A Visionary Way to Novel Process Optimization Techniques

The Transfer of a Process Modeling Language to the Neuronal Level

Norbert Gronau and Marcus Grum

Department of Business Informatics, esp. Processes and Systems, University of Potsdam,

August-Bebel-Strasse 89, 14482 Potsdam, Germany

ngronau@lswi.de, mgrum@lswi.de

Keywords:

Process Modeling, Artiﬁcial Intelligence, Machine Learning, Neuronal Networks, Knowledge Modeling

Description Language (KMDL), Process Simulation, Simulation Process Building, Process Optimization.

Abstract:

Modern process optimization approaches do build on various qualitative and quantitative tools, but are mainly

limited to simple relations in different process perspectives like cost, time or stock. In this paper, a new

approach is presented, which focuses on techniques of the area of Artiﬁcial Intelligence to capture complex

relations within processes. Hence, a fundamental value increase is intended to be gained. Existing modeling

techniques and languages serve as basic concepts and try to realize the junction of apparently contradictory

approaches. This paper therefore draws a vision of promising future process optimization techniques and

presents an innovative contribution.

1 INTRODUCTION

A great potential of Artiﬁcial Neural Networks (short:

ANN) is well known since nearly four decades. In

general, those techniques copy the capabilities and

working behavior of the brain in simulating a network

of simple nerve cells. Early ANN architectures go

back to the 1940s and numerous improvements can

be found in late 1980 - 2000 (Schmidhuber, 2015).

Because of their ability to learn non-linear relations,

to generalize correctly and to built biologically moti-

vated efﬁciently working structures, ANN have been

applied successfully in various contexts such as mu-

sic composition, banking issues, medicine, etc. Even

simple processes have been modeled on behalf of

ANN (Chambers and Mount-Campbell, 2000).

Nowadays, in times of big data, enormous

amounts of data are available and the computing

power has increased immensely and with this, the

possibility to create bigger and more complex net-

works. Although, the collection of processing data

has become easy, the neuronal modeling and decod-

ing of complex processes has not been realized, yet.

Hence, the following research will focus on deep

learning with ANN with the intention to answer the

following research question: ”How can the capabil-

ity to create efﬁciently working structures of ANN be

used for process optimizations?” This paper intends

not to draw an all-embracing description of concrete

technical realizations of those novel process optimiza-

tion techniques. It intends to set a ﬁrst step to real-

ize the conjunction of the process modeling and opti-

mization world on the one hand and the ANN world

on the other hand, such that a sub research question

is: ”How can a process modeling language be trans-

ported on a neuronal level?”

In the following, a Neuronal Process Modeling is

referred to as the modeling of processes on a neuronal

level with a common process modeling language,

the reinterpretation of the common process modeling

based on that understanding as well as their differ-

ence quantity. The Neuronal Process Simulation is

referred to as the process simulation of common pro-

cess models considering ANN as knowledge model of

process participants (persons and machines), the sim-

ulation of common process models reinterpreted as

deep neuronal network and their difference quantity.

The Neuronal Process Optimization is referred to as

common process optimization techniques that are re-

alized on a neuronal level (e.g. double-loop learning

on a neuronal level), process optimizations that can be

realized because of the learning capabilities of ANN

in the domain of common process models as well as

their difference quantity. Within this paper, a focus

lays on the Neuronal Process Modeling.

The research approach is intended to be design-

oriented as Peffers proposes (Peffers et al., 2006; Pef-

fers et al., 2007), such that the paper is structured as

follows: Section 2 presents underlying concepts; Sec-

tion 3 derives objectives for a Neuronal Process Mod-

eling; Section 4 provides the design, followed by its

demonstration (Section 5) and evaluation (Section 6);

Finally, Section 7 concludes the paper.

2 UNDERLYING CONCEPTS

Starting with the selection of a modeling approach

and the question, how processes can be optimized in

the ﬁrst subsection, the second subsection refers to

underlying knowledge generation concepts. A further

subsection introduces ANN.

2.1 Process Optimization

Following the fundamental procedure model for sim-

ulation studies of Gronau (2017), a model creation is

realized after the modeling purpose has been deﬁned,

analyzed and corresponding data has been collected.

Hence, the following starts with modeling issues. Af-

terwards, as the model is valid, simulation studies

are realized and results collected, analyzed and inter-

preted. As changes or optimizations are required, ad-

justments are deﬁned and simulations tested as long

as a sufﬁcient solution has been identiﬁed. This will

be realized.

The following starts with the understanding of

process models to be a homomorphous mapping of

a system that reduces the complexity of the real world

with respect to the modeling objectives (Gronau,

2016). According to Krallmann et al. (2001), a sys-

tem to be modeled consists of an amount of system

elements, that are connected with an amount of sys-

tem relations. As it is limited by a system border,

the system environment and the system are connected

with an interface to exchange system input and system

output.

For the modeling of systems, several process

modeling languages can be used. Considering or-

ganizational, behavior-oriented, informational and

knowledge-oriented perspectives, Sultanow et al.

(2012) identify the Knowledge Modeling Description

Language (short: KMDL) to be superior in the com-

parison of twelve common modeling approaches.

Because of the analogy with a human brain as

knowledge processing unit, especially a knowledge

process modeling is focused. Here, Remus gives an

overview of existing modeling methods and a com-

parison of their ability to represent knowledge (Re-

mus, 2002). ARIS, EULE2, FORWISS, INCOME,

PROMOTE and WORKWARE are only some repre-

sentatives. Again, the KMDL can be identiﬁed to

be superior because of its ability to overcome lacks

in visualizations and analyses through the combina-

tion of several views such as the process view, activ-

ity view and communication view (Gronau and Maas-

dorp, 2016).

This language has been developed over more than

ten years iteratively. Having collected experiences

in numerous projects of numerous application areas

such as software engineering, product development,

quality assurance and investment good sales, the evo-

lution of the KMDL can be found in (Gronau, 2012).

Currently, the development of a third version is in

progress (Gronau et al., 2016b). In addition to the

modeling language, the KMDL reaches a fully devel-

oped research method which is described by (Gronau,

2009) in detail.

With its strengths in visualization and the focus

of knowledge generation, the KMDL seems attractive

for a transfer to the neuronal level. To the best of

our knowledge, such a transfer has not been realized

yet in any other process modeling language. With its

intention to focus on the generation of knowledge fol-

lowing (Nonaka and Takeuchi, 1995) and to transfer

the learning potential of ANN, the KMDL enables the

modeling of tacit knowledge bases and single or nu-

merous knowledge transfers beside common process-

ing issues. Hence, the KMDL is selected as modeling

language for the demonstration in section 5. The cur-

rent paper builds on the wide spread KMDL version

2.2 (Gronau and Maasdorp, 2016).

Once, a valid process model has been created, a

dynamic process can be simulated. Aiming to gain

insights within a closed simulation system, the inten-

tion is to transfer them to reality. For this, the follow-

ing pre-conditions have to be fulﬁlled: process mod-

els have to provide completeness. This includes the

registration of input data such as time, costs, partici-

pants, etc. Further, process models have to provide in-

terpretability of decisions. Here, values of variables,

state change conditions and transfer probabilities are

included. Further, meta information have to be con-

sidered, as for example the number of process real-

izations within a simulation. Beneath further objec-

tives, the following can be evaluated quickly and at

low costs: current sequences of operations, as well as

plans and process alternatives. Those evaluations can

be realized before expensive adjustments within cur-

rent process models are carried out (Gronau, 2017).

2.2 Knowledge Representation

Nonaka and Takeuchi distinguish between ex-

plicit knowledge and tacit knowledge (Nonaka and

Takeuchi, 1995). While the ﬁrst can be verbalized

Seventh International Symposium on Business Modeling and Software Design

and externalized easily, the second is hard to detect.

On-building, the following four knowledge conver-

sion types can be distinguished:

• An internalization is the process of integration of

explicit knowledge in tacit knowledge. Here, ex-

periences and aptitudes are integrated in existing

mental models.

• A socialization is the process of experience ex-

change. Here, new tacit knowledge such as com-

mon mental models or technical ability are cre-

ated.

• An externalization is the process of articulation

of tacit knowledge in explicit concepts. Here,

metaphors, analogies or models can serve to ver-

balize tacit knowledge.

• A combination is the process of connection of

available explicit knowledge, such that a new ex-

plicit knowledge is created. Here, a reorganiza-

tion, reconﬁguration or restructuring can result in

new explicit knowledge.

With the intention to focus on the potentials

of human brains and its generation of knowledge,

the knowledge generation concepts of (Nonaka and

Takeuchi, 1995) seem attractive for the modeling on

a neuronal level. Further, the KMDL builds on them,

which is selected for demonstration purposes.

2.3 Neuronal Networks

Originally, neural networks were designed as mathe-

matical models to copy the functionality of biologi-

cal brains. First researches were done by (Rosenblatt,

1963), (Rumelhart et al., 1986) and (McCulloch and

Pitts, 1988). As the brain connects several nerve cells,

so called neurons, by synapses, those mathematical

networks are composed of several nodes, which are

related by weighted connections. As the real brain

sends electrical activity typically as a series of sharp

spikes, the mathematical activation of a node repre-

sents the average ﬁring rate of these spikes.

As human brains show very complex structures

and are confronted with different types of learn-

ing tasks (unsupervised, supervised and reinforce-

ment learning) various kinds of networking struc-

tures have established, which all have advantages

for a certain learning task. There are for example

Perceptrons (Rosenblatt, 1958), Hopﬁeld Nets (Hop-

ﬁeld, 1982), Multilayer Perceptrons (Rumelhart et al.,

1986), (Werbos, 1988), (Bishop, 1995), Radial Ba-

sis Function Networks (Broomhead and Lowe, 1988)

and Kohonen maps (Kohonen, 1989). Networks con-

taining cyclic connections are called feedbackward or

recurrent networks.

The following focuses on Multilayer Perceptrons

and recurrent networks being confronted with super-

vised learning tasks. Here, input and output val-

ues are given and a learning can be carried out in

minimizing a differentiable error function by adjust-

ing the ANN’s weighted connections. For this, nu-

merous gradient descent methods can be used, such

as Backpropagation (Plaut et al., 1986) and (Bishop,

1995), PROP (Riedmiller and Braun, 1993), quick-

prop (Fahlman, 1989), conjugate gradients (Hestenes

and Stiefel, 1952), (Shewchuk, 1994), L-BFGS (Byrd

et al., 1995), RTRL (Robinson and Fallside, 1987)

and BPTT (Williams and Zipser, 1995). As the weight

adjustment can be interpreted as a small step in an op-

timization direction, the ﬁx step size can be varied to

reduce great errors quickly. The learning rate decay

can be used to reduce small errors efﬁciently and a

momentum can be introduced to avoid local optima.

In this stepwise optimization, analogies to continuous

process optimizations can be found (see section 2.1).

Since neuronal networks model human brains and

model the knowledge of a certain learning task, the

following refers to neuronal networks as neuronal

knowledge models.

3 OBJECTIVES OF A NEURONAL

PROCESS MODELING

As one assumes to have a given process model and

one aims to consider a neuronal network as a process

participant’s knowledge model within the simulation

of that process model, the following objectives have

to be considered coming from a modeling side:

1. Neuronal knowledge models have to be integrated

within existing process models.

2. The same neuronal knowledge models have to be

able to be integrated several times within a pro-

cess model.

3. Neuronal knowledge models have to be integrated

within process simulations.

4. Modeled environmental factors (material such as

non-material objects) have to be integrated with

considered knowledge models.

5. Outcomes (materialized such as non-

materialized) of considered knowledge models

have to be considered within the process model.

Further, objectives have to be considered coming

from a neuronal techniques side:

6. Neuronal tasks have to be considered following its

neurons biological models.

A Visionary Way to Novel Process Optimization Techniques - The Transfer of a Process Modeling Language to the

Neuronal Level

7. Parallel neuronal task realizations have to be con-

sidered within neuronal networks.

8. Time-dependent neuronal behaviors have to be

considered within neuronal networks.

9. Sequential neuronal task realization have to be

considered within neuronal networks.

10. Different levels of neuronal task abstractions have

to be considered in the neuronal process modeling

and simulation.

11. Sensory information and knowledge ﬂows have to

be considered within the modeled neuronal net-

work.

12. Actuator information and knowledge have to be

considered as outcomes of neuronal networks.

Each identiﬁed objective of those domains is rel-

evant for the transfer of a process modeling language

and serves as input for the following sections.

4 DESIGN OF A NEURONAL

PROCESS MODELING

The following gives deﬁnitions of the concept of neu-

ronal modeling. For this, basic deﬁnitions are given

ﬁrstly and on-building deﬁnitions are given after-

wards.

Neuronal knowledge objects are deﬁned to be neu-

ronal patterns, that evolve as current over a certain pe-

riod of time that causes a speciﬁc behavior of consec-

utive neurons. Those patterns can reach from single

time steps to long periods of time.

Neuronal information objects are deﬁned to be

neuronal currents, that serve as interface from and to

the environment such as incoming sensory informa-

tion and outgoing actuator information. Here, stored

information is included as well.

Considering those objects, a neuronal conversion

is deﬁned to be the transfer of neuronal input objects

to neuronal output objects. In accordance to (Nonaka

and Takeuchi, 1995), the following neuronal conver-

sion types can be distinguished:

• A neuronal internalization is the process of inte-

gration of explicit knowledge (neuronal informa-

tion objects) in tacit knowledge. Here, experi-

ences and aptitudes are integrated in existing men-

tal models.

• A neuronal socialization is the process of expe-

rience exchange between neurons within a closed

ANN. Here, new tacit knowledge such as com-

mon mental models or technical abilities are cre-

ated.

• A neuronal externalization is the process of artic-

ulation of tacit knowledge (neuronal knowledge

objects) in explicit concepts (neuronal informa-

tion objects). Here, patterns can serve to verbalize

tacit knowledge.

• A neuronal combination is the process of con-

nection of available explicit knowledge (neuronal

information objects), such that a new explicit

knowledge is created. Here, a reorganization, re-

conﬁguration or restructuring can result in new

explicit knowledge.

Neuronal input objects are deﬁned to be sensory

information objects and knowledge objects.

Neuronal output objects are deﬁned to be actuator

information objects and knowledge objects.

An atomic neuronal conversion is deﬁned to be a

neuronal conversion considering only one input ob-

ject and only one output object.

Complex neuronal conversion are deﬁned to be

neuronal conversions considering at least three neu-

ronal objects of one neuron. Pure complex neuronal

conversions do consider only one neuronal conversion

type, while impure complex neuronal conversion do

consider several neuronal conversion types such that

one is not able to distinguish them.

Abstract neuronal conversion are deﬁned to be

neuronal conversions considering neuronal objects of

more than one transferring neuron such that one is not

able to identify participating neurons.

All together, those deﬁnitions are the basis for the

transfer of a process modeling languages to the neu-

ronal level.

5 DEMONSTRATION OF THE

NEURONAL PROCESS

MODELING

The following subsections show the realization of the

neuronal process modeling on behalf of the KMDL.

For this, theoretic examples and corresponding pro-

cess process models are given, that visualize basic

deﬁnitions. Then, practical examples follow.

5.1 Theoretical Examples

Deﬁnitions as they were given in section 4 are visual-

ized in the following three theoretical examples: First,

atomic knowledge conversions on a neuronal level

can be found in Figure 1.

In this Figure, one can see a neuronal socializa-

tion on the top left, a neuronal externalization on the

top right, a neuronal combination on the bottom right

Seventh International Symposium on Business Modeling and Software Design

Figure 1: Atomic neuronal conversions.

and a neuronal internalization on the bottom left. All

of them were visualized in the activity view of the

KMDL.

The entity of persons as process participants (yel-

low) was mapped to neurons who interact on a neu-

ronal level. In consequence, the entity of tacit knowl-

edge objects (purple) are connected to neurons. The

entity of the conversion (green) was mapped to the ac-

tivity of a neuron that generates new knowledge based

on the transfer of its input objects. The environment

as well as interaction possibilities with the environ-

ment are modeled with the entity of a database (white

rectangle). Further, neuronal information objects are

stored within a database. In consequence, the shape

of information objects (red) are connected to those

databases.

Second, complex neuronal conversions are visual-

ized in Figure 2.

Again, in this Figure, one can see a neuronal so-

cialization on the top left, a neuronal externalization

on the top right, a neuronal combination on the bot-

tom right and a neuronal internalization on the bottom

left. All of them were visualized in the activity view

of the KMDL.

Following the KMDL, conversions of the activity

view can be repeated without control ﬂow. Hence,

each neuron can develop several neuronal knowl-

edge objects or neuronal information objects over

time. Hence, modeled neuronal objects do represent

the identiﬁed current knowledge of a certain neuron.

Therefore, a strict sequence modeling therefore can

be realized with help of the listener concept or the

process view.

Third, an abstract neuronal conversion can be

found in Figure 3.

In this Figure, one can see several impure complex

conversions, which is the reason for the black color of

the visualized arrows, as the KMDL asks for. Since

more than one neuron (B1 and B2) are considered on

that process model, an abstract level of neuronal con-

versions has been visualized.

5.2 Practical Examples

Using basic deﬁnitions of a neuronal process model-

ing, their transfer to practical examples coming from

the industry is intended. The following gives four

practical examples. All of them serve as a fruitful

domain to visualize neuronal modelings, simulations

and optimizations.

A ﬁrst example focuses on the organization of

goods depots. Those can follow various strategies.

For example ﬁx places can hold reservations for cer-

tain goods. Alternatively, goods can get an arbitrary

place, which considers current free spaces. Here, the

human brain can serve as biological inspiration for

strategies to store memories and can optimize the de-

pot organization of goods.

A second example focuses on production pro-

cesses. Here, goods are not needed constantly. Mean-

while, they can be stored in goods depots or storage

areas. Once, they are needed, they can be brought

to the corresponding process step with help of trans-

portation elements (Gronau et al., 2016a). As they

are not required, a transportation element pauses and

buffers currently not needed goods. Alternatively,

materials can be considered as just-in-time inventory,

such that they do not have to be stored in expensive

goods depots. Here, the velocity of transportation el-

ements is adjusted in dependence to the production

order. Analogies can be found in the human brain. As

the storage of goods, the storage of memories can be

organized or vice versa. A short-term-memory (cur-

rent currencies) deals with neuronal knowledge ob-

jects similarly to just-in-time inventory. Here, neu-

ronal knowledge objects are used at consecutive neu-

rons as they are needed. Buffered goods are stored

within long-term-memories similar to goods depots.

Here, currencies are unlocked as they are needed

within the current process.

A third example focuses on specializations of pro-

duction machines. As production processes can be

considered as a single process network, machines are

A Visionary Way to Novel Process Optimization Techniques - The Transfer of a Process Modeling Language to the

Neuronal Level

Figure 2: Complex neuronal conversions.

part of them. Since machines can show high special-

izations, the organization of production processes can

be inspired by the organization of the human brain.

Here, certain areas are responsible for a certain task

and show high specializations as well. For example

the auditory cortex deals mainly with acoustic infor-

mation, the visual cortex mainly with optical informa-

tion, etc.

The best choice to realize the entire process model

is not always to realize all process parts in the own

company. As parts can be outsourced to external par-

ties, analogies can be found in the human brain as

well. Here, speed relevant actions can be initiated by

reﬂexes. This is efﬁcient since the realization of a full

cognitive task processing would be to slow. As an ex-

ample, one can imagine the start of a sprinkler system.

In case of a ﬁre, it was not sense full to create action

alternatives but start ﬁghting a ﬁre immediately like a

reﬂex.

6 EVALUATION

Faced with the demonstration artifacts of the previous

section, objectives of section 3 have been considered

as follows.

Objective 1 can be fulﬁlled as neuronal knowledge

models are modeled within the activity view charac-

terizing a certain person. Here, a decomposition rises

the process model granularity of the selected activity

and connects all neuronal process models with com-

Figure 3: Abstract neuronal conversion.

mon process models. Since the common activity view

characterizes a corresponding process task of the pro-

cess view, neuronal knowledge models are integrated

within existing process models. Since a neuronal net-

work characterizes entities of persons, a trained neu-

ronal network can be reused in any activity (objec-

tive 2). As neuronal knowledge models can be acti-

vated and can evolve over time, they can be integrated

within discrete process simulations easily (objective

3). From a common activity view modeled environ-

mental factors (material such as non-material objects)

serve as interface for the activity view on a neuronal

level. Hence, objective 4 and objective 5 are consid-

ered as well.

Further, objectives have been considered coming

from a neuronal techniques side as follows: As learn-

ing with neuronal networks is not affected by the

here presented concepts, neuronal tasks can follow

the neurons biological models (objective 6). A paral-

lel neuronal task realization within neuronal networks

has been considered (objective 7) as can be seen in

Figure 2 (neuronal socialization and neuronal exter-

nalization) and Figure 3. Here, at least two neurons

realize a parallel task processing. Objective 8 can

be met as soon as recurrent connections are consid-

ered within the neuronal process models. Then, time-

dependent neuronal behaviors are considered within

neuronal networks. A sequential neuronal task real-

ization within neuronal networks can be considered

within the neuronal process modeling (objective 9),

as presented activity views are characterizing corre-

sponding tasks of the process view. Since logical

control-ﬂow operators can be used here, a sequential

neuronal task processing can be modeled easily. Fur-

ther, a time-dependent behavior of a network mod-

eled within the activity view can result in a sequential

task processing. Objective 10 has been met as can

be seen in Figure 3. Here, the task ”Neuronal Per-

ception of Neuron Group B” models the activity of

Neuron B1 and Neuron B2 on an abstract level. Fur-

Seventh International Symposium on Business Modeling and Software Design

ther, knowledge objects, information objects, neurons

and databases can be grouped and visualized on an

abstract level. Sensory information and knowledge

ﬂows can be considered within the modeled neuronal

network as can be seen in for example in Figure 1

and Figure 2. In both Figures, possible sensory in-

formation ﬂows can be seen on the bottom (neuronal

internalization and neuronal combination). Possible

knowledge ﬂows can be seen in both Figures on the

top (neuronal socialization and neuronal externaliza-

tion). Objective 12 can be met as follows: Actuator

information and knowledge have been considered as

outcomes of neuronal networks, as can be seen in Fig-

ure 1 and Figure 2. In both Figures, possible actuator

information ﬂows can be seen on the right (neuronal

externalization and neuronal combination). Possible

knowledge ﬂows can be seen in both Figures on the

left (neuronal socialization and neuronal internaliza-

tion).

Considering the here presented evaluation of

given objectives, it becomes clear that an idea for ev-

ery objective has been identiﬁed. This supports the

functioning of the transfer of the KMDL to the neu-

ronal level, such that a neuronal process modeling, a

neuronal process simulation and a neuronal process

optimization can be built on base of that.

7 CONCLUSIONS

In this paper, a visionary way to novel process opti-

mization techniques has been drawn and the base has

been realized on behalf of the KMDL. Main contribu-

tions and scientiﬁc novelties are the following: Deﬁ-

nitions of a neuronal process modeling, neuronal pro-

cess simulation and a neuronal process optimization

have been created. Objectives of a transfer of a com-

mon process modeling language have been identiﬁed.

Further, deﬁnitions for those concepts have been cre-

ated and a modeling language has been transferred to

the neuronal world. This includes the reinterpretation

of existing shapes of the KMDL. On that base, the-

oretical examples have been visualized on behalf of

the KMDL. Further, analogies for the use of the here

presented concepts in the industry context have been

identiﬁed.

With this, the drawn transfer has been applied and

proven. Hence, the sub research question was an-

swered and the following potentials are suitable next

steps:

The concretion of the functioning of previously

presented concepts will be realized. Then, those

will be realized as quantitative neuronal process mod-

elings, simulations and optimizations. Further, the

comparison of the here presented concepts with tra-

ditional results was attractive as well. Still promising

is the rebuilding of common process model optimiza-

tion on behalf of the here presented concepts.

The application of the here presented concepts

are assumed to cause a fundamentally value increase.

As simple and complex relations in different process

perspectives like cost, time or stock can be consid-

ered, the prediction quality of process simulations is

strongly improved. Further, common optimization

potentials can be estimated efﬁciently. Additionally,

new optimization approaches and optimization poten-

tials can be identiﬁed.

REFERENCES

Bishop, C. (1995). Neural Networks for Pattern

Recognition. Oxford University Press, Inc., ISBN

0198538642, page 1.

Broomhead, D. and Lowe, D. (1988). Multivariate func-

tional interpolation and adaptive networks. Complex

Systems, 2:321–355.

Byrd, R. H., Lu, P., Nocedal, J., and Zhu, C. Y. (1995).

A limited memory algorithm for bound constrained

optimization. SIAM Journal on Scientiﬁc Computing,

16(6):1190–1208.

Chambers, M. and Mount-Campbell, C. (2000). Pro-

cess optimization via neural network metamodel-

ing. International Journal of Production economics,

79(2000):93–100.

Fahlman, S. (1989). Faster learning variations on back-

propagation: An empirical study. Proceedings of

the 1988 connectionist models summer school, In D.

Touretszky, G. Hinton and T.Sejnowski, editors, San

Mateo, Morgan Kaufmann, pages 38–51.

Gronau, N. (2009). Process Oriented Management of

Knowledge: Methods and Tools for the Employment

of Knowledge as a Competitive Factor in Organiza-

tions (Wissen prozessorientiert managen: Methode

und Werkzeuge f

ur die Nutzung des Wettbewerbsfak-

tors Wissen in Unternehmen). Oldenbourg Verlag

unchen.

Gronau, N. (2012). Modeling and Analyzing knowledge in-

tensive business processes with KMDL - Comprehen-

sive insights into theory and practice. GITO mbH Ver-

lag Berlin.

Gronau, N. (2016). Gesch

aftsprozessmanagement in

Wirtschaft und Verwaltung. Gito.

Gronau, N. (2017). Gesch

aftsprozessmanagement in

Wirtschaft und Verwaltung, volume 2. Gito.

Gronau, N., Grum, M., and Bender, B. (2016a). Determin-

ing the optimal level of autonomy in cyber-physical

production systems. Proceedings of the 14th Interna-

tional Conference on Industrial Informatics (INDIN).

Gronau, N. and Maasdorp, C. (2016). Modeling of or-

ganizational knowledge and information : analyzing

A Visionary Way to Novel Process Optimization Techniques - The Transfer of a Process Modeling Language to the

Neuronal Level

knowledge-intensive business processes with KMDL.

GITO mbH Verlag Berlin.

Gronau, N., Thiem, C., Ullrich, A., Vladova, G., and Weber,

E. (2016b). Ein Vorschlag zur Modellierung von Wis-

sen in wissensintensiven Gesch

aftsprozessen. Tech-

nical report, University of Potsdam, Department of

Business Informatics, esp. Processes and Systems.

Hestenes, M. R. and Stiefel, E. (1952). Methods of con-

jugate gradients for solving linear systems. Jour-

nal of Research of National Bureau of Standards,

49(6):409–436.

Hopﬁeld, J. J. (1982). Neural networks and physical sys-

tems with emergent collective computational abilities.

PNAS, 79(8):2554–2558.

Kohonen, T. (1989). Self-organization and associativ mem-

ory. Springer-Verlag New York, Inc., New York, NY,

USA, ISBN 0-387-51387-6, 3rd edition:1.

Krallmann, H., Frank, H., and Gronau, N. (2001). Sys-

temanalyse im Unternehmen. Oldenbourg Wis-

senschaftsverlag.

McCulloch, W. S. and Pitts, W. (1988). A logical calcu-

lus of the ideas immanent in nervous activity. MIT

Press, Cambridge, MA, USA, ISBN 0-262-01097-6,

pages 15–27.

Nonaka, I. and Takeuchi, H. (1995). The knowledge-

creating company: How Japanese companies create

the dynamics of innovation. Oxford university press.

Peffers, K., Tuunanen, T., Gengler, C. E., Rossi, M., Hui,

W., Virtanen, V., and Bragge, J. (2006). The design

science research process: A model for producing and

presenting information systems reseach. 1st Interna-

tional Conference on Design Science in Information

Systems and Technology (DESRIST), 24(3):83–106.

Peffers, K., Tuunanen, T., Rothenberger, M. A., and Chat-

terjee, S. (2007). A design science research method-

ology for information systems research. Management

Informations Systems, 24(3):45–78.

Plaut, D. C., Nowlan, S. J., and Hinton, G. E. (1986). Exper-

iments on learning backpropagation. Technical Report

CMU-CS-86-126, Carnegie-Mellon University, Pitts-

burgh, PA, page 1.

Remus, U. (2002). Process-oriented knowledge manage-

ment. Design and modelling. PhD thesis, University

of Regensburg.

Riedmiller, M. and Braun, H. (1993). A direct adap-

tive method for faster backpropagation learning: The

RPROP algorithm. Proc. of the IEEE Intl. Conf. on

Neural Networks, San Francisco, CA, pages 586–591.

Robinson, A. J. and Fallside, F. (1987). The utility driven

dynamic error propagation network. Technical Report

CUED/F-INFENG/TR.1, Cambridge University Engi-

neering Department, page 1.

Rosenblatt, F. (1958). The perceptron: A probabilistic

model for information storage and organization in the

brain. Psychological Review, 65:386–408.

Rosenblatt, F. (1963). Principles of Neurodynamics. Spar-

tan, New York, page 1.

Rumelhart, D. E., Hinton, G. E., and Williams, R. J. (1986).

Learning internal representations by error propaga-

tion. MIT Press, Cabridge, MA, USA, ISBN 0-262-

68053-X, pages 318–362.

Schmidhuber, J. (2015). Deep learning in neural networks:

An overview. Neural Networks, 61:85 – 117.

Shewchuk, J. R. (1994). An introduction to the conjugate

gradient method without the agonizing pain. Technical

Report, Carnegie Mellon university, Pittsburgh, PA,

USA, page 1.

Sultanow, E., Zhou, X., Gronau, N., and Cox, S. (2012).

Modeling of processes, systems and knowledge: a

multi-dimensional comparison of 13 chosen meth-

ods. International Review on Computers and Software

(IRECOS), (6):3309–3319.

Werbos, P. J. (1988). Generalization of backpropagation

with application to a recurrent gas market model. Neu-

ral Networks, page 1.

Williams, R. J. and Zipser, D. (1995). Gradient-based learn-

ing algorithms for recurrent networks and their com-

putational complexity. In Y. Chauvin and D.E. Rumel-

hart, editors, Lawrence Erlbaum Publishers, Hillsdale,

N.J. Back-propagation: Theory, Architectures and

Applications, pages 433–486.

Seventh International Symposium on Business Modeling and Software Design