Escolar Documentos
Profissional Documentos
Cultura Documentos
I. INTRODUCTION
XISTING research on fuzzy reasoning can roughly be
divided into three overlapping categories: fuzzy reasoning
methods and their analysis, logical foundation of fuzzy reasoning, and applications of fuzzy reasoning. Various fuzzy
methods have been proposed and are mainly based on three
different ideas. The rst idea is that of composition. This leads
to the Zadeh compositional rule of inference (CRI) method
[1] and its variants [2][4]. The second idea is that of analogy
and similarity [5][8]. The third idea is that of interpolation
[9], [10]. Analysis of fuzzy reasoning methods is concerned
consequences in fuzzy reasoning. The importance of the robustness issue can further be justied from other perspectives. In
fuzzy modeling of complex systems, ne representations often
lead to a vast number of fuzzy rules. In order to circumvent
the curse of dimensionality, fuzzy rule base reduction is invoked, which incurs undesirable errors in the mapping from
fuzzy premises to fuzzy consequences [27]. In fuzzy control
with fuzzy inputs, noises are often associated with the fuzzy inputs, and it is highly desirable that the noises can be attenuated
[28].
The robustness issue of fuzzy reasoning was studied in our
previous work [26], [29] in the setting of -equalities of fuzzy
sets1 and extended in the recent work reported in [30]. An important observation is that the popular Zadeh CRI method is
not sufciently robust and should be improved. In contrast, the
optimal fuzzy reasoning methods [31], [32], which treat fuzzy
reasoning as a process of optimization rather than logical inference, can be more robust. It is further shown that optimal fuzzy
reasoning methods can improve the robustness of fuzzy control
systems [33].
Inspired by the general idea in control engineering that feedback can help to attenuate the uncertainty of controlled objects
and the observation noises [34], in this paper we treat fuzzy reasoning as a control problem and embed an explicit feedback
mechanism into an optimal fuzzy reasoning method proposed
in our previous work [31], [32]. The resulting new fuzzy reasoning method is expected to be more robust than the optimal
fuzzy reasoning method, and thus to be much more robust than
the CRI methods and the like. Section II reviews the optimal
fuzzy reasoning method that serves as a reference method in this
paper. Section III introduces three distinct measures to capture
the robustness of the optimal fuzzy reasoning. Section IV formulates the new fuzzy reasoning method proposed in this paper.
Section V presents the results of Monte Carlo simulation to evaluate the robustness of the new fuzzy reasoning method in comparison with the optimal fuzzy reasoning. Section VI discusses
fuzzy reasoning in a general context of feedback control and
proposes the triple ingredients perspective for fuzzy reasoning
and logical inference that may lead to new forms of reasoning
such as robust reasoning and adaptive reasoning. Concluding remarks are contained in Section VII.
II. OPTIMAL FUZZY REASONING
Consider a set of fuzzy rules as follows:
If
If
is
then
then
is
is
If
is
then
is
where
is a fuzzy set dened on the universe of discourse
and
is a fuzzy set dened on the universe of discourse
. The problem of fuzzy modus ponens is concerned
with what is if is a fuzzy set dened on .
Let
and
be the membership functions
of
and , respectively. Suppose is with membership function
. The basic idea of the Zadeh CRI method
is a composition of the given fuzzy rules and
is that
1Two
601
the given fuzzy premise [1]. In doing so, all the fuzzy rules
are composed into a new fuzzy rule rst, which is then used
to generate a fuzzy consequence in accordance with the given
fuzzy premise. This is the so-called composition-based inference [20]. Alternatively, each of the fuzzy rules is individually used to generate a fuzzy consequence in accordance with
the given fuzzy premise. The resulting fuzzy consequences
are then composed into a new fuzzy consequence. This is the
so-called individual-rule based inference [20]. In this paper, we
follow composition-based inference.
More specically, consider the case of nite discrete uniand
.
verses with
Further, let
where
adopted in the fuzzy relation
is the Mamdani
to the fuzzy
implication operator from the fuzzy premise
composes the fuzzy rules into a single
consequence
fuzzy rule via the union operator, and
determines
the fuzzy consequence corresponding to the fuzzy premise
according to the Zadeh CRI method for fuzzy modus ponens.
It is implicitly assumed that the fuzzy relation incorporates
all the information underlying the fuzzy rules. With this assumption, the problem of fuzzy modus ponens can be restated
as follows: given fuzzy relation and fuzzy premise , what is
the corresponding fuzzy consequence ? The CRI method denes a single fuzzy consequence for each fuzzy premise and is a
generalized version of modus ponens in classical logical inference. The desired properties or goals of the CRI method are not
explicitly stated or used to derive the method.
In contrast, the optimal fuzzy reasoning methods proposed in
our previous work follow a dramatically different idea for fuzzy
modus ponens that can be stated as follows [31][33].
is gained from the
1) Suppose a fuzzy relation (matrix)
experience of experts (fuzzy rules) or data. Then we should
trust and treat it as a basis for evaluating the quality of a
fuzzy reasoning method.
2) If a fuzzy premise
is given, then a fuzzy reasoning
method will generate the corresponding consequence .
A fuzzy reasoning method is optimal if the matrix
is
in some sense, where
,
the closest to
denoting that implies .
A fundamental feature of optimal fuzzy reasoning is that an optimization goal is introduced, which species what is meant by
being the closest to . The process of optimal fuzzy reasoning is goal driven towards optimizing the objective function
given a priori. The resulting fuzzy reasoning process is no longer
dened a priori. Rather, it depends on the underlying goal for
reasoning and is derived. Different goals lead to different fuzzy
602
that minimizes
2) Obtain
Let
Suppose
reduces to
with
The value of
that minimizes
as follows.
. Then
is
if
if
if
(2.1)
For
This objective function measures the absolute distance between
the given fuzzy relation
and the new fuzzy
relation
. Other objective functions
were also adopted in our previous work [31][33]. The resulting fuzzy reasoning processes are optimal in the sense that
the given objective functions are optimized or minimized. The
corresponding to the fuzzy
fuzzy consequence
premise
is determined by minimizing the given
. This is dramatically different from the
objective function
is
CRI method. Note that in the CRI method,
determined by composing the fuzzy relation and the fuzzy
without considering any optimization
premise
or extra function. It is basically a kind of generalizations of
the classical logical modus ponens. No objective function or
reasoning goal is explicitly involved.
, the fuzzy
With the fuzzy relation
, and the objective function
premise
to be mini2
can be
mized, the resulting fuzzy consequence
determined as follows [31], [32]. Note that
where
We can easily see that if the extra objective function achieves
its minimum at
achieves its minimum
. Denote the corresponding values as
at
and
, respectively. In this way,
can be obtained in the
following procedure.
into
such that
1) Rearrange
. Rearrange the row ordering of
into
according to the
change of the row ordering of .
2 [b
convenience.
Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:438 UTC from IE Xplore. Restricon aply.
or
reduces to
In this way
and
3)
.
Symbolically, we can denote
, where denotes
a fuzzy reasoning method such as the one presented in this secthe corresponding fuzzy relation,
the given fuzzy
tion,
premise, and the resulting fuzzy consequence.
A few remarks should be made for the above algorithm. For
any pair of and , the above algorithm guarantees that there
is a consequence . In the case of
in Step 2) of the
above algorithm, we specically let
to avoid multiple choice of consequence. The choice of
asserts
that the maximal grade of membership of the resulting fuzzy
consequence should not be greater than that of the given fuzzy
premise. Intuitively, the maximal condence in the fuzzy consequence should not exceed that in the given fuzzy premise. How, then there must be
and
.
ever, if
The choice of
in Step 2) becomes superuous accordingly.
Now let us consider a simple example to illustrate the difference between the CRI method and the optimal fuzzy reasoning (OFR) method presented in this section. Suppose we obTHEN
; IF
THEN
tain two (crisp) rules: IF
TABLE I
1 0
THE REASONING RESULTS FOR R =
0 1
AND
A = (x 1) WITH TWO
DIFFERENT METHODS
. Then we obtain
603
only
604
where
denotes the optimal fuzzy reasoning
method presented in Section II and
Step 7) Stop.
. However,
Here we note that it is assumed that
this is not essential. Other values can be assigned to and .
The simulation is actually conducted for 50 trials, each of which
generates a value for measuring robustness of fuzzy reasoning.
Step 6) calculates the average of these values and treats it as the
index of robustness of fuzzy reasoning. Of course, more trials
of simulation can be conducted for calculating the robustness
measure.
The robustness measure calculated in Step 6) takes account
of errors in the resulting fuzzy consequences as well as their
causing errors in fuzzy premises and looks like a measure of
partial difference. This is different from the robustness measures
Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:438 UTC from IE Xplore. Restricon aply.
Step 4) Stop.
B. Robustness Measure II
The second category of robustness measures considers
only and ignores
. Symbolically, we denote
where
denotes certain norm operator.
As for
, we can have the following algorithm to
.
evaluate
Simulation Algorithm 3:
Step 1) Obtain the fuzzy consequence
corresponding to the fuzzy premise with fuzzy relation
in accordance with the optimal
fuzzy reasoning method presented in Section II.
Step 2) Generate 50 sets of random numbers
,
taking values in the interval
with
in accordance with a uniform
probability distribution.
Step 3) Let
if
if
if
Step 4) Let
.
Step 5) Obtain the fuzzy consequence
corresponding to the fuzzy premise
with fuzzy
relation
in accordance with the optimal
605
Step 3) Let
if
if
if
if
if
if
where
denotes the optimal fuzzy reasoning
method presented in Section II and
Step 7) Stop.
Similar to Simulation Algorithm 2, we can have the following
algorithm.
Simulation Algorithm 4:
Step 1) Generate 500 pairs of fuzzy relations and fuzzy
premises
, at random
as follows, where
. For each
and
are
each generated in accordance with the uniform
probability distribution dened over the interval
.
, perform Simulation
Step 2) For each pair
Algorithm 3 and obtain
.
Step 3) Let
Step 4) Stop.
C. Robustness Measure III
The third category of robustness measures considers
well as
as
Step 4) Let
.
Step 5) Obtain the fuzzy consequence
corresponding to the fuzzy premise
with
in accordance with the optimal
fuzzy relation
fuzzy reasoning method presented in Section II,
.
Step 6) Calculate
where
denotes the optimal fuzzy reasoning
method presented in Section II and
Step 7) Stop.
Similar to Simulation Algorithm 2, we can have the following
algorithm.
Simulation Algorithm 6:
Step 1) Generate 500 pairs of fuzzy relations and fuzzy
at random
premises
as follows, where
. For each
and
are
each generated in accordance with the uniform
probability distribution dened over the interval
.
Step 2) For each pair
, perform Simulation Algorithm 5 and obtain
.
Step 3) Let
Step 4) Stop.
606
D. Comparative Results
In order to judge which robustness measures should be
adopted for evaluating robustness of a given fuzzy reasoning
method, we perform Monte Carlo simulation according to the
following algorithm.
Simulation Algorithm 7:
Step 1) Generate 500 pairs of fuzzy relations and fuzzy
at random
premises
as follows, where
. For each
and
are
each generated in accordance with the uniform
probability distribution dened over the interval
.
Step 2) For each pair
, perform Simulation Algorithm 1 and obtain
. Let
.
, perform Simulation AlStep 3) For each pair
gorithm 3 and obtain
. Let
.
, perform Simulation AlStep 4) For each pair
gorithm 5 and obtain
. Let
.
Step 5) Stop.
Example 3.1: Consider the following fuzzy relation:
Further, let
.
Performing Simulation Algorithm 5, we obtain
. Suppose we ignore the contribution of the perturbations of fuzzy relations; then we obtain [refer to Step 6) of Simulation Algorithm 5]
TABLE II
ROBUSTNESS MEASURES OF OPTIMAL FUZZY REASONING METHOD =
607
(4.2)
replaces the initial
for optimal fuzzy reasoning. Given
fuzzy premise
at time 1, the corresponding fuzzy concan be determined by using according to the
sequence
optimal fuzzy reasoning method presented in Section II.
and
are then used to make up
that updates
and
accordingly. An explicit feedback mechanism is embedded
into the process of fuzzy reasoning.
and a single fuzzy
Note that given the fuzzy relation
premise , the fuzzy reasoning method presented in Section II
generates a single fuzzy consequence . The method is a mapping from the universe of single fuzzy premises to the universe
of single fuzzy consequences. However, for the fuzzy reasoning
method with the feedback mechanism presented in this section,
it does not make sense to state that a single fuzzy premise
generates a single fuzzy consequence individually. Rather,
it should be stated that a single sequence of fuzzy premises
if
if
Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:438 UTC from IE Xplore. Restricon aply.
(4.1)
608
be treated as an inner loop. These two loops should be complementary rather than conicting.
Similar differences can be observed between the feedback
mechanism presented in this section and those incorporated in
fuzzy neural networks that implement fuzzy reasoning. In the
adaptive-network-based fuzzy inference system (ANFIS) [42],
the network is updated by using a hybrid learning algorithm
to t crisp-in crisp-out data pairs. The feedback to the network occurs as new data pairs are given. In the compensatory
neurofuzzy systems [43], crisp-in crisp-out data pairs are
employed to dynamically rene the fuzzy operators adopted in
the networks. In [44], a new adaptive fuzzy inference neural
network (AFINN) is proposed to model crisp-in crisp-out
data pairs. The underlying feedback mechanism is similar to
that of ANFIS. There are two major differences in feedback
mechanism between these works and the one presented in
this section. First, these fuzzy neural networks serve as a
modeling formalism for crisp-in crisp-out data pairs, whereas
the fuzzy reasoning formalism presented in this section can
handle fuzzy-in fuzzy-out data pairs. Secondly, the feedback
loop occurs in the neural networks as learning algorithms are
employed to train the networks. However, the feedback mechanism in the fuzzy reasoning formalism presented in this section
is intrinsic in the sense that no learning algorithms are invoked.
On the other hand, the feedback mechanisms incorporated
in recurrent neural networks can be treated as intrinsic since
they link outputs of a recurrent neural network or the neurons
thereof backward to the inputs of the neural network or other
neurons thereof [45]. The recurrent neural networks are mainly
devoted to crisp-in crisp-out data pairs. The fuzzied radial
basis function networks proposed in [46] can handle fuzzy-in
fuzzy-out data pairs and do nonlinear regression analysis.
No intrinsic feedback mechanisms are involved in them. The
most related work should be the fuzzied recurrent neural
fuzzy network that can handle fuzzy-in fuzzy-out data pairs
[47]. However, the fuzzy reasoning formalism presented in
this section distinguishes itself from the fuzzied recurrent
neural fuzzy network in several aspects. First, the feedback
mechanism presented in this section is explicitly expressed
in terms of fuzzy rules. This makes it intuitively interpreted.
Secondly, no learning algorithms are involved in the feedback
mechanism presented in this section, and thus the potential
weaknesses of learning algorithms are avoided. Lastly, fuzzy
reasoning is treated as a control problem, which leads to the
triple ingredients perspective of fuzzy reasoning as will be
discussed in Section VI.
V. SIMULATION RESULTS
Denote the fuzzy reasoning method presented in Section IV
. Recall that can denote the fuzzy reasoning method preas
is more robust
sented in Section II. Now we need to judge if
than . If the answer is positive, then we can say that feedback
mechanisms do improve the robustness of fuzzy reasoning. Following the observation of Section III that the contribution of the
perturbations of fuzzy premises to the robustness measure becomes marginal if the contribution of the perturbations of fuzzy
relations is considered, here we consider perturbations of fuzzy
premises and those of fuzzy relations individually.
Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:438 UTC from IE Xplore. Restricon aply.
609
where
Step 7) Stop.
In the above, 20 trials of simulation are actually conducted. In each trial, a robustness measure
is calculated.
The average of these values is then treated as the robustness
with respect to
and
. Of
index of
course, more trials can be conducted. Note that
reduces to
610
TABLE III
ROBUSTNESS MEASURES OF DIFFERENT FUZZY REASONING METHODS
AGAINST PERTURBATIONS OF FUZZY PREMISES
if
. Therefore, the above algorithm is also applicable
for evaluating robustness of for a given pair of fuzzy relation
and the sequence of fuzzy premises
. The
corresponding output is
. In this
way, we can employ the following algorithm to evaluate and
and
against perturbations of
compare robustness of
fuzzy premises.
Simulation Algorithm 9:
Step 1) Generate 20 pairs of fuzzy relations and sequences
of fuzzy premises
at random as follows, where
.
For each
and
are each generated in accordance with the uniform probability distribution dened over the interval
.
, perStep 2) For each pair
form Simulation Algorithm 8 for
and obtain
Step 2) Generate
20
sets
of
random
numbers,
,
taking values in the interval
with
in accordance with a uniform
probability distribution.
Step 3) Let
if
if
if
Step 4) Let
.
Step 5) Obtain the fuzzy consequence
corresponding to the fuzzy premise
in accordance with
fuzzy relation
rameter
with
with pa-
.
Step 6) Calculate
. Let
.
Step 3) For each pair
Algorithm 8 for
obtain
. Let
, perform Simulation
(with parameter
) and
.
Step 4) Stop.
Example 5.1: Simulation Algorithm 9 was performed for
three times (trials 1, 2, and 3). Table III tabulates the resulting
is about half of
.
robustness measures. We can see that
This implies that the optimal fuzzy reasoning method
with
the feedback mechanism is much more robust against perturbations of fuzzy premises than the optimal fuzzy reasoning method
without a feedback mechanism. Feedback mechanisms do improve the robustness of fuzzy reasoning.
B. Robustness Against Perturbations of Fuzzy Relations
Similar to Simulation Algorithm 8, we can have the following
algorithm to evaluate the robustness of the fuzzy reasoning
method
against the perturbations of fuzzy relations with
respect to a given pair of fuzzy relation and the sequence of
.
fuzzy premises
Simulation Algorithm 10:
Step 1) Obtain the fuzzy consequence
corresponding to the fuzzy premise
with
with paramfuzzy relation in accordance with
eter
.
Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:438 UTC from IE Xplore. Restricon aply.
where
Step 7) Stop.
reduces to if
, the above algorithm is
Since
for a given
also applicable for evaluating robustness of
pair of fuzzy relation
and the sequence of fuzzy
premises
. The corresponding output is
. In this way, we can employ the
following algorithm to evaluate and compare robustness of
and
against perturbations of fuzzy relations. Here we
note that
is evaluated in comparison with . No other
existing fuzzy reasoning methods including the CRI method
are considered. This is because our previous work has shown
that is more robust than the CRI method [32].
Simulation Algorithm 11:
Step 1) Generate 20 pairs of fuzzy relations and
sequences of fuzzy premises
at random as
follows, where
. For each
and
are
each generated in accordance with the uniform
611
TABLE IV
ROBUSTNESS MEASURES OF DIFFERENT FUZZY REASONING METHODS
AGAINST PERTURBATIONS OF FUZZY RELATIONS
.
, perform Simulation
and obtain
. Let
(6.1)
denotes the state vector of the controlled object at
where
its transpose,
the control signal vector
. time and
matrices of appropriate dimensions, and
at time
Step 4) Stop.
Example 5.2: Simulation Algorithm 11 was performed for the terminal or nal time of concern. The corresponding control
three times (trials 1, 2, and 3). Table IV tabulates the resulting goal is to minimize the objective functional .
Control constraints may also be imposed. For example, there
is about a third of
. The feedrobustness measures.
should be an upper bound for the switching frequency of actidramatically improves the robustness
back mechanism in
of fuzzy reasoning. Recalling the results presented in Example vation/stopping of the outdoor condensing unit of an air condi5.1, we can see that the feedback mechanism is more effective tioner. The corresponding control signals cannot follow an arbito attenuate the perturbations of fuzzy relations than to attenuate trary law. In summary, a control system contains three distinct
classes of ingredients: controlled objects, controllers, and conthose of fuzzy premises.
trol goals/constraints.
With the triple ingredients perspective of control, we can reexamine various fuzzy reasoning methods, including those preVI. A CONTROL-THEORETIC PERSPECTIVE
sented in Sections II and IV. For the CRI method, each single
OF FUZZY REASONING
fuzzy premise generates a single fuzzy consequence, and no
Recall that in control engineering, a control system consists other fuzzy premises or fuzzy consequences are involved. The
of at least a controlled object and a controller [34], as depicted in fuzzy premise is applied to the underlying fuzzy rule base. Then
Fig. 1. The controller accepts input signals and delivers control the fuzzy rule base can be viewed as the controlled object and
signals to the controlled object. The output signals are often used the CRI method as the corresponding controller. The fuzzy rule
by the controller to deliver new control signals and thus make up base and the CRI method constitutes an open-loop reasoning
a feedback loop from the controlled object to the controller. For system, as depicted in Fig. 2. However, the underlying control
example, consider an air conditioner installed in a room.3 The or reasoning goals/constraints are not explicitly stated.
air in the room and compressor in the air conditioner serve as
presented in
The optimal fuzzy reasoning method
the controlled object, and the control clip in the air conditioner Section II also corresponds to an open-loop control system.
serves as the controller. The input signal is an expected voltage However, the reasoning goal is given in terms of (2.1), which
according to desired air temperature (e.g., 23 C) designed by species an objective function for each single reasoning step.
users. The output signal is the actual air temperature in the room. This contrasts with the global objective function given in
The feedback information is gained by the temperature sensor
(6.1), which takes account of the whole process of control
(reasoning). The fuzzy rule base serves as the rst ingredient of
3The description presented here is oversimplifying and inaccurate. It is only
fuzzy reasoning, the reasoning method serves as the second
for an explanation of the concept of feedback control.
Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:438 UTC from IE Xplore. Restricon aply.
612
ingredient of fuzzy reasoning, and the reasoning goals/constraints serve as the third ingredient of fuzzy reasoning. Instead
of dening a reasoning method a priori, a theory of fuzzy
reasoning should guide how to synthesize the desired reasoning
method to achieve or satisfy given reasoning goals/constraints.
This is partially exemplied in the optimal fuzzy reasoning
method . However, no feedback mechanisms are explicitly
involved in since the objective function given in (2.1) is local
rather than global.
presented in Section V inThe optimal fuzzy reasoning
volves an explicit feedback mechanism and corresponds to a
closed-loop reasoning system as depicted in Fig. 3. The rst
two ingredients of fuzzy reasoning are the same as those for
the reasoning method . Equation (2.1) still species a reasoning goal for the reasoning process. However, an additional
reasoning goal is to improve the robustness of the fuzzy reasoning process. The underlying robustness measures given in
Section III are global in nature in the sense that they take account
of the whole reasoning process. In this way, it is a reasonable
result to embed an explicit feedback mechanism into the reasoning process. On the other hand, it should be noted that how
to synthesize the required feedback mechanisms that guarantee
the desired goals of reasoning robustness is an open problem.
presented in Section IV, the
In the fuzzy reasoning method
reasoning results of the latest reasoning steps are employed
as feedback information. This only species one possible form
of feedback mechanisms. An alternative form is to employ the
reasoning result of a single reasoning step in the history of reasoning that is closest to the fuzzy relation . Multiple criteria
can be adopted to identify the required feedback information in
the reasoning process, depending on the underlying reasoning
goals/constraints.
No matter whether an explicit feedback mechanism is employed, the preceding discussions suggest that fuzzy reasoning
can be treated as a control problem. The underlying three ingredients should be explicitly identied. A sound theory of fuzzy
reasoning should clearly formulate the possible relationships
among the three distinct ingredients of fuzzy reasoning.
Treating fuzzy reasoning as a control problem opens a new
arena for applying various control principles of feedback control and optimization to fuzzy reasoning in particular and reasoning in general. For example, robust control is concerned with
how to analyze and synthesize a single controller that guarantees the required stability and performance for an uncertain
class of controlled objects in the presence of input/output noises
[48]. As a counterpart to robust control, we can talk about robust reasoning. A robust reasoning method should guarantee the
Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:438 UTC from IE Xplore. Restricon aply.
required reasoning robustness even if perturbations are associated with the underlying rule bases and premises. On the other
hand, adaptive control is concerned with how to dynamically
adapt the controllers to accommodate the underlying changes
occurring to the controlled objects online [49]. The principles of
adaptive control should lead to adaptive reasoning that adjusts
the required reasoning methods to accommodate the underlying
changes occurring to the knowledge bases during reasoning. Of
course, optimal fuzzy reasoning can serve as the optimal control counterpart in fuzzy reasoning and should be further investigated.
Another direction to extend the underlying principle (treating
fuzzy reasoning as a control problem) of the reasoning methods
and
is to apply various control principles of feedback control and optimization to classical logical inferences. Here we
, no logical lannote that in the reasoning methods and
guages are involved. They are not based on logic and cannot be
treated as an extended form of logical inference. However, this
does not mean that the triple ingredients perspective cannot be
applied to logical inference. For a logical system, the underlying
axioms and the given knowledge base can serve as a controlled
object, and the reasoning method can serve as the corresponding
controller. Suppose a reasoning goal can be identied a priori to
serve as the third ingredient of reasoning; then the resulting reasoning process generates a consequence for each given premise.
The consequence is not only a function of the axioms, the given
knowledge, and the premise but also a function of underlying
reasoning goals/constraints. This should lead to new logics that
are dramatically different from various logics that ignore the underlying reasoning goals/constraints.
VII. CONCLUDING REMARKS
Robustness of fuzzy reasoning is concerned with the effects
of perturbations associated with given fuzzy rule bases and/or
fuzzy premises on fuzzy consequences. This problem is important since fuzzy rule bases can hardly be extracted from human
expertise or observation data accurately, and quantifying fuzzy
premises inevitably incurs errors. Our previous work shows that
the popular Zadeh CRI method of fuzzy reasoning is not as
robust as desirable, and the optimal fuzzy reasoning methods,
which generates fuzzy consequences for given fuzzy premises
such that given objective functions are minimized or optimized,
may be more robust. Inspired by the idea in control engineering
that feedback may help to attenuate the uncertainty associated
with the controlled object and the noises associated with the inputs/outputs, in the preceding sections we treat fuzzy reasoning
as a control problem and embed feedback mechanisms into optimal fuzzy reasoning methods to further improve the robustness
of fuzzy reasoning. More specically, the given fuzzy rule base
serves as the controlled object, and the fuzzy reasoning method
serves as the corresponding controller. The fuzzy rule base and
the reasoning method constitute a control system, which may be
open-loop or closed-loop, to achieve or satisfy given reasoning
goals/constraints. The fuzzy rule base, the reasoning methods,
and the reasoning goals/constraints specify the three distinct ingredients of fuzzy reasoning. In the preceding sections, a new
fuzzy reasoning method with an explicit feedback mechanism
is proposed and evaluated. The fuzzy consequence obtained in
a specic reasoning step is not only a function of the underlying
fuzzy rule base and the fuzzy premise applied in the reasoning
step but also a function of the history of fuzzy reasoning. It is
also a function of reasoning goal of concern. The new fuzzy
reasoning method is a mapping from the universe of sequences
of fuzzy premises to the universe of sequences of fuzzy consequences rather than from the universe of single fuzzy premises
to the universe of single fuzzy consequences. The Monte Carlo
simulation shows that the new fuzzy reasoning method is much
more robust than the optimal fuzzy reasoning methods, and the
feedback mechanism does signicantly improve the robustness
of fuzzy reasoning. The work presented in this paper sets a
new starting point to develop a sound theory of fuzzy reasoning
that formulates the relationships among the three ingredients
of fuzzy reasoning. It also offers new opportunity for various
principles of feedback control and optimization to be applied
in fuzzy reasoning in particular, and reasoning in general. New
forms of reasoning such as robust reasoning and adaptive reasoning can be explored in the future, in which logical or nonlogical languages can be adopted. Since a fuzzy reasoning method
actually denes a class of pure fuzzy systems that are different
from Mamdami fuzzy systems and T-S fuzzy systems, it can
be expected that the new fuzzy reasoning method presented in
this paper can be used for modeling and control of complex
systems and for decision-making under complex environments.
Applications of the ideas and methods presented in this paper to
real-world problems should be also explored in the future.
ACKNOWLEDGMENT
The valuable comments of the anonymous reviewers are
gratefully appreciated.
REFERENCES
[1] L. A. Zadeh, The concept of a linguistic variable and its applications
to approximate reasoning: Parts IIII, Inf. Sci., vol. 8, 9, pp. 199249,
301357, 4393, 1974, 1975.
[2] D. Dubois and H. Prade, Fuzzy sets in approximate reasoning, Part 1:
Inference with possibility distributions, Fuzzy Sets Syst., vol. 40, pp.
143202, 1991.
[3] H. Maeda, S. Asaoka, and S. Murakami, Dynamical fuzzy reasoning
and its application to system modeling, Fuzzy Sets Syst., vol. 80, pp.
101109, 1996.
[4] D. S. Yeung and E. C. Tsang, Weighted fuzzy production rules, Fuzzy
Sets Syst., vol. 88, pp. 299313, 1997.
[5] B. Bouchon-Meunier, R. Mesiar, C. Marsala, and M. Rifqi, Compositional rule of inference as an analogical scheme, Fuzzy Sets Syst., vol.
138, pp. 5365, 2003.
[6] E. S. Lee and Q. Zhu, Fuzzy and Evidence Reasoning. Berlin, Germany: Physica-Verlag, 1995.
Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:438 UTC from IE Xplore. Restricon aply.
613
614
Authorized licensd use limted to: IE Xplore. Downlade on May 10,2 at 19:438 UTC from IE Xplore. Restricon aply.
University, London, U.K., and a Visiting Scholar with the City University of
Hong Kong; Swinburge University of Technology, Australia; University of
Technology, Sydney, Australia; and Purdue University, West Lafayette, IN. He
has published more than 35 research papers in international journals and is the
author of Software Defect and Operational Prole Modeling (Boston, MA:
Kluwer, 1998), Introduction to Fuzzy Reliability (Boston, MA: Kluwer, 1996),
and Elements of Software Reliability Engineering [Beijing, China: Tshinghua
Univ. Press, 1995 (in Chinese)]. His main research interests include software
reliability and testing, intelligent systems and control, and software cybernetics.
He is a member of the Editorial Board of Fuzzy Sets and Systems and Editor of
the Kluwer International Series on Asian Studies in Computer and Information
Science. He was Program Committee Cochair for the Fifth International
Conference on Quality Software (Melbourne, Australia, September 2005), the
First International Workshop on Software Cybernetics (Hong Kong, September
2004), and the Second International Workshop on Software Cybernetics
(Edinburgh, UK, July 2005); and General Cochair for the Third International
Workshop on Software Cybernetics (Chicago, September 2006) and the Second
International Symposium on Service-Oriented System Engineering (Shanghai,
China, October 2006). He also was Guest Editor for Fuzzy Sets and Systems
(1996), the International Journal of Software Engineering and Knowledge
Engineering (2006), and the Journal of Systems and Software (2006).