Sie sind auf Seite 1von 7

A Hybrid Fuzzy-Neural Expert System for Diagnosis 

Christoph S. Herrmann
Intellektik, Informatik, TH Darmstadt
Alexanderstrae 10, D{64283 Darmstadt, Germany

Abstract recognition in medical data, neural network based ap-

proaches have led to quite remarkable results, for exam-
Fuzzy Logic, a neural network and an expert ple in processing MRI pictures [Hall et al., 1992] or EEG
system are combined to build a hybrid diag- traces [Mamelak et al., 1991; Jando et al., 1993]. For the
nosis system. With this system we introduce task of acquiring knowledge bases, which is a part of our
a new approach to the acquisition of knowl- hybrid approach, neural networks have been proposed
edge bases. Our system consists of a fuzzy ex- recently [Thrun and Mitchell, 1993].
pert system with a dual source knowledge base. Fuzzy logic [Zadeh, 1965] also makes its appearance in
Two sets of rules are acquired, inductively from medicine, dealing with the uncertainty of verbal expres-
given examples and deductively formulated by sions [Kuncheva, 1991; Nishimura et al., 1991]. Terms
a physician. A fuzzy neural network serves to like many, few or probably are hard to model with con-
learn from sample data and allows to extract ventional logic. The linguistic variables o ered by fuzzy
fuzzy rules for the knowledge base. The diag- representations allow pseudo-verbal descriptions close to
nosis of electroencephalograms by interpreta- natural human expressions.
tion of graphoelements serves as visualization All of the above methods bear advantages as well
for our approach. Preliminary results demon- as disadvantages as will be seen in Section 2. Com-
strate the promising possibilities o ered by our bining these methods not only sums up the advan-
method. tages but also avoids some of the disadvantages. Up
to now, only few approaches in medical diagnosis com-
bine multiple methods of Arti cial Intelligence, although
1 Introduction good results have been made by these means, model-
Repetitively applied cognitive tasks of recognizing and ing a physician's decision process [Kuncheva et al., 1993;
evaluating certain phenomena, called diagnostic tasks, Orsier et al., 1994].
are among the main applications for Arti cial Intelli- Here, we will describe a hybrid system consisting of
gence (AI). As there exists a vast variety of such diag- a fuzzy expert system for rule-based reasoning with a
nostic tasks in medicine, it has always belonged to the fuzzy neural network for acquiring case-based knowledge
spectrum of potential users of Arti cial Intelligence. in addition to the explanation-based knowledge from an
Most popular among AI methods in medicine are expert (Section 3). The automatic acquisition of rules by
knowledge based systems [Buchanan and Shortli e, the network is implemented in parallel to the classical
1985], modeling the diagnostic behaviour of experts. A formulation of expert rules. Two modes of processing
variety of such expert systems is being used in everyday result for the hybrid system: A learning mode to feed
practice of physicians since Shortli e introduced MYCIN the knowledge base and an execution mode diagnosing
[Shortli e, 1976], an expert system designed to diagnose patient data. All components of the system are based on
infections of the human blood. One of the greatest di- a fuzzy representation, serving as an interface notation
culties in designing a convenient expert system is acquir- between the components and making a fuzzi cation of
ing the knowledge base. We introduce a new approach input data necessary.
where a dual source knowledge base is generated by de- In Subsection 3.2, a very e ective mapping technique
ductive and inductive learning. will be introduced, transforming fuzzy variables into a
Neural networks have also made their way into di- neural representation.
agnosis. They are able to learn relationships between To visualize the processing of real medical data in such
data sets by simply having sample data represented to a system, we chose the diagnosis of electroencephalo-
their input and output layers. In the eld of pattern grams (EEGs) for demonstration. This type of medical
data, measured and stored electronically, is very well
 published in C.S. Mellish: International Joint Confer- suited for automatic processing since it need not be con-
ence on Arti cial Intelligence (IJCAI), pp. 494{500, Morgan verted to an electronic representation any more. Other
Kaufmann, 1995 types of data will also be appropriate for diagnosis in
our system. We will describe our sytem apart from the 2.2 Neural Networks
application as far as possible and propose related topics Neural networks have been applied successfully to pat-
that o er similar processing. tern recognition in medical data [Hall et al., 1992;
With our preliminary results (Section 4) we demon- Jando et al., 1993; Mamelak et al., 1991], being a subtask
strate how the automatic system outperforms the accu- of our diagnosis, and o er the following advantages:
racy of a human in one topic of EEG diagnosis. In addi-  The network learns to solve a given classi cation
tion, we found out that the inductive acquisition of rules task autonomously. If the physician knows how to
by learning from examples generates not only quantita- label a given data sample but cannot describe his
tively more rules but also rules with qualitatively higher reasons for doing so, a neural net can learn to assign
precision than the deductive method of formulating ex- the same labels without explicit rules.
pert rules.
 Neural networks are very fault tolerant with respect
2 Comparing and Combining to noisy data. Little change of input causes only lit-
tle change of output, since the data has topographic
In this section we will compare the mentioned methods, characteristics. This is important for medical pat-
expert systems, neural networks and fuzzy logic, with tern recognition, since the net is trained with a -
respect to medical diagnosis and show the prospects of nite number of patient samples. New patients may
a combination that overcomes some of the major disad- have slightly di ering phenomena but an identical
vantages of its components. diagnose.
2.1 Expert Systems When neural networks are applied without the hy-
brid combination shown here, the following disadvan-
Two of the advantages of using expert systems in the tages arise:
context of medical diagnosis are:  After completed training, the network serves its pur-
 The knowledge can be represented in simple 'IF pose but it remains hidden how it does so (black box
premise THEN conclusion' type rules [Bibel, 1993]. phenomenon). Our fuzzy neural net is easily inter-
Since physicians use similar descriptions for their pretable due to a limited but logically de ned ar-
medical decision process, this makes the formula- chitecture. Of course this is bought with a decrease
tion of the rule base simple. of generalization ability.
 In case of malfunction, faulty rules can easily be  It is usually impossible to interact with the learning
changed by a human knowledge engineer, due to process of the network. There is no possibility of
the adequate representation. interfering with its activity except for new training.
The following disadvantages may arise from expert Since every neuron in our network represents the
system usage: premise of a rule, it is possible to weaken, strengthen
or freeze a premise or a whole rule.
 Usually the initially acquired set of rules does not  For the purpose of supervised learning, which is re-
fully satisfy the processing of empiric data and needs quired for our application, a teacher is needed to la-
re nement [Ourston and Mooney, 1994]. This is bel the data. Application speci c, this implies high
avoided in our hybrid system by splitting the rule e ort in labeling EEG traces.
base into two parts, one formulated by a human ex-
pert, and another one generated by machine learn- 2.3 Fuzzy Logic
ing. A fuzzy neural network with the possibility of Fuzzy logic is an excellent means to combine Arti cial
extracting knowledge in the form of rules achieves Intelligence methods [Zadeh, 1993]. The advantages of
the automatic acquisition. fuzziness dealing with imprecision t ideally into medical
 Classical expert systems approach the imprecision diagnosis:
of human expressions solely by means of assigning  The vagueness and uncertainty of human expres-
probabilities to rules, so-called certainty factors 1 sions is well modeled in the fuzzy sets, and a pseudo-
[Buchanan and Shortli e, 1985]. If the verbal ex- verbal representation, similar to an expert's formu-
pression of a rule-premise already contains vague- lation, can be achieved.
ness, these systems fail to model the given problem.  Fuzzy logic avoids the abrupt change from one
The fuzzy expert system we use can handle those discrete output state to another when the input
problems by applying linguistic fuzzy variables. is changed only marginally. This is achieved by
Among these disadvantages, the rst one is of great a quantization of variables into membership func-
importance for rule-based systems, while the second one tions.
touches uncertainty of human reasoning, both of them Fuzzy logic also has some disadvantages that have to
being of particular interest for diagnosis. be taken care of:
1 Even though MYCIN's certainty factor model has proven  High accuracy is needed in tuning the membership
inconsistent, the system outperforms a human expert and values in order to get plausible results. For this
performs nearly una ected by minor changes to the certainty reason we oriented our membership design on actual
factors [Heckerman, 1986]. medical parameters rather than creating own ones.
 Compared to classical expert systems, fuzzy ones inference engine. The fuzzy expert system then comes
require higher computational e ort, due to the com- up with a diagnose.
plex inference mechanisms needed. We `solved' this The following three sections describe the implemen-
problem by taking a SUN workstation to crack the tation of the components of the hybrid system in more
diagnosis nut, resulting in reasonable run times. detail and show the advantages and avoidance of disad-
3 Hybrid Combination
In order to combine the advantages and overcome some 3.1 Fuzzy Representation
of the disadvantages, as sketched above, we designed a It is important for our hybrid approach that all compo-
hybrid system of all three approaches. In Figure 1 the nents are able to deal with fuzzy representations, since
order of processing is visualized. that is the interface between the components. For this
reason, data has to be transformed into a fuzzy repre-
Digital EEG
sentation before it can be diagnosed by our system. This
process depends on the type of data and shall be demon-
strated via our EEG example.
In order to obtain a fuzzy description of EEG patterns,
EEG spectrum some preprocessing has to be carried out with the EEG
traces. First, the brain currents are digitized to allow
computerized manipulation (see gure 1, `Digital EEG').
In the next step the discrete signal is transformed from
the time domain into the frequency domain (see Figure
1, `EEG spectrum'). In its frequency domain represen-

tation each point of an EEG spectrum consists of two


dimensions, namely frequency and amplitude.

The frequency axis is divided into four terms of the
fuzzy variable frequency. Those four terms delta, theta,
alpha and beta correspond to the medical standard of
classifying EEG frequencies [Niedermeyer and Lopes da
Silva, 1993]. In the same manner the amplitude is fuzzi-
Inference Diagnose ed into the terms zero, low, mid, and high of the fuzzy
variable amplitude. It is now possible to formulate de-
scriptions like a wave with beta frequency and low ampli-
Dual tude. This pseudo-linguistic description is very close to
how signal experts or physicians would classify certain
knowledge graphoelements in EEGs.
Physician’s base
3.2 Mapping Fuzzy Variables to Neural
Figure 1: Schematic drawing of the hybrid system In order to learn the diagnosis process in a neural net-
work, the fuzzy features have to be mapped into the in-
The rst two preprocessing steps are application spe- put layer of the net. Here, we demonstrate a new method
ci c (light arrows) and serve as a basis for the extraction of two-dimensional mapping, minimizing the amount of
of fuzzy features. The hybrid system itself is the ser- neurons in the network and keeping their number con-
ial/parallel combination of the lower components where stant.
grey arrows indicate a learning or knowledge acquisition All terms of the two fuzzy variables frequency and
phase and the black arrow indicates the execution phase. amplitude must be mapped into the input layer neu-
During the learning phase two sets of rules are merged to rons of our neural net. In a straight-forward manner
form the dual source knowledge base of the fuzzy expert each out of eight neurons would be assigned the mem-
system. On the one hand rules are gathered from physi- bership value of one of the eight fuzzy terms. This
cians and on the other hand rules are extracted from the would allow the representation of exactly one spectral
fuzzy neural network. For the latter purpose the net is at phenomenon. But, in every EEG time-slice there ex-
rst trained to perform the physician's diagnosis task on ist various di erent phenomena. For example, the main
sample data [Herrmann, 1995b]. After completed train- rhythm should be present in any healthy subject, de-
ing, rules describing the diagnosis process are extracted scribed as a wave with an alpha frequency and a mid
from the net. When the acquisition of knowledge is n- amplitude. Also, there might exist artifacts at the same
ished, the executing phase can be run, actually diagnos- time|mostly bulbus artifacts resulting from eye move-
ing medical data. To achieve this, the execution path ments, being waves with delta frequency and high am-
is switched from the grey to the solid arrows. Now the plitude. If more than one feature were encoded in eight
fuzzy features of a sample (e.g. EEG trace) are com- neurons one could no longer decide which amplitude be-
pared to the formerly acquired knowledge base via the longs to which frequency (binding problem [Rumelhart
and Mc Clelland, 1986]). Therefore, to represent multi- delta zero 0

ple phenomena eight neurons would be required for each delta low 0
phenomenon. Since the number of features contained Or-terms

in every time-slice varies through the EEG (not every

delta mid 0

sample is deranged by an artifact) a representation is delta high 1

needed which is capable of coding multiple phenomena theta zero 0

in a constant number of neurons. theta low 1
We have developed a mapping scheme that brings two unary

fuzzy variables into a network suited representation by theta mid 0 terms

calculating the cross product, which is described in de- theta high 0 bulbus artifact
tail in [Herrmann, 1995a]. The two four-term fuzzy vari- alpha zero 0
1 (ba)

ables result in 16 neurons Nfrequencyamplitude (see g-

ure 2). Each neuron represents the conjunction of two
alpha low 1

fuzzy terms of each variable, thus overcoming the bind- alpha mid 0

ing problem. The activation value of a neuron is calcu- alpha high 0 AND-

lated via the algebraic product of the two represented

beta zero
membership functions:

beta low 1
Ndeltazero = delta  zero = 0:6  0:6 = 0:36 beta mid 0
The sum of activation, resulting from one spectral phe-
nomenon always sums up to 1 for sake of reinterpretabil-
beta high 0

ity of the rules learned by the network. Figure 2: Neural network after completed training
3.3 Fuzzy Neural Network
The fuzzy features, presented to the neural network by Rules (1) to (3) are extracted from the three shaded
the two-dimensional mapping method, are then trained neurons of the network and build up part of the knowl-
to be detected by the net. In order to extract the ac- edge base for the fuzzy expert system.
quired knowledge, a fuzzy-neural network, called FuNe
[Halgamuge and Glesner, 1993; Halgamuge et al., 1993], (1) IF delta mid OR delta high THEN ba
is used in our three layer network. The special multi- (2) IF delta zero THEN NOT ba
layer perceptron architecture is trained with a gradient (3) IF delta high THEN ba
descent algorithm. There exist three types of neurons These rules do not all contribute to the result in the
in the middle layer grouped together topographically. same way, which is expressed by di erent rule strengths.
One group of neurons can perform only the or function The strength of a rule is computed from the weight of
of multiple inputs while another group only performs the connection between the neuron representing the rule
the and function. As there may as well be unary rules, and the output neuron. Rule (2) has a negative strength
composed of simple one-term-premises, there is a third indicating a negation of the conclusion (dashed line in
group of neurons having single inputs and single outputs. Figure 2).
The output neuron simply acts as an or function of all
middle neurons. In the initial state the fully intercon-
nected network represents all possible logical functions
3.4 Fuzzy Expert System
of or premises, and premises, and the 16 unary premises. The heart of our hybrid system is a fuzzy expert sys-
During the learning process all connections below a cer- tem, called FuzzyCLIPS [Knowledge Systems Labora-
tain threshold are eliminated. This pruning method has tory, 1994], derived from CLIPS [Arti cial Intelligence
been proposed by Le Cun in 1990 [LeCun et al., 1990] in Section, 1993]. This expert system consists mainly of
order to increase learning speed|but, it is also useful for three components (see Figure 1, 'Fuzzy expert system'),
limiting the number of resulting rules. This is of special the dual source knowledge base containing the combined
interest to us, because we will extract exactly these rules rule sets from inductive and deductive learning, the user
after the training and want to avoid rules with negligible knowledge base where the actual phenomena of a patient
rule strength. are entered, and the inference mechanism concluding a
A net trained to detect bulbus artifacts is shown in diagnosis from the comparison of the knowledge base.
its nal state in Figure 2. As an example, the pattern When acquiring rules for the knowledge base of an ex-
at the input neurons represents a bulbus artifact (ba) in pert system two major learning paradigms apply: deduc-
a simpli ed manner 2 . The delta frequency component tive learning and inductive learning. Deductive learn-
is high while all other frequency components are low. ing mostly is carried out by explanation-based learn-
The artifact is detected in the output neuron, shown ing [Minton et al., 1990], meaning that the system is
by its activation of 1. Some of the inputs no longer tought which rules lead to the desired decision capa-
contribute to the detection task at all, since their low- bility. Inductive learning is applied when this knowl-
weight connections have been pruned. edge is incomplete but examples may serve as a teacher
for machine learning techniques [Michalski, 1983]. Our
2 For the sake of simplicity we used only values 0 and 1, knowledge base consists of two seperately acquired sets
although they will not occur in reality. of rules. The rst one is acquired deductively from an
expert by explanation-based learning. The second one
is acquired inductively by machine learning in a neural

network. The resulting dual source knowledge base in-

tegrates two autonomous rule sets. They might contain 80

rules that are equal in premises and conclusion but dif-

ferent in the degree of certainty, which results in contra-
diction in the worst case. Like Holland proposed in 1986

[Holland, 1986], we do allow these contradictory rules,
postulating it as a natural feature of decision processes 40

and thus being well suited to model those. The following

example might shed some light upon this matter: 20

Wife rule: IF evening THEN go home

Boss rule: IF evening THEN stay in oce 0

If one equally obeys to wife and boss (equal rule

0 0.2 0.4 0.6 0.8 1

strengths) one certainly needs a third rule to make a

decision, like Figure 3: False positive BAs (dashed) and false nega-
IJCAI rule: IF deadline close THEN stay in oce tive BAs (solid) of a patient record containing no other
artifacts than bulbus artifacts
which contributes to one of the former conclusions. The
same is supposed for our dual source knowledge base.
There may very well be phenomena that exist in sample false positives and false negatives (0.3 - 0.6). A thresh-
data but are described di erently in the expert rules. In old centered in this gap will assure a high probability for
this case the resulting uncertainty has to be solved by neither diagnosing false positives nor false negatives.
di erent rule strengths or a further rule, manifesting one
of the possible conclusions.
Besides an explanation component it is also important

for this medical application that a don't know conclusion

exists, telling the user if abnormal phenomena were de- 80

tected but could not be interpreted.

4 Preliminary Results

When talking about results in medical diagnosis, we

would rst like to mention one major point in where

to put the main focus. Other than in common detec-

tion tasks it is not sucient to simply watch the overall 20

performance of the system in means of average error,

like it would be appropriate for character recognition, 0

for instance. The diagnosis performance must be split

0 0.2 0.4 0.6 0.8 1

into the missing of phenomena, called false negatives,

and the accidental nding of phenomena which are not Figure 4: False positive BAs (dashed) and false negative
actually present, called false positives. With these two BAs (solid) in a patient record with additional electrode
measures it is possible to decide whether the two can be artifacts
kept apart or will be overlapping. For certain diagnosis
types, it is very important to have absolutely no false The example of Figure 4 nicely demonstrates the de-
positives whereas a few false negatives were still accept- gree of accuracy of the system. The diagram shows only
able, like in diagnosing cerebral death. Vice versa, in a narrow gap for the threshold, proposing possible misses
other cases it is important to have no false negatives, in future data. The system can correctly diagnose all
like in diagnosing epileptic seizures for emergency pur- existing bulbus artifacts in the patient record with ease,
poses. (Both examples were chosen from the class of shown in the narrow band of false negatives (rightmost
EEG diagnosis tasks.) peak in Figure 4). Nevertheless, the wide scope of false
Figure 3 shows a typical diagram illustrating false pos- positives gives rise to the assumption that the diagnosis
itives (dashed line) versus false negatives (solid line) as performance is of minor quality, since with a threshold
a function of the threshold  (FP/FN-diagram) for the of  = 0:5 , 10% diagnosed BAs would not actually be
detection of bulbus artifacts (BAs). Since we apply fuzzy such. Retrospectively examined, it showed that those
logic, we get fuzzy results rather than discrete ones. (We false positives were indeed BAs, but of minimal ampli-
do not use standard defuzzi cation.) The example shows tude and thus overlooked in the visual diagnosis. Only
the diagnosis of bulbus artifacts in EEGs. The decision the 1.8% false positives that remained up to a threshold
of the system, whether it is a bulbus artifact or not, is of  = 0:94 were not BAs but electrode artifacts, be-
represented by an analog value in the range from 0 to 1. ing graphoelements that are somewhat related to BAs in
For a binary decision a threshold is needed, above which terms of frequency and amplitude and can only be kept
a BA is assumed. In Figure 3 there is a wide gap between apart due to their shape. Thus, without explicitely com-
paring the system and a human expert yet, we showed This o ers the full range of bene ts from neural net-
with these preliminary results that the system's precision work learning. Additionally, the automatically ac-
is of competing accuracy. quired rules can be supervised and adapted in the
The comparison of the two rule sets, serving as ba- expert system, eliminating the black box problem
sis for the dual source knowledge base, brought up two (see Subsection 2.2).
interesting results.  The coexistence of the two resulting rule bases, with
 The inductively generated rule set, extracted from the possibility of competing knowledge, is not only
the network, contained more rules than the one de- permitted but desired. The competition of rules and
ductively acquired from the expert. These extra the mechanism of taking further rules into account
rules seem redundant at rst glance, because the is part of the cognitive decision task.
expert can easily distinguish between di erent phe- Besides the integration of the existing fuzzy neural
nomena on the basis of more obvious criteria. But network and fuzzy expert system, some intelligent inter-
for the electronic system, designed to diagnose a facing techniques are introduced, that might as well be
vast variety of phenomena, it may very well be of used for di erent applications:
great importance to 'know' every extra description
neglected by the expert.  A two-dimensional mapping technique, used to as-
sign the membership values of linguistic variable
 The extracted rules were more precise in their de- terms to input layer neurons of a network. By
gree of fuzzy membership, since they were actually this means, it is possible to input existing fuzzy
calculated from examples. representations into neural networks in order to
For our example of the bulbus artifact, the physician autonomously acquire case-based knowledge from
formulated the rule: sample data (see Subsection 3.2).
(4) IF frequency=delta and amplitude=high THEN ba  An application-speci c fuzzi cation of spectral EEG
data that will work for most other multidimensional
The automatically generated rule base contained a data, especially other frequency domain data, like
whole group of rules with di erent rule strengths. The voice spectra in speech recognition, as well as all
premises, conclusions and rule strengths are shown in the kinds of medical images (see Subsection 3.1).
following table (compare to the network connections in
gure 2). Only the rules with rule strenghts (RS) above Our preliminary results (see Section 4) point out some
1.0 are listed, while rules with lower rule strengths were promising features:
pruned.  We proved the possible precision of our system, be-
PREMISE CONCLUSION RS ing more accurate than the human expert, when
delta mid OR delta high ba 1.75 applied to the task of diagnosing bulbus artifacts in
delta zero NOT ba 1.51 EEGs.
delta high ba 2.73  Comparing the two rule sets of our dual source
This detailed rule base is the reason, why the system was knowledge base it showed that the inductively ac-
able to detect 10% low-amplitude artifacts (see gure 4). quired set was more extensive and of higher preci-
This would not have been possible with the mere expert sion than the one deductively acquired.
rule (4). In future we plan to investigate the interference ef-
fects of a dual source knowledge base. They occur when
5 Discussion the two bases contain contradictory or similar rules with
di erent rule strengths. For the case of perfect contra-
By combining three major methods of Arti cial Intelli- diction of two rules, possibly resulting in no conclusion,
gence into a single hybrid system we managed to combine a third rule could be generated by a supervising mech-
most of their advantages, avoiding some of their disad- anism to indicate to the user that such a contradiction
vantages at the same time. occurred.
The hybrid system, described in this article, intro-
duces the following new paradigms of modeling the cog-
nitive task of diagnosis:
6 Acknowledgements
The author wishes to thank Andre de Waal for his last
 Instead of either acquiring the whole knowledge polish of the English, as well as Frank Reine and Michael
base automatically from examples, being an induc- Thielscher for their valuable comments on an earlier ver-
tive learning method [Michalski, 1983], or re n- sion of the paper.
ing a rough knowledge base [Ourston and Mooney,
1994], being an enhanced explanation-based learn-
ing method [Minton et al., 1990], we are using a
dual source knowledge base. This knowledge base [Arti cial Intelligence Section, 1993] Arti cial In-
consists of two sets of rules, coexisting with equal telligence Section. CLIPS Reference Guide. NASA,
importance, one generated deductively from rules 1993.
formulated by an expert and another one generated [Bibel, 1993] W. Bibel. Wissensreprasentation und In-
inductively by machine learning in a neural network. ferenz. Vieweg, 1993.
[Buchanan and Shortli e, 1985] B.G. Buchanan and [Kuncheva, 1991] L.I. Kuncheva. Evaluation of comput-
E.H. Shortli e. Rule-based expert systems. Addison erized medical diagnostic decisions via fuzzy sets. In-
Wesley, 1985. ternational Journal on Biomedical Computing, 28:91{
[Halgamuge and Glesner, 1993] S. K. Halgamuge and 100, 1991.
M. Glesner. The fuzzy neural controller FuNe II with a [LeCun et al., 1990] Y. LeCun, J. S. Denker, and S. A.
new adaptive defuzzi cation strategy based on CBAD Solla. Optimal brain damage. In D. S. Touretzky,
distributions. In European Congress on Fuzzy and editor, NIPS, volume 2, pages 598{605, 1990.
Intelligent Technologies, pages 852{855. Verlag der [Mamelak et al., 1991] A.N. Mamelak, J.J. Quattrochi,
Augustinus-Buchhandlung, 1993. and J.A. Hobson. Automated staging of sleep in cats
[Halgamuge et al., 1993] using neural networks. Electroencephalography and
S.K. Halgamuge, W. Pochmuller, S. Ting, M. Hohn, Clinical Neurophysiology, 79:52{61, 1991.
and M. Glesner. Identi cation of underwater sonar [Michalski, 1983] R.S. Michalski. A theory and method-
images using fuzzy-neural architecture FuNeI. In In- ology of inductive learning. In R.S. Michalski,
ternational Conference on Arti cial Neural Networks, J.G. Carbonell, and T.M. Mitchell, editors, Machine
pages 922{925. Springer, 1993. Learning, chapter 4, pages 83{134. Tioga, 1983.
[Hall et al., 1992] L.O. Hall, A.M. Bensaid, L.P. Clarke, [Minton et al., 1990] S. Minton, J.G. Carbonell, C.A.
R.P. Velthuizen, M.S. Silbiger, and J.C. Bezdek. A Knoblock, D.R. Kuokka, O. Etzioni, and Y. Gil.
comparison of neural network and fuzzy clustering Explanation-based learning: A problem solving per-
techniques in segmenting magnetic images of the spective. In J.G. Carbonell, editor, Machine Learning:
brain. IEEE transactions on neural networks, 3(5), Paradigms and Methods, pages 63{118. MIT/Elsevier,
1992. 1990.
[Heckerman, 1986] D. Heckerman. Probabilistic inter- [Niedermeyer and Lopes da Silva, 1993]
pretatioin for mycin`s certainty factors. In [Kanal and E. Niedermeyer and F. Lopes da Silva. Electroen-
Lemmer, 1986]. 1986. cephalography, Basic Principles, Clinical Applications
[Herrmann, 1995a] C.S. Herrmann. Fuzzy logic as inter- and Related Fields. William & Wilkins, 1993.
facing technique in hybrid ai-systems. In A.L. Ralescu, [Nishimura et al., 1991] H. Nishimura, M. Kambe, Fu-
editor, Fuzzy Logic in Arti cial Intelligence. IJCAI
Workshop, 1995. to appear. tagami, K. K. Morishita, and Tsubokura T. Fuzzy
[Herrmann, 1995b] C.S. Herrmann. A fuzzy neural net- realization in clinical test database system. Interna-
tional Journal on Biomedical Computing, 28:289{296,
work for detecting graphoelements in EEGs. In Super- 1991.
computers in Brain Research. World Scienti c Pub- [Orsier et al., 1994] B. Orsier, I. Iordanova, V. Rialle,
lishing Company, 1995.
[Holland, 1986] J.H. Holland. Escaping brittleness: The A. Giacometti, and A. Villa. Hybrid systems for ex-
possibilities of general-purpose learning algorithms pertise modeling: From concepts to a medical appli-
to parallel rule-based systems. In R.S. Michalski, cation in electromyography. Computers and Arti cial
J.G. Carbonell, and T.M. Mitchell, editors, Machine Intelligence, 13(5):423{440, 1994.
Learning, volume 2, chapter 20, pages 593{623. Mor- [Ourston and Mooney, 1994]
gan Kaufmann, 1986. D. Ourston and R. J. Mooney. Theory re nement
[Jando et al., 1993] G. Jando, R.M. Siegel, Z. Horvath, combining analytical and empirical methods. Arti -
and G. Buzsaki. Pattern recognition of the electroen- cial Intelligence Journal, 66:273{309, 1994.
cephalogram by arti cial neural networks. Electroen- [Rumelhart and Mc Clelland, 1986] D.E. Rumelhart
cephalography and Clinical Neurophysiology, 86:100{ and J.L. Mc Clelland. Parallel Distributed Process-
109, 1993. ing: Explorations in the Microstructure of Cognition.
[Kanal and Lemmer, 1986] L.N. Kanal and J.F. Lem- MIT Press, 1986.
mer, editors. Uncertainty in Atrti cial Intelligence. [Shortli e, 1976] E. H. Shortli e. Computer-based med-
North-Holland, 1986. ical consultation: MYCIN. American Elsevier, 1976.
[Klement and Slany, 1993] E.P. Klement and W. Slany, [Thrun and Mitchell, 1993]
editors. Fuzzy logic in arti cial intelligence. 8th Aus- S. Thrun and T.M. Mitchell. Integrating inductive
trian Arti cial Intelligence Conference, LNAI 695, neural network learning and explanation-based learn-
Springer, 1993. ing. In Proceedings IJCAI-93, pages 930{936. Morgan
[Knowledge Systems Laboratory, 1994] Knowledge Sys- Kaufmann, 1993.
tems Laboratory. FuzzyCLIPS User`s Guide Version [Zadeh, 1965] L.A. Zadeh. Fuzzy sets. Information and
6.02A. National Research Council Canada, 1994. Control, 8:338{353, 1965.
[Kuncheva et al., 1993] L.I. Kuncheva, R.Z. Zlatev, S.N. [Zadeh, 1993] L.A. Zadeh. The role of fuzzy logic and
Neshkova, and H. Gamper. A combination scheme soft computing in the conception and design of intel-
of arti cial intelligence and fuzzy pattern recognition ligent systems. In [Klement and Slany, 1993], page 1,
in medical diagnosis. In [Klement and Slany, 1993], 1993.
pages 157{164, 1993.