INTELLIGENT SOLUTION EVALUATION BASED ON
ALTERNATIVE USER PROFILES
Georgios Bardis
Laboratoire Méthodes et Structures Informatiques – MSI, Faculté des Sciences, Université de Limoges
83, rue d’Isle, 87060 Limoges cedex, France
Technological Education Institute of Athens, Department of Computer Science,
Ag.Spyridonos St., 122 10 Egaleo, GREECE
Georgios Miaoulis
Technological Education Institute of Athens, Department of Computer Science,
Ag.Spyridonos St., 122 10 Egaleo, GREECE
Dimitri Plemenos
Laboratoire Méthodes et Structures Informatiques – MSI, Faculté des Sciences, Université de Limoges
83, rue d’Isle, 87060 Limoges cedex, France
Keywords: Machine Learning, Multicriteria Decision Making, User Modeling
Abstract: The MultiCAD platform is a system that accepts the declarative description of a scene (e.g. a building) as
input and
generates the geometric descriptions that comply with the specific description. Its goal is to
facilitate the transition from the intuitive hierarchical decomposition of the scene to its concrete geometric
representation. The aim of the present work is to provide the existing system with an intelligent module that
will capture, store and apply user preferences in order to eventually automate the task of solution selection.
A combination of two components based on decision support and artificial intelligence methodologies
respectively are currently being implemented. A method is also proposed for the fair and efficient
comparison of the results.
1 INTRODUCTION
The continuously increasing performance of modern
computer hardware has made available software
features that used to be prohibitive in terms of
required time and complexity a few years ago.
System developers are now not only willing but also
able to design and implement powerful
environments, rich in characteristics, capable of
producing vast numbers of results in limited time.
Nevertheless, the diversity of the user basis as well
as the increased complexity and power of software
systems call for intelligent features that will adapt
the environment to each user’s characteristics.
Personalized system behavior with respect to an
individual user’s profile facilitates its use, increases
both user and system efficiency and improves
quality of the results.
Adoption of user preferences for intelligent
sy
stem response has been presented in numerous
efforts in the area of hypermedia and the WWW,
e.g. (Brusilovsky 01), (Chen 02), (Pazzani 97),
(Soltysiak 98). Incorporation of user preferences in
geometric representations is presented in (Essert-
Villard 00), where the user submits a set of constants
together with a sketch from which the system
extracts additional solution restrictions as well as in
(Joan-Arinyo 03) where a genetic algorithm is
periodically aided by the user to produce solutions
closer to the latter’s preferences.
On the other hand, the notion of multicriteria
ev
aluation of building assemblies has also been
74
Bardis G., Miaoulis G. and Plemenos D. (2005).
INTELLIGENT SOLUTION EVALUATION BASED ON ALTERNATIVE USER PROFILES.
In Proceedings of the Seventh International Conference on Enterprise Information Systems, pages 74-82
DOI: 10.5220/0002523900740082
Copyright
c
SciTePress
Figure 1: A Typical MultiCAD Session
presented in (Nassar 03), based on AHP calculated
weights and a heuristic evaluation algorithm, but
machine learning and user preferences have not been
discussed.
Our work proposes a component for intelligent
solution evaluation according to user’s preferences
in a declarative description environment. The
proposed component combines multicriteria decision
support and machine learning techniques for user
modeling requiring only qualitative feedback on
behalf of the user instead of exact geometric
properties.
2 THE MULTICAD
ENVIRONMENT
The MultiCAD system, presented in detail in
(Miaoulis 02), was introduced as a platform
supporting declarative object modeling (Plemenos
95), thus assisting the transition from the intuitive to
the geometric object representation. The described
system is a complete design environment including
modules for validation of the object description,
storage and maintenance of the solutions produced,
etc. Subsequent works have successfully
implemented most of the described modules as well
as additional ones that have evolved from this initial
design – solution generation using CSP (Bonnefoi
02) or genetic algorithms (Vassilas 02), concept
modeling and ontology (Ravani 04), incorporation of
architectural styles (Makris 03) and collaborative
design (Golfinopoulos 04). Solution evaluation
based on user preferences in this context was first
introduced in (Plemenos 02) proposing a system
based on the representation of each scene by a
dedicated neural network. A new approach towards
a user profile module was presented in (Bardis 04)
describing two methods for user modeling and
solution evaluation: a method based on the
multicriteria nature of the problem and one relying
on a neural network for the representation of each
user’s preferences.
The current work continues towards this
direction by presenting the final design of the
specific module – an approach incorporating
multicriteria decision support and machine learning
techniques – and the current stage of the
implementation. Moreover, a set of criteria
regarding alternative methods performance is
introduced, that will serve as the basis for future
testing and adjustment of the implemented module.
3 THE INTELLIGENT USER
PROFILE MODULE
Figure 1 shows a typical session of the current
implementation of MultiCAD where the declarative
description of a scene appears together with the
INTELLIGENT SOLUTION EVALUATION BASED ON ALTERNATIVE USER PROFILES
75
3. Solution
Visualisation
5. Attribute Values
Extraction
4. Human
Solution
Evaluation
6. Automatic
Solution
Evaluation
User Profiles
Database
7. Intelligent
User Profile
Update
1. Scene
Description
Delarative
Representation
2. Solution
Generation
Geometric
Representation
0. User
Profile
Initialisation
3. Solution
Visualisation
5. Attribute Values
Extraction
4. Human
Solution
Evaluation
6. Automatic
Solution
Evaluation
User Profiles
Database
7. Intelligent
User Profile
Update
1. Scene
Description
Delarative
Representation
2. Solution
Generation
Geometric
Representation
0. User
Profile
Initialisation
Figure 2: User Profile Module - Block Diagram
visualization of one of the corresponding solutions.
The diagram in Figure 2 concentrates on the
integration of the User Profile Module to the
MultiCAD platform.
This version of the system applies user profile
information only after the geometric representations
of the described objects have been generated, i.e. not
during their generation. The current stage of our
work focuses on the construction of those modules
that are immediately connected to the user profile
component of the system.
Each declarative scene description may lead to a
few thousands geometric representations complying
with this description: the solutions. However, not all
solutions are equally preferred by the user. Our
intention is to eliminate those solutions not
conforming to the user's preferences. Ideally, this
has to happen with minimal or no user intervention.
In particular, optimal user profile incorporation to
the solution visualization process will maximize
solution visualization throughput (SVT) (Bardis 04)
and minimize the user intervention at later stages. In
order to achieve this we have to resolve the
following inter-connected problems:
Solution representation and evaluation.
User preferences modeling and representation.
4 SOLUTION REPRESENTATION
The geometric representation of each solution is
translated to a set of attributes. This is a need that
arises mainly from the fact that we have to reduce
the complexity of the representation of each solution
in order to be able to submit it as input to a neural
network. In addition, this approach complies with
the multicriteria decision methodologies (Vincke
92), (Goodwin 98), in particular, their requirement
to request the user’s evaluation through a limited set
of object attributes instead of numerous geometric
properties. We choose to observe a minimal set of
attributes (Fribault 03), (Bardis 04). This set will be
extended or revised in the future, since, for the
moment, we care more for the development of a
prototype covering all stages of the MultiCAD cycle
(Miaoulis 02) instead of capturing all possible
aspects of user preferences with respect to a building
assembly. The attributes we have chosen are based
on geometric characteristics of each solution and,
therefore, can be easily extracted by its geometric
representation.
In particular, the observed attributes for any
solution S
i
are:
BD
i
= Number of bedrooms
BT
i
= Number of bathrooms
NA
i
= Night-zone area
DA
i
= Day-zone area
NDS
i
= Night-zone / Day-zone separation
SWB
i
= Existence of at least one south-western
bedroom
Therefore, each solution S
i
is represented by a
vector of values:
S
i
= ( BD
i
, BT
i
, NA
i
, DA
i
, NDS
i
, SWB
i
),
for example
S
i
= ( 2, 3, 52.4, 40.8, Partial, No).
ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS
76
USER
USER
DECLARATIVE
DESCRIPTION
DECLARATIVE
DESCRIPTION
SOLUTION
SOLUTION
ATTRIBUTE
ATTRIBUTE
PROJECT
TYPE
PROJECT
TYPE
SUBMITS
SUBMITS
YIELDS
YIELDS
REPRESENTED
BY
REPRESENTED
BY
BELONGS
TO
BELONGS
TO
DS PROFILE
DS PROFILE
WEIGHT
WEIGHT
DATE
DATE
IMPORTANCE
IMPORTANCE
VALUE
VALUE
GENERATOR
GENERATOR
MIDDLE
MIDDLE
1
N
1
N
N
N
M
1
N
GRADES
GRADES
1
M
K
N
INTUITIVE
INTUITIVE
AUTOMATIC
AUTOMATIC
ML PROFILE
ML PROFILE
NEURAL
NETWORK
PARAMETERS
NEURAL
NETWORK
PARAMETERS
M
N
K
USER
USER
DECLARATIVE
DESCRIPTION
DECLARATIVE
DESCRIPTION
SOLUTION
SOLUTION
ATTRIBUTE
ATTRIBUTE
PROJECT
TYPE
PROJECT
TYPE
SUBMITS
SUBMITS
YIELDS
YIELDS
REPRESENTED
BY
REPRESENTED
BY
BELONGS
TO
BELONGS
TO
DS PROFILE
DS PROFILE
WEIGHT
WEIGHT
DATE
DATE
IMPORTANCE
IMPORTANCE
VALUE
VALUE
GENERATOR
GENERATOR
MIDDLE
MIDDLE
1
N
1
N
N
N
M
1
N
GRADES
GRADES
1
M
K
N
INTUITIVE
INTUITIVE
AUTOMATIC
AUTOMATIC
ML PROFILE
ML PROFILE
NEURAL
NETWORK
PARAMETERS
NEURAL
NETWORK
PARAMETERS
M
N
K
Figure 3: User Profile Module - Entity Relationship Model
The exact range of values for each observed
attribute will be affected by the design of the ML
component described in Section 6 as well as the
solution generator used. Nevertheless, it is important
to observe that solution generation of the existing
system is not based on observed attributes. This
implies that the number of generated solutions is not
restricted by the range of values of the observed
attributes. The observed attributes map each
generated solution to a vector of values of restricted
range. Thus, it may be the case that, at the present
stage, two or more different solutions, i.e. of
different geometric representations, are mapped to
the same vector of observed attribute values.
5 USER MODELING
Figure 3 presents the database model that has been
developed in order to store and maintain user profile
information. Notice that only a few representative
properties of entities and relationships appear in the
ER graph. Apart from the entities directly connected
with the user profile module, the database model
also includes entities representing scene descriptions
and geometric representations of the corresponding
solutions, namely the DESCRIPTIONS and
SOLUTIONS entities. Specialized databases have
already been developed as part of other work, taking
place in the context of the MultiCAD platform
(Ravani 03). The database model proposed here is
flexible enough to cooperate with these already
existing structures and yet able to incorporate
alternative implementations in case these become
available in the future.
Figure 2 offers insight regarding the choice of
the specific entities and relationships appearing in
the database model. In particular,
USER. Example fields for the properties for the
specific entity are the (unique) User Name,
Password, First/Last Name, etc.
PROJECT TYPE. A text name plus extra
information connecting the specific project type with
the corresponding description prototypes contained
in alternative databases of the MultiCAD
environment.
DESCRIPTION. This entity represents the
Prolog-like description of a scene. Importance
represents the influence of the results of this session,
i.e. the properties of the approved solutions during
the specific this session, to the overall user profile
for the specific project type.
SOLUTION. Full geometric representation of
each solution.
INTELLIGENT SOLUTION EVALUATION BASED ON ALTERNATIVE USER PROFILES
77
Figure 4: Decision Support Component - Weight Assignment using AHP
ATTRIBUTE. Attributes used to map solution to
a smaller space that will allow further processing
with respect to user profiles. The fields include
name, type (int, real, scalar – big, medium, small,
max, min, etc.)
NEURAL NETWORK PARAMETERS. The
values of all neural network construction variables.
In addition to the aforementioned entities, a set
of relationships will maintain the information about
the active interconnection of the entities during user
sessions. In particular,
submits. Users submit scene descriptions for
processing. Each scene is connected with a certain
project type.
yields. Each description, using one of the
alternative solution generators that are available,
results in a set of solutions complying with the
description.
is mapped to. Each solution is mapped to a set of
values for the attributes we have chosen to observe.
decision support profile. This relationship
contains all information regarding the initial user
profile as obtained by the Decision Support
component. In particular, for each attribute of a
specific project type, the user has already provided
his/her personal interpretation of its importance for
solution evaluation. This importance is represented
by the corresponding weight. An additional personal
parameter is that of the middle value for any
attribute. The user is requested to suggest the middle
value for all attributes, i.e. the actual value of the
attribute that represents 50% performance of a
solution with respect to the specific attribute.
machine learning profile. This is the
relationship that interconnects the user, in the
context of (any) specific project type, with the
Machine Learning component, i.e. all values needed
to fully describe the neural network used to
represent the specific user’s dynamic profile.
grades. Solutions are evaluated by the user
through visual inspection, thus yielding the intuitive
grade (approved/rejected or a number in case an
alternative grading scheme is used). In addition,
solutions are also evaluated based on the specific
user’s profile for the specific project type. This
information may be regenerated based on the
contents of the database. Nevertheless, solution
evaluation is a crucial and time-consuming task;
hence, once the results are available they are stored
in the database.
6 SOLUTION EVALUATION
Two alternative approaches are used for solution
evaluation based on user preferences forming two
independent components of the user profile module.
Nevertheless, their concurrent operation and results
affect the overall behavior of the system, as it will
become apparent in the following section.
ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS
78
Figure 5: Decision Support Component - Weight Assignment Using SMART
6.1 Decision Support Component
Each user is requested to assign weights representing
the importance of each one of the observed
attributes. Two alternative methods for weight
assignment are available to the user based on the
corresponding stages of SMART (Goodwin 98) and
AHP (Saaty 90) multicriteria decision algorithms
respectively. An interesting discussion regarding
dynamic weight assignment versus fixed weight
values is presented in (Roberts 02).
The weights are then used to produce a score for
each generated solution using a fitness function.
Alternative functions may be used in the future but
we currently use the inner product of the user
weights with the attribute values for each solution
(Bardis 04). These scores are then used in order to
sort the set of solutions as a list of descending user
preference order: from the most to the least
preferable solution. This Decision Support (DS)
component has already been implemented and
currently operates on solutions in the form of
attribute vectors. An example execution of the DS
component for user profile initialization is shown in
Figures 4-5. Notice the inconsistency index in the
AHP method, signifying discrepancies in the user’s
answers, as well as the normalized weights
appearing at the top row of the sample solution set
evaluation of the SMART method. The user may
rely on this component in order to obtain a set of
automatically selected solutions based on the
aforementioned weights. In this case solutions will
be automatically visualized and presented to the user
in descending preference order. However, the user
may choose to manually select the preferred
solutions thus contributing to the training of the
neural network described in the next section.
6.2 Machine Learning Component
The ML component will be based on a neural
network of six inputs one for each observed
attribute, at least one hidden layer and a single
output, representing the approval/rejection of a
solution by the user. Alternative structures of the
network will be implemented and tested according to
the criteria presented in the next section. This
process may lead to the selection of more than one
structures for systematic use as alternative ML
components, similar to the use of two alternative
methods for attribute weight assignment in the DS
component.
The user, having submitted a scene description,
will evaluate the solutions that are generated and
visualized for the specific scene. This set of
approved/rejected solutions will serve as one of the
training sets for the network(s). In particular, each
example in a training set represents a correct input-
INTELLIGENT SOLUTION EVALUATION BASED ON ALTERNATIVE USER PROFILES
79
output mapping. In the present context, the input
part, for any given solution, is comprised by the
attribute values representing the specific solution.
The output part simply contains the user’s approval
or rejection for the specific solution.
U
It is important to notice that the aforementioned
process will be a completely transparent system task:
the user will not have to submit any additional
information with respect to the network(s) training
and therefore, does not have to be aware of it.
Automatic solution evaluation will be available from
the very first session via the Decision Support
component. In general, automatic solution
evaluation will be at the user’s discretion and the
choice of the most appropriate component for this
purpose will be based on the criteria presented in the
next section.
M
2
M
1
G
Figure 6: Example Methods Performance
7 PERFORMANCE COMPARISON
Either during the testing period or during the regular
use of the system, at least two alternative methods
will have to be compared with respect to their
performance. In particular, solutions approved by
each method will be compared to the solutions
approved by the user.
7.1 Performance Indices
In order to be able to compare these methods we
must provide the means to measure their
performance. We concentrate on the solution
selection stage that takes place after solution
generation. Therefore, in the following, we will
assume that solution generation has already taken
place and the methods have been applied to the
results. The application of each method to the
solutions yields the corresponding subset of
approved solutions. For simplicity, we mention only
two methods in the following whereas, in practice,
more alternatives – due to alternative weight
assignment, alternative network adjustments, etc. –
may be concurrently evaluated.
In particular, let us define the following sets:
G = The solutions generated based on the
specific description of a scene.
U = The preferred solutions, i.e. solutions in G
that comply with the user preference. These
solutions represent the user preference in the current
context. Formally, U G. In the following we may
also refer to the members of U as approved
solutions.
G-U = The discarded solutions, i.e. the
generated solutions that are not preferred by the
user.
M
1
= solutions in G approved by Method 1.
Formally, M
1
G.
M
2
= solutions in G approved by Method 2.
Formally, M
2
G.
|S| = the number of members of any set S.
Therefore we may now define the hit rate of each
method as:
}21{ ,
||
||
,i
U
UM
HR
i
i
= ,
i.e. the percentage of approved solutions
captured by the specific method.
The ratio of approved vs. total solutions selected
by each method could also be used as measurement
of their performance:
}2,1{,
||
||
= i
M
UM
PR
i
i
i
There are more than one ways to define a miss
rate. We may define it as the percentage of discarded
solutions that are selected by the method, expressed
as:
}2,1{,
||
||
= i
UG
UM
MR
i
i
However, we expect that, only a small number of
the generated solutions will fulfill the user
preference. This is mainly due to time limitations
posed by the requirement for human visual
inspection. On the other hand, |G| greatly depends on
the description and can vary significantly. Therefore,
we need to relate the size of the error for each
method with |U| instead of a quantity including |G|.
Hence, we could alternatively define
:
}2,1{,
||
||
= i
U
UM
MMR
i
i
ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS
80
and interpret a lower value as a better performance.
Intuitively, this interpretation implies that a method
should not select many discarded solutions when
only a few preferred solutions exist. For example,
when this rate is more than 1 the method gives more
discarded solutions than the total number of
approved solutions. Instead of 1, another value may
be selected to reflect a specific performance
threshold.
The above are clarified in the example Venn
diagram of Figure 6, representing a general case (i.e.
no intersection is empty, no two sets are equal). For
the sake of simplicity of the picture, the total number
of solutions is rather small, i.e. |G| is only 35
whereas this is generally not the case. Nevertheless,
for the specific example, we have the following
numbers:
|G| = 35, |U| = 10, |M
1
| = 6, |M
2
| = 10, |M
1
U| = 3,
|M
2
U| = 4, |M
1
– U| = 3, |M
2
– U| = 6
Therefore for M
1
we have:
HR
1
= 3/10, MR
1
= 3/25, MMR
1
= 3/10, PR
1
= 3/6
and for M
2
we have:
HR
2
= 4/10, MR
2
= 6/25, MMR
2
= 6/10, PR
2
= 4/10
Table 1: Example Methods Performance Indices
Metho
d
Hit Rate
Performance
Ratio
Miss Rate
Modified
Miss
Rate
Method 1 3/10 = 30% 3/6 = 50% 3/25 = 12% 3/10 = 0.3
Method 2 4/10 = 40% 4/10 = 40% 6/25 = 24% 6/10 = 0.6
Extreme
Case 1
1/10 = 10% 1/1 = 100% 0/25 = 0% 0/10 = 0.0
Extreme
Case 2
10/10 = 100% 10/35 = 28.6%
25/25 =
100%
25/10 = 2.5
M
2
could be considered a worse (because of the
higher miss rate) or a better (because of the higher
hit rate) method than M
1
depending on the
interpretation of these numbers. Ideally, hit rate
should be equal to 1, miss rate equal to 0 and
performance ratio equal to 100%. Notice, however,
that a method selecting only one preferred solution
every time it is invoked (Extreme Case 1) would
yield a performance ratio of 100% without
necessarily representing an optimal method as
shown by the low hit rate. On the other hand, simply
selecting all produced solutions (Extreme Case 2)
maximizes the hit rate but yields a low performance
ratio. Extreme Case 1 appears to represent a method
with acceptable performance whereas Extreme Case
2 represents a trivial approach of no practical use.
Therefore, a high Performance Ratio appears to be a
necessary, although not sufficient, indication of an
efficient method and it becomes apparent that we
need a combination of these indices in order to
accurately evaluate the performance of each method.
7.2 Method Integration
The training process of the neural network will have
to continue, with alternative scene descriptions, until
the ML component is considered ready to support
automatic solution evaluation. We may define this
threshold of ML component based on the values of
the performance indices we described above for the
two alternative DS components as well as for the
ML component itself. In the following we will
assume that the user has initialized his/her profile
giving answers to the DS components that
reasonably represent his/her preferences.
In particular, we will be able to rely on the
results given by the network, and therefore adopt
fully automated solution selection, as soon as the
ML component performs consistently better than
both of the DS alternatives. We can state that the
ML component is mature, and therefore ready to
take over automatic solution selection iff:
PR
ML
> PR
DS1
PR
ML
> PR
DS2
HR
ML
> HR
DS1
HR
ML
> HR
DS2
MMR
ML
< MMR
DS1
MMR
ML
< MMR
DS2
for an (adjustable) number of recent descriptions
submitted by the user.
The strictness of this set of conditions may be
relaxed by omitting some of the inequalities. In any
case, it is important to observe that, if the complete
set of conditions is repeatedly true, this implies that
the ML component will be capturing preferences
better than the weight vectors submitted by the user
himself/herself. In such a case, it will also be
interesting to explore the possibility of capturing
additional criteria that the user is not fully aware of,
i.e. sub-conscious criteria. This could become
apparent through the examination of experimental
results and users’ comments regarding system
performance with respect to their preferences.
8 FUTURE WORK
This stage of our work will conclude with the
detailed design and implementation the ML
component. Subsequent performance comparison of
the two components will lead to further refinement
of their properties. Extending the solution evaluation
to grade assignment instead of plain
approval/rejection will also be considered. In that
case performance indices will have to be modified to
also reflect the quality of the selected set of
solutions.
INTELLIGENT SOLUTION EVALUATION BASED ON ALTERNATIVE USER PROFILES
81
The next major stage of our work will focus on
the enhancement of the user profile model with
information originating from the connection between
the declarative description and the corresponding
approved solutions. Such an association will offer
insight regarding the specific user’s interpretation of
declarative properties and relations. Successful
modeling of user preferences at the declarative as
well as the geometric level will allow incorporation
of user profile information to the process of solution
generation, thus significantly improving system
performance.
ACKNOWLEDGMENTS
This study was co-funded by 75% from the
European Union and 25% from the Greek
Government under the framework of the Education
and Initial Vocational Training Program –
‘Archimedes’.
REFERENCES
Bardis G., Miaoulis G., Plemenos D., 2004. An Intelligent
User Profile Module for Solution Selection Support in
the Context of the MultiCAD Project. 7
e
Infographie
Interactive et Intelligence Artificielle (3IA), Limoges,
France.
Bonnefoi P.-F., Plemenos D., 2002. Constraint satisfaction
techniques for declarative scene modeling by
hierarchical decomposition, 3IA, Limoges, France
Brusilovsky P., 2001. Adaptive Hypermedia, User
Modeling And User-Adapted Interaction 11: 87-110.
Chen C.C., Chen M.C., 2002. PVA: A Self-Adaptive
Personal View Agent, Journal Of Intelligent
Information Systems, 18:2/3, 173–194.
Essert-Villard C., Schreck P., Dufourd J.-F. 2000. Sketch-
Based Pruning of a Solution Space within a Formal
Geometric Constraint Solver, Artificial Intelligence
124, 139-159.
Fribault P., 2003. Modelisation Declarative d’Espaces
Habitable, (in French), Doctoral dissertation,
University of Limoges, France.
Golfinopoulos V., Dragonas J., Miaoulis G., Plemenos D.,
2004. Declarative Design in Collaborative
Environment, 7
e
3IA, Limoges, France.
Goodwin P., Wright G., 1998. Decision Analysis for
Management Judgement, Second Edition, Wiley.
Joan-Arinyo R., Luzon M.V., Soto A., 2003. Genetic
Algorithms for Root Multiselection in Constructive
Geometric Constraint Solving. Computers and
Graphics 27, 51-60.
Makris D., Ravani I., Miaoulis G., Skourlas C., Fribault
P., Plemenos D., 2003. Towards a domain-specific
knowledge intelligent information system for
Computer-Aided Architectural Design, 3IA
conference, Limoges, France.
Miaoulis G., 2002. Contribution à l'étude des Systèmes
d'Information Multimédia et Intelligent dédiés à la
Conception Déclarative Assistée par l'Ordinateur – Le
projet MultiCAD (in French), Professorial dissertation,
University of Limoges, France.
Miaoulis G., Plemenos D., Skourlas C., 2000. MultiCAD
Database: Toward a unified data and knowledge
representation for database scene modeling, 3IA,
Limoges, France.
Nassar K., Thalet W., Beliveau Y., 2003. A Procedure for
Multicriteria Selection of Building Assemblies,
Automation in Construction 12, 543-560.
Pazzani M., Billsus D., 1997. Learning And Revising User
Profiles: The Identification Of Interesting Web Sites,
Machine Learning 27, 313–331.
Plemenos D., 1995. Declarative modeling by hierarchical
decomposition. The actual state of the MultiFormes
project, Communication in International Conference
GraphiCon'95, St Petersburg, Russia.
Plemenos D., Miaoulis G., Vassilas N., 2002. Machine
learning for a General Purpose Declarative Scene
Modeler. International Conference GraphiCon'2002,
Nizhny Novgorod, Russia.
Plemenos D., Tamine K., 1997. Increasing the efficiency
of declarative modeling. Constraint evaluation for the
hierarchical decomposition approach. International
Conference WSCG’97, Plzen, Czech Republic.
Ravani I., Makris D., Miaoulis G., Constantinides P.,
Petridis A., Plemenos D., 2003. Implementation of
Architecture-oriented Knowledge Framework in
MultiCAD Declarative Scene Modeling System, 1
st
Balcan Conference in Informatics, Thessaloniki,
Greece.
Ravani J., Makris D., Miaoulis G., Plemenos D., 2004.
Concept-Based Declarative Description Subsystem for
Computer Aided Declarative Design (CADD). 7
e
3IA,
Limoges, France.
Roberts R., Goodwin P., 2002. Weight Approximations in
Multi-attribute Decision Models, Journal of
Multicriteria Decision Analysis 11: 291-303.
Saaty, T.L., 1990. The Analytic Hierarchy Process, RWS
Publications, Pittsburgh, USA.
Soltysiak S.J., Crabtree I.B., 1998. Automatic learning of
user profiles — towards the personalization of agent
services, BT Technology Journal Vol. 16, No 3.
Vassilas N., Miaoulis G., Chronopoulos D., Konstantinidis
E., Ravani I., Makris D., Plemenos D, 2003.
MultiCAD-GA: A System for the Design of 3D Forms
Based on Genetic Algorithms and Human Evaluation,
SETN 203-214, Thessaloniki, Greece.
Vincke P., 1992. Multicriteria Decision-aid, Wiley (1992).
ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS
82