Sei sulla pagina 1di 11

FUI2Y

sets and systems


ELSEVIER Fuzzy Sets and Systems 79 (1996) 163-173

Using meta-rules for fuzzy inference control


Zhengxin Chen
Department oJ' Computer Science, University of Nebraska at Omaha, Omaha, NE 68182-0500. USA
Received March 1993; revised May 1995

Abstract
Recent cross-fertilization between fuzzy set theory and rule-based expert systems has resulted in successful fuzzy
expert systems. A fuzzy expert system is an expert system which incorporates fuzzy sets and/or fuzzy logic into
its reasoning process and/or knowledge representation scheme. In this paper, we investigate the issue of employ-
ing fuzzy set techniques into inference process itself and its impact on the architecture of rule-based expert sys-
tems. In particular, we discuss the use of fuzzy set concepts to achieve flexible inference control of expert systems.
An extended model of expert systems is introduced, in which fuzzy techniques are used to deal with conflict res-
olution through meta-rules. A top level algorithm is provided which supports flexible inference control based on
different user environments or user stereotypes. The algorithm is illustrated by two examples. First we use a sim-
ple example (a travel agency) to describe in some detail how this algorithm works. In the second example, we
use an expert system for orthopedics to explain how the key ideas illustrated in the first example can be ap-
plied to real-world problems in the medical domain. In addition, we also provide a comparative study and a tax-
onomy which categorizes different ways of incorporating fuzzy set theory into expert systems.
Keywords: Expert system architecture; Efficient search of knowledge base; Meta-rules; Fuzzy inference control; Fuzzy pri-
ority vector; User environments and user stereotypes

I. Introduction reasoning (including the reasoning of human experts)


is at least somewhat fuzzy in nature, then this direction
Recent cross-fertilization between fuzzy set theory of study is extremely important. Since inference en-
and rule-based expert systems has resulted in suc- gine is the heart of an expert system, this consideration
cessful fuzzy expert systems [13, 15,27]. A fuzzy leads to the suggestion that inference engine should be
expert system is an expert system which incorporates fuzzified. In this paper, we investigate the issue of em-
fuzzy sets and/or fuzzy logic into its reasoning pro- ploying fuzzy techniques into inference process itself
cess and/or knowledge representation scheme [13]. and its impact on the architecture of rule-based expert
Studies in fuzzy expert systems typically apply fuzzy systems.
set techniques to conventional expert systems. Less Early expert systems simply dealt with uncertainty
attention has been paid in the other direction of cross- issue in an ad hoc manner (such as using certainty
fertilization, namely, starting from the intrinsic fea- factors in MYCIN [7]). Later developments have put
tures (or inherent problems) of expert systems, how more emphasis on theories of uncertainty, but the bulk
to incorporate fuzzy set techniques to improve the of study has been on the uncertainty of domain know-
architecture of expert systems. If we agree that human ledge rather than the inference itself. Although

S0165-0114/96/$15.00 (~) 1996 Elsevier Science B.V. All rights reserved


PII S0165-0114(95~00179-4
164 Z Chen/Fuzzy Sets and Systems 79 (1996) 163-173

there are indications of incorporating fuzzy set systems (by examining real-time systems and systems
theory to generalize individual inference rules in adapted to user environments or stereotypes), and
classical logic, such as modus ponens [8], little the role of meta-level reasoning in realizing flexible
work has been done on how the factor of uncer- inference control. In Section 3, we provide a revised
tainty has changed the overall architecture of expert model of expert systems, which extends conventional
systems. architecture of expert systems and supports flexible
Conventional expert systems typically include a inference control. In Section 4, we further discuss how
domain knowledge base, a domain-independent infer- to use meta-rules to support fuzzy inference control.
ence engine, as well as other components, such as the In Section 5, a top level algorithm which supports
working memory which holds the case-specific data, flexible inference control based on user environments
an explanation unit which provides explanations to the or stereotypes is provided, and is illustrated by two
user, and a user interface. In a rule-based expert sys- examples. The first example is concerned with a
tem, the inference engine searches the domain know- travel agency. Since the example is simple, we
ledge base to determine which rule should be fired can use it to explain the key ideas in our ap-
next. Conventional expert systems employ a fixed in- proach in detail. Note that this example serves the
ference control scheme to fire the rules stored in the similar purpose as the Travel-Expert does [1], al-
knowledge base using backward reasoning or forward though the actual problems to be solved by these
reasoning. A problem associated with the conven- two travel expert systems differ. In the second ex-
tional approach is that in many applications a more ample, we use an expert system for orthopedics
efficient search of knowledge bases is desirable. In to explain how the flexible inference control as il-
order to support efficient search, a kind of flexibility is lustrated in the simple example can be used in
needed in inference control. Starting from these prob- a real-world problem in the medical domain. In
lems inherent to expert systems, in this paper we in- Section 6, we provide a comparative study and
vestigate what fuzzy set theory can offer. As a special a taxonomy which categorizes different ways
case of the flexible inference control, we consider the of incorporating fuzzy set theory into expert
issue of conflict resolution which deals with the systems. The paper is concluded in Section 7,
case when there are several rules satisfying a cer- where our contribution and some future work are
tain condition and competing for firing. The tradi- summarized.
tional way of conflict resolution is to develop some
fixed strategies, such as more specific rules (namely,
rules have a greater number of conditions) are al- 2. Meta-level reasoning for flexible inference
ways preferred to more general rules with fewer control
conditions. If flexible inference is supported, con-
flict resolution should be handled in a more flexible In this section and next section, we temporarily
manner. put aside the issue of fuzzy set theory. Instead, we
As summarized in [9], uncertainty associated with will take a look on one problem inherent in rule-
conflict resolution can be divided into two cate- based expert systems, namely, that related to flexible
gories: explicit priority of rules and implicit priority inference.
of rules. Implicit priority may be due to specificity
of pattems, recentness of facts matching patterns, 2.1. The need for flexible inference control
ordering of pattems, or the order in which rules
are entered. In this paper, we address the issue of As noted by [2], there is a need for more flexible
dealing with conflict resolution using explicit pri- styles of inference and control over the strategies
ority; fuzzy operators are used to control the pri- guiding the order of inferences. One choice is to
ority for more efficient search in the knowledge attach a scheduler in inference engine that enables
base. explicit decisions to be made about which actions
The paper is organized as follows. In Section 2, we are to be taken (e.g., which rules to apply, whether
discuss the need for flexible inference control in expert to use forward or backward chaining, etc.) [2]. The
Z. Chen/Fuzzy Sets and Systems 79 (1996) 163 173 165

main consideration here is to provide flexible in- model. To support such kind of system adaptability,
ference control to reduce searching in knowledge flexible inference control is needed. A more radical
bases. To get a more concrete idea about the impor- idea on the role of users can be found in [22] where
tance of flexible inference control, in the following the issue of user participation of reasoning was dis-
we briefly examine two different application areas: cussed. In addition, meta-level reasoning can be used
real-time systems and systems adaptable to various to incorporate user modeling for flexible inference
users. control [6].
Flexible control in real-time systems. There are
some common considerations behind real-time expert 2.2. The role of meta-level reasoning
systems. In order to provide timely response, most
approaches have relied on a more restricted manner The discussion on the need for flexible inference
of searching, namely, to restrict the portion of know- control and the employment of meta-level reason-
ledge base to be searched. The difference is only in ing to realizing flexible inference control can be
how to get this done. In the designated inference traced back to early expert systems such as Meta-
engine approach [20], each inference engine only pro- DENDRAL and TEIRESIAS (the preprocessor of
cesses a particular kind of data; although this seems to MYC1N). Associated with these studies is a discus-
be performing data partitioning, the effect is that only sion from Davis [7], where a framework of view-
part of knowledge base will be searched by a parti- ing issues of knowledge indexing and retrieval was
cular inference engine, because the particular data summarized. Strategies were viewed as a means
processed by it determines the portion of knowl- of controlling invocation in situations where tra-
edge base to be searched. In contrast to partioning ditional selection mechanisms become ineffective.
knowledge base, the approaches taken by [12] and Several ways of effecting such control were dis-
[23] rely on meta-level reasoning and changing of cussed; in particular, meta-rules were used as a means
focus in searching. The term "meta-level inference" of specifying strategies which offer a number of
means "inference performed at one level which is advantages.
about another level." To assure flexibility, some According to a more recent study on meta-level in-
degree of domain-dependent control has also been ference systems [25], the terms "domain knowledge"
introduced [23]. and "control knowledge" are often used to distinguish
User environments, user modeling, system adapt- what a system knows from how the system uses what
ability and perturbation models. The need for flexible it knows. One problem with this terminology is that
inference control can also be justified from the dif- it suggests that control knowledge is domain indepen-
ferent circumstances or environments in which users dent. However, control knowledge can be either do-
use the system (which will be referred to as user en- main dependent or domain independent, depending on
vironments), or from the perspective of different user whether it refers to the contents of particular elements
groups of user models [18]. Our viewpoint bears a of domain knowledge, as in "use rules that mention
similarity with the perturbation models as presented in cheap blood-tests before rules that mention expensive
[14] where the role of user models in cooperative in- blood-tests", or whether it only refers to the general
teractive systems was analyzed. According to the dis- form of the domain knowledge, without referring to
cussion provided there, a user model refers to a sys- its domain-specific contents, as in "use cheap rules
tem knowledge source that contains explicit assump- before expensive rules".
tions on all aspects of the user that may be relevant Meta-level reasoning thus plays an important role
for the dialogue behavior of the system. Among our between the domain knowledge in the knowledge base
interest is the notion of perturbation models which and the control knowledge used by the inference en-
can represent information about the user beyond that gine. If reasoning is fuzzy in nature, then the inference
of the system model, while maintaining a close link procedure itself should be fuzzified. The fuzziness of
between the two. In a perturbation model, the user inference can be achieved by using meta-level reason-
model is assumed to be similar to the domain model, ing in an extended expert system model as shown in
differing only in certain perturbations to the domain the next section.
166 Z. Chen/Fuzzy Sets and Systems 79 (1996) 163-173

3. Incorporating flexible inference controller into explanation component, and U to denote user inter-
expert system architecture face, then a classical expert system can be denoted
as (K, LE, W, U).
In order to describe the extended model, we use After incorporating the discussions in the previous
the following notations. If we view the inferencing sections, the above classical model can be enriched
process in the architecture of a classical expert sys- when meta-level flexible control is introduced. The
tem as an abstract operator ! (the inference engine) revised model consists of an additional component at
operating on its unary operand K (the knowledge the conceptual level, namely, C, a flexible searching
base), then we can write controller. As to be explained in the next section, the
flexible inference controller discussed in this paper is
I(K). a fuzzy controller. Putting these things together, the
resulting model can be denoted as (K, (L C),E, W, U).
To make the inference process more flexible, the func-
The relationships between these components are
tionality of the conventional inference engine should
depicted in Fig. 1.
be expanded. According to the consideration of meta-
level reasoning, we may introduce a new component
setting in between the inference engine and the know-
4. Using meta-rules for fuzzy inference control in
ledge base; this component contains domain-
expert systems
dependent meta-rules for flexible searching. The
inference engine with expanded functionality will
We can now address the issue of fuzzy inference
be denoted as an extended operator I', and the in-
control. Intuitively, the role that fuzzy set theory can
ferencing process in an expert system with such an
play in flexible inference control can be described as
inference engine can be denoted as
follows. As already indicated earlier, flexible inference
t'(K). control means to restrict the search in the knowledge
base. One way is to restrict the search to a particu-
Since the expanded inference engine can be realized lar region (namely, a semantically interrelated group
by attaching on it a flexible inference controller (which of rules) in the knowledge base (instead of the en-
can be considered as an operator C), the extended tire knowledge base). Alternatively, we may increase
operator I ~ can be viewed as the result of the com- the difference between the firing priority of the rules
bination ! o C (where o denotes combination of two when dealing with a particular search request, so that
operators). Replacing I ~ by ! o C, we have fewer rules will remain in the conflict set. Since these
two approaches are related (for instance, different re-
(I o C)(K). gions in the knowledge base may be given differ-
ent priority for searching), in the following, we will
But since the flexible inference controller C can examine the issue of how to use fuzzy operators to
directly work on the knowledge base K, the above
control the priority. In order to explicitly express the
formula can be rewritten as priority, each rule in the knowledge base has an ad
Jr o ( C ( K ) ) .
hoc number ranging from 0 to 1 (or a vector of such
numbers) associated with it to indicate the likeliness
This formula expresses an alternative view on the role of rule firing (or different conditions of the likeliness
of C: the flexible inference controller itself is part when a vector is used). The number (or more gen-
of the operand on which a more general operator I erally, the vector) discussed here to indicate the pri-
operates; if C consists of a set of rules, then these rules ority of a rule will be referred to as the fuzzy prior-
are apparently recta-rules. ity number (or more generally, fuzzy priority vector)
If we use W to denote working memory (which of that rule. The use of a vector (instead of a single
stores case-specific data), K to denote the knowl- number) is an extremely important idea in our ap-
edge base in the problem domain, I to denote a proach, because it provides a chance to allow the sys-
domain-independent inference engine, E to denote tem to fire these rules in multiple ways. For instance,
Z. Chen/Fuzzy Sets and Systems 79 (1996) 163 173 167

Knowledze Base

Controller

Worklng Memor~ Us~ lnt~v.e Uler

] .gxpln,tion unit

Fig. 1. Expertsystem with fuzzy inferencecontrol.

when user stereotypes are considered, the priority of tions on how to activate the rules stored in the domain
rules firing for one type of users may be different for knowledge base. Each meta-rule takes the following
some other type of users. If we use one dimension to format:
indicate one type of users, then a k-dimensional vec-
tor can take care of all k types of users. Obviously, the if condition
numbers associated with the rules represent the pri- then fire rules in knowledge domain with certain
ority of using these rules (a consideration from tradi- features.
tional knowledge engineering), rather than the uncer-
When the fuzzy priority vectors are used, a meta-
tainty of these rules. The flexible inference engine (as
rule can take a more concrete form such as
discussed in the previous section) contains meta-rules
which provide fuzzy operators to operate on the prior- if the ith factor is involved in current request
ity (thus change the priority) of the rules, so that rules then form a meta-level vector whose ith dimension
can be fired according to the specification of the meta- is CON
rules. and all the other dimensions take the value of 0.
One possible scheme is described below.
( 1) Instead of attaching a single number to indicate (The meaning of CON is discussed below.)
the priority of a rule, each rule in the domain know- Fuzzy inference controller stores various meta-
ledge base is associated with a f u z z y priority vector rules which indicate how to employ fuzzy operators.
with k factors (namely, a k-dimensional vector) to Various fuzzy operators exist [9, l 1]; of particular
indicate k different ways of firing rules (so that flexible interest tons are concentration and intensification
inference control can be realized). [9]. For instance, meta-rules may use concentration
(2) Meta reasoning is achieved through meta-rules CON(A),
fired by the inference engine. Meta-rules are stored in
the flexible inference controller; they provide instruc- ~CON(A)(X) = (re(X)) 2.
168 Z Chen/Fuzzy Sets and Systems 79 (1996) 163-173

The use of concentration operator will not change the 3. perform fuzzy operations indicated by the meta-
relative relationship (namely, relative priority), but inference rule on domain rules;
will increase the difference between priorities. If cut- 4. perform backward chaining using activated do-
off threshold remains unchanged, the number of rules main rules;
in the conflict set will be reduced. For instance, sup- 5. display result to the user.
pose we use s-cut = 0.5, and the original conflict set
In the above algorithm, step 1 takes place in the
consists of three rules with fuzzy priority numbers 0.5,
user interface component U, step 2 takes place in the
0.6, and 0.8, respectively. After the concentration op-
inference engine I and fuzzy inference controller C,
erator is applied, the fuzzy priority numbers associated
step 3 takes place in fuzzy inference controller C and
with these three rules become 0.25, 0.36 and 0.64, re-
knowledge base K, step 4 takes place in inference
spectively; if c~-cut remains unchanged, then only the
engine I, knowledge base K and working memory
last rule (with fuzzy priority number 0.64) will remain
W, and finally, step 5 again takes place in the user
in the conflict set.
interface component U.
Another useful operator is intensification INT(A),
We use a fuzzy inference controller to reduce the
which is defined as
size of the conflict set. The form of the meta-rules
~INT(A)(X) = provided by the fuzzy inference controller follow what
was described in the previous section; the following
2(IrA(X))2 for 0 ~</~A(x) ~ 0.5,
is an example:
1-- 2(1-- /~A(x))2 for 0.5 < IrA(X) < 1.
if user belongs to ith type,
The INT operation is like contrast intensification of then form a meta-level reasoning vector
a picture: it raises the membership grade of those (0, 0, ..., CON . . . . . 0).
elements within the crossover points and reduces (namely, a vector with ith dimension set to fuzzy
the membership grade of those outside the crossover concentration operator, while all the other dimen-
points, thus increasing the contrast grade between two sions set to 0).
kinds of elements. In our case, since each element
represents a priority, the INT operation as specified in The vector as specified by the meta-level rule op-
the meta-rules will increase the differences between erates on rules (denoted as rKB'S) in the knowledge
the priority values associated with different rules, thus base. This can be denoted as
reducing the number of rules in the conflict rule set.
maX(rmeta(rKB,))
(here we use bold face to emphasize that meta-rule
5. A top level algorithm and examples rmeta is an operator). The specified operator is then
further carried out through vector multiplications:
5.1. A top level algorithm
m a x ( Vmeta • VtKBI ),
The purpose of this paper is to discuss the important where Vmetaand VKBI are two fuzzy priority vectors
role of fuzzy set theory for expert systems through associated with the meta-rule and a rule in domain
meta-rules, rather than to provide the exact form of the knowledge base, respectively; and a symbol ' is used
meta-level reasoning. Nevertheless, to illustrate our
to denote vector transposition.
idea, in this section, we provide a concrete version
of a top level algorithm for fuzzy flexible inference 5.2. A simple example for travel agency
control. This algorithm uses meta-rules to deal with
conflict resolution by incorporating user environments To illustrate how fuzzy inference control can im-
or stereotypes. prove conflict resolution, as well as how the top
while request from user do level algorithm works, we now provide a simple yet
1. categorize the user environment or stereotype; still (somewhat) comprehensive example from user
2. fire a meta-rule; modeling.
Z. Chen/Fuzzy Sets and Systems 79 (1996) 163-173 169

Table 1

Lef~-hand side Right-hand side


R1 Season Interest Destination Fuzzy priority vector (B Y F)

rl-1 Spring for kids Orlando (1.0, 0.3, 1.0)


rl-2 Spring for fun Orlando (1.0, 0.9, 0.5)
rl-3 Spring for sun Orlando (1.0, 0.9, 0.6)
rl-4 Winter for sun Orlando (1.0, 1.0, 0.9)
rl-5 Summer for sightseeing Orlando (1.0, 0.7, 0.5)

Table 2

Left-hand side Right-hand side


R2 Destination Cost Package Fuzzy priority vector (B Y F)

r2-1 Orlando Luxury 1 (0.8, 0.8, 0.4)


r2-2 Orlando Economy 2 (1.0, 1.0, 1.0)
r2-3 Orlando Medium 3 (1.0, 0.9, 0.6)

For convenience of discussion, the example is umn) for the right-hand side (RHS). It also contains
simplified (thus it is somewhat hypocritical), but the a rule number (r-#). Each tuple (row) in a relation
idea illustrated here can be generalized into more re- table represents the meaning as "if attributes in LHS
alistic situations. The expert system mimics a travel then take the value of attribute at the RHS". For in-
agency which selects appropriate trip packages (from stance, for the relation R2, the first row in Table 2
a set of existing packages) for the users (potential represents rule r2-1: "if destination = Orlando and
travelers). A package is selected based on such fac- cost = Luxury then package = 1 with fuzzy priority
tors as destination and the cost the user is willing to vector (0.8, 0.8, 0.4)".
pay. Three types of users have been identified: Type To simplify our discussion, competitive rules in-
I (business people, denoted as B), type II (young volving the same set of variables (attributes) will be
couple vacationers, denoted as Y), and type III (fam- represented as tuples of the same relation (as dis-
ily vacationers, denoted as F). In case the destination cussed in the context of relational databases), with
is not known (this may happen when vacationers those shared attributes as relation schemes. In Ta-
just need a break, but do not know where to go), bles 1 and 2, some tuples (which involve destination
the agency may first suggest a destination for the Orlando) in two relations are shown.
user based on his or her interest and the season the Suppose backward chaining is used for reasoning,
trip is to be made, then use this destination to find and our goal is "package = X" without known desti-
a package. For business type travelers we assume nation. If we use 0.6 as the threshold (namely, c~-cut),
that destination is always known, the system thus there are 5 x 3 possible combinations of rule firing.
will not select destination for this type of user. Con- In general, if there is no meta-rule used, there may
sequently, the fuzzy priority number is assigned to be up to nl x n2 number of combinations of possible
be 1. firings; each combination is associated with a fuzzy
When a user enters a session, the user interface priority vector which is calculated from the involved
module will ask questions so that the user will be rules. For instance, the firing of rule r2-1 followed by
categorized into one of these three categories. rule rl-1 produces a fuzzy priority vector
Each relational scheme (i.e., the variable names
appearing at the table headline, see Tables 1 and (rain(1.0,0.8), rain(0.3,0.8), min(1.0, 0.4))
2) contains at most two attributes (columns) for the
left-hand side (LHS) of rules and one attribute (col- = 0.8,0.3,0.4).
170 Z. Chen/Fuzzy Sets and Systems 79 (1996) 163-173

Now let us describe how the fuzzy inference con- the operations. This system is to be used in a clinic or
troller is used in the vacation example. Suppose a user emergency room.
is a family vacationer. A meta-rule stored in the fuzzy There is a bulk of knowledge which should be
inference controller will be fired, which forms a fuzzy acquired to build the knowledge base. For example,
priority vector Vmetaz(0, 0, CON) (since family va- to diagnose knee pain, it can be one of the follow-
cationer is the 3rd dimension in the vector). The rule ing six categories: medial meniscus tear, ligament
r2-1 provides its own fuzzy priority vector Vr~, = tear, fracture, arthritis, Patella femural pain, Os-
(0.8, 0.8, 0.4). The result of vector multiplication gives good Schlauter's disease, and cartilage injury. The
size of the knowledge requires an efficient search-
max(O 0 1)(0.8 0.8 0.4)' = max(O 0 0 . 4 ) = 0.4. ing mechanism. In addition, it has been noticed
that for the domain of medicine and diagnosis, spe-
Meta rules can be applied to other domain rules in
cial considerations need to be taken into account in
a similar manner. If we still use threshold (a-cut) =
knowledge acquisition. Experts gather evidence for
0.6, then after meta-rules are fired and fuzzy concen-
diagnosis over time, reevaluating current informa-
tration operator (CON) is applied on the rules, only
tion on a patient's condition at various points in the
the following rules survive:
workup. In different cases, the physician performs
different sequences of actions. In attempting to build
R1 Season Interest Destination Fuzzy priority for F a knowledge-based system for this task, general-
purpose control strategies fall short [ 10]. In our ortho-
rl - 1 Spring for kids Orlando 1.0
pedics system this problem is handled by introducing
rl -4 Winter for sun Orlando 0.8
multiple user environments so that a kind of flexibi-
lity of inference control can be supported using a
fuzzy inference controller. When the system performs
a consultation, the hypothesis examined can be tai-
R2 DestinationC o s t Package Fuzzypriority for F lored to the population by which this system is served.
r2- 2 Orlando Economy 2 1.0 For example, if the system is used at a ski resort town,
since the anterior cruciate ligament tear and fracture
are common injuries, rules associated with these dis-
In addition to the concentration operator CON, other eases should be given higher priority. On the other
operators (such as INT) may also be used by meta- hand, if the system is used in a retirement commu-
rules. nity, the arthritis diagnoses would be considered first.
Three user environments have been identified: resort
5.3. An expert system for orthopedics town, retirement community, ~,nd others. Meta-rules
have been developed to support flexible inference
As an example for real-world problems, we now control so that domain rules with different priorities
briefly describe an expert system for orthopedics in the will be fired in different user environments.
medical domain. The emphasis of our discussion will The inference engine uses backward reason-
be on how the use of fuzzy inference controller can ing. Each rule may have up to six conditions and
provide efficient search in real-world applications. Or- one conclusion. Each rule is attached with a reg-
thopedics is the surgical specialty dealing with bones ular certainty factor (similar to what was used in
and joints. The need for having such a system can MYCIN [7]), along with the priorities of each
be explained below. There are two camps of physi- rule in three different user environments in the
cians: those who practice medicine and those who order of: (resort town, retirement commu-
practice surgery. Not many expert systems have been nity, o t h e r s ) . The following are some sample
developed for those who practice surgery, because the rules:
diagnosis part of the surgeons' work is seen as mini-
mal. Being as it is, an expert system would assist the If the patient has a swollen knee,
surgeon to diagnosis so that he or she can focus on the patient is having fevers,
Z. Chen/Fuzzy Sets and Systems 79 (1996) 163-173 171

the knee has been aspirated, from imprecise knowledge. In effect, in fttzzy logic
bacteria was grown on the culture, everything is allowed to be a matter of degree [26].
then the patient has septic arthritis. (1.0) (0.4, 0.9, 0.6) A more detailed model of fuzzy expert system was
given by Buckley et al. [3] where a fuzzy expert sys-
If the patient has a ligament tear,
tem (based on OPS5) was described. Such a system
the patient has an MRI scan,
accepts as input a fuzzy vector and produces as out-
the patient has positive pivot shift,
put a fuzzy set of conclusions; internally, all data are
then the patient has an anterior cruciate ligament tear.
considered fuzzy. A fuzzy set of fireable rules is de-
(0.8) (0.9, 0.4, 0.4)
fined for each state of the system's working memory;
The meaning of these rules should be clear. For this set is executed recursively until a null c~-cut set of
example, in the second sample rule, there are three fireable rules is encountered. In the fuzzy rule based
conditions and one conclusion. The uncertainty factor expert system, an action around the inference engine
is 0.8, while the fuzzy priority vector is (0.9, 0.4, 0.4). may be make or modify data (thus modifying working
The system works in a very similar manner as in the memory), make modify rules (thus updating know-
first example (namely, the travel agency), although ledge base), or producing output (fuzzy set of con-
the knowledge base is now much more complicated. clusions). To obtain crisp output, a threshold is used.
Since we have already used the first example to explain A particular form of system studied has been the
the basic ideas of the top level algorithm, details are fuzzy controller. Buckley and Ying [4] argued that
omitted. An examination on the contents of the know- there are three reasons for basing a fuzzy controller
ledge base indicates that the portion of the knowledge on a fuzzy expert system shell: fuzzy controllers are
base to be searched is now reduced to about half when special purpose fuzzy expert systems; more complex
flexible inference control is used. Although the use of fuzzy control rules are possible; the fuzzy expert sys-
fuzzy inference control itself introduced some over- tem shell can be used in designing an optimal fuzzy
head, it will be paid offby the efficient search. Overall, controller. According to Buckley [5], in the theory of
a 30% of saving of average response time is expected. the fuzzy controller what one is interested is the func-
tional relationship between the defuzzified output and
all inputs to the fuzzy controller. The basic notation is
6. Relationship with other approaches: to fuzzify the input, from the output to form rules and
A comparison and a taxonomy then defuzzify it to get crisp result. More discussion
can be found in Maeda and Murakami [16] and Matia
In order to explain how our consideration of et al. [17].
applying fuzzy set theory to expert systems is related Recent progress on fuzzy expert systems can be
to and different from existing approaches, here we found in a special issue of this joumal [19], as well
provide a brief summary of some related work. The as a monograph [13]. Turksen [24] suggested that in
notion of fuzzy expert systems was first discussed by second generation expert systems, both fuzzy knowl-
Zadeh [26] and Negoita [21]. Fuzzy logic forms the edge representation and fuzzy inference methods will
basis for fuzzy expert systems. Zadeh emphasized be used. Instead of considering a set of rules of the
that the employment of fuzzy logic as a framework form " I f X is Ak then Y is Bk" (where Ak's and Bk's
for the management of uncertainty in expert systems are linguistic values of linguistic variables), we can
makes it possible to consider issues like the following: consider its fuzzified version " I f X is A' then Y is B " ;
(1) the fuzziness of antecedents and/or consequences since A~ and B' can represent many linguistic vari-
in rules, (2) partial match between the antecedent ation of linguistic values and since the approximate
of a rule and fact supplied by the user, and (3) the reasoning can cope with such variations, we need to
presence of fuzzy quantifiers in the antecedent and/or identify only a set of basic anchor rules in the design
the consequent of a rule. Fuzzy logic provides a nat- and implementation of the second generation expert
ural framework for the management of uncertainty in systems, thus limiting the rule explosion on the one
expert systems because its main purpose is to pro- hand and providing an added form of intelligence that
vide a systematic basis for representing and inferring can handle various linguistic variations.
172 Z. Chen/Fuzzy Sets and Systems 79 (1996) 163 173

In contrast to the work summarized above, the ap- I × K or I x K denotes Turksen's second generation
proach to be described in this paper is mainly driven expert systems.
from the considerations of conflict resolution and more
efficient searching in knowledge bases. The theme of Other combinations can be established to denote yet
our study is to show that fuzzy set theory can be ap- different types of expert systems. In addition, although
plied to attack intrinsic (inherent) problems of expert in the present paper we have only discussed fuzzy in-
systems. Instead of dealing with fuzzy controller as ference control on conventional (namely, non-fuzzy)
mentioned before, we have set our task to investigate expert systems, the basic idea employed here may be
the fuzzy inference controller in expert systems. generalized to fuzzy expert systems.
We can now summarize several different ways of
employing fuzzy set theory in expert systems by in-
troducing a taxonomy. We are not trying to provide 7. Concluding remarks
a complete listing of all possible ways of applying
fuzzy set theory into expert systems. Instead, we just Flexible inference control for complex problem
want to illustrate how diverse the application may be, solving for expert systems is an important issue in
because at least the following different perspectives applied artificial intelligence (AI). As the result of
exist. this study is a revised model of expert systems. Our
Following the notations used in an earlier section, study has been motivated by enhancing conventional
we use I for inference control and K for knowledge architecture of expert systems; and our approach has
base (since I and K are now treated as system com- focused on the issue of using fuzzy techniques to deal
ponents, we will not use bold face character to dis- with problems inherent in expert systems, such as
tinguish an operator from an operand, as we did ear- those related to conflict resolution. In particular, we
lier). We also use overline with one horizontal bar to examined the aspect of using meta-rules to provide
indicate a partly fuzzified component, and overlines fuzzy inference operator.
with two horizontal bars to indicate a fully Juzzified This paper has addressed one fundamental issue of
component. A crisp component is thus indicated by a reasoning: the uncertainty of the inference itself and
character without overlines. Here, by partly fuzzified its impact on the architecture of expert systems. In
we mean that the knowledge base consists of rules particular, the contribution of this paper can be jus-
which are fuzzified while by fully fuzzified we mean tified from the following features as involved in our
that the entire unit is by itself fuzzified (as discussed work: an analysis on fuzzy set theory for flexible in-
in Turksen's second generation expert systems). We ference control; a more advanced model of expert
also use the notation × to denote Cartesian product. systems incorporating such a kind of flexible con-
These notations provide a convenient way to describe trol; as well as a comparative study and a taxon-
how fuzzy set theory is used in expert systems. For omy of employing fuzzy set theory in expert systems.
instance, Since our motivation has been to seek help from fuzzy
set theory to cope with inherent problems in expert
I × K denotes a conventional (namely, non-fuzzy) systems rather than finding applications or extending
expert system; fuzzy set theory to expert systems, our research is
complementary to existing approaches in fuzzy ex-
pert systems. Therefore, although this paper does not
I × K or I × K denotes a fuzzy expert system with a make direct contribution to the fuzzy set theory it-
fully or partly fuzzified knowledge-base, and a partly self, it does introduce a novel approach to extend the
fuzzified inference engine; use of fuzzy set theory, which may have profound
x K or ~ x K denotes fuzzy inference on a conven- impact on the architecture of the rule-based expert
tional expert system or a conventional expert system systems.
with a partly fuzzified knowledge base (this approach A new concept or a new approach may need many
is exemplified by the model presented in the current years of research and practice before it reaches its
paper); maturity. The history of fuzzy set theory itself has
Z. ChenlFuzzy Sets and Systems 79 (1996) 163 173 173

proved this. For the approach introduced in this paper, [5] J.J. Buckley, Theory of the fuzzy controller: an introduction,
extensive research and i m p l e m e n t a t i o n are needed. Fuzzy Sets and Systems 51 (1992) 249-258.
In addition, we can offer the following remarks for [6] Z. Chen, User modeling for flexible inference control and its
relevance to economics and management, Comput. Econom.
future directions o f study. These remarks extend the 6 (1993) 163-175.
importance o f flexible inference control in real-world [7] R. Davis, Meta-rules: reasoning about control, Artificial
applications. Intelligence 15 (1980) 179-222.
(1) Instead o f using discrete variables to form fuzzy [8] B. De Baets and E.E. Kerre, The generalized modus ponens
priority factor to achieve fuzzy inference control, con- and the triangular fuzzy data model, Fuzzy Sets and Systems
59 (3) (1993) 305-317.
tinuous variables can be used to distinguish different [9] J. Giarratano and G. Riley, Expert Systems: Principles and
user e n v i r o n m e n t s or stereotypes. W h e n continuous Programming (PWS-KENT, Boston, 1989).
variables are used, the vectors will be replaced by con- [10] T. Gruber, Learning why by being told what, 1EEE Expert
tinuous functions. 6 (4) 65-75 (1991).
(2) Furthermore, flexible inference control m a y be [11] J.W. Grzymala-Busse, Managing Uncertainty in Expert
Systems (Kluwer, Boston, 1991).
needed not due to the different types or e n v i r o n m e n t
[12] F.F. Ingrand, MP. Georgeff and A.S. Rao, An architecture
o f h u m a n users, but due to different types o f other for real-time reasoning and system control, IEEE Expert 7
factors. For example, an expert system developed for (1992) 34-44.
diagnosing automobile problems m a y be adjusted by [13] A. Kandel, Fuzzy Expert Systems (CRC Press, Boca Roton,
the different makes or models o f the cars. The basic FL, 1992).
[14] R. Kass and T. Finin, The role of user models in cooperative
idea introduced in this section can be generalized to
interactive systems, Int..L InteL Sys., 4 (1989) 81-112.
these scenarios. [15] K.S. Leung and W. Lam, Fuzzy concepts in expert systems,
(3) So far we have focused on the use o f fuzzy in- IEEE Computer 21 (1988) 43-56.
ference control in conflict resolution. More generally, [16] M. Maeda and S. Murakami, A self-tuning fuzzy controller,
other forms o f fuzzy inference control m a y be used. Fuzzy Sets and Systems 51 (1992) 29-40.
[17] F. Maria, A. Jimenez, R. Galan and R. Sanz, Fuzzy
controllers: lifting the linear-nonlinear frontier, Fuzzy Sets
and Systems 52 (1992) 113 128.
Acknowledgements [18] K.R. McKeown, User modeling and user interfaces, in: Proc.
AAAI-90 (1990) 1138-1139.
Critical c o m m e n t s from a n o n y m o u s reviewers [19] M. Mizumoto (Guest ed.), Special issue on fuzzy expert
on an earlier version o f this paper have helped the systems, Fuzzy Sets and Systems 44 (1991) 339-468.
[20] T. Murayama, K. Kushima and S. lshigaki, An inference
author to improve the presentation. The author thanks
mechanism suited for real-time control, Proc. IEA/AIE-89,
S. Arroyo for knowledge acquisition and partial im- Vol. I (1989) 245-253.
plementation in the orthopedic example. [21] C. Negoita, Expert Systems and Fuzz), Systems (Benjamin
Cummings Press, Menlo Park, CA, 1985).
[22] M.E. Pollack, J. Hirschberg and B. Webber, User participation
References in the reasoning processes of expert systems, Proc. AAAI-
1982, 358 361.
[1] P. Biswas and A. Kandel, imprecise object models: A key [23] J.C. Sanborn, Coping with control in rule-based programs,
to conceptualizing knowledge based management systems, Heuristics. J. Knowledge Eng. 3 (1990) 50-58.
Fuzzy Sets and Systems 58 ( 1993) 143 - 153. [24] I.B. Turksen, Fuzzy expert systems for IE/OR/MS, Fuzz),
[2] B.G. Buchanan and R.G. Smith, Fundamentals of expert Sets and Systems 51 (1992) 1-27.
systems, in: A. Barr, P.R. Cohen and E.A. Feigenbaum, [25] F. van Harmelen, Meta-level Inference Systems (Pitman,
Eds., The Handbook of Artificial Intelligence, Vol. IV, London, 1991).
149-192. [26] L.A. Zadeh, The role of fuzzy logic in the management of
[3] J.J. Buckley, W. Siler and D. Tucker, A fuzzy expert system, uncertainty in expert systems, Fuzzy Sets and Systems 11
Fuzzy Sets and Systems 20 (1986) 1-16. (1983) 199-227.
[4] J.J. Buckley and H. Ying, Expert fuzzy controller, Fuzzy [27] H.-J. Zimmermann, Fuzzy Sets, Decision Making, and
Sets and Systems 44 (1991) 373-390. Expert Systems (Kluwer, Boston, 1987).

Potrebbero piacerti anche