Computational
Personality Type Classification (1999-2001)
Rresearch funded by the Ministry of Education of Czech Republic under the research program ‘Transdisciplinary Biomedical Engineering Research’, No. MSM 210000012.
This project was focused on application of artificial intelligence and pattern recognition methods to the biological data processing in the area of computational personality type classification. Five approaches to the computational personality type classification problem. In contrast to the standard personality type classification process involving psychological investigation, our approach is based on the human operator’s physiological parameters measurement (pulse frequency, systolic and diastolic blood pressures, skin resistance, muscular tonus). Personality type classification consists in neuroticism and tendency to risk behaviour classification. Five approaches were compared one to one – expert system, machine learning, combination of expert system and machine learning, fuzzy clustering, and fuzzy preference expert systems complemented by an aggregation mechanism. The set of tested persons consisted of 60 persons in age between 18 an 30. This set embodies persons with a tendency to risk behaviour, persons with the tendency to neuroticism, and persons without any of the tendencies mentioned above. Two new classification methods were suggested.
Staff:
Dr. Ing. Lenka Lhotska*
Dr. Ing. Vladimir
Eck
Ing. Ladislava Janku
Ing. Milan Sorf, Ph.D.
*Project manager
Publications:
-
Janku, L. - Šorf, M. - Lhotska, L. - Eck, V.: Five
Approaches to the Computational Personality Classification Problem.
In: Proceedings of CIMA 2001, ICSC/NAISO conference, UK, 2001
-
Janku, L. - Eck, V.: Fuzzy Distributed Reasoning
about Multi-Dimensional Medical Data. In: Proceedings: MEDICON
2001. Zagreb: University of Zagreb, 2001, vol. 1, p. 441-444. ISBN
953-184-023-7.
-
Janků, L.- Lhotska, L.: Aggregation of
Fuzzy Preferences as an Efficient Approach to the Computational Personality
Type Classification Problem In: N. Mastorakis (Editor): Problems in
Applied Mathematics and Computational Intelligence, WSES Press, 2001
-
Šorf, M. - Lhotská, L. - Janku, L. - Eck, V.: Application
of the Methods AI Approaches to Personality Type Classificaion.
In: Modelling, Identification and Control - volume I. Anaheim : Acta Press,
2001, vol. 1, p. 403-408. ISBN 0-88986-316-4.
-
Janků, L. - Šorf, M. - Eck, V. - Lhotská, L. -
Fejtová, M. Multiple Criteria Choice Mechanism for Intelligent
Man-Machine Interface. In: Proceedings of INES 2000. Vol. 1. IEEE
Industrial Electronics Society. 2000. ISBN 961-6303-23-6
-
Šorf, M. - Janků, L. - Lhotská, L. - Eck, V. Applications
of Expert System and Machine Learning Approach to Intelligent Man-machine
Interface. In: Intelligent Techniques and Soft Computing in Nuclear
Science and Engineering. Vol. 1. Singapore: World Scientific. 2000. p. 191-197.
- ISBN 981-02-4356-1
Project Overview:
1. Introduction
Intelligent interface, which could be able to monitor and estimate human operator’s psychical state and predict his/her future behaviour, could minimize system failings caused by human factor.
Investigations on the human’s operator
psychical state monitoring and estimation have recently received remarkable
attention particularly for the following applications: intelligent interface,
real-time measurement and processing of physiological data, on-line human
operator monitoring, and extraction of information from selected measured
physiological parameters.
Other research in the area of intelligent
man-machine interface has been devoted to the design of special multi-agent web
based systems, which appears as a suitable solution for large internet and
e-commerce systems, but which don’t provide a satisfying solution for the
mentioned problem of human operator’s psychological state estimation.
The necessity of estimation and prediction of
the human’s operator psychical state implies a necessity of utilization of an
operator’s internal model. A task of internal model selection is closely
related to the problem of human operator’s personality type classification.
Particular personality type is characterized via the different mental model. It
appears much more suitable to provide this classification automatically than to
follow a standard way consisting in a special kind of psychological
investigation.
This approach sets a new requirement to
intelligent man-machine interface. Interface must be able to classify human’s operator personality type.
To solve this problem, we designed a way and a methodology how to do it. This
methodology consists in several physiological parameters measurement while the
tested person performs psychical or physical exercises putting a psychical stress
to her or him. This methodology is
called a computational personality type classification.
Our research in essence involves application on
artificial intelligence methods to the computational personality type
classification problem. The major advantage of this approach towards the
standard way consisting in psychological investigation was the higher
objectivity and independence on the expert presence.
2. Computational Personality Type
Classification
Computational personality type classification
is based on assumption that it’s possible to estimate a personality type from
the data obtained by monitoring the tested person’s physiological responses
(pulse frequency, systolic and diastolic blood pressures, skin resistance,
muscular tonus) to the different psychical load. We selected two factors on
which the personality type depends - neuroticism and tendency to risk behaviour
[6].
A classification problem can be formulated by
the following question: how we can estimate the monitored features on condition
we have information about the relationships between measured data and estimated
features on condition, that there is no exactly given mathematical formula
expressing relationship between
measured data and estimated features; only some
vague rules are given.
3. Applied Artificial Intelligence and Soft Computing Methods
3.1. Expert System
The first approach consists in application of
the expert system with knowledge base developed in a standard way. This
knowledge base [1] uses as input information reaction of heart frequency,
muscular activity, skin galvanic reaction, and systolic and diastolic blood
pressures. The knowledge has been acquired from measured data, questionnaires
and theoretical knowledge. In the process of development of the knowledge base
we have started from theoretically known reactions of individual physiological
quantities on psychic load. The knowledge base has been developed in
traditional way – using expert knowledge. At first unstructured interview has
been performed when an expert has commented the problems. Then structured
interview has followed when concrete questions concerning influence of psychic
load on physiological parameters for individual psychological groups have been
asked. In this phase the expert has confirmed influence of psychic load on
reaction of physiological quantities and has not assigned any significance to
average values of physiological quantities at psychic load. Finally, we have
used the method of repertoire table for knowledge acquisition. In this method objects
are individual psychological groups and constructs are reactions of heart
frequency, muscular activity, systolic and diastolic pressure, and skin
galvanic reaction. Based on the repertoire table we have formulated rules with
corresponding values of certainty measures.
The most difficult task after development of
the knowledge base has been assigning values of prior probabilities to nodes
and measures of necessity and sufficiency to rules. The medical doctors have not
been usually able to determine more precisely how probable the given statement
is. Therefore after consultations with the medical doctors, we have tried to
estimate the probabilities of the nodes and rules by ourselves. The first
consultations with the expert system have shown that prior probabilities of
nodes and measures of necessity and sufficiency of rules have been estimated
quite well; only some of them have been overestimated.
After each response from the database we have
inspected updated values of probabilities of all nodes. From these values we
have determined after which question and response incorrect changes of
probabilities of goal hypotheses have followed and which rule has caused this
change. Then we have changed values of measures of that rule and started new
compilation. It has showed whether the change of the value has been correct. If
not, the process has been repeated. After compilation of the knowledge base we
have determined resulting success rate of the knowledge base.
Assigning values to nodes has preceded
assigning values to rules that have been done during the phase of tuning the
knowledge base. At first prior probabilities of goal hypotheses have been
determined. Then prior probabilities of all remaining nodes have been determined.
The designed knowledge base contains 25 nodes
out of which 10 are quantitative and 15 qualitative. The quantitative nodes are
of 1-value type and are determined for input of physiological parameters
values. Qualitative nodes are of Bayesian type – 13 are askable nodes and 2 are
goal hypotheses. In leaf nodes there are questions about heart frequency
reactivity, systolic and diastolic blood pressure, muscular tonus and skin
resistance. Top nodes (goal hypotheses) characterize risk groups of persons (neurotics
and persons with risk type of behaviour). The knowledge base contains 5 rules
expressing relations between individual nodes that represent given statements
on respondent’s state. Low number of rules is caused by the fact that during
consultation with an expert influence of average values of physiological
parameters has been underestimated and on the other hand influence of reaction
of physiological parameters on psychic load has been overestimated.
There has appeared necessity to include 20
context links into the knowledge base so that the questions follow in the right
order (first questions on input parameters and then on their values).
The knowledge base structure has been divided
into several substructures. The substructures associate attributes that are
linked together and frequently give evidence for several types of risk groups.
In particular, there are three substructures: cardiovascular parameters,
muscular tonus, skin resistance.
3.2. Machine Learning
This chapter describes an application of machine
learning approach. For generation of the decision tree, algorithm ID3 [3] has
been used. This algorithm is of TDIDT type (Top-Down Induction of Decision
Tree) that constructs the decision tree recursively from the root to leaves.
Original training set is thus divided into smaller and smaller subsets.
Application of learning algorithm has two
phases:
·
generation
of the decision tree
·
decision
tree pruning
We have applied classification system [4] based
on the ID3 algorithm. It does not use any prior knowledge and objects are
formally represented by a list of attributes.
The system contains methods for discretization,
determination of the most informative attribute and pruning.
Attributes are both discrete with low number of
values and continuous. The continuous attributes have been discretized using
R-binarization method. It is based on the following principle: when using
binarized attribute for branching in a certain node this attribute is not
excluded from the list of attributes applicable for branching in the sub-tree
of the given node, but if the attribute has tendency to divide cases into more
than two intervals it may appear in one branch of the decision tree several
times, each time with different discretization value. In this way we reach similar
effect as with direct division of the set into more subsets. Data for the
training set have been selected randomly. Mántaras method [5] of determination
of distance between two partitions (further called distance gain) is used for
selection of the most informative attribute. We determine distance gain for all
attributes. The attribute with the highest value of distance gain is placed on
the position of the root node. Homogeneous branches are ended with the leaf
node that is assigned name of the class. In non-homogeneous branches, the
attribute with the next highest value of distance gain is looked for.
Partial goal of this work has been application
of machine learning to respondent classification. Input into the classification
system is the training set that forms interface between this system and the
environment. The training set is composed of header and training examples. In
the header, there are defined the classes into which examples will be
classified and the attribute names.
3.3. Combination of Machine Learning and Expert System
Since the success rate of the knowledge base
developed in the classical way is not high, we have developed new methodology
of knowledge base development utilizing machine learning algorithms. The
database is divided into two sets: training and testing. The knowledge base is
developed using the training set and afterwards it is tested using data from
the testing set. This approach is based on measured data and data acquired from
the questionnaires in contrast to the classical approach where the knowledge is
delivered by the expert. The rules for the knowledge base are acquired from a
generated decision tree. These rules are consulted with experts, modified and
completed with rules expressing influence of reactivity of physiological
parameters on risk groups. Addition of further rules leads to improved
behaviour of the knowledge base. We have used the same methodology for tuning
and verification of the knowledge base and determination of prior probabilities
of the nodes and rules as with the previous knowledge base using only expert
knowledge.
This knowledge base contains 32 nodes. All
nodes are qualitative, out of which 21 are logical and 11 Bayesian (9 askable
and 2 goals). Askable nodes contain questions on level of average values of
heart frequency, muscular tonus, skin resistance and systolic and diastolic
blood pressure, completed with case-history information (coffee, alcohol and
cigarette consumption, current health problems, medicament use and movement
activity). The goal hypotheses are neuroticism diagnosis and risk type of
behaviour diagnosis. The knowledge base contains 36 rules that express
relations between individual nodes that represent given statements on
respondent’s state. Rules acquired by machine learning express relations
between individual risk groups and average values of physiological parameters
and case-history information.
Structure of this knowledge base is composed of
three substructures that associate attributes linked together and giving
evidence for different types of risk groups. In the first substructure
attributes connected with systolic and diastolic blood pressure are grouped, in
the second substructure attributes connected with remaining physiological
parameters are grouped and in the third substructure case-history information
is expressed.
3.4. Fuzzy Clustering
This approach requires several stages: data
pre-processing & normalisation, data fuzzyfication, application of fuzzy
clustering algorithm.
Normalisation – five signals have been measured
and each of them has its units. The
absolute normalisation has been used (the same value has been assign to the
specific value of the particular signal in each case, e.g. the value 0.6 is
assigned to pulse frequency in each case).
Case History Data Normalisation – during this
phase, the information obtained from the case history questionnaire is included
into normalised data. This normalisation is based on the experimental rules
provided by experts. The data are re-counted in accordance to these rules to be
comparable with the other ones. The aim of this step is excluding of non-linear
influence of medicaments etc.
Normalised average values setting – average
values have been counted in periodic intervals of several seconds. These
average values are fuzzificated.
Normalised average difference setting – average
value of the particular normalised signal measured during the first rest time
is set. This value is subtracted from the results obtained in the previous
step. Result values are fuzzyficated.
Correlation – If the signal is valid, it has to
be something correlated with the other ones.
This procedure computes correlation between
pairs of signals to exclude no longer valid signal from the further processing.
A measured data fuzzyfication consists in the
assigment of the one membership function grade or more ones to the each
measured value. The universe set has been divided into five carriers. The
universe set overlapping is not equal; the fuzzy set density is higher close to
null.
The membership function has been estimated from
the measured data. Well-known Fuzzy ISODATA method was applied [8].
3.5. A Set of Fuzzy Preference Expert Systems and Fuzzy Aggregation Mechanism
When the expert system is applied to the
problem of computational personality type classification, the knowledge base
inconsistency is inadmissible and has to be solved during knowledge base
construction. The simplest approach could be missing that information, which is
regarded as the least important. Then, the rate of successfulness of the
designed expert system depends something on creator’s intuition to exclude the
least important information. The second great disadvantage of this quite simple
approach to information inconsistency is that usually each expert is not able
to give an expression to each measured physiological parameter.
To find an efficient solution, we set a little
more different problem formulation towards the standard approach described in
the previous text and presented on ‚Figure 1‘. Standard approach involves a construction
of one knowledge base consisting of rules concerning with information given by
all experts.
Figure
1: A Standard way of knowledge base construction
We decided not to design this well-known kind
of knowledge base, but to use a few of preference expert systems [9, 10]. Standard diagnostic expert system gives an
answer consisting of one alternative. Preference expert system is a special
kind of diagnostic expert system, which gives the evaluation of some relation
defined on the set of all possible alternatives. This relation can be fuzzy or
non-fuzzy and depends mainly on structure of this preference expert system.
Figure
2: The change of the problem formulation: a set of preference expert systems
and aggregation mechanism.
In this case, the fuzzy relation ‘Alternative
(model, personality type) A is not worse than alternative (model, personality
type) B for the given situation’ was used. Knowledge base of each expert system
was designed to provide evaluation of this fuzzy relation; it involves
knowledge and information obtained by one of the experts. The rules each
knowledge base consists of needn’t be consistent with the ones included in the
other knowledge bases. Input information of an individual expert system may
differ from the input information of the other ones. As we said above, each
preference expert system provides an evaluation of the specific – in this case
fuzzy – relation. These evaluations are aggregated with respect to relation
characterizing mutual relative importance of used preference expert systems
(See ‚Figure 2‘).
Described approach should be formulated
mathematically. We implemented a non-dominance aggregation algorithm suggested
by Orlowski [12]. For detail system structure see [9].
4. Experiments and Results
4.1.
Measurement Methodology
Psychical load has been presented to the
operator both in the visual and acoustic forms. As psychical load, standard diagnostic
and load tests are used. Short test description follows. The first test
consists of a set of the questions investigating verbal, numeric and perceptual
logic, space perception and analytic and technical skills.
Human operator is invited to answer these
questions as quickly as he/she can. Naturally, the correctness is also
required. The second test consists in the task of ‘number seven subtracting’. A
random number between 900 and 1000 is selected. Operator is invited to subtract
seven from this number, then to subtract seven from the obtained result, etc.
The necessity to answer in the periodic intervals assures psychical stress
genesis. The selected physiological parameters are measured during the both
rest and load test phases. To obtain basic information about tested person life
style, we use case history data (coffee and alcohol drinking, medicaments
taking, movement activity, etc.).
Examples are classified into four classes:
neurotics, persons with risk type of behaviour, neurotics and persons with risk
type of behaviour, persons with negative neuroticism and persons without risk
type of behaviour.
To
assure a possibility to compare and evaluate results obtained by application of
the described approach, each tested person passed a psychological investigation
to diagnose his/her personality type, tendency to psychotic disorder, and
tendency to neurotic behaviour. This diagnosis was used as an absolute standard
during the comparative evaluation. The success rate of each of designed
algorithms for human operator’s personality type classification refers to this
diagnosis.
4.2.
Structures of the Training and Testing Sets
The set of tested persons consisted of 60
persons in age between 18 and 30. Each person could be tested several times
(reason: to obtain more data). This set embodies persons with a tendency to
risk behaviour, persons with the tendency to neuroticism, and persons without
any of the tendencies mentioned above. 600 testing sets, each containing about
30 elements, were selected randomly from the measured data. The relative
representations of neurotics, persons with the tendency to risk behaviour, and
persons without any of the tendencies mentioned above were very similar in all
training sets – 1% neurotics, 24% persons with the tendency to the risk
behaviour, 15% neurotics with the tendency to the risk behaviour, 60% persons
without any of the tendencies mentioned above. Methods, which don‘t required
existence of the training set, e.g. expert system, were tested on the same test
set as the other ones.
4.3.
Expert System
The knowledge base utilizing knowledge about
changes and reactions of physiological parameters has reached the success rate
of 70 per cent for neuroticism diagnosis and 68 per cent for diagnosis of risk persons
according to Bortner’s scale. The total success rate of the expert system with
this knowledge base is 60 per cent.
The expert system with this knowledge base has
manifested lower reaction of heart frequency and systolic blood pressure on
psychic load at persons with risk type of behaviour in comparison with persons
without this type of behaviour.
The results of this knowledge base have not
confirmed our assumptions which has been caused by the fact that during
consultation with an expert influence of average values of physiological
parameters has been underestimated and on the other hand influence of reaction
of physiological parameters on psychic load has been overestimated.
4.4.
Machine Learning
The success rate of generated decision trees
has been between 67 and 72 per cent. The best decision tree has the success
rate of 79 per cent. The classification results have confirmed our assumptions
about possibility to use decision trees for respondents’ classification into
psychological groups with results similar to results of psychological
questionnaires (Bortner scale, Eysenck questionnaire).
Figure
3: A rate of successfulness for the system trained on the particular training
set (20 cases).
4.5.
Combination of Machine Learning and Expert System
The success rate of the knowledge base
developed by application of machine learning is for neuroticism diagnosis 94.9
per cent and 83 per cent for diagnosis of risk persons according to Bortner’s scale.
The total success rate of the expert system with this knowledge base is 81.4
per cent.
Since this knowledge base is relatively small
for verification of the developed methodology, we have repeated the process of
knowledge base division into
training and testing sets with the knowledge
base generated by machine learning algorithm for respondents’ classification
into psychological groups with results similar to results reached by
psychological questionnaires (Bortner scale, Eysenck questionnaire).
Figure 4: A rate of
successfulness for the system trained on the particular training set (20
cases).
4.6. Fuzzy Clustering
The success rate of the model choice provided
by the intelligent interface is about 48 % in a case of using average values of
measured parameters and 55 % in a case of using values of an average
difference. These values are relatively low in comparison with the success rate
of another approaches. We try to suggest several possible reasons for these
results obtaining. The first one is a case history data normalisation. It’s
provided according to several rules obtained by experts, but there were no
comparison study performed to prove validation of these rules. This step could
be one of the possible error sources. The second possible error source consists
in rates of the relevance of the measured signals. For example, skin-galvanic
response is a better stress indicator than muscular tonus. We considered the
same relevance of all measured parameters.
4.7 A
Set of Fuzzy Preference Expert Systems and Fuzzy Aggregation Mechanism
Firstly, three preference expert systems were
applied to all selected testing sets. Table 1 gives an overview of inputs of
these systems. In this example, we consider nine possible personality types
characterized by the different grade of the neuroticism and by the different
grade of tendency to risk behaviour.
Let’s note, the used fuzzy relation is
understood as ”not less important than”. The average success rate of this
system was about 93 percent for neuroticism classification and about 90 percent
for tendency to risk behaviour classification. The average success rate for the
classification of personality type was about 90 percent. The successfulness of
the presented approach depends mostly on the qualitative parameters of the
preference expert system knowledge bases.
Table 1:
An overview of the inputs of the applied preference expert systems
|
Input |
ES1 |
ES2 |
ES3 |
|
Systolic blood pressure |
YES |
- |
|
|
Diastolic blood pressure |
YES |
- |
- |
|
Pulse frequency |
YES |
YES |
- |
|
Skin-galvanic Response |
- |
YES |
- |
|
Muscular Tonus |
- |
- |
YES |
Table 2: Our relative evaluation of the
importance of the applied preference expert systems, the relation ”expert system
A is not less important than expert system B”
|
|
ES1 |
ES2 |
ES3 |
|
ES1 |
1 |
0.8 |
0.9 |
|
ES2 |
0.9 |
1 |
0.9 |
|
ES3 |
0.4 |
0.5 |
1 |
If some of the expert systems outputs of which are
aggregated may not for any reason provide relation for any pair of
alternatives, and other relations are evaluated as the two alternatives are the
same, it can ”blur” a result. Several alternatives can be evaluated as optimal.
5. Conclusion
Results acquired by individual methods do
not differ very significantly. The differences are caused by application of
different methods of knowledge and data processing. The advantage of machine
learning approach is that the decision tree is generated only from measured
data and information from case-history questionnaire. No expert knowledge is
required for these methods. However, the expert knowledge is necessary when
developing a knowledge base. The expert system is fully based on expert
knowledge and theoretical assumptions. Combination of expert system and machine
learning joins knowledge acquired from the database of measured data, data from
case-history questionnaire and knowledge acquired from experts in the given
problem area. Maximum utilization of all these possibilities brings the best
results. From these reasons, machine learning is more suitable for data sets
with great number of attributes than expert system. If we have a data set with
small number of attributes, advantages of machine learning are decreasing. In
such case expert system is more advantageous.
However, in both cases combination of
expert system and machine learning is the best solution because it reaches the
same quality as better of both methods and more over it is strengthened by
advantages of the other method.
Machine learning need for credibility of
results large volume of data, of which one part is used for generation of the
decision tree, and the other one for testing. The advantage of expert system is
its applicability with small data sets, because expert knowledge is used for
development of the knowledge base and data set is used for testing. Combination
of expert system and machine learning is applicable for both large and small
data sets thanks to combination of advantages and properties of both methods.
Fuzzy clustering appears as the least
suitable among the compared approaches. The reason for the relative low rate of
successfulness seems to be normalisation process or considering the same
relevance of all measured parameters.
Last approach is based on an application
of the aggregation of the individual fuzzy preferences into one fuzzy relation
with respect to information about relative importance of individual attributes
characterizing each of the alternatives. These attributes are obtained from the
fuzzy preference expert systems that are able to provide independent
comparative evaluation of all alternatives. There is also available relation
describing importance of the expert systems and thus expressing significance of
individual attributes. Input information of an individual expert system is
different from the input information of the other ones. The rules each
knowledge base consists of needn’t be consistent with the ones included in the
other knowledge bases. For each pair of models A and B, each expert system
supplies evaluation of the relation "model A is not worse than model B for
the given situation". These evaluations are aggregated with respect to
relation characterizing mutual relative importance of used expert systems. This
approach enables not only to perform choice for a set of alternatives but also
to include views of several experts into decision making while incorporation of
a new expert system into the current system can be done more or less without
any problems.
References
[1]
Y.
Kodratoff, Introduction to Machine
Learning (London: Pitman, 1988).
[2]
J. R.
Quinlan, Probabilistic Decision Trees.
Machine Learning: An Artificial Intelligence Approach, Volume III (San Mateo: Morgan-Kaufmann, 1990).
[3]
R. S.
Michalski, Learning Flexible Concepts: Fundamental Ideas and a Method Based on
Two-Tiered Representation. Machine
Learning: An Artificial Intelligence Approach, Volume III (San Mateo:
Morgan-Kaufmann, 1990).
[4]
R.
Mántaras, A Distance-Based Attribute
Selection Measure for Decision Tree Induction (Boston: Kluwer Academic
Publishers, 1991).
[5]
Eysenck,
H.J, The Structure of Human Personality, 2nd edition, London 1960
[6]
Fodor,
J., Ovchinnikov, S.: On aggregation of t-transitive fuzzy binary relations, in Fuzzy
Sets and Systems, 72:135-145, 1995
[7]
Gath,
I., Geva, A.: Unsupervised fuzzy clustering, IEEE Transactions PAMI, 11,
7, 1989, 773-781
[8]
Janků,
L., Lhotská, L..:
Aggregation of Fuzzy Preferences as an Efficient Approach to the Computational
Personality Type Classification Problem,
in Masorkais, N. (Ed.): Problems in Applied
Mathematics and Computational Intelligence, ISBN 960-8052-30-0
[9]
Janků,
L., Šorf, M., Eck, V.: Preliminary Phase of Co-operation between Human
Operator and Intelligent Interface - Computational Personality Type
Classification, research report, BIO 333-09/00. ČVUT FEL, Praha, 38 pp., 2000
[10]
Ovchinnikov,
S.: Means and social welfare functions in fuzzy binary relation spaces. In
Kacprzyk, J. and Fedrizzi, M.: Multiperson Decision Making Using Fuzzy Sets
and Possibility Theory, 143/154, Kluwer
[11]
Orlovski,
S.: Calculus of decomposable properties, fuzzy sets and decisions,
Allerton Press, 1994
[12]
Šorf,
M., Janků, L., Lhotská, L., Eck, V.: Applications of Expert System and
Machine Learning Approach to Intelligent Man-machine Interface. In: Intelligent
Techniques and Soft Computing in Nuclear Science and Engineering. Vol. 1.
Singapore: World Scientific. p. 191-197. - ISBN 981-02-4356-1, 2000