseudo-code in Algorithm 4 describes the algorithm for the case, that parallel
algorithm instances are used. If the parallel algorithm instances should only
compensate the differences to the main model then the handling has slightly to
be adopted. Instead of Step 8, during the initialization, the coefficients are not
copied. Instead, in Step 9 only the residual error from the main model is passed
to the specific model algorithm.
The MSE performance with different handling of categorical variables is exempli-
fied using the datasets Width and StripForce in Figure 7.4. Its very interesting to
see that for the width dataset the performance of the different coding strategies has
only minimal impact on the performance, while the performance on the StripForce
dataset is highly dependent on the coding strategy. For reference, also the classical

96
 7.1. Passive Aggressive

(a) Width

(b) StripForce

Figure 7.4.: MSE over the different datasets and samples with different dummy coding
strategies. The rolling datasets Width 7.4(a) and StripForce 7.4(b) are shown
to illustrate the behavior. For the Width dataset also the strategies for
offline algorithms are shown as comparison. For the StripForce only the
online usable dummy coding strategies Numeric, Remove, Parallel 1
and Parallel 2 are shown.

97
7. Algorithms

dummy coding and the dummy coding with all combinations of other features are
shown in Figure 7.4(a).
It seems, that the categorical variables in the width dataset have almost no or
even a negative impact on the performance so that the relationship between our
target value and the numeric features can be best described with just neglecting
the categorical variables. The reason why the parallel algorithm instances has
the worst performance is obvious. The basic correlation has to be learned by
each of the algorithm instances and if dynamic changes are occurring it has to be
recognized by each instance. This will take much more time than if this change
has only to be learned by one instance. If, on the other hand, some changes occur
specifically within one level of a categorical variable, then this concept can highly
benefit from the implementation.
In the StripForce dataset however, the categoricals and also the correlations be-
tween them and the numerical features have a greater impact. The strategy for
using a correction algorithm for each combination of categorical achieves the
best performance while removing the categorical has the worst performance. The
reason why the algorithm for each combination of level, i.e. the parallel version,
performs worse is based on the fact that this dataset consists of two categorical
variables with 24 and 4 different levels. In total, this will result in 96 different
instances of the algorithm. Since they will occur already at the beginning the
material independent relationship between the numerical variables and the target
value has to be compensated by each single instance before the accuracy improves.
This can be seen when looking at the performance on the first 10 000 samples.
Here, the parallel version achieves almost the worst performance among all dummy
coding strategies. This clearly indicates that this strategy should not be used for
a warm-start application since it will take too long to achieve a certain performance.

7.1.4. Performance
The default datasets were chosen in order to compare the implementation against
publicly available performance achievements. Table 7.1 lists the performance of
the PA variants, an offline linear model and the results achieved by an online
regression tree described in [Iko12]. The best online performance is marked bold.
It can be seen that the PA algorithms are in most cases superior to the offline
version. Furthermore, also the MSE performance is better than the achievements
with online regression trees in four cases.
Result for the rolling dataset are compared with the other implemented online
algorithms after their discussion.

98
 7.1. Passive Aggressive

Table 7.1.: Performance on the default dataset for the PA algorithm variants. As
comparison also the offline linear model and the results achieved by an online
regression tree described in [Iko12] is shown. Only dummy coding strategies
which are applicable are used for comparison. The online algorithm which
achieves the best performance is marked bold.
Problem Lin PA PA-I PA-II IKO
Abalone 8.4 6.5 4.3 4.1 5.7
Cal housing 7.5e9 3.3e9 2.6e9 2.4e9 5.1e9
Elevators 1.0e-5 5.3e-6 5.3e-6 4.9e-6 2.2e-5
House 8L 1.8e9 4.0e9 2.5e9 2.3e9 1.1e9
House 16H 2.1e9 4.4e9 2.7e9 2.7e9 1.6e9
Mv delve 1.5 9.0 8.6 6.6 17
Pol 9.9e2 7.2e3 2.4e3 2.1e3 2.3e2

99
7. Algorithms

7.2. Recursive Least Squares

The least squares method is a very popular method used in many fields for
regression analysis and dates back to 18th century. The main idea of the least
squares method is to estimate the n coefficients βi (i = 1, . . . , n) in the linear
regression formula,

y(t) = β1 x1 (t) + β2 x2 (t) + · · · + βn xn (t) (7.7)

where xt = (x1 (t), x2 (t), · · · , xn (t))T represents the input data at time t. The
coefficients β = (β1 , β2 , · · · , βn )T are estimated in such a way that the squared
sum of residuals for the whole dataset of size N is minimized.
N (︂ )︂2
E (β) = yi − xiT β = (y − Xβ)T (y − Xβ) (7.8)
∑︂

i=1

If Equation 7.8 is minimized with respect to β it follows:

(︂ )︂−1
β̂ = XT X XT y. (7.9)

This solution is only valid if all rows of X are linear independent and thus XT X is
positive definite and the inverse matrix exists. For the calculation of n coefficients
at least n observations of x and y are required. The estimation of β̂ requires the
computation of the inverse of a matrix product and has therefore a high complexity
of O(N 3 ). Furthermore, with each new observation the complete matrix has to
be inverted and with each observation the input matrix X grows. The Recursive
Least Squares (RLS) algorithm computes the coefficients recursively without the
requirement of a matrix inversion. The idea for using a recursive variant should
be illustrated on the following example taken from [Str10].
Example 3. Let us assume we have received 99 values of variable y, i.e. y1 , . . . , y99 .
The average value is m99 = 99 i=1 yi . Now we want to update the calculated
1 ∑︁99

average on arrival of a new sample y100 , without calculating the total sum again.
We can use m99 for the calculation of m100 and get:

99 1 1
m100 = m99 + y100 = m99 + (y100 − m99 ) . (7.10)
100 100 100
This can be seen as an recursive update instruction. It contains the old value m99
and an innovation term which expresses the new information in sample y100 .
This method gains popularity because of its low complexity and is widely used
in the field of control systems [AW94]. Let us assume that β̂ i are the estimated

100
 7.2. Recursive Least Squares

coefficients at time i and that Xi is the input matrix at time i. Then the prediction
at time t is:

yt+1 = Xt+1 β̂ t+1 . (7.11)

With equation 7.9 and with

[︄ ]︄
Xt
Xt+1 =
xt+1

we obtain:

(︂ )︂−1
β̂ t+1 = XTt+1 Xt+1 XTt+1 yt+1
(︄ [︄ ]︄)︄−1 [︄ ]︄
[︂ ]︂ Xt [︂ ]︂ yt
= XTt xt+1 XTt xt+1
xt+1 yt+1
(︂ )︂−1 (︂ )︂
= XTt Xt + xt+1 xt+1 XTt yt + xt+1 yt+1

With equation 7.11,

P−1 = XTt Xt (7.12)

and the matrix inversion lemma [PTVF92]

(︂ )︂−1
(A + BCD)−1 = A−1 − A−1 B C −1 + DA−1 B DA−1 (7.13)

we get:
T
Pt xt+1 xt+1 Pt (︂ −1 )︂
β̂ t+1 = Pt − P β̂ + x y
t+1 t+1 (7.14)
1 + xt+1
T
Pt xt+1 t t

Pt xt+1 (︂ )︂
= β̂ t + y t+1 − x T
β̂ (7.15)
1 + xt+1
T
Pt xt+1 t+1 t

Equation 7.15 represents the determination of the coefficients without the expensive
matrix inversion. The update of the coefficients is expressed as a scalar correction
factor which is multiplicated by an error vector. The only problem is that this
equation still contains the matrix P but this matrix can also be expressed in a
recursive form [You14]:
Pt xt+1
Pt+1 = Pt − xT Pt (7.16)
1 + xt+1
T
Pt xt+1 t+1

101
7. Algorithms

With equations 7.15 and 7.16 all coefficients can easily be updated without taking
into account previous samples. The memory consumption is only dependent on
the size of each sample.
One problem which arises is that the initialization of P is only possible if the
matrix XTt Xt is invertible and this will only be the case after the number of
samples is greater than the number of features received with each sample. A
workaround described in [AW94] is to initialize the matrix P = P0 where P0 is a
positive definite matrix with sufficient large values on the diagonal.
If the process is time variant and the coefficients may therefore be changed
over time it is beneficial to introduce a decay factor or forgetting factor λ with
0 ≤ λ ≤ 1. This factor is interpreted as weighting factor for the new data. While
the actual data are weighted with λ0 = 1, old data is weighted with λi , i > 0. The
matrix Pt+1 and the coefficients β̂ t+1 can then be calculated by:
Pt xt+1 (︂ )︂
β̂ t+1 = β̂ t + yt+1 − x T
β̂ (7.17)
λ + xt+1
T
Pt xt+1 t+1 t

1
(︄ )︄
Pt xt+1
Pt+1 = Pt − T
x Pt (7.18)
λ λ + xt+1
T
Pt xt+1 t+1
A major advantage of the RLS algorithm is its low computational complexity. The
number of features which should be taken into account have to be defined prior to
its usage. Adding or removing of features requires resizing of the matrix P. When
the exponential forgetting factor is incorporated, the algorithm can also handle
time varying effects like concept drift in the data or abrupt changes of features.
However, the forgetting factor λ is normally fixed so the factor should be chosen
as a trade off between taking into account new data and keeping old information.
A further advantage is that the sample data doesn’t have to be stored because
the coefficients can be calculated sequentially.
Several variants of the RLS algorithm were developed. Some of them are are using
different weighting functions, like linear weighting of the last n samples, others
are making other assumptions about the noise term. A very good overview and
detailed analysis over all these algorithms can be found in [You14].

7.2.1. Influence of Exponential Decay

To analyze the impact of the exponential decay λ, Figure 7.5 shows the influence
of the exponential decay factor on sample data generated with the gap5 function
introduced in Equation 7.6.
After half of the total samples have been seen by the algorithm, i.e. at x = 5,
the slope is instantly halved and an offset is introduced. This data is used to
train three different RLS variants with exponential decay λ set to 0.9,0.98 and

102
 7.2. Recursive Least Squares

Figure 7.5.: Influence of exponential decay on test data. Shown is the prediction on
generated data according to Equation 7.17. y denotes the true training sam-
ple, RLS(λ) denotes the prediction of the RLS algorithm with exponential
decay factor set to λ.

1.0. Until the sudden change in data occurs all three RLS variants have almost
the same prediction performance. At the point of the sudden change in data, the
variant with lowest decay (0.9) is adapting to that change much faster than the
other RLS variants. The slowest adaptation is done when no decay is (1.0) is
used.
To analyze the impact on real-world applications, the performance for the rolling
data set width and the default dataset calhousing is analyzed. The performance of
the Width dataset is shown in Figure 7.6(a) and the performance of the Cal housing
dataset is shown in Figure 7.6(b) respectively. Experiments were conducted with
different exponential decays λ from the interval [0.9, 1.0] and the two best results
are shown in Figure 7.6. Some of the experiments had numerical instabilities. This
happens if the arriving data are linearly dependent and the matrix XTi Xi becomes
almost singular. Then, the corresponding matrix P tends towards infinity. To
overcome this problem a proper variable selection is required. Using only a small
set of variables is therefore preferred if a cold-start problem has to be optimized.
The performance for the three RLS variants in the Width dataset shows the huge
impact of the exponential decay. For the decay factor λ = 1, the impact on the
event between samples 35 0000 and 50 000 of the dataset can clearly be seen. If

103
7. Algorithms

(a) Width dataset (b) Cal housing dataset

Figure 7.6.: MSE over the whole dataset for the Width dataset 7.6(a) and the Cal housing
dataset 7.6(b) for different values of the exponential decay. The two best
experiments are shown for both datasets. To demonstrate the sensitivity of
the decay, an additional experiment is also shown for the Width dataset.

lower values are used this dynamic change can easily be corrected and almost no
dynamic effect in the residual MSE will be visible.
The Cal housing dataset on the other hand indicates that there is some dynamic
behavior in the data which can be reduced with a lower exponential decay, but
not completely corrected. Experiments with lower values of the exponential decay
showed instabilities during prediction and are therefore not shown. The choice of
the exponential decay should therefore be made carefully.

7.2.2. Handling Categorical Variables

For the handling of categorical variables, all online methods mentioned in Section
5.3.6 are compared. For the parallel algorithm instances the categorical RLS
algorithm will look for an already existing model for the specific group and will
instantiate a new one if no group has been found. The pseudo code for this case is
depicted in Algorithm 5. The baseline model will always be fed with the available
sample and represents the general correlation. If the parallel algorithm instances
are used to correct a difference to the baseline model, then the initialization in
Step 8 is modified. Instead of copying the coefficients from the main model, the
initialization is done from scratch because it is assumed that the correlations are

104
 7.2. Recursive Least Squares

Algorithm 5: RLS algorithm for categorical variables.

1 let C be a vector of k categorical variables;
2 let i be the new sample with Ci = (C1 (i), C2 (i), · · · , Ck (i))T ;
3 let S be a set of models M (C) for each combination of the k categoricals;
4 foreach new sample i do
5 update matrix P and coefficients β̂ of main model with all numerics in
current sample i;
6 if M (Ci ) ∈ / S then
7 create new instance of algorithm;
8 initialize the exp. decay, matrix P , and estimated coefficients β̂ from
the main model.;
9 update P and β̂ of specific model M (Ci ) ;

very specific for each instance and no general error has to be compensated. Then,
in Step 9, the residual error of the main model is used to train the specific models.
Two of the rolling datasets, i.e. Width and StripForce, had also at least one
categorical variable. Their performance is depicted in Figure 7.7.
The Width dataset contains only the furnace as categorical variable and according
to the performance in Figure 7.7(a) it will have no positive effect on the perfor-
mance. Yet, it will decrease the performance, if it will be included in the model.
This is the reason why the strategy to remove the variable achieves the highest
performance. If the categorical variable is used with all possible correlation terms,
the performance will drop significantly. This is shown within the performance for
the strategy of dummy coding with all possible correlations (Dummy 2). In this
case, the model will have many features. Due to the fact, that the drift in the
data cannot be explained by the categorical variable, the performance will drop.
This is especially visible within strategy Parallel 1, where for each level of the
categorical a separate model instance is used.
The RLS performance on the StripForce dataset is showing a completely different
behavior. Inclusion of the categorical variables in the online algorithm is increasing
the performance significantly. Since the number of levels within the categorical
is very high for this data set, the parallel treatment of those variables will take
some time to learn the basic relationships. Furthermore, if dynamic changes
occur within the main process and not within a certain category, the change
has to be learned by each instance of the algorithms. This can clearly be seen
when looking at the performance development at approximately sample number
9 000. The performance when using a separate version for each occurrence of the
categorical (Parallel 1) is continuously getting worse, while the other strategies
are almost constant (Numeric, Parallel 2) or have a significantly less performance

105
7. Algorithms

(a) Width dataset

(b) StripForce dataset

Figure 7.7.: Mean absolute error (MAE) for the RLS algorithm for different categorical
handling strategies on two different real world datasets. For the Width
dataset 7.7(a) additionally to the online applicable coding strategies (Nu-
meric, Remove, Parallel 1 and Parallel 2), the dummy coding and full
dummy coding strategy is shown. Figure 7.7(b) shows the performance for
the StripForce dataset.

106
 7.2. Recursive Least Squares

Table 7.2.: Performance on the default datasets for the RLS algorithm. On the cate-
gorical variable, only dummy coding strategies which are applicable online
are used for comparison. Shown is the mean squared error (MSE) for the
complete dataset. As reference the performance of a linear model which was
trained on the first 10% of the data is also shown. Column IKO shows the
performance achieved with online regression trees by Ikonomovska [Iko12].
Problem Lin RLS IKO
Abalone 8.4 4.6 5.7
Cal housing 7.5e9 3.0e9 5.1e9
Elevators 1.0e-5 8.9e-6 2.2e-5
House 8L 1.8e9 1.9e9 1.1e9
House 16H 2.1e9 2.0e9 1.6e9
Mv delve 1.5 2.0 17
Pol 9.9e2 1.9e3 2.3e2

drop (Remove). The usage of a base model with separate model instances which
are calculating corrections for the specific material group seems to be clearly the
most favorable solution.
Based on these observations, we recommend to neglect the categorical variables
for cold-start problems. After some samples have been received the impact of the
categorical should be analyzed. Then a decision can be made if the categorical
should be considered in the algorithm or not. Depending on the levels of the
categorical variables a proper strategy can be selected.

7.2.3. Performance
A summary of the RLS performance in terms of MSE on the default datasets
is given in Table 7.2. As a reference, also the performance of an offline linear
model which was trained on the first 10% of the data is shown. Furthermore, a
comparison to the results achieved by an online regression tree from [Iko12] is
made. The comparison on the rolling datasets Width, StripForce and PlateForce is
made after the discussion of the next online algorithm.

107
7. Algorithms

7.3. Online Support Vector Regression

Support Vector Machines (SVMs) have been developed in the late nineties at
AT&T Bell Laboratories by Vapnik and others [SS04] as a nonlinear generalization
of the Generalized Portrait algorithm [VL63]. The Generalized Portrait algorithm,
also originally developed by Vapnik, Lerner and Chervonenkis in the sixties, was
grounded in the statistical learning theory. Most of the literature about SVM is
focusing on classification rather than regression problems with SVMs. Support
Vector Regression (SVR) is considered as a special case of SVMs with engineering
as main application field [SS04].
Cauwenberghs and Poggio[CP01] developed an exact solution for recursively
training of SVMs for classification. Ma and Theiler [MTP03] and Martin [Mar02]
extended this algorithm for the regression case. Although this algorithm would
allow its online usage, there are still some problems which need to be solved and
understand. An important aspect of this is the potential optimization possibility
of the corresponding SVR parameter and the determination of a suitable storage
size.
In this chapter first the idea of standard SVR is presented and afterwards the
incremental and decremental variant of SVR, based on the Accurate Online
Support Vector Regression (AOSVR) Algorithm presented in [MTP03], is shown.
After demonstrating the effect of the storage size and the management of it, two
variants for optimization are introduced and the handling of categorical variables
are compared. The chapter is concluded with recommendations for different
real-world scenarios.
A more detailed introduction to support vector regression is given in [SS04].

7.3.1. Introduction to Support Vector Regression

The support vector algorithm has its origin in the statistical learning theory and
dates back to the sixties but was further developed at AT&T Laboratories [SS04].
Due to this, the main focus has always been real-world applications although the
application field was more in the area of optical character recognition. There are
two variants of support vector regression available: ϵ-SVR and ν-SVR. ϵ-SVR
was originally developed 1995 by Vapnik [Vap95] for pattern recognition whereas
ν-SVR was developed three years later by Schölkopf and Smola [SSWB00].
The main advantage of ν-SVR is the fact, that it eliminates the usage of the
parameter ϵ in case of regression and the regularization constraint C in case of
classification. This is done by reformulation of the optimization problem and
controlling the amount of support vectors with the parameter ν. For real world
applications the target is to establish a best possible prediction. The usage and

108
 7.3. Online Support Vector Regression

optimization of parameter ϵ is of great interest and can easily be interpreted

physically. The number of support vectors used in the algorithm is somehow a
parameter which is hard to translate and therefore the focus in this thesis will be
on ϵ-SVR. A detailed comparison between both variants is given in [CL01].
The main idea for support vector regression is to find a function that fits some
training data xi ∈ RN , i = 1, . . . , l with a maximum deviation of ϵ, i.e. :

f (x) = wT Φ(X) + b. (7.19)

|yi − f (xi )|≤ ϵ, ∀i (7.20)

Function 7.19 can be seen as linear combination of basis functions Φ(X) with
weighting factors w and some additional offset or bias b. The basis function
Φ maps the inputs to a feature space F and is known as kernel function in
literature. The usage of this kernel functions allows the regression to be extended
to nonlinearity. Valid kernel functions should in general be symmetric, continuous
and their correlation matrix should be positive semi-definite. All kernel functions
which fulfill these conditions are called Mercer kernels. Somehow, there have also
been found some kernels which are not satisfying these conditions but still may
be used. A prominent example is the sigmoid kernel with mapping:
(︂ )︂
Φ(x, z) = tanh ϑ + κxT z

Here, ϑ is an intercept constant and κ is the slope constant. It is also possible to

define a linear combination of two kernels because it can be shown that a linear
combination of two Mercer kernels is again a Mercer kernel. Throughout this
section the SVR with linear kernel, polynomial kernel and radial basis function
(RBF) kernel are used. They are defined as follows:

Φ(x, z) =xT z (Linear)

(︂ )︂p
Φ(x, z) = xT z + γ (Polynomial)
−γ∥x−z∥2
Φ(x, z) =e (RBF)

The parameter γ for the polynomial case defines whether the kernel should be
homogeneous (γ = 0) or inhomogeneous (γ > 0). Parameter p defines the degree
of the polynomial. The RBF kernel has a parameter γ which will initially set to
γ = 2σ1
in this thesis where sigma is the variance of the target value. All those
parameters have to be tuned in order to get the best results for each kernel and
each dataset. The optimization problem itself will be the same for each kind of
kernel and therefore we will not specify the kernel in the following discussion of
the SVR algorithm.
Equation 7.19 should approximate our function with an accuracy of ϵ. This means

109
7. Algorithms

Figure 7.8.: ϵ-insensitve loss function for SVR. Prediction errors with deviation less
than ϵ will not be penalized. A deviation above will be penalized linearly
dependent on the parameter C

that every sample which has a prediction error of less than ϵ will not affect the
regression algorithm. This is known as ϵ-insensitive loss which is shown in Figure
7.8. Here, the loss for a deviation ≤ ϵ is zero. Clearly, if the parameter ϵ is chosen
too high, than the approximation would be very poor. A too small parameter ϵ
would sometimes be infeasible because of the presence of noise. Therefore, this
parameter has to be optimized for each dataset. A good starting point can be
selected if the root cause of a random error, e.g. for an physical experiment, is
known. Typical values can also be derived from the measuring accuracy with
which the data have been stored. The regression function should be as flat as
possible which means that the weighting factors w should be as small as possible.
This can be represented as a convex optimization problem:

1 l
{︄ }︄
minimize ||w||2 + C (ξi + ξi∗ ) (7.21)
∑︂
2 i

− wT Φ(xi ) − b ≤ ϵ + ξi
⎧
⎪yi
⎪
⎨
subject to: wT Φ(xi ) + b − yi ≤ ϵ + ξi∗ (7.22)
⎪
ξi , ξi∗ ≥ 0, i = 1, · · · , l.
⎪
⎩

The slack variables ξ and ξi∗ in Equation 7.21 are introduced because the problem
is not always feasible and a solution would not always exist. The constant C
penalizes the data points which lie outside of the desired ϵ band and can therefore
been seen as an optimization variable for our problem. It also defines how flat
our function will be and how big the penalty on positive and negative deviations

110
 7.3. Online Support Vector Regression

greater than ϵ can get. The underlying loss function which was used here is known
as the ϵ insensitive loss function shown in Figure 7.8. Points within a prediction
error of ϵ will have no associated loss while the loss on errors above |ϵ| will have a
loss increasing linearly with parameter C. This is the most typical loss function for
SVR but also other loss functions may be used. In [SS04] the authors are deriving
solutions for other loss functions, like Laplacian or Gaussian loss functions.
After introducing Lagrange multipliers αi , αj∗ and defining Qij = Φ(xi )T Φ(xj ) as
a general kernel function, the optimization leads to the Lagrange formulation:
l (︂ )︂ l l
L = 1
Qij αj − αj∗ (αi − αi∗ ) + ϵ (αi + αi∗ ) − yi (αi − αi∗ )
∑︁ ∑︁ ∑︁
2
i=1 i=1 i=1
l l
− (δi αi + δi∗ αi∗ ) + (ui (αi − C) + u∗i (αi∗ − C))
∑︁ ∑︁
i=1 i=1
l
+b (αi − αi∗ )
∑︁
i=1

Here, δj , δj∗ , uj , u∗j are Lagrange multiplier. If we optimize this Lagrange formula-
tion we get the Karush Kuhn Tucker (KKT) conditions:
l (︂ )︂ l
∂L
= Qij αj − αj∗ + ϵ − yi + b = Qij βi + ϵ − yi + b
∑︁ ∑︁
∂αi
j=1 j=1
l (︂ )︂ l
∂L
=− Qij αj − αj∗ + ϵ + yi − b = − Qij βi + ϵ + yi − b
∑︁ ∑︁
∂α∗i
j=1 j=1
l (︂ )︂ l
∂L
= αj − αj∗ = βj = 0
∑︁ ∑︁
∂b
j=1 j=1
u∗i (αi∗ − C) = 0, ui (αi − C) = 0
δi αi = 0 δi∗ αi∗ = 0

with the simplification βi = (αi − αi∗ ). Analogues to [MTP03] a margin function

h (xi ) is defined
l
h(xi ) = f (xi ) − yi = Qij βj − yi + b (7.23)
∑︂

j=1

and each sample can be assigned to one out of three subsets.

⎧
i) ≥ ϵ, βi = −C
⎨h(x
Error Set:
⎩h(xi ) ≤ −ϵ, βi = C
⎧
i)
= ϵ,−C ≤ βi ≤ 0
⎨h(x
Support Set:
⎩h(xi ) = −ϵ, 0 ≤ βi ≤ C

{︂
Remaining Set: −ϵ ≤ h(xi ) ≤ ϵ, βi = 0

111
7. Algorithms

Example 4. To illustrate the SVR principle an easy example is shown in Figure

7.9. Samples were created using a linear combination of a sine function with a
linear dependency and some random noise σe , i.e.:

f (x) = 0.3sin(5x − 2) + 0.4(x − 1) + σe .

The true function is drawn without noise and is depicted as blue solid line. 20
samples which are equally distributed along the x-axis are used for training and
marked green. After training of the SVR - we used here already the online SVR
algorithm without loss of generality - the prediction of the SVR along the x-axis
is shown as red solid line. The maximum allowed deviation to this prediction is
shown as a dashed solid line. The margin functions for support vectors have to be
exactly ϵ or −ϵ and are exactly on the dashed lines. These samples are marked
”S” for Support Set Samples. Samples outside the desired ϵ band belong to the
error set and are therefore marked ”E”. All other samples lie in the ϵ band and
are belonging to the remaining set. Note that these samples have no influence on
the prediction because their weight is zero.

A more detailed introduction to SVR can e.g. be found in [VGS97].

7.3.2. Incremental Support Vector Regression

The standard SVR requires to solve a convex optimization problem which might
take too much time for the usage within an online system, where a continuous
stream of data is received. A possibility to use this offline version of the SVR is
to window the dataset and to calculate the solution every time from scratch. But
this is only feasible with a very small amount of data and dependable variables.
For larger datasets and fast incorporation of new samples an iterative solution is
needed.
The first incremental algorithm for SVMs for classification was developed by
Cauwenberghs and Poggio [CP01]. This algorithm was further developed for
the regression case by Martin [Mar02] and Ma and Theiler [MTP03]. The main
difference to standard SVR is its incremental and decremental update mechanism
to calculate the support vector and the corresponding sets. Regularly, the quadratic
equation given in Equation 7.21 has to be solved for the whole data set. Incremental
SVR ensures that with each new sample which is added, the KKT conditions are
still fulfilled for all samples rather than computing everything from the beginning.
Therefore, a suitable value for the βc for a new sample c has to be determined.
Also, while changing the coefficient βc for the new sample, all other coefficients
and the offset b must be updated. With ∆b, ∆β and ∆h specifying the change of

112
 7.3. Online Support Vector Regression

Figure 7.9.: Prediction example to illustrate the SVR principle. The blue solid line
is the true function and the green dots represent the samples which were
created based on the true function with additional noise added. The red
solid line is the prediction of the SVR after training of all examples. The
dashed lines represent the maximum allowed deviation ϵ. Every sample
which is directly on the dashed lines belong to the support set (S). Samples
with higher deviations are belonging to the error set (E), samples with lower
deviation than ϵ are belonging to the remaining set (R).

113
7. Algorithms

the offset, the coefficients and the margin vector and

l
βj = 0, (7.24)
∑︂

j=1

we get:

∆h(xi ) = Qic ∆βc Qij ∆βj + ∆b. (7.25)

∑︂

j∈S

For samples which are in the support set S, h(xi ) will only change if the samples
migrate to another set. Therefore for samples which stay in the support set
h(xi ) = 0 and it follows ∀i ∈ S:

Qij ∆βj + ∆b = −Qic ∆βc (7.26)

∑︂

j∈S

∆βj = −∆βc (7.27)

∑︂

j∈S

With
⎤−1
0 1 ··· 1
⎡

⎢1 Qs1 ,s1 ··· Qs1 ,sns ⎥

⎢
Q=⎢
⎢ .. .. .. ⎥ (7.28)
⎥
... ,
⎣. . . ⎦
⎥

1 Qsns ,s1 · · · Qsns ,sns

and s1 · · · sns specifying all support vectors it follows that:

∆b 1
⎡ ⎤ ⎡ ⎤
⎢
∆β ⎥
s ⎥
⎢
Q
⎢ s c⎥
⎥
⎢ . 1 ⎥ = −Q ⎢ . 1 ⎥ ∆βc (7.29)
⎢
⎢ . ⎥ ⎢ . ⎥
⎣ . ⎦ ⎣ . ⎦
∆βsns Qsns c

For samples within the Error or Remaining Set, h(xi ) will change according
Equation 7.25. ∆βj and ∆b can be calculated using Equation 7.29 and ∆h(xi )
can therefore be easily calculated.
Up to now only the case where all samples will stay in their corresponding set was
considered. The idea for the migration of samples is, that if ∆βc is too high and
samples would migrate to another set, than the change of the weight is done in
multiple steps. The maximum allowed change ∆βcmax will be calculated such that
all samples will just stay in their sets. Therefore for each sample a corresponding
maximum allowed change of their weight has to be calculated.

114
 7.3. Online Support Vector Regression

Migration of Samples in Support Set

Samples within the Support Set have a deviation of ±ϵ and a weight of 0 < βi < |C|.
The samples are migrating to the remaining set if their weight is approaching 0.
They are migrated to the error set if βi gets C. With Equation 7.27 and

1
⎡ ⎤

⎢ Qs1 c ⎥
⎢ ⎥
ϑ = −Q ⎢
⎢ .. ⎥
⎥ (7.30)
⎣ . ⎦
Qsns c

it follows from equation 7.29 that:

⎪− ϑi , sign (ϑ∆βc ) > 0, h(xi ) = ϵ

⎧ β
⎪ i
⎪
⎨− C+βi , sign (ϑ∆βc ) < 0, h(xi ) = ϵ
⎪
⎪
max∆βc = ϑi
(7.31)
⎪
⎪
⎪
⎪
C−βi
ϑi
, sign (ϑ∆βc ) > 0, h(xi ) = −ϵ
⎩ βi
− ϑi , sign (ϑ∆βc ) < 0, h(xi ) = −ϵ
⎪

Migration of Samples in Remaining Set

Samples from the remaining set have a deviation h(xi ) < |ϵ| and can only migrate to
the support set. This will happen when the deviation h(xi ) changes to h(xi ) = ±ϵ.
With {r1 . . . rn } representing the samples in remaining set, {s1 . . . sn } representing
the samples in support set, c representing the actual sample and

1 Qr1 s1 · · · Qr1 sn
⎡ ⎤ ⎡ ⎤
Qr1 c
⎢ Qr2 c ⎥ ⎢1 Qr2 s1 · · · Qr2 sn ⎥
⎢ ⎥ ⎢
⎢ .. ⎥ + ⎢ ..
γ=⎢ .. .. ⎥ (7.32)
⎥
⎥ ⎢
... ⎥ϑ
⎣ . ⎦ ⎣. . . ⎦
Qrn c 1 Qrn s1 · · · Qrn sn

the maximum allowed change of ∆βc can be calculated before a sample migrates
to the support set:
⎧
⎨ ϵ−hi , sign(γ∆βc ) > 0
max∆βc = γi
(7.33)
⎩ −ϵ−hi , sign(γ∆βc ) < 0
γi

Migration of Samples in Error Set

Samples within the error set have a deviation h(xi ) > |ϵ| and can only migrate to
the support set. The calculation of the maximum allowed change is similar to the

115
7. Algorithms

Algorithm 6: online SVR algorithm.

1 let S be the a vector of samples in Support Set, E a vector of samples in Error
Set, and R be a vector of samples in Remaining Set ;
2 let c be the new sample which should be incorporated;
3 calculate margin (error) hc for current sample;
4 if hc ≤ ϵ then
5 add sample c to remaining set and exit
6 while c not added to S or E do
7 calculate minimum ∆βc until sample c migrates to set S or E;
8 foreach sample i in R do
9 calculate minimum allowed change of βc until sample i migrates from R
to S;
10 foreach sample i in S do
11 calculate minimum allowed change of βc until sample i migrates from S
to R or E
12 foreach sample i in E do
13 calculate minimum allowed change of βc until sample i migrates from E
to S
14 set ∆βc as minimum allowed change for all stored samples;
15 change weight of current sample c by ∆βc and migrate samples if necessary.
update weight β for all samples and calculate new margin h;
16 if min ∆βC was determined by sample c then
17 add sample c to E or S

remaining set and we get:

⎧
⎪ −ϵ−h
⎪
⎨ γi
⎪
i
, sign (γ∆βc ) > 0, h(xi ) < 0
max∆βc = hi −ϵ
, sign (γ∆βc ) < 0, h(xi ) > 0 (7.34)
⎪ γi
otherwise
⎪
⎩∞,
⎪

Then the sample, which defines the minimum change, will be migrated into another
set and the weight βc is further changed until all samples meet the KKT criteria.
A pseudo-algorithm of this procedure is shown in Algorithm 6.
A detailed description of this procedure can be found in [Mar02, MTP03, CP01].
The removal of a sample is done in the same way: The weight of the specific
sample is reduced in multiple steps until it is zero. If it is already in the remaining
set it directly can be removed.

116
 7.3. Online Support Vector Regression

Figure 7.10.: Training time in ms over sample size for incremental SVR. The storage
size is assumed to be indefinite and no sample was removed.

Storage Management
The incremental behavior for the SVR algorithms enables it for applications
where the time for a complete training cannot be afforded. For incorporation of a
new samples, the algorithm is not forced to learn from scratch. So the expected
time to incorporate a new sample is significantly reduced. When a new sample
arrives, it has to be guaranteed that all samples in storage still fulfills the KKT
condition, while the weight of the new sample is incrementally modified. So the
time is strongly correlated to the storage size of the SVR. Figure 7.10 shows the
training time of an industrial problem with 35 dependable variables, although the
computational power of industrial machines are much higher than on notebooks.
This test was done on a regular i5 notebook and should only give an indication
on training time.
The plot shows, that the training time continuously increases but also the
variation of training time is increasing. For this dataset and parameter setting,
up to 4 000 samples can easily be integrated with training times below 100ms.
Above this value, the variation of the training time is too high, which might lead
to infeasible computing times. It can also be seen that some of the samples have
training time of almost 0. This is the case for samples which are predicted within
the tolerance of ϵ and can directly be added to the remaining set.
Let us assume a storage limit of 4 000 and that already 4 000 samples have been

117
7. Algorithms

received. For the next sample the algorithm has to decide, which samples are
stored and which will not be considered anymore. This can be achieved with
the decremental behavior of the online SVR algorithm which will work similar
to the incremental behavior. The weight of the sample, which should be deleted,
is continuously decreased while all other samples still have to fulfill the KKT
condition. If the weight is zero, it can be removed.
The question is now how to select the proper candidates which should be stored
and how to decide which should be removed? Cauwenbergs et al.[CP01] suggest
to only take so-called ”reserve” vectors into account which have a high chance to
migrate to support vectors. But this would still only reduce the memory usage
and not limit it within an infinite dataset. Another issue is caused by the dynamic
behavior of the optimization problem. Actual samples with high prediction errors
might end-up in the error set with no or only small chance to migrate to support
vectors. But they are reflecting at best the current process with all its states. In
[TL03] for statically incremental support vector machines also the least relevant
values will be kept instead of always removing the last one.
Four different strategies to select the samples which should be removed are
considered. The most intuitive solution would be to balance the usage of new
samples and old ones in order to adapt to new situations very fast without loosing
the capability for the prediction of rare situations. In general, also the storage
size should be considered. In detail, the following strategies are used:

118
 7.3. Online Support Vector Regression

1. Oldest: The strategy Oldest is just removing the oldest sample in the
storage and is based on the idea, that the best description of
the actual process state is obtained by the most
recent samples. This will ensure that the algorithm will always
adapt to new situations.
2. Distance: The distance based strategy (Distance)will compare on arrival
of a new sample the Euclidean distance to the stored samples.
The sample with the closest distance will be removed first and
afterwards the new sample will be added. To ensure that the
latest sample are not removed on arrival of the next sample,
the most recent 30 % of the samples are protected. The idea
behind this strategy is to balance between rare situations and
dynamic changes.There will be additional costs for the
distance calculation or the calculation of the least promising
candidate which should be considered when parameters like
storage size are chosen.
3. Random: Strategy Random will just remove a random sample from the
dataset.
4. Reserve: When using the strategy of removing the least promising
candidate (Reserve), the possibility for migration of samples
belonging to the remaining set or error set will be considered.
The sample with the lowest chance to migrate into the support
set will be removed.

There will be additional costs for the distance calculation or the calculation of
the least promising candidate which should be considered when parameters like
storage size are chosen.

7.3.3. Stability
Stability is a central aspect for real-world applications. It is essential that the
algorithm will not show any side effects or unexpected behavior. There are certain
situations where the iteration might become unstable, e.g., if duplicate samples or
negative eigenvalues occur or if the process limits are violated. These problems
will be discussed in the following.

119
7. Algorithms

Duplicates
The first reason for instability which should be considered are duplicate samples.
Duplicate samples may cause the iteration process to run into an endless loop.
Assume a sample which is already stored in one of the three sets is fed again to the
online algorithm. If the stored duplicate is not in the remaining set, it will have
a weight unequal to zero. But this would mean, that the new arriving sample,
which starts with weight zero, will also not belong to the remaining set since it
will have the same prediction error as the already evaluated duplicate. The new
sample will start with a weight of zero and has to incrementally update its weight.
With steps 7 through 15 of the Algorithm 6, both weights will be incrementally
updated and the iteration process may oscillate. Duplicates in hot rolling may
occur if the online process is not properly designed. An example of such a scenario
can be seen if the measurement values are send twice to the process which triggers
the online algorithm.
The implemented online SVR algorithm will check each new sample and will
reject any duplicate. The check is done based on the euclidean distance of the
new sample to the set of stored samples. A disadvantage of this check is that
also updates of the error are not possible since the check is only based on the
independent variables. However, this feature can easily be implemented. The
stored sample can be removed and afterwards the algorithm will be fed with the
updated version.

Negative Eigenvalues
Laskov et al. [LGK+ 06] demonstrated that the phenomena of immediate cycling is
only possible if the kernel matrix is not positive semi-definite. Immediate cycling
refers to a situation where a sample migrates from one set to another and will
immediately migrate back to its origin within the next iteration.
To prevent the kernel matrix from becoming not positive semi-definite, a small
value, i.e., a regularization constraint, is added to the kernel matrix. The con-
straint should express the highest possible negative eigenvalue and will be added
to the diagonal of the kernel matrix. Since it is not known how high this value
might get, two situations are addressed within the code. The first one regularly
introduces a small regularization constraint for the kernel matrix calculations.
The second one is detecting problems during iterations. The detection is basically
an iteration counter which assumes a maximum number of possible migrations
during the procedure. Once such a situation was detected the algorithm will
recalculate the complete kernel matrix and add a higher regularization constraint.
For some cases even this will not help because there is another possible reason for
an instability which can be found in the numerical representation of the data.

120
 7.3. Online Support Vector Regression

Figure 7.11.: Histogram of the target value from the House 16H dataset.

Double Precision
Every system has certain precision boundaries. The programming languages
R and C++ are both using the norm IEEE 754 which defines the standard for
floating point computations. It is important that both are using the same standard
because otherwise data would already be lost or precision would be wasted at
the interface between both languages. To be able to properly use the online SVR
algorithm, the data is scaled in advance as described in Section 5.3.5. The scaling
limits are either set by a process expert or, in case of a warm-start usage, based
on the samples already received. If the samples have the same magnitude and
ranges the scaling procedure can also be skipped. The question is now what would
be the best way to scale the data? A usual naive approach would be just to scale
them to minimum and maximum. This would lead to a range between zero and
one or minus one and one. If the data contains extreme outliers, this scaling
method can cause unexpected effects. Consider for example the distribution of
the P1 value from the House 16H dataset. Figure 7.11 shows the distribution of it
with a logarithmic x-axis. The minimum value of P1 is zero and its maximum is
approximately 5.5e5. When the scaling to minimum and maximum is applied here
most of the values fall in the range between 5e-5 and 6e-4. This is unfortunate
for the algorithm and the precision ranges can easily be exceeded. Clearly, if this
information is known in advance a different scaling to this value should be used.
The online SVR approach is vulnerable to floating point errors. A huge number
of computations have to be made and the scaling of the input and output values
is of great importance. With an improper kernel choice and scaling the double
precision limit can easily be violated. This has to be prevented. Therefore the

121
7. Algorithms

recommendation is not to scale the values between zero and one but to scale
according to specified percentiles. A profound and robust choice is to scale to the
first and third quartile, i.e., the 25th and 75th percentile.

7.3.4. Parameter Influence

Regular SVR performance as well as the incremental SVR performance are
strongly dependent on the specific parameters of the algorithm. These are ϵ for
the maximum tolerable error, C which describes the penalty term and the kernel
parameter(s). While the traditional offline SVR algorithm can directly optimize
the parameter for any given dataset, the online algorithm lacks this possibility.
Strategies which are successfully applied for optimization, e.g., cross-validation,
can not easily be transferred to the online variant. The dynamic behavior of
datasets would be completely neglected for traditional cross-validation.
For cold-start optimization problems, where no data have been seen, a good initial
parameter set is hard to find. There are some parameter settings which can be
seen as ”default settings” but are definitely not the best parameters for every data
set. This is one of the consequences of the No-Free Lunch theorem [Haf16]. Once
the parameters have been chosen there is no strategy for switching over to new
parameters as this would require a completely recalculation of every single data
point.
The initial parameter settings for all considered SVR algorithms throughout
this thesis were optimized prior to their usage. The algorithms were used with
optimized parameter settings in dependence on their storage size. This means, that
an online SVR algorithm with a storage size of 200 samples was optimized with
parameters that achieve the best performance on the first 200 data. For higher
storage sizes more samples were used, i.e., it was assumed that the algorithm will
process at least the number of samples corresponding to the maximum capacity
of the storage. The parameter optimizations were performed with the R package
nloptr [Joh18] with parameters shown in Table 7.3.
The initial parameters for γ and the penalty C were chosen to be 1 since it turned
out to be a reasonable start value for most of the datasets. The value for ϵ is set
to represent the half of the standard deviation of the target value.

7.3.5. Influence of the Storage Size and Storage Management

The sample storage for the SVR is a central aspect and will significantly influence
the performance in terms of accuracy but also properties like time needed for

122
 7.3. Online Support Vector Regression

Table 7.3.: Parameters passed to the optimization function. The package nloptr[Joh14]
was used to conduct the optimization.
Name Value Description
A variant of the Nelder-Mead algorithm
which is operating on sequences of sub-
algorithm NLOPT LN SBPLX
spaces. It was originally developed by
Tom Rowan [Row90].
Start values for insensitivity loss ϵ,
x0 (0.5σ, 1, 1)
penalty C and RBF kernel parameter γ.
lb (0.0001, 0.0001, 0.0001) Lower bound for (ϵ, C, γ).
ub (0.5, 10, 10) Upper bound for (ϵ, C, γ)
maxeval 50 Function evaluations for nloptr

learning and ability to predict rare process states. Mainly, two parameters
determine these properties: Storage size Ns and the storage management strategy
Ms . The storage size Ns defines the maximum number of samples which may be
stored internally and the management strategy Ms defines a way to select samples
which should be removed once this maximum is reached. Both parameters are
discussed in the following.

Storage Size
The storage size Ns within the online SVR determines how much samples are
stored and therefore also how much memory of the previous measurements is
kept. It will have a great impact on both, performance and run-time. Special
materials may be rolled only on rare occasions and keeping them in memory would
probably increase the performance. But since the system might be very dynamic
this cannot be guaranteed and an adaptation to the new learned samples can
be favorable. With a proper choice of the storage size the process experts have
therefore to balance between fast adaptation and a more robust performance but
slower adaptation.
The choice of the storage size will further highly correlate with the training time
which can be seen as a limit to the storage. The performance for the online SVR
with storage sizes of 200, 500 and 1 000 is depicted for two rolling datasets in
Figure 7.12. On the left side the performance for the Width dataset is shown, and
on the right side the performance of the StripForce dataset is shown. Performance
in both cases is the mean absolute error (MAE). For industrial applications, the
MAE is favorable because of its interpretability. The algorithms were used without
considering the categorical variable. To have a fair comparison, all versions are
using the same kernel (RBF) and the same strategy for the deletion of samples

123
7. Algorithms

(a) Width (b) StripForce

Figure 7.12.: Mean absolute error of the roll force for the two rolling datasets Width
and StripForce with three different storage sizes. The categorical variables
were not considered (removed) and the RBF kernel was chosen. Further,
the oldest sample was removed (Oldest).

(Oldest).
It can be seen that the performance in terms of MAE is increasing with a higher
value of storage size. For the first part of the StripForce dataset, it seems that the
performance of the SVR with 500 and 1 000 samples is almost the same. Only
after a while they are continuously drifting away from each other. The dataset has
no high dynamic and almost all of the changes are occurring because of material
changes.
The Width dataset has a strong dynamic impact at approximately sample 40
000 (see Figure 6.1). Here, the performance of the SVR with 200 samples will
be reacting faster than the variants with more storage. The curves will get
closer together. In order to proof this, Figure 7.13 shows a the mean absolute
performance over a fixed windows size for 5% of the data. Therefore, the total
achieved performance at the end will also only present the last 5% of the data but
the dynamic behavior can be analyzed much better. Between samples 30 000 and
50 000, it can be seen, that the algorithm with storage size of 200 is achieving an
even better performance than both other version at this point of the dataset (see
Figure 7.13(b)), although the improvement is only marginal.

124
 7.3. Online Support Vector Regression

(a) Full dataset (b) Samples 30 000 to 50 000

Figure 7.13.: Windowed mean absolute error of the Width dataset. The values are
representing the mean MAE of the last 5% of the data. This is done in
order to analyze the dynamic behavior of the data. Figure (a) shows the
full dataset and Figure (b) focuses on samples 30 000 to 50 000

Storage Management Strategy

The storage management strategy Ms is deciding which samples should be kept
and which should be removed. This might be a simple choice a posteriori, but the
decision is difficult if no information about future samples can be retrieved. Storing
really rare samples will not be beneficial if this working area is not contained
in some future samples. However, storing only similar samples is also not very
promising when sudden changes occur. To analyze the impact on the selection of
the strategy, Figure 7.14 depicts the performance for rolling datasets Width and
StripForce in terms of mean absolute error. All categorical variables have been
removed and only the RBF kernel performance is shown. Storage size was chosen
to be 200, 500 and 1 000. Only the best performance for each strategy is shown.
For the Width dataset in Figure 7.14(a) the strategies to remove the oldest and
to remove a random sample are performing almost identical. The distance-based
strategy and the strategy to keep the reserve vector are slightly worse. The reason
for this is the high dynamics within this dataset. All strategies which are keeping
old samples in the dataset would cause a decrease in performance since they will
not be suitable to describe the current state. A different behavior can be seen in
Figure 7.14(b). The dataset contains some dynamic behavior but this is mainly
caused by the categorical variable. To remove this effect, a special instance of
the algorithm was used to correct the behavior to the main model (Parallel).

125
7. Algorithms

(a) Width (b) StripForce

Figure 7.14.: Mean absolute error of the width (a) for the Width dataset and the rolling
force (b) for the StripForce dataset with four different storage management
strategies. The decision which sample should be removed from storage is
either done randomly (random), based on the sample distance (distance),
the age (oldest) or on the possibility to migrate into the support vector set
(reserve). The categorical variables were removed for the Width dataset.
For the StripForce dataset a separate model instance was used (Parallel 2).

126
 7.3. Online Support Vector Regression

Therefore, the algorithm can benefit from storing some more rare sample rather
than deleting them. The total mean absolute error of the distance based strategy
has the best performance, followed by the random deletion of samples. Remov-
ing the oldest sample or keeping the reserve samples achieve the worst performance.

7.3.6. Selecting the Kernel Function

The kernel selection is a very difficult choice since it usually requires some knowl-
edge about the data structure and the distribution of it. If those information are
available prior to application, there are some general rules which can be taken over
from regular SVR. Linear kernels and polynomial kernels may suffer from some
numerical difficulties, especially if the kernel parameters are chosen inappropriate.
The selection of an adequate degree of the polynomial, i.e., p,(︂is a crucial
)︂p issue. A
too large degree can cause the polynomial kernel, Φ(x, z) = x z + γ , to go to
T

infinity if xT z + γ > 1. This is also another reason why scaling is so important

for kernel methods.
RBF and polynomial kernels are highly dependent on the settings of the hyper
parameter of the kernel. The linear kernel does not need to be parameterized.
The RBF kernel and the homogeneous polynomial needs tuning of one parameter
and the inhomogeneous polynomial requests two parameter to be tuned.
The computational complexity of the online SVR algorithm will be highly depen-
dent on the kernel choice. The kernel choice will define which correlations of the
dependable variables can be taken into account for the optimization. The linear
model allows only a linear separation in hyperplanes while polynomial of order
d > 1 and RBF kernel allows the features to have correlations of higher order. So
ideally, the choice would be to choose the kernel with the lowest computational
complexity, which satisfies the desired performance.
The authors of libsvm software package advise to use the linear kernel instead of
the RBF or polynomial kernel when the number of features is much higher than
the number of instances [HCL16].
Figure 7.15 shows the MSE performance over the whole datasets for the rolling
dataset PlateForce and StripForce. The polynomial kernel was (︂
a polynomial
)︂5
of
degree five with the linear offset set to one, i.e. Φ(x, z) = x z + 1 . For the
T

rolling datasets it seems that the RBF kernel has a very good performance. The
kernel choice is closely related to the optimization problem and therefore it cannot
be assumed that this behavior will always be the same. Interestingly, the kernels
behavior at the beginning of the dataset is quite different. Within the Width
dataset the polynomial kernel has huge inaccuracies at the initial stage of the
prediction. This is again something which is on concordance to our expected per-

127
7. Algorithms

(a) PlateForce (b) StripForce

Figure 7.15.: Performance for datasets PlateForce and StripForce for different kernel
choices. All dummy variables have been removed from the dataset. All
other parameters are varied as described in previous sections. Only the
best performing kernel algorithm is shown for each dataset.

Table 7.4.: Percentage of training time which was spent in the corresponding kernel
calculation routine.
Kernel Function Consumption[%]
(︂ )︂
Linear Φ(x, z) = xT z 11
(︂ )︂5
Polynomial Φ(x, z) = xT z + 1 56
RBF Φ(x, z) = e−∥x−z∥ 12
2

formance and closely related to the extrapolation difficulties. This will separately
be addressed in Section 9.1. Beside the CPU power and the number of selected
features, the kernel choice has also a great impact on the training time. The kernel
choice is significantly consuming a great chunk of the whole training time for each
individual sample. Table 7.4 shows a profiling result for all three different kernel
selections with a storage size of 2 000 samples on the width problem with five
features. The percentages are representing the amount of time for the training of
one sample which was used for the calculation of the kernel products.
This means, that the choice of the kernel should be chosen carefully, especially
for time critical applications and large storage sizes.
Based on the previous analysis, the best initial choice of the kernel function
seems to be the RBF kernel. It offers a robust initial prediction accuracy with

128
 7.3. Online Support Vector Regression

a moderate complexity in terms of time consumption. After an initial phase of

learning the kernel choice has to be rethought. This can be done by evaluating the
performance of other kernels on the already seen data. After the discussion of the
different treatment of the categorical variable, the training time will be discussed.

7.3.7. Handling Categorical Variables

For real-world applications and especially for hot rolling it is important to have
strategies for handling categorical variables. One of the major problems is the
uncertainty regarding the number of levels for the categorical variable. Following
strategies can be applied for the online SVR algorithm:

1. Removing categorical
2. Converting to a numeric value according to its level
3. Creating a separate algorithm instance for each unique combination (two
variants)
The first two strategies can directly be implemented. The separate model instances
are more complex to implement than the simpler algorithms PA and RLS. While
the separate model instances in case of PA and RLS are initialized by copying the
coefficient of the main model, the SVR algorithms lacks this feature. Copying of
the whole stored matrices and vectors would be applicable but this will require too
much time. The amount of memory which has to be copied is dependent on the
number of training samples stored in the main model. Instead of initialization with
parameters from the base model, the new algorithm instance will be initialized
without any knowledge about the latest samples. Furthermore, the usage of a
base model which is corrected by the special models will also be considered.
Figure 7.16 shows the MAE on the rolling datasets StripForce and PlateForce.
All algorithms were using the RBF kernel and the oldest strategy for removing of
samples and were using a storage size Ns of 200.
Especially the performance for the StripForce dataset in Figure 7.16(a) is highly
dependent on the strategy how to handle categorical variables. As already pointed
out in Section 6.2 the target value showed strong correlations to the material
which was rolled. Therefore, it is not surprising that the strategy to use an own
model instance for each categorical variable achieves the best performance. The
PlateForce dataset shows a different behavior. The performance is depicted in
Figure 7.16(b). At the initial learning the separate model instances show the best
performance but the average performance over the whole dataset is almost the same

129
7. Algorithms

(a) StripForce (b) PlateForce

Figure 7.16.: Mean absolute error for datasets StripForce and PlateForce for different
handling of the categorical variable. For comparison, only RBF kernel was
used. The oldest samples have been removed. For both algorithms 200
samples have been stored (Ns = 200).

for all strategies. The worst performance is clearly achieved when a base model is
corrected by the specialized model for each material (Parallel 2). The reason for
this might be that the SVR parameters for the specialized model instances are not
optimized. They are just copied from the base model. The base model parameter
are optimized on the first samples as described in Section 7.3.4. Since the com-
plexity and the memory requirements are higher if more algorithm instances are
used the best choice would be to remove the categorical variable. Unfortunately,
this behavior might change quickly, if new materials are rolled. This risk has to
be considered when selecting the strategy for the treatment of categorical variables.

7.3.8. Training Effort

The main idea of the online SVR is to incrementally update parameters instead
of solving the convex optimization problem directly. Therefore it is not surprising
that the time for incorporation of new samples is much faster than calculating
the solution for the complete dataset. One of the first papers which describes the
online SVR method shows also a comparison between the libsvm package and
their implementation of the online SVR algorithm[MTP03].
The exact training time will be a result of the complexity of the dataset, its

130
 7.3. Online Support Vector Regression

properties and the concrete implementation. Additionally the hardware, i.e. the
processor, has a significant impact on the absolute time. A huge amount dur-
ing training will be used for the handling of matrices and datasets. Comparing
training times achieved with current state of the art machines with training times
achieved in 2003 would be unfair. Therefore, this comparison is not made at this
point. This section aims more for the discussion of the differences between various
strategies and settings rather than the implementation itself. Some details about
the implementation is described in Section 7.3.10.
The incremental update time for learning a new sample is of huge interest for online
learning. The available time depends on the specific task and the application field
of the algorithm. For reversing mills it is important to improve the prediction
between two consecutive passes. This time should be as short as possible but has
physical limitations due to the drives of the mill and some time for adjusting to
the new pass. The complete incorporation of all measurements should not exceed
one second. Only a small fragment of that time is available to the SVR algorithm.
Obviously the time to incorporate a new sample will depend on the storage size of
the SVR. A higher storage size leads to a higher chance of migration for already
learned samples into other sets during the update procedure. So, more incremental
steps to determine the weight of the new samples might be needed.
The chosen initial values for the SVR parameter have also a huge impact on the
training time. A high value for the allowed deviation ϵ will force more samples
into the remaining set. Samples in the remaining set however have a weight of
zero and can directly be added without any migration. A too low value for ϵ will
force more samples into the error set with maximum weight. When a new sample
is added to the error set, chances for other samples to migrate into a different set
is higher.
Figure 7.17 depicts the training time on the StripForce dataset with three different
storage sizes. Parameter ϵ was chosen to be half of the standard deviation σ of
the target value. It can be seen that the training time will continuously rise until
the storage is completely filled. After this point, the training time will almost be
constant. As already pointed out during the discussion of the storage size, the
differences during training for each sample will be higher with higher storage size.
The choice of the kernel can have also a significant impact on the training time.
The major part of this influence is founded in the problem description and their
relationship. Some problems will require linear kernel and some may require the
selection of RBF or polynomial kernel. The complexity for the calculation of the
kernel itself is only one part. Another part with a major impact on the complexity
is caused by the optimization problem itself, i.e., the dataset. This phenomena
is depicted in Figure 7.18 where the training time for three different kernels are
shown for datasets StripForce and PlateForce. For all experiments, a storage size
of 1 000 and the strategy oldest was used. Further, ϵ was set to 0.5σt and the

131
7. Algorithms

Figure 7.17.: Training time for one samples over the total number of samples already
trained. Three different storage sizes Ns , i.e. 200, 1 000 and 2 000 are
used. The parameter ϵ is chosen to be half of the standard deviation of
the target value for the whole dataset.

penalty parameter C was set to 1.

It can be concluded that there is no significant difference during learning between
the linear kernel and the RBF kernel. The polynomial kernel however has in
general a much higher training time than the others. Another interesting effect
can be seen during training on the StripForce dataset. The training time with
the polynomial kernel is increasing until sample number 1 200. For the next 1
000 samples the training time is almost stable before it is decreasing continuously.
This can only be explained with some structural changes within the dataset which
is very hard to analyze. During training of the PlateForce dataset (Fig. 7.18)
the training time continuously rises, until the storage reached the maximum
capacity. Beside the training time, the selected parameter for the SVR will also
determine the memory consumption which will briefly be discussed in the following.

7.3.9. Memory Requirement

The memory usage for the online SVR algorithm is an important property al-
though it is not as critical as the training time. It will be more a limitation for the
settings of the algorithm, i.e. a limitation to the storage size and the possibility
of using the optimization. The total memory consumption will highly depend on

132
 7.3. Online Support Vector Regression

(a) StripForce

(b) PlateForce

Figure 7.18.: Training time per sample over the total number of samples already trained
for different choice of the kernel function. The maximum storage size
Ns has been set to 1 000 and the oldest sample will be removed for all
variants. A parameter optimization was not used. The depicted kernels
are linear (0), polynomial (1), or RBF (2) based kernels. On the top side
the training time for the StripForce dataset is shown while on the bottom
side the training time for the PlateForce dataset is sown.
133
7. Algorithms

Table 7.5.: Memory consumption on rolling datasets PlateForce, StripForce and Width
for a storage size of 1 000 and 3 000. The amount of instances required for
the parallel treatment of the categorical variable is also shown. Basically it
is a multiplication of the levels for each individual categorical in the dataset.
The maximum memory is a result based on a storage size Ns of 3 000 with
every instance of the algorithm completely filled.
Memory Comb. of Levels Max. Memory
Dataset 1 000 3 000
PlateForce 15MB 80MB 12 960MB
StripForce 19MB 81MB 96 (24x4) 7.7 GB
Width 53MB 82MB 4 328 MB

the storage size and the number of features. The storage size is usually much
larger than the number of features and thus it will be the dominating parameter
for memory.
For hot rolling problems, the online SVR algorithm will be integrated into a
prediction process for the mill. If the dataset contains categorical variables the
major memory consumption will be defined by the choice of the strategy how the
categorical variable is treated. When for each combination of the categorical vari-
ables a separate instance of the algorithm is used, then the memory consumption
is much higher. Although not all of these data have to be kept in memory, also
the total memory consumption has to be taken into account.
To get an idea about the consumption Table 7.5 lists the memory consumption
with different storage sizes in dependence of the categorical treatment for the
rolling datasets. It can easily be seen that the required memory allocation will
quickly reach certain boundaries when using larger storage sizes and an improper
handling of categorical variables. While all dataset will require less than 100 MB
without special treatment of the categorical variables, the memory will tremen-
dously increase when multiple instances are used. The StripForce dataset consists
of 24 different materials and is using four stands which results in 96 different
combination of the categorical. Consequently, the maximum storage size would be
approximately 7.7 GB. When more material is rolled this value can easily increase
to hundreds of GB or even TB. This has to be prevented, since this amount would
require to store a large amount of data on hard drives. Instead, the amount of
data which is used in the specialized algorithm instances has to be reduced or
different strategies for handling of those data have to be chosen.

134
 7.3. Online Support Vector Regression

7.3.10. Implementation
The online SVR algorithm requires a huge amount of matrix and kernel calculations
which would take too long in R. Because of speed, the whole implementation of
the algorithm is done in C++. Although the time and memory measurements
were done with OS specific implementations, the algorithm itself is implemented
OS independent and was tested on Windows and MacOS. The C++ template
library Eigen[GJ+ 10] was used to improve algebraic calculations. Different CPU
profiler were used to detect bottlenecks and to optimize the matrix handling. The
algorithm can be used completely within C++ but also provides an interface to
R. This was created with the help of the package Rcpp [EF11]. The handling
and managing of the different datasets, the matrix calculations and the different
strategies for removal and optimization are implemented in C++. Every control
parameter for the algorithm can either be passed through the R or C++ interface.
Following functions are provided:
1. Training of one or more samples
2. Prediction of one or more samples
3. Controlling the size of the storage
4. Controlling the removal strategy
5. Control of the optimization strategy
6. Initial settings of the parameters
7. Removing of a sample
Removing of a specific sample can be used to update the target value of the
corresponding sample. This is helpful if a second measurement is made and
the reliability of this measurement is higher than the first one. Furthermore,
sometimes problems with measurements are detected after learning already took
place. Then this function can be used to remove those erroneous samples. The
sample has to be passed by its dependable variables and will be detected through
calculation of the euclidean distance. If the sample is found in one of the three
sets it will be removed.
The function to control the optimization strategy will select a specific strategy
how the SVR parameter may be updated online. This optimization is closely
related to dynamic changes within the dataset which will be discussed in the
following chapter.

135
8. Online Parameter Optimization
The previous chapters were describing the different online algorithms and their
properties. Until now, all those algorithm parameters and changes have been
done statically in dependence of the dataset. Parameters can been tuned when a
warm-start problem has to be optimized. Without proper parameter tuning those
online algorithms were achieving reasonable good performance. The idea is now
to sequentially update the parameters during training in order to achieve an even
better performance. This update has to be achieved without violating the time
constraint. Furthermore, it is assumed that the requirement for these updates are
bounded at least to some change within the data stream, i.e. some discrete events
or drifts. Summarizing, answers to the following questions are analyzed:

Necessity: Is online parameter optimization necessary?

Time: How can parameter tuning with a time constraint be achieved?
Detection: How to detect events and drifts?

The answer to these questions are the main topic for this chapter.

8.1. Necessity
If parameter optimization is required depends on many aspects. First and fore-
most the used algorithms define the possibility to optimize certain parameters.
For the RLS algorithm the exponential decay can be modified and for the PA
variants the regularization constraint can be used to optimize the predictions.
The SVR algorithm is using the insensitivity loss, the kernel parameters and
the violation constraint for prediction. The difference for those three algorithms
is, that parameter changes can only be tested for the SVR algorithm. Here,
different parameters can be used directly on the internally stored parameters for
the online SVR algorithm. RLS and PA do not store any samples and additional
implementations would be needed. So the indication if a parameter update should
take place cannot be answered by the algorithm itself and an independent instance
is required. The online SVR algorithm can test different parameter settings on
the internal storage to calculate the performance.

137
8. Online Parameter Optimization

Secondly, the requirement to update parameters is defined by the dynamic prop-

erties of the dataset. If the stream is almost static then clearly the necessity
to perform an online parameter optimization is not given. But as soon as the
data contains some time dependency the optimization can be beneficial. See e.g.
Figure 7.5 which includes an event at a certain time. Clearly, a parameter update
to adapt as fast as possible to the new situation will cause the performance to
increase.
Finally, if the data contains categorical variables with differences in occurrence,
chances are high to benefit from parameter optimization. Different occurrence of
the categorical variables may show completely different relationships among the
independent variables.

8.2. Time Constraint

The only algorithm where an optimization can be made without additional costs,
is the online SVR algorithm. For all other described algorithms the parameters
may be updated to external knowledge. The optimization of online SVR will take
some time since the new parameters have to be validated on the internal storage.
All common optimization algorithms have constraints to limit the number of
iterations and some have also a time limit. During hot rolling with full production
it may happen that the only spare time available is the transport time from
a slab or ingot to the mill while the previous product has just finished rolling.
The production is continuously trying to improve the capability of the mill and
therefore minimizing those gaps. Usually, these gaps are below ten seconds. This
will not be sufficient for many optimization procedures. If new parameters are
tested they have to be added to the SVR algorithm with new parameters, i.e.
the training of the whole storage has to be repeated. The time required for the
parameter optimization of the SVR is therefore highly correlated to the storage
size and the chosen kernel time. See Sections 7.3.8 where the training time is
analyzed.
To be able to execute a parameter optimization, more than this time is needed.
Each mill has some scheduled downtime for maintenance or roll change. If these
downtimes can be detected, they can be used to trigger a more expensive optimiza-
tion. The optimization will cause the algorithm to achieve a better performance
in terms of RMSE on the internal stored samples. Since the parameter will also
have an effect on the training time, the optimization also has to consider this
aspect to ensure that the maximum training time is not violated.

138
 8.3. Drift Detection

8.3. Drift Detection

Dynamic changes are occurring inherently in real-world applications. Changes can
affect the distribution over time of the dependent variables, the output variable
or both. The change within the distribution of the dependent variables is referred
in the literature as concept drift and the change of the conditional distribution
of the output variable is called real concept drift [SG86, WK96]. Another type
of drift is called virtual drift. According to [GvB+ 14] this term is not exactly
specified but usually defines a situation where the distribution of the dependable
variables changes but is not affecting the distribution of the target value.
For the hot rolling problems, the only drift where the algorithms need to be
adjusted is the real concept drift. The regular concept drift, i.e., the change of
the conditional distribution of input values may signal a change in production
which cannot be avoided. The virtual concept drift is not affecting the output
variable and therefore no action is required. In most real-world scenarios also
hidden context drift can occur. Hidden context drift defines a situation where the
drift is caused by variables not included in the dataset. Some of the strategies
dealing with those kind of drift, e.g. to store concept descriptions, are described
by Widmer et al. [WK96] for classification problems.
The real concept drift may manifest in different forms. Table 8.1 gives a summary
of their time behavior analogues to [GvB+ 14] and lists possible reasons in the hot
rolling context. Drift can occur in many different contexts and might seem to be
random. Most of the time, the process state is not completely known and therefore
analysis of the real-drift reasons can get very tedious. For the remainder of this
thesis its assumed that the drift is visible in the input data which is available in
the dataset.

139
8. Online Parameter Optimization

Table 8.1.: Different type of drifts with corresponding time behavior and possible reasons.
Type of drift Time behavior Reasons

Usually event based, e.g.:

• Roll change
• Calibration
Sudden / Abrupt • Measurement problems.
• Mechanical issues (clear-
ance)

Random or event-based start.

End behavior may or may not
be limited.
• Sensor problems
Incremental
• deterioration
• incremental predictive er-
rors (roll wear, roll crown)

Different reasons with complex

analysis
Gradual (over time) • Sensors
• Mechanical

Usually also reoccuring process

state, e.g. categoricals not used
or monitored.
• Hidden context change
(variable not in dataset or
Reoccurring concepts
not used for prediction).
• Material supplier
• Furnace
• Material

Various reasons
• Sensors
• Insufficient predictive capa-
Outlier bilities
• Wrong assumptions on pro-
cess state / product state

140
 8.4. Online Parameter Optimization for Support Vector Regression

8.4. Online Parameter Optimization for Support

Vector Regression

Detection of drift and events is important when parameters can be optimized to

improve the prediction accuracy. One of the first algorithms which were capable of
drift detection were developed in the late eighties of the last century by Schlimmer
and Granger[SG86]. They are using ensemble based descriptions most relevant for
the current context. Other methods are using only a subset of the last instances
[WK96]. Usually this subset is defined by a fixed or dynamically sized window,
i.e. all of the recently received samples are used[Kli04].
Other methods for the detection of drift are usually monitoring the prediction
accuracy. For classification problems methods are available which are using the
error probability or the distance between two errors to detect changes [JAF+ 06].
The described online SVR method has already different methods for selecting or
discarding samples since the storage size has to be limited to a certain amount of
samples. A remaining issue is the optimization of parameters which is necessary
on the occurrence of drift in the data.
Therefore the implementation was extended to be able to optimize the most
relevant parameters. Dependent on the application, two different implementations
can be used: discrete and continuous.
The discrete optimization is triggered through an event. Because the online pa-
rameter may be changed during optimization the algorithm is not able to receive
any new training data in that time. This situation reflects typical real-world
scenarios, e.g. if there is a scheduled maintenance or some predictable delay in the
processes. This spare time can then be used to trigger the discrete optimization.
A pseudo code of the discrete optimization is shown in Algorithm 7.
This discrete optimization will cause a simple grid search for the two main SVR
parameter ϵ and C. Because the optimization time should be as short as possible,
it has to finish as fast as possible. For each optimization call only four different
parameter variations are tested and compared to the current performance. The
new parameters are determined through multiplication of the current parameters.
For sake of simplicity, the actual parameter are just multiplied by 0.9 and 1.1.
This means that the parameter variation is 10% in both directions. Therefore
only four new optimization threads are created. In principle also more threads
and combinations may be tested. Table 8.2 shows the different settings, which are
tested within one optimization call. The execution of each parameter variation is
done in parallel in different threads (Step 2 in Algorithm 7). For reference also
the original setting is shown.
Each thread is trained with all samples of the internal storage incrementally and
before training of each sample, the prediction is stored and compared with the

141
8. Online Parameter Optimization

Algorithm 7: Discrete online SVR optimization.

1 let No be the number of total optimizations loops; for n = 1 to No do
2 start 4 threads with different parameters and collect results ;
3 start optimization with parameters based on regression of all RMSE results;
4 let RM SEi determine the RMSE of thread i and i = 0 reflect the RMSE
with current settings;
5 if RM SEi < 0.95RM SE0 , i = 1 · · · 5 then
6 start 2 threads with new parameter set and variation of kernel
parameter;
7 Switch parameter of main algorithm to the best performing parameter
combination
8 else
9 stop optimization

Table 8.2.: Parameter variation for optimization of C and ϵ. After the performance of
all variants are determined a linear regression model is build and another
variant is tested.
Thread C ϵ
Main C0 ϵ0
Thread 1 C1 = 1.1C0 ϵ1 = 1.1ϵ0
Thread 2 C2 = 0.9C0 ϵ2 = 1.1ϵ0
Thread 3 C3 = 1.1C0 ϵ3 = 0.9ϵ0
Thread 4 C4 = 0.9C0 ϵ4 = 0.9ϵ0
Main (C5 , ϵ5 ) = f (C0..4 , ϵ0..4 , RM SE0..4 )
Main min RM SE {{ϵ0 , C0 } , · · · , {C5 , ϵ5 , }}

true value.
For comparison of the different variants the performance in terms of RMSE is used.
The MAE can also be used for the internal evaluation and will yield similar results.
If a thread achieves a better result than the original one the algorithm creates a
2D linear regression based on both parameters and all individual thread results .
Then the maximum gradient is used to calculate an optimum parameter set which
is then used and tested again (Step 3 in Algorithm 7). If the best result out of
all four threads and the last parameter variation achieves a significantly better
performance than the original parameter set the kernel parameter variation is
tested in two variants (Step 6 in Algorithm 7). Significance can be set in terms of
RMSE improvement. The parameters of the best performing variant is transferred
to the main algorithm and the optimization stops.
The background optimization procedure will continuously search for better pa-

142
 8.4. Online Parameter Optimization for Support Vector Regression

rameters in background, i.e., in another thread. Therefore, the thread will first
collect some samples and validate the performance against different parameters
than used in the main thread. If the background optimizer found some better
parameters it will trigger an event to switch to the new settings. This is simply
realized with the usage of a mutex in C++. This mutex has to be used because,
for a short time, the main algorithm will not be able to receive and handle any
new data.
The background optimization can be used if information regarding scheduled
downtimes or delay are not available and memory and computational power is
sufficient for another algorithm continuously running in a separate thread.
The supply of training samples duplicated, i.e. the background thread will receive
the same samples like the main algorithm. In principle any kind of parameter
optimization algorithm can be used but for sake of simplicity the current imple-
mentation also uses a grid search as described in the case of discrete optimization.
The main advantage of this optimization is that the main algorithm still can be
used and is not affected during determination of the best parameter. Another
advantage is that it will not rely on some delay detection of the process. The
pseudo code of the procedure is shown in Algorithm 8. After a sufficient amount
of samples have been collected, the background optimizer starts searching for
better parameters. During this search, it will not be able to incorporate any new
samples. If the optimizer found parameters with higher performance, they have to
be verified against the latest data only received by the main algorithm. Only if the
performance on the latest data has also improved, then the parameters of the main
algorithm have to be changed. Due to the fact that already a duplicate version
was trained in background the transfer is very fast and should not introduce a
long disturbance of the process.
A problem when using the background optimization are the synchronization calls
to the main thread. If better parameters have been found then the background
process will retrieve all currently stored data from the main SVR algorithm in-
stance and feed these samples to a new instance. At some point a lock has to be set
for a short period of time in order to swap parameters from background and main
thread. Clearly, the background optimization will require additional resources like
memory and computational power. This will cause the incremental training to be
slower than without optimization. Figure 8.1 shows a comparison of training time
per sample for different storage sizes with and without the background optimiza-
tion. Another reason for the high training times for the background optimization
is the available CPU power. Multiple threads are requesting processor resources
at the same time. This will influence the main process since no prioritization of
the different thread is used. Another problem which might occur when using the
background optimization is determined by the number of categorical variables
and the strategy to handle those. When for each combination of categorical a

143
8. Online Parameter Optimization

Algorithm 8: Background online SVR optimization.

1 start thread with parameters of main model;
2 start collecting samples;
3 let n be the number of samples received;
4 if n > 10 then
5 store arriving samples in buffer (stop training);
6 use grid search to check for better parameter settings;
7 if better parameter found then
8 train algorithm with buffered samples;
9 check latest performance of improved variant against main model
performance;
10 if performance improved then
11 lock main model and initialize switch to new parameters;
12 else
13 continue grid search for improved parameter settings;

14 else
15 train with buffered samples and start collecting new samples;
16 continue after a certain time with parameter search;

17 else
18 continue collecting samples;

separate model instance is used, then the required resources and memory will
quickly be too high. Consider the example of the StripForce dataset with two
categorical variables, i.e. material and stand. The material has 24 different values
and the stand variable four different values. This would mean, despite already 96
different instances of the algorithm are created, that also 96 separate threads will
be created and will consume processor resources. Therefore, both strategies can
only be used together if the number of different levels of the categorical variables
are low and are assumed to be known a priori.
An example of the impact on the optimization is given in Figure 8.2 where the
Width dataset is used for the comparison on cold-start and warm start optimization
problems. Shown is the performance when the SVR is used with default parame-
ters without any optimization (cold-start) and with the two online optimization
methods. For comparison, also the performance of the SVR algorithm with opti-
mized parameters on the first 200 samples is shown (warm-start). It can be seen,
that the online optimization yields significant improvements in comparison to the
default parameter settings. Both optimization methods achieve almost the same
performance throughout the complete dataset. When samples are known, the
SVR parameters can be optimized. The achieved performance is therefore much

144
 8.4. Online Parameter Optimization for Support Vector Regression

Figure 8.1.: Training time per sample for the StripForce dataset. Three different storage
sizes Ns , i.e., 200, 1 000 and 2 000 with and without the background
optimization is analyzed. For all variants the oldest sample was always
deleted and the RBF kernel was used.

Figure 8.2.: Comparison of the MAE performance on the Width dataset for different
usage of the optimization. Default SVR parameters (cold-start) without
parameter optimization is compared with the performance when background
optimization (BG) or optimization on event (Event) is used. Additionally,
the performance for optimized parameters on the first 200 datasets is shown
(warm-start).

145
8. Online Parameter Optimization

better on the first 20 000 samples. It is interesting to see that the optimization can
only marginally improve the overall performance for the width problem. In both
cases, the abrupt drift and the incremental drift can be compensated. The drift is
inherently compensated by integrating new samples into the SVR algorithm. The
reason, why the optimization in case of the warm-start problem will only have a
small impact, may lie in the bias-variance trade off which is described in detail
in [JWHT14, HTF01]. Initial optimization will yield great improvement of the
prediction error, but further optimization may only improve the prediction on
the samples stored internally. For new samples, the prediction error might even
increase.
As a consequence, the parameter optimization is highly recommended for cold-start
problems. For warm-start problems, the optimization should be triggered if high
dynamic changes are detected in the data.

146
9. Performance
After discussing the different algorithms and their parameter impact on various
scenarios this chapter will now compare their performance. First, a comparison
on the extrapolation behavior is made which is especially important on cold-start
problems. Afterwards, the comparison on the default datasets is made before the
chapter closes with the discussion about the performance achievements on the
rolling dataset.

9.1. Extrapolation
An important criteria for online algorithm is their behavior for unseen areas, i.e.,
if they have to extrapolate. Offline algorithms will usually have enough data
available and therefore, this behavior is not of great importance. If the online
algorithm is used for cold-start problems, the prediction is always in regions where
no prior information is available. Therefore, the algorithm has to extrapolate from
regions, where initial data has been received. The question is, how the different
algorithms are handling such scenarios? Extrapolation will also occur, if rare
materials or rare constellations are considered. Figure 9.1 shows the prediction
of all three online algorithms when they were trained in the region of 0 ≤ x ≤ 1.
The prediction over the whole area is very bad in all three cases. Clearly, the
extrapolation with online SVR regression with RBF Kernel is not producing any
prediction values much higher than the training region. The typical application
scenario for the online algorithms in the hot rolling is to correct analytical models.
Therefore, the assumption is that no huge errors are present. Therefore such
extrapolation behavior as shown for PA-II and RLS are not desirable and have to
be avoided. This might be accomplished with different strategies like a maximum
correction term for regions without any data.
This chapter compares all discussed algorithms and their performance on the
described datasets. Further, for the default dataset, a comparison with state-of-
the-art models is made. To answer the question if online models are really needed,
different offline algorithm performances are also shown.

147
9. Performance

Figure 9.1.: Comparison of PA, RLS and SVR with RBF Kernel prediction. All
algorithms were trained with training samples of a 1D test function
f (x) = 0.3 sin (5 × x − 2) + 0.4 × (x − 1) + σ between 0 ≤ x ≤ 1. The
plot shows the prediction in the region of −10 ≤ x ≤ 10 and the true
function (violet) for the RLS algorithm (green), PA-II (red) and SVR with
RBF Kernel (cyan).

148
 9.2. Performance on Default Datasets

Table 9.1.: Mean squared error (MSE) for different algorithms on the default datasets.
The first 3 columns contain offline algorithms for comparison which are using
the first 10% of the data as training data. The following algorithms are
used: Linear regression model (Lin), Random uniform forest(RUF), Support
Vector Machine (SVM), Passive Aggressive in two variants (PA-I and PA-II),
Recursive Least Squares (RLS), Support Vector Regression (SVR) and the
best algorithm described in [Iko12] (IKON). Values marked bold achieve the
best performance.
Problem LIN RUF SVM PA-I PA-II RLS SVR IKON
Abalone 8.4 7.9 7.0 4.9 4.1 4.6 4.9 5.7
Cal housing 7.5e9 6.4e9 9.7e9 2.6e9 2.3e9 3.0e9 5.6e10 5.1e9
Elevators 1.0e-5 2.2e-5 1.2e-5 5.3e-6 4.9e-6 8.9e-6 8.0e-6 2.2e-5
House 8L 1.8e9 9.1e8 1.2e9 2.5e9 2.3e9 1.8e9 1.5e9 1.1e9
House 16H 2.1e9 1.2e8 1.6e9 2.7e9 2.7e9 2.1e9 1.7e9 1.6e9
Mv delve 1.5 2.6e-2 0.5 8.6 6.6 1.9 0.3 1.7e1
Pol 9.9e2 6.0e1 2.7e2 2.4e3 2.1e3 1.9e3 2.5e3 2.3e2

9.2. Performance on Default Datasets

To compare the implemented online algorithms the default datasets described in
Section 6.1 are used. In addition to the best result obtained from [Iko12] other
simple algorithms are used as baseline results. To have also a comparison to
offline models three simple offline algorithms are used which are trained by the
first 10% of the data. The first one is a simple linear regression model (LIN) , the
second one is a random uniform forest model (RUF) and the third one uses the
support vector machine implemented in the e1071 package which is based on the
popular libsvm (SVM) [CL11]. The tested online algorithms are the two variants
of the PA algorithms, i.e. PA-I and PA-II, the RLS algorithm and the Online
SVR algorithm. The performance value is taken from the mean squared error
(MSE) of the complete data. This is used to directly compare the results with
the results obtained in [Iko12]. Only dummy coding strategies which are suitable
for online algorithms are used here. This means, that traditional dummy coding
schemes, which causes the feature vector to increase, are not considered.

9.3. Performance on Rolling Datasets

The main application of the presented algorithms are analyzed in the context of the
hot rolling process. The performance of the best algorithms on the corresponding
rolling problem is analyzed in the following part.

149
9. Performance

Figure 9.2.: The plot over the residuals of a static liner model is used to detect the
dynamic changes like discrete events or drifts in the dataset.

Width Problem
The Width dataset is an excellent example to test algorithms on their capability
to deal with drift in the data. At some point of the dataset which occurs ap-
proximately at sample 40 000 a drift in the dataset is visible. Using the different
drift types presented in Table 8.1 it can easily be identified as a sudden or abrupt
drift followed by an incremental drift. The later one is not easily visible on the
target data itself but on the residuals of an offline linear regression model. The
residuals of a linear model which was trained on the first 5 000 samples is shown
in Figure 9.2. Before training, the dataset was scaled to values between 0 and 1.
The abrupt drift and the incremental drift is clearly visible.
The impact of the algorithm parameters has already been shown in the cor-
responding sections of the algorithm. The interesting question is now, which
algorithm achieves overall the best performance. To get an idea about how good
the predictions are the first comparison is done based on the MAE. Figure 9.3
shows the performance of the different algorithms. For all online algorithms, only
the best variant is shown. The corresponding offline variants, i.e. a linear model
and a SVM model, are also shown for comparison. It clearly turns out, that
there is a high dynamic within the datasets since both offline variants are the
two worst algorithms. Especially around at approximately half of the dataset the
performance of the linear model drops significantly. All online variants are almost

150
 9.3. Performance on Rolling Datasets

Figure 9.3.: Comparison on the best algorithm variants for the Width dataset. Shown is
the mean absolute error over the whole dataset. For each algorithm variant
only the best one is shown. Beside the three online algorithms PA, RLS
and SVR also a linear regression model and an offline SVR algorithm is
shown.

151
9. Performance

Table 9.2.: Summary of the mean absolute error (MAE) for the best algorithms over
the whole dataset.
Rank Algorithm Details Coding MAE
Size Ns : 1 000, Mgmt. Ms : Reserve
1 SVR online Numeric 1.89
Kernel: RBF
2 RLS Decay: 0.98 Remove 2.04
3 PA2 Aggressive Parameter: 0.1 Remove 2.27
4 SVR Offline Libsvm trained on first 10% Numeric 3.40
5 Linear Offline Trained on first 10% Numeric 9.50

Table 9.3.: Top 10 online SVR variants for the Width dataset.
Rank Size Ns Kernel Storage Mgmt. Coding MAE
1 1 000 RBF Reserve Numeric 1.89
2 1 000 RBF Random Remove 1.89
3 1 000 RBF Distance Remove 1.92
4 1 000 RBF Oldest Numeric 1.93
5 1 000 RBF Oldest Numeric 1.94
6 500 RBF Distance Numeric 1.94
7 1 000 RBF Distance Numeric 1.95
8 1 000 RBF Distance Remove 1.95
9 1 000 RBF Oldest Remove 1.95
10 1 000 RBF Oldest Remove 1.96

stable over the whole dataset with SVR achieving the best performance followed
by the RLS algorithm. The PA 2 variant is the best among the algorithms of class
Passive Aggressive but are the worst online algorithm. Nevertheless, the result is
only slightly worse than the RLS variant.
The achieved performance over the whole dataset is summarized in Table 9.2. An
interesting observation is, that the dummy coding scheme is either the conversion
to numerical or the removal. The categorical factor has almost no impact on the
performance.
For the online SVR algorithm the best performance was achieved with minimum
storage size which clearly indicates the high dynamic of the dataset. It is not
necessary to store any old information. The fact that the random removal of
samples achieves the best performance is also nothing special and was already
discussed in [JZBBR17]. Tables 9.3 shows the top ten performance of the different
online SVR variants.

It can be seen from Table 9.3 that all online SVR variants are achieving a better

152
 9.3. Performance on Rolling Datasets

performance than any of the other algorithms. Dominating kernel was the RBF
kernel which was used in all of those algorithms. The best linear kernel algorithm
achieves a MAE performance of 1.96 and the best polynomial kernel has a perfor-
mance of 2.19, which is still better than all offline variants and also all PA variants.
Clearly, the storage size of 1 000 is also dominating the best performing algorithms.
Only one algorithm with smaller storage size has a similar performance. The best
algorithm with storage size of 200 achieves a MAE performance of 2.04.

PlateForce Dataset
The PlateForce dataset is the only dataset for hot steel rolling. The flow curves
for steel are much more established than those for aluminum. Additionally, the
grouping for steel is very fine coarse in comparison to aluminum. Therefore it is
assumed, that there are no huge differences between the material groups. Still
the questions arises, if online algorithms can significantly improve those kind
of problem. The handling of the material group as categorical variable with a
potential high number of levels is further of great importance.
To answer these questions the best variants for all online algorithms and two
corresponding offline algorithm are used and the MAE is plotted over the whole
dataset. The result can be seen in Figure 9.4. The figure shows, that the perfor-
mance of the offline SVR and the linear model can significantly be improved with
the usage of their online versions. It is interesting to see, that the best PA and
the best RLS algorithm have almost the same performance. After sample number
10 000, some drift in the data occurs. At this point, both offline algorithms have
a tremendous decrease in performance. Either some new material was rolled at
this point or there is some event-based drift occurring. Somehow, this drift can
be better compensated by the RLS algorithm than the PA algorithm.
The best performance can be achieved by the online SVR algorithm which has far
the best performance in terms of MAE. While the total MAE for the online SVR
is ≈ 850 kN, the other online algorithms achieve a performance of approximately
≈ 2 700 kN (RLS) and ≈ 3 000 kN (PA-II).
Surprisingly, the offline SVR which was only trained on the first 10% of the data
achieve a better performance than PA and RLS. The online algorithms achieve
a more or less stable performance, whereby the performance of the offline SVR
algorithms is almost continuously decreasing and is expected to further decrease.
It seems that the properties of the SVM are ideally suited to deal with this kind
of problem.
In general, a force deviation of 1 000 kN can be seen as a very good prediction
which is a perfect prerequisite for a robust stable production with highest quality
demands.

153
9. Performance

Figure 9.4.: Mean absolute error (MAE) for the PlateForce dataset. Shown is the MAE
over the whole dataset for the best variants of all analyzed online algorithms,
i.e. Passive Aggressive (PA), Recursive Least Squares (RLS) and online
SVR. For reference and also to see whether an online algorithm is necessary
two offline algorithms are also shown which were trained on the first 10%
of the data.

154
 9.3. Performance on Rolling Datasets

Table 9.4.: Top 10 online SVR variants for the PlateForce dataset.
Rank Size Ns Kernel Storage Mgmt. Ms Coding MAE
1 500 RBF Oldest Numeric 842
2 1 000 RBF Reserve Remove 851
3 1 000 RBF Oldest Remove 854
4 1 000 RBF Oldest Numeric 854
5 500 RBF Oldest Remove 855
6 500 RBF Oldest Numeric 855
7 1 000 RBF Reserve Numeric 862
8 1 000 RBF Oldest Remove 871
9 500 RBF Oldest Remove 874
10 1 000 RBF Random Remove 878

A further analysis is summarized in Table 9.4 where the top ten online SVR
variants are listed with the corresponding parameter settings. The dominating
kernel was again the RBF kernel. The dummy variable should be removed or
converted to a numerical value and the oldest samples should be removed.

StripForce Dataset
The StripForce dataset contains data from aluminum mills where the flow curve is
very sensitive to variation in the chemical composition. As already discussed, the
target value is strongly correlated to the categorical variable. The comparison of
the MAE performance is shown in Figure 9.5 where the best variants of all three
online algorithms are compared to two offline variants, which were trained on the
first 10% of the data.
It can be seen, that the performance can be significantly improved with online
algorithms and the best performance was achieved with the online SVR algorithm,
followed by the RLS and PA algorithm. Table 9.5 summarizes the top ten SVR
algorithm variants. The best results were achieved when the categorical variables
were treated separately and the RBF kernel was used. It is also interesting to
see, that despite both other rolling datasets, the distance based strategy to select
samples achieves the best performance.

155
9. Performance

Figure 9.5.: Mean absolute error (MAE) of the StripForce over the whole dataset. The
two offline algorithms, i.e. a linear model and the offline SVR were trained
on the first 10% of the data for comparison. For the online algorithms only
the best performance is shown.

Table 9.5.: Top 10 online SVR variants for the StripForce dataset.
Rank Size Ns Kernel Storage Mgmt. Ms Coding MAE
1 1 000 RBF Distance Parallel 1 647
2 500 RBF Distance Parallel 1 647
3 1 000 RBF Distance Parallel 1 647
4 500 RBF Distance Parallel 1 648
5 200 RBF Distance Parallel 1 650
6 200 RBF Distance Parallel 1 653
7 1 000 RBF Random Parallel 1 654
8 200 RBF Random Parallel 1 656
9 1 000 RBF Oldest Parallel 1 658
10 500 RBF Reserve Parallel 1 658

156
10. Conclusion

Chapters 5-9 introduced online learning for real-world problems, focusing on hot
rolling problems. It was shown that classical, offline optimization is not suitable
for this kind of optimization problem. For dynamic real-world problems, online
algorithms should be used. The section showed that all presented online algo-
rithms outperform their offline variants and may improve the predictive accuracy
of the process. This will ensure that the process will be more robust and is able
to produce products of high quality. For all considered real-world examples the
online SVR has proven to be the best among the three algorithms.
It was further shown that the algorithms are also applicable for cold-start optimiza-
tion problems with some limitations. Parameters should be guessed or set prior to
known values of similar problems. Since no information is available, the categorical
variables should be neglected for cold-start problems. This will allow the model
to learn the base problem constellations and the main effects on the error variable.
With online parameter optimization a significant improvement could have been
achieved and depending on the exact application the background optimization
process or the event-based trigger can be used for prediction improvements.
The usage of categorical variables has a significant effect on the achievable per-
formance of the problems. However, it is wise to use feature selection techniques
to determine, if the categorical variables have an impact on the predictive accu-
racy at all. In general, if the categorical has an impact on the predictions using
parallel model instances, i.e. either parallel model instances or just corrections
are beneficial. However, if there is a high dynamic in the dataset with or without
the presence of drift, categoricals may cause the learning of that drift to be much
slower. In such cases, especially if an abrupt drift is present or cannot be excluded,
the usage of categoricals has to be rethought.
The size of the stored samples is highly dependent on the properties of the dataset.
High dynamic problems will benefit from low storage sizes. Otherwise the size
should be as high as possible while keeping the learning time constraint in mind.

157
10. Conclusion

10.1. Recommendations for Other Real-World

Problems
The algorithms presented and discussed in previous chapters are not restricted to
hot rolling. They can be used on any kind of real-world problem. The following
recommendations can be derived from the previous results.
• For cold start problems: Start with PA or even simpler algorithms. Also
SVR with only few dependable variables can lead to reasonable results.
Do not consider categorical variables directly. Use online optimization of
parameters for SVR.
• Feature selection with various algorithms to reveal the most important
variables. Do not use standard methods for this screening since most of
them are not considering dynamic effects. Instead, use online algorithms
and adequate performance measurements.
• If the problem is to reduce the residual prediction errors of other prediction
models, the usage of categorical variables is highly dependent on the internal
structure of those models. Using the categorical as ordinal is most often not
desirable and will usually not improve the performance.
• If the problem is known to be highly dynamic, a low storage size is benefi-
cial. Otherwise use as much storage as suitable without violating the time
constraint.
• For highly dynamic data the best strategy for selecting the samples is to
remove the oldest one. Otherwise the distance based strategy achieves the
best result.
• If processing power, memory and time is not violated use either background
optimization or the discrete events for optimization of the SVR parameter.
• The RBF kernel seems to be a very robust kernel for most of the problems.
However if some data are available other kernels should also be tested and
compared to the RBF kernel.

10.2. Outlook
There are still some items which could not fully been analyzed in this thesis. The
optimization of kernel parameter is one of them. It might be, that polynomial of
different degree and offsets might be more suitable than the polynomial used here.
Further, also other kernel like the sigmoid kernel should be tested and compared
to other kernels.
A problem for categorical variables is still the optimization of all parallel instances
of the algorithm. In the currently presented approach only the base models were

158
 10.2. Outlook

optimized but surely also the specific models can benefit from online parameter
optimization. Due to the high CPU usage, the background optimization of multi-
ple instances is currently not recommended. A possibility could be to instantiate
one background optimization process which sequentially optimizes the parameters
of all instances. Then, only one additional thread would be used for optimization
and computation.
Another more general problem is the variable selection which still has to be made
prior to the usage of these online algorithms. A solution might be to integrate the
feature selection into the background optimization process. Then, the full samples
would always be stored but the algorithms would work on a subset selected by a
separate process.

159
Part III.

Summary

161
11. Summary
The previous chapters described optimization in context of real-world applications
with focus on hot rolling of aluminum and steel. The complex rolling process,
which was introduced in Chapter 2, can highly benefit from the presented ideas
and algorithms if it is combined with knowledge from process experts. Without
an in-depth understanding of the process on the one-hand side and the analytical
models on the other hand side the potential improvement is limited.
Data-driven surrogate based optimization, introduced in Chapter 4 is an excellent
method for improving even complex processes. The method for optimization of
individual analytical models was generated especially for the hot rolling context
but can be transferred to any kind of complex process. The prediction process
itself was included as black box. Knowledge about the internal design of the
process models however was determining the requirements for the optimization
framework.
The first optimization problem was using data of an aluminum roughing mill to
optimize flow curve parameters for an unknown material. It was shown in Section
4.4, that a tremendous improvement can be achieved. These improvements were
validated with real process data. The optimization consisted of two consecutive
optimizations. After an initial parameter screening, the five most promising points
were selected and used for a local optimization with a Nelder-Mead algorithm.
The second optimization example, which was discussed in Section 4.5, illustrated,
that it is sufficient to use only the local optimization if the region for the initial
flow curve parameters are known in advance.
Afterwards it was shown in Section 4.6, that the concept is also applicable to other
mill types. To transfer the result from one mill to another, certain conditions
and limitations have to be made. In general it would be good to extend the
simulation to be able to cover multiple mill types with wide ranges of different
product geometries to achieve the best performance.
Although the concept was applied only for flow curve parameter optimization it is
suitable for any kind of prediction for the rolling process. The only restriction is,
that parameters can be modified by the optimization framework. The simulation
was already used to optimize the friction calculation in aluminum mills. The
friction will also influence the roll force calculation and therefore first the flow
curves have to be optimized. It can be seen as a logical next step if the flow curve

163
11. Summary

parameters have been optimized. As mentioned already during the introduction

of the rolling process, the friction is dependent on the emulsion which is almost
unique to every rolling mill.
Prediction residuals will always occur for various reasons. Most often these reasons
are not easy to detect and will have a high dynamic behavior. These residuals
can be reduced with help of online algorithms as demonstrated in Chapter 5.
Three different online algorithms were compared with state-of-the-art algorithms
on default datasets but also on data coming from different rolling mills. It was
demonstrated, that all three algorithms can significantly reduce the residuals
for the rolling datasets. The online algorithms were used here to improve the
predictions coming from various analytical models. The most promising candidate
for all rolling datasets was the online SVR algorithm, which was introduced in
Section 7.3. The online SVR algorithm was extended to fit the requirements
for real-world processes. The storage size of the SVR has to chosen in a way
to balance between robustness and speed. It was shown in Section 7.3.5 and
Chapter 9, that for high dynamic optimization problems the strategy to remove
the oldest data is the best choice. Only for static problems, other distance-based
strategies seems to be beneficial. The handling of categorical variables is of great
importance for online processes since most of the common strategies will only work
properly for offline algorithms. It turned out, that depending on the optimization
problem, the usage of separate models for each unique combination of categorical
variables can highly improve the performance. This was shown in Section 7.3.7.
For cold-start optimization problems, i.e., when no information about the residuals
is available, the categorical has to be neglected since it would only slow down the
initial learning process. Another extension for the online SVR was made to enable
online parameter optimization. Online parameter optimization as introduced in
Chapter 8 has to be used in case of high dynamic processes and if no information
about the dataset is known. Both presented optimization strategies will ensure
that the performance will significantly improve. If initial samples are available
then the optimization will be only beneficial for highly dynamic datasets.
Online algorithms represent a model class which are of great importance for
all kind of real-world application. The presented extension to the online SVR
algorithm will enable multiple future application possibilities across industrial
branches. The online optimization of SVR parameters will ensure that the model
will adapt as fast as possible to any kind of problems regardless of the dynamic
behavior of the dataset.

164
List of Figures
1.1. Hot Rolling Optimization Overview . . . . . . . . . . . . . . . . . 9

2.1. Layout of Hot Rolling Mills . . . . . . . . . . . . . . . . . . . . . 14

2.2. Width and Thickness Reduction in Reversing Mills . . . . . . . . 17
2.3. Dog Bone Shape . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4. Mode interactions for the description of a single pass . . . . . . . 21

4.1. Data-driven surrogate model . . . . . . . . . . . . . . . . . . . . . 30

4.2. Extraction of Data for Simulation Framework . . . . . . . . . . . 35
4.3. Selection Process for Suitable Flow Curves . . . . . . . . . . . . . 38
4.4. Simulation Flow Chart for Optimization of Flow Curve Parameter 46
4.5. Comparison of the Impact of Expected Improvement on the Simu-
lation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.6. Overview of Simulation Results with Various Number of Local
Optimizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.7. Contour Plot of the Best Solution . . . . . . . . . . . . . . . . . . 52
4.8. Roll Force Prediction on a Real Rolled Product With Old Parameters 54
4.9. Roll Force Prediction on a Real Rolled Product With Optimized
Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.10. Relative Roll Force Deviation . . . . . . . . . . . . . . . . . . . . 55
4.11. Developement of the Root Mean Squared Error Over The Number
of Evaluations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.12. Prediction Error When Parameter are Optimized on Another Mill 59
4.13. Prediction Error For Parameter Optimized on Another Mill Type 60

5.1. Comparison of Warm-Start and Cold-Start Behavior . . . . . . . 68

5.2. Pre-Processing of Measurements in Hot Rolling . . . . . . . . . . 73
5.3. Handling of Categorical Variables in own Algorithm Instances . . 78

6.1. Target Value in the Width . . . . . . . . . . . . . . . . . . . . . . 86

6.2. Boxplot of the Target Value over all Rolled Materials in the Plate-
Force Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.3. Boxplot of the Target Value over all Rolled Materials in the Strip-
Force Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

165
List of Figures

7.1. Performance Comparison for Passive Aggressive . . . . . . . . . . 94

7.2. Parameter Influence for Passive Aggressive on Learning of Drifts . 95
7.3. Influence of Aggressiveness Parameter on Prediction Performance 95
7.4. Performance Impact for Different Handling of Categorical Variables
with Passive Aggressive . . . . . . . . . . . . . . . . . . . . . . . . 97
7.5. Influence of Exponential Decay on Learning . . . . . . . . . . . . 103
7.6. Performance Impact of Exponential Decay . . . . . . . . . . . . . 104
7.7. Performance on Rolling Dataset with Different Handling of Cate-
gorical Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.8. Epsilon Insensitive Loss Function . . . . . . . . . . . . . . . . . . 110
7.9. Prediction Example to Illustrate Support Vector Regression . . . 113
7.10. Training Effort for Incremental Support Vector Regression . . . . 117
7.11. Histogram of the Target Value from the House 16H Dataset . . . . 121
7.12. Influence of the Storage Size on the Online Support Vector Regres-
sion Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.13. Windowed Performance with Different Storage Sizes . . . . . . . . 125
7.14. Influence of the Storage Managment on the Online Support Vector
Regression Performance . . . . . . . . . . . . . . . . . . . . . . . 126
7.15. Performance for Different Kernel on Rolling Datasets . . . . . . . 128
7.16. Performance for Different Handling of Categoricals on Rolling Datasets130
7.17. Training Effort on StripForce Dataset . . . . . . . . . . . . . . . . 132
7.18. Training Time per Sample for Different Kernels . . . . . . . . . . 133

8.1. Training Effort Variance for Different Storage Sizes . . . . . . . . 145

8.2. Comparison of Warm-Start and Cold-Start With and Without
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

9.1. Comparison of all Online Algorithms on Extrapolation Behavior . 148

9.2. Residual Prediction Error on Width Dataset for Linear Regression
Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
9.3. Performance of Best Algorithms on Width Dataset . . . . . . . . . 151
9.4. Performance of Best Algorithms on PlateForce Dataset . . . . . . 154
9.5. Performance of Best Algorithms on StripForce Dataset . . . . . . . 156

166
List of Tables
2.1. Products Geometries for Various Mill Types . . . . . . . . . . . . 13
2.2. Summary of Hot Rolling Mill Components . . . . . . . . . . . . . 20

4.1. Overview of Flow Stress Equations . . . . . . . . . . . . . . . . . 38

4.2. Boundaries for Flow Curve Parameters . . . . . . . . . . . . . . . 44
4.3. Settings Used for the Optimization with the Sequential Parameter
Optimization Framework . . . . . . . . . . . . . . . . . . . . . . . 45
4.4. Detailed Results for Simulation with Various Number of Local
Optimizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.5. Description of Experiments Conducted to Analyze the Impact of
the Initial Design and Local Optimization . . . . . . . . . . . . . 51
4.6. Parameter Comparison of Best Local Optimization for Different
Seeds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.7. Summary of the Roll Force Prediction Error With Parameter Op-
timized on Another Mill Type . . . . . . . . . . . . . . . . . . . . 60

5.1. Dummy Coding Example . . . . . . . . . . . . . . . . . . . . . . . 76

6.1. Overview of Samples and Data Types in the Default Datasets . . 84

6.2. Overview of Samples and Data Types for the Rolling Datasets . . 84

7.1. Performance Comparison for Passive Aggressive on Default Datasets 99

7.2. Performance Summary for Recursive Least Squares on Default
Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
7.3. Parameter Settings for Warm-Start Optimization Used During
Optimization with nloptr . . . . . . . . . . . . . . . . . . . . . . . 123
7.4. Influence of Kernel Selection on Training Time . . . . . . . . . . . 128
7.5. Memory Requirement for the Rolling Datasets . . . . . . . . . . . 134

8.1. Drift Types and Possible Reasons in the Context of Hot Rolling . 140
8.2. Parameter Variations Used to Determine Optimum Parameters
Settings for one Optimization Cycle . . . . . . . . . . . . . . . . . 142

9.1. Performance Summary for all Online Algorithms on Default Datasets149

9.2. Detailed Summary of Best Algorithms on Width Dataset . . . . . 152

167
List of Tables

9.3. Best Support Vector Regression Algorithms on Width Dataset . . 152

9.4. Best Support Vector Regression Algorithms on PlateForce Dataset 155
9.5. Best Support Vector Regression Algorithms on StripForce Dataset 156

168
Bibliography

[Agr07] A. Agresti. An Introduction to Categorical Data Analysis. Wiley

Series in Probability and Statistics. Wiley, 2007.
[AW94] Karl Johan Astrom and Bjorn Wittenmark. Adaptive Control.
Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA,
2nd edition, 1994.
[AY01] Kazuya Asano and Kazuhiro Yamamoto. Parameter estimation for
the tension model in hot strip mills. IFAC Proceedings Volumes,
34(18):345 – 350, 2001. 10th IFAC Symposium on Automation in
Mining, Mineral and Metal Processing (MMM 2001), Tokyo, Japan,
4-6 September, 2001.
[BB03] Thomas Bartz-Beielstein. Experimental Analysis of Evolution
Strategies—Overview and Comprehensive Introduction. Reihe CI.
SFB 531 157/03, University Dortmund, November 2003.
[BBLP05] Thomas Bartz-Beielstein, Christian Lasarczyk, and Mike Preuss.
Sequential Parameter Optimization. In B McKay et al., editors,
Proceedings 2005 Congress on Evolutionary Computation (CEC’05),
Edinburgh, Scotland, pages 773–780, Piscataway NJ, 2005. IEEE
Press.
[BBPV04] Thomas Bartz-Beielstein, Konstantinos E Parsopoulos, and
Michael N Vrahatis. Design and analysis of optimization algorithms
using computational statistics. Applied Numerical Analysis and
Computational Mathematics (ANACM), 1(2):413–433, 2004.
[BBZ12] T Bartz-Beielstein and M. Zaefferer. A gentle introduction to sequen-
tial parameter optimization. Technical Report TR 01/2012, CIplus,
2012.
[BHKP10] Albert Bifet, Geoff Holmes, Richard Kirkby, and Bernhard Pfahringer.
Moa: Massive online analysis. J. Mach. Learn. Res., 11:1601–1604,
August 2010.

169
Bibliography

[BYHB10] Salah Bouhouche, Laib Laksir Yazid, Sissaoui Hocine, and Jürgen
Bast. Evaluation using online support-vector-machines and fuzzy
reasoning. Application to condition monitoring of speeds rolling
process. Control Engineering Practice, 18(9):1060 – 1068, 2010.

[CCWA03] J. Cohen, P. Cohen, S.G. West, and L.S. Aiken. Applied Multiple
Regression/Correlation Analysis for the Behavioral Sciences. Applied
Multiple Regression/ Correlation Analysis for the Behavioral Sciences.
L. Erlbaum Associates, 2003.

[CDK+ 06] Koby Crammer, Ofer Dekel, Joseph Keshet, Shai Shalev-Shwartz,
and Yoram Singer. Online passive-aggressive algorithms. J. Mach.
Learn. Res., 7:551–585, December 2006.

[CDSSS03] Koby Crammer, Ofer Dekel, Shai Shalev-Shwartz, and Yoram Singer.
Online passive-aggressive algorithms. In Proceedings of the 16th
International Conference on Neural Information Processing Systems,
NIPS’03, pages 1229–1236, Cambridge, MA, USA, 2003. MIT Press.

[CL01] Chih-Chung Chang and Chih-Jen Lin. Training ν-support vector

classifiers: Theory and algorithms. Neural Comput., 13(9):2119–2147,
September 2001.

[CL11] Chih-Chung Chang and Chih-Jen Lin. LIBSVM: A library for

support vector machines. ACM Transactions on Intelligent Systems
and Technology, 2:27:1–27:27, 2011. Software available at http:
//www.csie.ntu.edu.tw/˜cjlin/libsvm.

[CP01] G. Cauwenberghs and T. Poggio. Incremental and decremental

support vector machine learning. In Advances in Neural Information
Processing Systems (NIPS*2000), volume 13, 2001.

[DSP+ 19] Jifei Deng, Jie Sun, Wen Peng, Yaohui Hu, and Dianhua Zhang.
Application of neural networks for predicting hot-rolled strip crown.
Applied Soft Computing, 78:119 – 131, 2019.

[DT97] J. Dennis and Virginia Torczon. Managing approximation models in

optimization. In N.M. Alexandrov, M.Y. Hussaini, I.C.A.S. Engineer-
ing, and L.R. Center, editors, Multidisciplinary Design Optimization:
State of the Art, Proceedings in Applied Mathematics Series, pages
330–347. SIAM, 1997.

[EF11] Dirk Eddelbuettel and Romain François. Rcpp: Seamless R and

C++ integration. Journal of Statistical Software, 40(8):1–18, 2011.

170
 Bibliography

[Enc08] The Concise Encyclopedia of Statistics. Springer New York, New

York, NY, 2008.
[FF09] Andreas Vollmer Frank Feldmann, Mark Gerdau. Adaptive setup
models for cold rolling mills. https://library.e.abb.com/
public/77efb4c9677e46c5c125759100274942/43-48%201M924_
ENG72dpi.pdf, 2009. Accessed: 2019-02-18.
[FSK07] Alexander Forrester, András Sóbester, and Andy Keane. Multi-
fidelity optimization via surrogate modelling. Proceedings of the
Royal Society A: Mathematical, Physical and Engineering Science,
463(2088):3251–3269, 2007.
[FSK08] Alexander Forrester, András Sóbester, and Andy Keane. Engineering
Design via Surrogate Modelling. Wiley, 2008.
[GB00] V.B. Ginzburg and R. Ballas. Flat Rolling Fundamentals. Manu-
facturing Engineering and Materials Processing. Taylor & Francis,
2000.
[GE03] Isabelle Guyon and André Elisseeff. An introduction to variable and
feature selection. J. Mach. Learn. Res., 3:1157–1182, 2003.
[GG01] Jacek Gondzio and Andreas Grothey. Reoptimization with the
primal-dual interior point method. SIAM Journal on Optimization,
13:842–864, 01 2001.
[GJ+ 10] Gaël Guennebaud, Benoı̂t Jacob, et al. Eigen v3. http://eigen.
tuxfamily.org, 2010.
[GKR+ 01] Martin Grötschel, Sven O. Krumke, Jörg Rambau, Thomas Winter,
and Uwe T. Zimmermann. Combinatorial Online Optimization
in Real Time, pages 679–704. Springer Berlin Heidelberg, Berlin,
Heidelberg, 2001.
[Gon98] Jacek Gondzio. Warm start of the primal-dual method applied in the
cutting-plane scheme. Mathematical Programming, 83(1):125–143,
Jan 1998.
[GvB+ 14] João Gama, Indṙe Žliobaiṫe, Albert Bifet, Mykola Pechenizkiy, and
Abdelhamid Bouchachia. A survey on concept drift adaptation. ACM
Comput. Surv., 46(4):44:1–44:37, 2014.
[H+ 72] Milan Hajduk et al. Effect of improper selection of the rpm of
vertical and horizontal drives on balanced rolling force distribution
in a universal rolling mill. Hutnicke Listy, 27(8):259, 1972.

171
Bibliography

[Haf16] Raphael T. Haftka. Requirements for papers focusing on new or

improved global optimization algorithms. Structural and Multidisci-
plinary Optimization, 54(1):1–1, 2016.
[HCL16] Chih-Wei Hsu, Chih-Chung Chang, and Chih-Jen Lin. A practical
guide to support vector classification. Technical report, Department
of Computer Science, National Taiwan University, 2016.
[HELR08] Manu Hietaniemi, Ulla Elsilä, Perttu Laurinen, and Juha Röning.
Defect prediction in hot strip rolling using ANN and SVM. In
Tenth Scandinavian Conference on Artificial Intelligence, SCAI 2008,
Stockholm, Sweden, May 26-28, 2008, pages 44–51, 2008.
[Hin03] Rolf Hinkfoth. Massivumformung. Wissenschaftsverlag, Aachen,
2003.
[HMR96] C.A. Hernandez, S.F. Medina, and J. Ruiz. Modelling austenite
flow curves in low alloy and microalloyed steels. Acta Materialia,
44(1):155 – 163, 1996.
[HNK14] U. K. Hassan, N. M. Nawi, and S. Kasim. Classify a protein do-
main using sigmoid support vector machine. In 2014 International
Conference on Information Science Applications (ICISA), pages 1–4,
2014.
[HS78] Arno Hensel and Thilo Spittel. Kraft- und Arbeitsbedarf bildsamer
Formgebungsverfahren. Verlag Grundstoffindustrie, 1978.
[HTF01] Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The Ele-
ments of Statistical Learning. Springer Series in Statistics. Springer
New York Inc., New York, NY, USA, 2001.
[HZEK72] Milan Hajduk, Milan Zidek, Jiri Elfmark, and Svatopluk Kopec.
Derivation of mean values of inherent deformation resistance in hot
rolling of tonnage steel. Hutnicke Listy, 27(8):567, 1972.
[Iko12] Elena Ikonomovska. Algorithms for Learning Regression Trees and
Ensembles on Evolving Data Streams. PhD thesis, Jožef Stefan
International Postgraduate School Ljubljana, Slovenia, 2012. http:
//kt.ijs.si/elena_ikonomovska/00-disertation.pdf.
[JAF+ 06] Manuel Baena-Garca Jose, Jose Del Campo Avila, Raul Fidalgo,
Albert Bifet, Ricard Gavalda, and Rafael Morales bueno. Early drift
detection method. http://www.lsi.upc.edu/˜abifet/EDDM.pdf,
2006.

172
 Bibliography

[JBBR18] C Jung, T Bartz-Beielstein, and G Rudolph. Extending support

vector regression for on-line learning methods on real-world data.
(To appear), 2018.
[Jin03] Y Jin. A comprehensive survey of fitness approximation in evolu-
tionary computation. Soft Computing, 9(1):3–12, October 2003.
[Joh14] Steven G. Johnson. The nlopt nonlinear-optimization package. 2014.
R package version 1.0.4.
[Joh18] Steven G. Johnson. The NLopt nonlinear-optimization package.
hhttp://ab-initio.mit.edu/nlopt, 2018. Accessed: 2018-10-20.
[Jon01] Donald R Jones. A Taxonomy of Global Optimization Methods Based
on Response Surfaces. Journal of Global Optimization, 21:345–383,
2001.
[JSW98] D R Jones, M Schonlau, and W J Welch. Efficient Global Op-
timization of Expensive Black-Box Functions. Journal of Global
Optimization, 13:455–492, 1998.
[JWHT14] G. James, D. Witten, T. Hastie, and R. Tibshirani. An Introduction
to Statistical Learning: with Applications in R. Springer Texts in
Statistics. Springer New York, 2014.
[JZBBR17] Christian Jung, Martin Zaefferer, Thomas Bartz-Beielstein, and
Günter Rudolph. Metamodel-based optimization of hot rolling pro-
cesses in the metal industry. The International Journal of Advanced
Manufacturing Technology, 90(1):421–435, 2017.
[Kar25] T.V. Karman. Beitrag zur Theorie des Walzvorganges. Zeitschrift
für Angewandte Mathematik und Mechanik, 5:139–141, 1925.
[Kar92] Richard M. Karp. On-line algorithms versus off-line algorithms: How
much is it worth to know the future? In Proceedings of the IFIP 12th
World Computer Congress on Algorithms, Software, Architecture -
Information Processing ’92, Volume 1 - Volume I, pages 416–429,
Amsterdam, The Netherlands, The Netherlands, 1992. North-Holland
Publishing Co.
[Kat17] Gauri Katiyar. Off-line handwritten character recognition system
using support vector machine. American Journal of Neural Networks
and Applications, 3:22, 01 2017.
[Kli04] Ralf Klinkenberg. Learning drifting concepts: Example selection vs.
example weighting. Intell. Data Anal., 8(3):281–300, 2004.

173
Bibliography

[Kru02] Sven O. Krumke. Online optimization: Competitive analysis and

beyond. (02-25), 2002.
[KW99] R. Kopp and H. Wiegels. Einführung in die Umformtechnik. Ver-
lagsgruppe Mainz, 1999.
[LGK+ 06] Pavel Laskov, Christian Gehl, Stefan Krüger, Klaus-Robert Müller,
Kristin Bennett, and Emilio Parrado-Hern. Incremental support
vector learning: Analysis, implementation and applications. Journal
of Machine Learning Research, 7:1909–1936, 2006.
[LMZZ15] Yuan-Ming Liu, Geng-Sheng Ma, Dian-Hua Zhang, and De-Wen
Zhao. Upper bound analysis of rolling force and dog-bone shape
via sine function model in vertical rolling. Journal of Materials
Processing Technology, 223:91 – 97, 2015.
[LRT01] Perttu Laurinen, Juha Röning, and Harri Tuomela. Steel slab temper-
ature modelling using neural and bayesian networks. Soft computing
and intelligent systems for industry (SOCO/ISFI 2001). Paisley,
Scotland, UK, 2001.
[LZHM12] Jega Laxmi, Jie Zhang, Mohd Hussain, and Julian Morris. Batch-
to-batch iterative learning control using updated models based on a
moving window of historical data. Procedia Engineering, 42:206–213,
09 2012.
[Mar02] Mario Martin. On-line support vector machine regression. In Tapio
Elomaa, Heikki Mannila, and Hannu Toivonen, editors, Machine
Learning: ECML 2002: 13th European Conference on Machine
Learning Helsinki, Finland, August 19–23, 2002 Proceedings, pages
282–294, Berlin, Heidelberg, 2002. Springer Berlin Heidelberg.
[Mis18] M. Miskowicz. Event-Based Control and Signal Processing. Embed-
ded Systems. CRC Press, 2018.
[MRS+ 09] Sumantra Mandal, V. Rakesh, P.V. Sivaprasad, S. Venugopal, and
K.V. Kasiviswanathan. Constitutive equations to predict high tem-
perature flow stress in a ti-modified austenitic stainless steel. Mate-
rials Science and Engineering: A, 500(1):114 – 121, 2009.
[MTP03] Junshui Ma, James Theiler, and Simon Perkins. Accurate on-line
support vector regression. Neural computation, 15:2683–2703, 2003.
[NM65] J A Nelder and R Mead. A Simplex Method for Function Minimiza-
tion. Computer Journal, 7:308–313, 1965.

174
 Bibliography

[OANY80] M Okada, T Ariizumi, Y Noma, and Y Yamazaki. On the behavior

of edge rolling in hot strip mills. In International Conference on
Steel Rolling, volume 1, pages 275–286, 1980.
[OJB11] Olufemi Omitaomu, Myong Jeong, and Adedeji Badiru. Online
support vector regression with varying parameters for time-dependent
data. Systems, Man and Cybernetics, Part A: Systems and Humans,
IEEE Transactions on, 41:191 – 197, 2011.
[OMBH07] Olufemi Omitaomu, Jeong Myong, Adedeji Badiru, and J. Hines.
Online support vector regression approach for the monitoring of
motor shaft misalignment and feedwater flow rate. Systems, Man, and
Cybernetics, Part C: Applications and Reviews, IEEE Transactions
on, 37:962 – 970, 2007.
[ÖÖ00] Can Özsoy and Ertan Öznergiz. A self-tuning thickness control in
plate hot-rolling. IFAC Proceedings Volumes, 33(22):233 – 238, 2000.
IFAC Workshop on Future Trends in Automation of the Mineral and
Metal Processing (MM 2000), Finland.
[Pow88] M J D Powell. A Review of Algorithms for Nonlinear Equations and
Unconstrained Optimization. In Proceedings ICIAM, pages 220–232,
1988.
[Pow94] M. J. D. Powell. A direct search optimization method that models the
objective and constraint functions by linear interpolation. In Susana
Gomez and Jean-Pierre Hennart, editors, Advances in Optimization
and Numerical Analysis, pages 51–67, Dordrecht, 1994. Springer
Netherlands.
[PPP17] Kedar Potdar, Taher Pardawala, and Chinmay Pai. A comparative
study of categorical variable encoding techniques for neural network
classifiers. International Journal of Computer Applications, 175:7–9,
2017.
[PTVF92] William H. Press, Saul A. Teukolsky, William T. Vetterling, and
Brian P. Flannery. Numerical Recipes in C (2nd Ed.): The Art of
Scientific Computing. Cambridge University Press, New York, NY,
USA, 1992.
[R C17] R Core Team. R: A Language and Environment for Statistical
Computing. R Foundation for Statistical Computing, Vienna, Austria,
2017.
[RH95] Brian D. Ripley and N. L. Hjort. Pattern Recognition and Neural

175
Bibliography

Networks. Cambridge University Press, New York, NY, USA, 1st

edition, 1995.
[Row90] Thomas Harvey Rowan. Functional Stability Analysis of Numerical
Algorithms. PhD thesis, Austin, TX, USA, 1990. UMI Order No.
GAX90-31702.
[RPS17] Matruprasad Rout, Surjya Kanta Pal, and Shiv Brat Singh. Finite
element analysis of cross rolling on aisi 304 stainless steel: Prediction
of stress and strain fields. Journal of The Institution of Engineers
(India): Series C, 98(1):27–35, 2017.
[RZD12] R. Ren, X. P. Zhang, and X. Z. Du. FEM Simulation of the Slab
Edge Rolling in Unsteady Rolling Stage. Applied Mechanics and
Materials, 152:297–300, January 2012.
[SG86] Jeffrey C. Schlimmer and Richard H. Granger. Incremental learning
from noisy data. Machine Learning, 1(3):317–354, 1986.
[SK96] Dirk FJ Staalman and Arend Kusters. On-line slab temperature
calculation and control. Manufacturing Science and Engineering,
4:307–314, 1996.
[SK08] Sudipta Sikdar and Sabita Kumari. Neural network model of the
profile of hot-rolled strip. International Journal of Advanced Manu-
facturing Technology, 42:450–462, 2008.
[SLSJ02] Amnon Shirizly, J. Lenard, J. Sauer, and K. Januszkiewicz. Lu-
bricant capture during hot rolling of an aluminum alloy. Tribology
Transactions - TRIBOL TRANS, 45:205–210, 04 2002.
[Søn03] Jacob Søndergaard. Optimization Using Surrogate Models—by the
Space Mapping Technique. PhD thesis, Technical University of Den-
mark, Richard Petersens Plads, DK-2800 Kgs. Lyngby, 2003.
[SS04] Alex J. Smola and Bernhard Schölkopf. A tutorial on support vector
regression. Statistics and Computing, 14(3):199–222, 2004.
[SS12] Shai Shalev-Shwartz. Online learning and online convex optimization.
Found. Trends Mach. Learn., 4(2):107–194, 2012.
[SS16] Marlene Spittel and Thilo Spittel. Influence of chemical composition
and forming conditions on flow stress. In Hans Warlimont, editor,
Part 3: Non-ferrous Alloys - Heavy Metals: Subvolume C: Metal
Forming Data - Volume 2: Materials - Group VIII:Advanced Materi-

176
 Bibliography

als and Technologies - Landolt-Börnstein New Series, pages 18–64,

Berlin, Heidelberg, 2016. Springer Berlin Heidelberg.
[SSWB00] Bernhard Schölkopf, Alex J. Smola, Robert C. Williamson, and
Peter L. Bartlett. New support vector algorithms. Neural Comput.,
12(5):1207–1245, 2000.
[ST17] Leslie N. Smith and Nicholay Topin. Super-convergence: Very fast
training of residual networks using large learning rates. CoRR,
abs/1708.07120, 2017.
[Str10] Gilbert Strang. Wissenschaftliches Rechnen. Springer, Berlin, Hei-
delberg, 2010.
[SWMW89] J Sacks, W J Welch, T J Mitchell, and H P Wynn. Design and
analysis of computer experiments. Statistical Science, 4(4):409–435,
1989.
[TC14] R. R. Draxler T. Chai. Root mean square error (rmse) or mean
absolute error (mae)? – arguments against avoiding rmse in the
literature. hhttps://www.geosci-model-dev.net/7/1247/2014/gmd-7-
1247-2014.pdf, 2014.
[THW+ 83] Masahiro Takeuchi, M Hoshiya, K Watanabe, O Hirata, T Kikuma,
and S Sadahiro. Heavy width reduction rolling of slabs. Nippon steel
technical report. Overseas, (21):235–246, 1983.
[TL03] David M. J. Tax and Pavel Laskov. Online SVM learning: From
classification to data description and back. In Proc. Neural Network
and Signal Processing, pages 499–508, 2003.
[TLWM13] S. Tang, Z.Y. Liu, G.D. Wang, and R.D.K. Misra. Microstructural
evolution and mechanical properties of high strength microalloyed
steels: Ultra fast cooling (ufc) versus accelerated cooling (acc). Ma-
terials Science and Engineering: A, 580:257 – 265, 2013.
[TNR81] A.I. Tselikov, G.S. Nikitin, and S.E. Rokotyan. The Theory of
Lengthwise Rolling. Mir Publishers, 1981.
[TOY+ 82] Hiromitsu Takei, Yoshihiro Onishi, Yoshimasa Yamasaki, Atsuhisa
Takekoshi, Masaharu Yamamoto, and Masaru Okado. Automatic
width control of rougher in hot strip mill. Nippon Kokan Technical
Report 34, Computer Systems Development Department Fukuyama
Works, 1982.
[TWW+ 18] Y. Tian, H. T. Wang, Z. D. Wang, R. D. K. Misra, and G. D.

177
Bibliography

Wang. Microstructural evolution and the precipitation behavior in

x90 linepipe steel during isothermal processing. Journal of Materials
Engineering and Performance, 27(4):1494–1504, 2018.
[Vap95] Vladimir N. Vapnik. The Nature of Statistical Learning Theory.
Springer-Verlag New York, Inc., New York, NY, USA, 1995.
[VGS97] Vladimir Vapnik, Steven E. Golowich, and Alex J. Smola. Support
vector method for function approximation, regression estimation and
signal processing. In Michael Mozer, Michael I. Jordan, and Thomas
Petsche, editors, Advances in Neural Information Processing Systems
9 — Proceedings of the 1996 Neural Information Processing Systems
Conference (NIPS 1996), pages 281–287, Dever, CO, USA, December
1997. MIT Press, Cambridge, MA, USA.
[VL63] V Vapnik and A Lerner. Pattern recognition using generalized
portrait method. Automation and Remote Control, 24:774–780,
1963.
[Web73] Karheinz Weber. Grundlagen des Bandwalzens. VEB Deutscher
Verlag fuer Grundstoffindustrie, Leipzig, 1973.
[Wen04] Craig Wendorf. Primer on multiple regression coding: Common
forms and the additional case of repeated contrasts. Understanding
Statistics, 3:47–57, 2004.
[WK96] Gerhard Widmer and Miroslav Kubat. Learning in the presence of
concept drift and hidden contexts. Machine Learning, 23(1):69–101,
1996.
[WM97] David H. Wolpert and William G. Macready. No free lunch theorems
for optimization. IEEE Transactions on Evolutionary Computation,
1(1):67–82, 1997.
[WSDQS98] Harald Wehage, Ulrich Skoda-Dopp, Uwe Quitmann, and Wolfgang
Sauer. Beispiele für die Simulation von Stichplänen für das Warm-
flachwalzen. Stahl und Eisen, 11:103–111, 1998.
[WZZ16] Xiaoshu Wang, Zhijun Zhang, and Peng Zhang. Development and
technology of large thickness tmcp steel plate with 390mpa grade
used for engineering machinery. In HSLA Steels 2015, Microalloying
2015 & Offshore Engineering Steels 2015, pages 961–966, Cham,
2016. Springer International Publishing.
[YL08] Jue Zhong Y.C. Lin, Ming-Song Chen. Prediction of 42crmo steel

178
 Bibliography

flow stress at high temperature and strain rate. Mechanics Research

Communications, 35:142–150, 2008.
[You14] Peter C. Young. Recursive Estimation and Time-Series Analysis: An
Introduction for the Student and Practitioner. Springer Publishing
Company, Incorporated, 2nd edition, 2014.
[YS16] Hirohisa Kawamura Hiroyuki Yoshino Ryuji Yamamoto Shigeru
Ogawa Keiji Oogushi Yoshihiro Serizawa, Yasuyuki Takamachi. Heat
transfer technology for steel rolling process. http://www.nssmc.com/
en/tech/report/nssmc/pdf/111-14.pdf, 2016. Accessed: 2019-02-
13.
[Zei12] Matthew D. Zeiler. ADADELTA: an adaptive learning rate method.
CoRR, abs/1212.5701, 2012.
[Zhi10] Leonid V. Zhigilei. Introduction to the science and engineering
of materials - chapter 6. http://people.virginia.edu/˜lz2n/
mse209/Chapter6.pdf, 2010. Accessed: 2019-02-13.

179