KREM: A Generic Knowledge-based Framework for Problem Solving in
Engineering
Proposal and Case Studies
Cecilia Zanni-Merk
INSA de Strasbourg / ICube Laboratory (BFO Team), Illkirch, France
Keywords:
Knowledge Technologies, Ontologies, Reasoning, Experience, Meta-knowledge.
Abstract:
This article presents a generic knowledge-based framework for problem solving in Engineering, in a broad
sense. After a discussion about the drawbacks of the traditional architecture used for deploying knowledge-
based systems (KBS), the KREM (Knowledge, Rules, Experience, Meta-Knowledge) architecture is presented.
The novelty of the proposal comes from the inclusion of experience capitalization and of meta-knowledge use
into the previously discussed traditional architecture. KREM improves the efficiency of classic KBSs, as it
permits to deal with incomplete expert knowledge models, by progressively completing them, learning with
experience. Also, the use of meta-knowledge can steer their execution more efficiently. This framework has
been successfully used in different projects. Here, the architecture of the KREM model is presented along
with some implementation issues and three case studies are discussed.
1 INTRODUCTION
Nowadays, organizations are becoming increas-
ingly knowledge-intensive and collaborative (Ouer-
tani et al., 2011). However, large parts of useful
knowledge are hidden and not readily available. A
growing number of companies have realized that a
tool that effectively enables the capture, representa-
tion, retrieval, and reuse of knowledge is the key to
supporting various organizational decisions (Daven-
port and Prusak, 2000; Leistner, 2010). And this is
one of the main interests of Knowledge Technologies.
Knowledge Technologies are computer-based
techniques and tools that provide a richer and more
intelligent use of information. Much of their power
comes from the way they combine ideas and applica-
tions from a number of fields: Psychology, Philos-
ophy, Artificial Intelligence, Engineering, Business
Studies, Computer Science and Web Technologies
(Milton, 2008).
Knowledge Technologies, as the name suggests,
are about doing things with knowledge. For example:
Identifying what knowledge is important to an or-
ganisation;
Deciding what knowledge needs to be captured
to provide an appropriate solution to a real-world
problem;
Capturing and integrating knowledge from ex-
perts or repositories (such as texts or databases,
for example);
Representing and storing knowledge in ways that
provide ease of access, navigation, understanding,
maintenance and re-use;
According to (Milton, 2008), Knowledge Tech-
nologies include: Knowledge Based Systems, Nat-
ural Language Processing, Data Mining, Semantic
Technologies, Case Based Reasoning and Intelligent
Agents, among others. We are interested in a subset
of the technologies presented by (Milton, 2008), but
use them in a slightly different way.
One of the main advantages of using Knowledge
Technologies for problem solving in Engineering is
that they facilitate decision-making. Better decisions
can be made when people have the right information
at the right moment. When information is out-of date,
incoherent, irrelevant, hard to find and hard to inte-
grate, then the decision-making process is made more
difficult. The information required to make good de-
cisions often exists but tends to be scattered in vari-
ous locations and stored in various formats (such as
databases or plain text). Knowledge Technologies
provide ways of finding, merging and exposing the
information required to make informed decisions.
Knowledge Technologies are dealt with methods
Zanni-Merk, C..
KREM: A Generic Knowledge-based Framework for Problem Solving in Engineering - Proposal and Case Studies.
In Proceedings of the 7th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2015) - Volume 2: KEOD, pages 381-388
ISBN: 978-989-758-158-8
Copyright
c
2015 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
381
coming from the field of Knowledge Engineering,
whose main goal is the development of Knowledge-
Based Systems (KBSs).
The classical architecture of a KBS (discussed in
Section 2) does not always provide satisfactory results
when solving a problem. In fact, the knowledge mod-
els needed for their implementation are often incom-
plete. Therefore, we propose here a novel architec-
ture to deal with these drawbacks by the inclusion of
the capitalization of the experience learnt with the use
of the KBS. The use of meta-knowledge is also inte-
grated to deal with the whole.
As stated above, Section 2 presents an introduc-
tion to knowledge-based systems. Section 3 deals
with the capitalization of experience. Sections 4 and 5
presents the KREM model with some notes about its
implementation. Section 6 describes three different
applications of KREM; and finally, Section 7 gives
our conclusions.
2 KNOWLEDGE-BASED
SYSTEMS
One of the main goals of using an approach based on
Knowledge Engineering methods for problem solving
is the development of KBSs. They have a compu-
tational model of some domain of interest in which
symbols serve as surrogates for real world domain
artefacts, such as physical objects, events, relation-
ships, etc (Sowa, 2000).
A knowledge-based system maintains a knowl-
edge base which stores the symbols of the com-
putational model in form of statements about the
domain, and performs reasoning by manipulating
these symbols. Applications can base their decisions
on domain-relevant questions posed to a knowledge
base. They are programmed to solve problems in a
similar way to that of an expert practitioner, e.g. make
inferences based on the case at hand and adopt the
right strategy to solve the problem; they can deal with
incomplete information by making requests for fur-
ther information or by making intelligent guesses at
the answer (just as an expert has to do when there
is limited information); they have a user interface that
makes intelligent requests for information and can ex-
plain how it has arrived at its answers; and finally, the
development can be made cost-effective by re-using
generic structures, rules and problem-solving meth-
ods.
The components usually found in a KBS are: a
knowledge base that contains domain-specific infor-
mation, structures and rules; a working memory (also
known as a blackboard) which holds case-specific
data (e.g. facts about the initial problem and inter-
mediate results); an inference engine (or reasoning
engine) that controls and directs the solving of prob-
lems by making inferences, i.e. it uses the knowledge
base to alter the contents of the working memory; an
interface to other computer systems and/or to human
users; and an editor that allows a knowledge engineer
or domain expert to inspect and update the knowledge
base.
The knowledge-base of a KBS needs a more ex-
plicit description. One of its components is a formal
conceptual model of the domain of interest. Among
the existing formal conceptual models, ontologies are
formalized representations of vocabularies that are
specific to a certain area. The dominating definition of
an ontology is the following: An ontology is a formal
explicit specification of a shared conceptualisation of
a domain of interest, based on (Gruber, 1993).
Technically, the principal constituents of an ontol-
ogy are concepts, relations and instances. They are
commonly used with a set of rules that are chained
to simulate the reasoning of a human expert. Rules
come in the form of ”if-then” constructs and allow to
express various kinds of complex statements and al-
low different types of reasoning about the facts in the
working memory based on the domain knowledge in
the ontology.
In summary, a knowledge based system contains a
conceptual yet executable model of an application do-
main. It is made machine-interpretable by means of
knowledge representation techniques and can there-
fore be used by applications to base decisions on rea-
soning about domain knowledge.
The traditional architecture that we have just dis-
cussed has drawbacks, associated, mainly, with the
difficulties that appear during the knowledge elici-
tation process from experts; and also with the non-
completeness of the formal conceptual model ob-
tained after the elicitation (Milton, 2008).
In fact, as the knowledge base can be incomplete,
there could be problems that this traditional architec-
ture cannot solve. Reasoning and analysis of this in-
complete knowledge implies that it is needed to take
advantage of the experience acquired from the inter-
ventions of human experts when the traditional sys-
tem does not lead to satisfactory results; that is, some
kind of capitalization of experience is needed.
3 ON THE CAPITALIZATION OF
EXPERIENCE
Set of Experience Knowledge Structure (SOEKS) is
an experience-based knowledge representation that
KEOD 2015 - 7th International Conference on Knowledge Engineering and Ontology Development
382
Figure 1: A SOEKS.
can store uncertain and incomplete data and make
qualitative and quantitative extractions of knowl-
edge from the available information, which can of-
ten be unstructured, semi-structured, fuzzy, and vague
(San
´
ın and Szczerbicki, 2009; San
´
ın et al., 2007).
Additionally, SOEKS can be shaped in an exten-
sive understandable and transportable language such
as Extensible Markup Language (XML) or Ontology
Web Language (OWL) (Grau et al., 2008). XML
and OWL representation of SOE allows knowledge
to be exchanged quickly and securely between appli-
cations and systems. The SOEKS has been designed
to collect experiences and knowledge from multiple
applications that are assembled as formal decision
events (San
´
ın and Szczerbicki, 2009). This collected
knowledge assists organizations in making precise de-
cisions, predictions, and recommendations, and it is a
dynamic structure that is dependent on the informa-
tion and data that it has received.
A SOEKS has four components (Figure 1) (San
´
ın
and Szczerbicki, 2009): variables, functions, con-
straints, and rules. Each formal decision event can
be stored in a combined structure of those four com-
ponents of the SOEKS.
Variables usually involve representing knowledge
using an attribute-value language. This is a tradi-
tional approach from the origin of knowledge repre-
sentation. Variables are related among them in the
shape of functions. Functions, the second component,
describe associations between variables. Therefore,
the set of experience uses functions and establishes
links among the variables constructing multi objec-
tive goals. Constraints are another form expressing
relationships among the variables. A constraint is a
restriction of the feasible solutions in a decision prob-
lem, and limits the performance of a system with re-
spect to its goals. Finally, rules are suitable for repre-
senting inferences or for associating actions with con-
ditions under which the actions should be performed.
They are conditional relationships of the universe of
variables.
4 THE KREM MODEL
Taking into account the points discussed in the pre-
vious sections, we propose a modular architecture,
called KREM (Knowledge, Rules, Experience and
Meta-Knowledge), to manage the complexity of de-
veloping KBSs, to try to solve the aforementioned
problems and to incorporate the capitalization of ex-
perience with the goal of improving decision-making.
Because to be effective, decision-making must re-
sult from reasoning and analysis of this knowledge,
also taking into account the experience and expertise
of decision-makers. As a consequence, it is needed
to capitalize them to take advantage of the experi-
ence acquired from the interventions of human ex-
perts when the traditional system does not lead to sat-
isfactory results.
The use of meta-knowledge to steer the ex-
ecution of the whole system is also necessary.
Meta-knowledge is knowledge about domain knowl-
edge, about rules or about experience. This meta-
knowledge can take the form of context, culture or
protocols to use this knowledge. Context is infor-
mation that characterizes a situation in relation to in-
teraction among human-beings, applications and their
environment, and can be of four types: identity, place,
status or time (Dey et al., 2001). Culture meta-
knowledge tries to take into account the fact that de-
cisions are made differently depending on the coun-
try or culture (Meyer, 2014). And finally protocols
may include strategies or problem-solving heuristics
for the task to be done (for example, in the case of
medical diagnosis, the protocols used by physicians
change according to the type of symptoms or the sus-
pected illness).
Therefore, the proposed components of the archi-
tecture are (Figure 2):
The Knowledge component that contains the do-
main knowledge to operate, by means of different
domain ontologies to be developed.
The Rules component that allows different types
of reasoning (monotone, spatial, temporal, fuzzy,
or other) depending on the application.
The Experience component that allows the capi-
talization and reuse of prior knowledge.
The Meta-knowledge component, including
knowledge about the other three bricks that
depends on the problem. For example, in a
medical diagnosis problem, this component may
include clinical protocols used by physicians.
The way the domain knowledge is formalized
will shape the way the rules are expressed. Experi-
ence will come complete the available knowledge and
KREM: A Generic Knowledge-based Framework for Problem Solving in Engineering - Proposal and Case Studies
383
Figure 2: The KREM architecture with its four interre-
lated components: knowledge, rules, experience and meta-
knowledge.
rules. Finally, meta-knowledge will directly interact
with the rules and the experience to indicate which
rules (coming from experience or from the initial rule
set) need to be launched according to the context of
the problem to solve.
A modular architecture, such as KREM, is one
of the main architectural design pattern for large and
complex systems. In this pattern, each module or
component has a specific functionality providing sep-
aration of concerns that, in turn, support reuse or re-
placement (i.e. changes in a single module would
not affect the others, permitting the continuous op-
eration of the system). Moreover the communication
between modules needs to be based on well defined
interfaces to provide low coupling.
The KREM model has already been successfully
used for very specific applications; the semantic anal-
ysis of urban satellite images, the diagnosis of small
and medium enterprises or the formalization of the
Lean enterprise. We will discuss these applications in
Section 6.
5 SOME NOTES ABOUT THE
IMPLEMENTATION
The implement the KREM architecture, it is neces-
sary, at least, to combine rules formalisms with de-
scription logics (this is the classic architecture of a
KBS). There are several possibilities to accomplish
this. But our choice has been to have hybrid systems
interfacing logic programming systems and descrip-
tion logic systems.
As explained before, the knowledge module con-
tains the domain ontologies that have been developed
for each project. These developments are made in
OWL (Grau et al., 2008). The Web Ontology Lan-
guage (OWL) is an integral component of the Seman-
tic Web, as it can be used to write ontologies or formal
vocabularies which form the basis for semantic web
data mark-up and exchange.
The Prot
´
eg
´
e
1
editor in several versions has been
our preferred choice for the design of the ontologies.
Prot
´
eg
´
e is usually used to construct domain mod-
els and knowledge-based applications with ontologies
because of possessing a Java-based Application Pro-
gramming Interface (API) and being easily extended
by a plug-in architecture.
Concerning the rules, we have also tested several
approaches. The first natural choice was the use of
SWRL (Semantic Web Rule Language).
SWRL is a proposal by the W3C to extend OWL
with rules. It is based on a combination of some
sub-languages of OWL with the Unary/Binary Data-
log RuleML sub-languages of the Rule Markup Lan-
guage
2
. The proposal extends the OWL abstract syn-
tax to include the syntax of these rules and the OWL
model-theoretic semantics to provide a formal mean-
ing for ontologies that include rules written in this
syntax (Golbreich, 2004).
However, some of the constraints of SWRL
(mainly the lack of input/output features and the im-
possibility to create new individuals in the ontologies
“on the fly”) made us begin investigating other possi-
bilities. And our choice went to Jess.
Jess (Java Expert System Shell) is a rule engine for
the Java platform, developed by Ernest Friedman-Hill
of Sandia National Labs since 1995. Jess supports the
development of rule-based expert systems which can
be tightly coupled to code written in Java language.
Compared with the traditional matching mechanism -
only one loop for all - the Jess rule engine continu-
ously matches facts against rules to infer conclusions,
which result in actions. Thus rules can modify the
collection of facts (Hill, 2003).
Unfortunately, Jess also suffers from some issues,
mainly regarding its licensing, so we are currently
investigating some other open-source inference en-
gines, such as Drools
3
.
Concerning the implementation of experience, we
are using the SOEKS in its ontology form. Figure
3 shows part of the SOEKS class hierarchy with the
four components involved in decision making events:
variables, functions, constraints and rules.
What is important to remark concerning the inter-
action between the experience module and the others
is that the reasoning process in the rules component
enables the evolution of the initial set of rules with
experience, as it was proposed by (Toro et al., 2012).
More precisely, every time the native reasoner is ex-
ecuted, the system will store both the output of the
reasoner and the final decision, if any, made by the
1
http://protege.stanford.edu
2
http://www.ruleml.org.
3
http://www.drools.org
KEOD 2015 - 7th International Conference on Knowledge Engineering and Ontology Development
384
Figure 3: The SOEKS ontology.
expert. Decisions that do not follow the proposition
of the reasoner produce an evolution of the set of ex-
perience rules.
Finally, concerning Meta-Knowledge, we are cur-
rently investigating the implementation of context, in
a broad sense. Context has been studied carefully for
a long time in areas like knowledge-based and ubiqui-
tous systems, either for handling the complex knowl-
edge in a dynamic manner (McCarthy, 1993; Sowa,
1995; Giunchiglia et al., 2012) or to provide smarter
human interfaces (Dey et al., 2001). However, there
is no agreement on a concise definition of context in
these areas. (Bazire and Br
´
ezillon, 2005), for exam-
ple, proposed the components of context as the user
and the observer involved, as well as the items, the en-
vironment, and other related contexts. (Porzel, 2010)
states that context helps when detecting semantic re-
lations to provide extra information and correct inter-
pretations for applications.
Diverse representations of context exist in differ-
ent research areas. (McCarthy, 1993) uses a term c
representing context, and the formal representation
of a proposition p is true in the context c is repre-
sented as ist(c, p). As described above, (Dey et al.,
2001) define context as any information that char-
acterizes a situation for context-aware applications,
while (Porzel, 2010) refers to a context work, where a
model of context contains components and the differ-
ent relations of the components. The components are
the user, an item, and the observer in the environment.
Relations here include not only the relations between
the components, but also the relations to other con-
texts.
We choose to represent context in the meta-
knowledge module in a way similar to the one used
by (McCarthy, 1993). An ontology enables the iden-
tification of the context for a certain agent in the ap-
plication, and this agent will reason with a subset of
all the rules in its Rules or Experience modules, that
will be chosen with the help of the above mentioned
predicate ist, once a judicious interpretation of it has
been chosen.
6 SOME APPLICATIONS OF
KREM
In this section, we will present some applications of
KREM in diverse fields, showing the degree of devel-
opment of the components and giving pointers to our
publications for the interested reader.
6.1 Diagnosis of SMEs
The MAEOS project (modeling of business advice for
organizational and strategic development of SMEs in
Alsace) is the result of a collaboration between AEM
Conseil (Alsatian SME working on business advice
to small and very small companies) and our labora-
tory. Its main goal was to develop a software pro-
totype for acquisition and processing of information
collected by AEM Conseil from its clients to improve
its business advice.
To do this, a set of knowledge bases relating to
business management has been built and used to char-
acterize the client companies. This characterization
was then augmented by expert knowledge provided
by AEM Conseil obtained during its professional
practice. As this is a multidisciplinary project about
Management Science and Knowledge Engineering,
one of the main obstacles is the combined use of di-
verse and sometimes contradictory knowledge, this is
inherent to models that come from Management Sci-
ence; different authors (or the same author in different
publications) may express different views on the same
situation.
Several domain ontologies were developed in the
Knowledge module of KREM and each ontology rep-
resents a conceptualization of the knowledge of a spe-
cific field in the context of SME (e.g. organization,
KREM: A Generic Knowledge-based Framework for Problem Solving in Engineering - Proposal and Case Studies
385
production, strategy, finance, etc.). Although an on-
tological study was conducted to provide the neces-
sary theoretical foundations (Renaud et al., 2009; Re-
naud et al., 2010), the majority of the existing ontolo-
gies provide formal exhaustive models that could be
applied to our purposes, but not many among them
include reasoning rules to permit analysis or diagno-
sis. Consequently, our own ontologies and rules bases
needed to be developed.
This Rules brick of KREM manages different
types of rules to allow to reason about the instances
of the classes in the domain ontologies of the knowl-
edge module. To exploit these rules, a prototype of
a multi-agent system (Zanni-Merk et al., 2011), capa-
ble of manipulating modular ontologies and reasoning
with the Jess inference engine, has been developed.
Three types of rules appear in this component:
Rules for diagnosis to obtain the analysis of the
current situation of the SME.
Recommendation rules providing hints to im-
prove the current situation of the company.
”Bridge” rules to ensure the semantic equivalence
among concepts belonging to different ontologies.
Our latest works permitted the formalization of
the first three bricks of the KREM architecture. We
have also begun to explore meta-knowledge that
could be included into the fourth module (Gartiser
et al., 2014).
In fact, making a diagnosis of the situation of
SMEs involves identifying a set of characteristics that
are ”present” in its current state and also identifying
another set of ”missing” features (the desired state).
In this project, we need to effectively assess the cur-
rent state and decide on actions necessary to achieve
the desired state. The current and desired states are
formalized as a SOEKS, and specific similarity mea-
sures to compare the current and ideal situations of
the SME have been developed. The Meta-Knowledge
module includes diagnosis models (e.g., about correct
or incorrect behavior of companies) and criteria rec-
ommendation according to the experience of the con-
sultant. In addition, in their daily tasks, consultants
follow certain unwritten protocols (theoretical or not)
to diagnose businesses that are also formalized here.
6.2 Semantic Analysis and
Interpretation of Remote Sensed
Images
The availability of very high resolution satellite im-
ages allows to envisage the identification of complex
urban objects such as composite objects (e.g. neigh-
borhoods, residential areas).
To extract and classify these urban objects, it is
relevant to design image analysis methods based on a
”thematic” modelling of these objects to extract and
label them. Using formal ontologies seems a judi-
cious choice. Our previous works (Cravero et al.,
2012; di Sciascio M. et al., 2013; de Bertrand de Beu-
vron et al., 2013) show that this can limit the involve-
ment of experts in time-consuming and unrewarding
tasks such as labelling a large set of objects examples.
Taking advantage of those previous works, we
have proposed the formalization of a new semi-
automatic semantic approach for classification of ur-
ban objects from satellite images, using the KREM
architecture (Zanni-Merk et al., 2015).
The Experience and Meta-Knowledge compo-
nents of KREM are particularly useful for this appli-
cation. The Experience component capitalizes the ex-
perience acquired from experts in the knowledge do-
main. After running the classification step and fail-
ing to identify some segmented regions, the state-
ments contained in this brick should be queried in
order to derive non-usual membership. For instance,
if an unidentified region corresponds to a swimming
pool in a backyard, then some additional knowledge
based on the expert’s experience could help to rec-
ognize it in this last instance. The Meta-Knowledge
module includes knowledge about the use of the other
components. In fact, according to the context of
each image (latitude, longitude, season), different sets
of rules could be launched to take into account that
context. For example, the radiometric characteristics
that allow the identification of vegetation in an image
change according to the season.
6.3 Towards a Formal Model of the
LEAN Enterprise
The goal of this project is to describe the character-
istics of the Lean Enterprise and make the case for
modelling it in order to reproduce its successful prac-
tices more easily.
The literature contains many good descriptions of
the Toyota Production System and Lean in general
(Womack et al., 1991; Byrne and Womack, 2012),
but no formal model that we can build upon. We
propose, then, to follow the KREM model: the K
(Knowledge) component will include domain knowl-
edge about Lean in the form of several ontologies, the
R (Rules) component will be expressed by probabilis-
tic rules, the E (Experience) component will describe
the practices (Kata) and the M (Meta-knowledge)
component will describe the context of the applica-
tion of Lean (different types of companies or cultural
environments, for example).
KEOD 2015 - 7th International Conference on Knowledge Engineering and Ontology Development
386
Our first works (Masai et al., 2015) permitted
to verify the feasibility of the approach, by devel-
oping two complementary ontologies of Lean: the
HOT (House Of Toyota production system or House
Of TPS) ontology, which structures the core issues
in Lean Management, and the ontology for Hoshin
Kanri, which represents the entities necessary to
model this specific process. The Hoshin Kanri pro-
cess is of particular interest because it displays the
behaviour of the agents (the employees in the house
of Toyota production system) at various levels in the
organisation going back and forth.
Concerning the Rules component, because the be-
haviour of enterprises experiences a high level of vari-
ability according to the current context while comply-
ing with specific management rules, they should be
modelled as stochastic processes. In the previously
cited work, we have chosen to model these stochastic
processes with graphical representations of Markov
Chains (Norris, 1998). We are working on the transla-
tion of these probabilistic graphical models into prob-
abilistic rules (Malhotra, 2001), in order to be exe-
cuted by a probabilistic inference engine.
We are also currently developing the Meta-
Knowledge brick of KREM, because in the present
world, it is not rare that not only two, but multiple cul-
tures coexist in large, multinational companies and in
global projects spanning multiple geographies. The
misunderstandings arising from these cultural differ-
ences are responsible for many failures, which are
then blamed by each nationality on the others, with-
out trying to understand those differences first and ad-
dressing them later.
We have developed a strategic ontology of cul-
ture, based on the works of (Meyer, 2014), who iden-
tifies eight dimensions: Communicating, Evaluating,
Persuading, Leading, Deciding, Trusting, Disagree-
ing and Scheduling.
From the specific point of view of Knowledge
Engineering, this ontology is being formally aligned
with the Upper Ontology of Culture or UOC (Blan-
chard et al., 2010) as a framework. The cultural con-
cepts formalized in this way in the Meta-Knowledge
component of the KREM model can be used in prac-
tice using the HOT ontology. Cultural aspects in the
Meta-Knowledge component will also steer the be-
haviour of the agents that is modelled in the Rules
component of KREM.
7 CONCLUSIONS
These article proposes a modular architecture called
KREM (Knowledge, Rules, Experience, Meta-
Knowledge) to integrate the capitalization of previ-
ous experience in the deploying of a knowledge-based
system. In fact, as in general, knowledge elicitation
from experts produces incomplete expertise models,
sometimes, knowledge-based systems are not able to
provide pertinent results. In this case, it is necessary
to call in an expert to provide an answer. It is inter-
esting, therefore, to capitalize that answer along with
the cognitive process (rules or others) that the expert
carries out to develop it.
Well known technologies exist to implement the
knowledge and the rules components of KREM
(OWL + SWRL). But the notion of “experience” has
not been fully addressed in the literature. The SOEKS
(set of experience knowledge structure), initially pro-
posed by the KERT team from the University of New-
castle (Autralia)
4
is the natural choice for formalizing
the experience capitalization.
Modern knowledge engineering practices support
the early criteria held in the eighties, based on a mod-
ular architecture for the development of intelligent
systems, which separates domain knowledge from the
reasoning mechanism. This division is thus held in
the proposal of the KREM architecture, where the
K module corresponds to knowledge stored in the
form of logic statements, and the R module con-
tains the necessary rules to manage and exploit such
knowledge. To achieve a better performance of the
knowledge-based system, we have incorporated an
additional brick, which represents the experience of
experts in the knowledge domain (this experience
capitalization permits, to a certain extent, to have ac-
cess to some tacit knowledge from the experts). Fi-
nally, the use of Meta-Knowledge allows the rational
use of the other three components.
The main interest of the KREM architecture is that
it is flexible and modular. Hence, each project pre-
sented in the last section was described within the ar-
chitecture according to the degree of its development
and maturation, showing its adaptability to cope with
different Engineering application domains in a broad
sense.
REFERENCES
Bazire, M. and Br
´
ezillon, P. (2005). Understanding context
before using it. In Modeling and using context, pages
29–40. Springer.
Blanchard, E. G., Mizoguchi, R., and Lajoie, S. P. (2010).
Structuring the cultural domain with an upper on-
4
http://www.newcastle.edu.au/research-and-innovation/
centre/engineering-built- environment/
knowledge-engineering-research-team.
KREM: A Generic Knowledge-based Framework for Problem Solving in Engineering - Proposal and Case Studies
387
tology of culture. The Handbook of Research on
Culturally-Aware Information Technology: Perspec-
tives and Models, pages 179–212.
Byrne, A. and Womack, J. (2012). The Lean Turnaround:
How Business Leaders Use Lean Principles to Create
Value and Transform Their Company. BusinessPro
collection. McGraw-Hill Education.
Cravero, M., de Bertrand de Beuvron, F., Zanni-Merk, C.,
and Marc-Zwecker, S. (2012). A description logics
geographical ontology for effective semantic analysis
of satellite images. In KES, pages 1573–1582.
Davenport, T. H. and Prusak, L. (2000). Working Knowl-
edge: How Organizations Manage What They Know.
Harvard Business School Press, Boston, Mass.
de Bertrand de Beuvron, F., Marc-Zwecker, S., Puissant, A.,
and Zanni-Merk, C. (2013). From Expert Knowledge
to Formal Ontologies for Semantic Interpretation of
the Urban Environment from Satellite Images. KES
Journal - Innovation in Knowledge-Based and Intelli-
gent Engineering Systems, 17:55 – 65.
Dey, A. K., Abowd, G. D., and Salber, D. (2001). A
conceptual framework and a toolkit for supporting
the rapid prototyping of context-aware applications.
Hum.-Comput. Interact., 16(2):97–166.
di Sciascio M., Zanni-Merk, C., C., W., S., M.-Z., and
de Beuvron F. (2013). Towards a semi-automatic se-
mantic approach to satellite images analysis. In KES,
pages 1388–1397.
Gartiser, N., Zanni-Merk, C., Boullosa, L., and Ana (2014).
A semantic layered architecture for analysis and diag-
nosis of sme. In Elsevier, editor, Procedia Computer
Science, pages 1165–1174. KES 2014.
Giunchiglia, F., Maltese, V., and Dutta, B. (2012). Domains
and context: first steps towards managing diversity in
knowledge. Web Semantics: Science, Services and
Agents on the World Wide Web, 12:53–63.
Golbreich, C. (2004). Combining rule and ontology reason-
ers for the semantic web. In Lecture Notes in Com-
puter Science (including subseries Lecture Notes in
Artificial Intelligence and Lecture Notes in Bioinfor-
matics), volume 3323, pages 6–22.
Grau, B. C., Horrocks, I., Motik, B., Parsia, B., Patel-
Schneider, P., and Sattler, U. (2008). OWL 2: The
next step for OWL. Web Semantics: Science, Services
and Agents on the World Wide Web, 6(4):309 322.
Semantic Web Challenge 2006/2007.
Gruber, T. R. (1993). Toward Principles for the design of
Ontologies used for Knowledge Sharing. Int. J. Hum.-
Comput. Stud., 43:907 – 928.
Hill, E. F. (2003). Jess in Action: Java Rule-Based Systems.
Manning Publications Co., Greenwich, CT, USA.
Leistner, F. (2010). Mastering Organizational Knowledge
Flow: How to Make Knowledge Sharing Work. Wiley.
Malhotra, Y. (2001). Expert systems for knowledge man-
agement: crossing the chasm between information
processing and sense making. Expert Systems with
Applications, 20(1):7–16.
Masai, P., Parrend, P., and Zanni-Merk, C. (2015). Towards
a formal model of the lean enterprise. In Elsevier, ed-
itor, Procedia Computer Science (Proceedings of KES
2015), volume 60C, pages 226–235.
McCarthy, J. (1993). Notes on formalizing context.
Meyer, E. (2014). The Culture Map: Breaking Through the
Invisible Boundaries of Global Business. PublicAf-
fairs.
Milton, N. (2008). Knowledge Technologies. Polimetrica,
Milano, Italy.
Norris, J. R. (1998). Markov Chains. Cambridge University
Press.
Ouertani, M.-Z., Ba
¨
ına, S., Yesilbas, L. G., and Morel, G.
(2011). Traceability and management of dispersed
product knowledge during design and manufacturing.
Computer-Aided Design, 43(5):546–562.
Porzel, R. (2010). Contextual computing: models and ap-
plications. Springer Science & Business Media.
Renaud, D., Zanni-Merk, C., Bouche, P., Gartiser, N., and
Rousselot, F. (2010). A strategy for reasoning with
close knowledge bases. Business Informatics - Data
Mining and Business Intelligence, 16(104):25–33.
Renaud, D., Zanni-Merk, C., and Rousselot, F. (2009).
A compound strategy for ontologies combining. In
KEOD, pages 200–205.
San
´
ın, C. and Szczerbicki, E. (2009). Experience-based
knowledge representation: Soeks. Cybernetics and
Systems, 40(2):99–122.
San
´
ın, C., Szczerbicki, E., and Toro, C. (2007). An owl
ontology of set of experience knowledge structure. J.
UCS, 13(2):209–223.
Sowa, J. F. (1995). Syntax, semantics, and pragmatics of
contexts. In Conceptual Structures: Applications, Im-
plementation and Theory, pages 1–15. Springer.
Sowa, J. F. (2000). Knowledge Representation: Logi-
cal, Philosophical and Computational Foundations.
Brooks/Cole Publishing Co., Pacific Grove, CA, USA.
Toro, C., Sanchez, E., Carrasco, E., Mancilla-Amaya, L.,
San
´
ın, C., Szczerbicki, E., Gra
˜
na, M., Bonachela, P.,
Parra, C., Bueno, G., et al. (2012). Using set of expe-
rience knowledge structure to extend a rule set of clin-
ical decision support system for alzheimer’s disease
diagnosis. Cybernetics and Systems, 43(2):81–95.
Womack, J. P., Jones, D. T., and Roos, D. (1991). The
Machine That Changed the World: The Story of Lean
Production. The MIT international motor vehicle pro-
gram. HarperCollins.
Zanni-Merk, C., Almiron, S., and Renaud, D. (2011).
A multi-agents system for analysis and diagnosis of
smes. In KES-AMSTA, pages 103–112.
Zanni-Merk, C., Marc-Zwecker, S., Wemmert, C., and
de Beuvron, F. d. B. (2015). A layered architecture
for a fuzzy semantic approach for satellite image anal-
ysis. International Journal of Knowledge and Systems
Science (IJKSS), 6(2):31–56.
KEOD 2015 - 7th International Conference on Knowledge Engineering and Ontology Development
388