The 7th International Conference on INFOrmatics and Systems (INFOS 2010) – 28-30 March
Applied Optimization and Metaheuristics Track
Fuzzy Cognitive Map with Dynamic Fuzzification
and Causality Behaviors
Ahmed M. Gadallah #1, Hesham A. Hefny#2
# Department of Computer and Information Science, Institute of Statistical Studies and Research (ISSR), Cairo University.
Institute of Statistical Studies and Research (ISSR), 5 Ahamed Zewal Street, Orman, Giza, Egypt
1
2
ahmgad10@yahoo.com
hehefny@hotmail.com
Abstract— This paper introduces a proposed enhanced model of
fuzzy cognitive map (FCM) called dynamic fuzzy cognitive map
(DFCM). The aim is to make both of fuzzification mapping and
causal relationships between concepts are environment sensitive.
Accordingly, knowledge representation and inference process
will be more flexible, reliable, and dynamic especially when
dealing with complex and dynamic environments. A special set of
environment dominant variables (V) is proposed that represents
the environment prevailing conditions. The behavior of DFCM
depends on V. Hence, the same fuzzy cognitive map knowledge
and concept values will infer deferent outcomes in response to
changes in V. Accordingly, it will be very appropriate to apply
DFCM when modeling a wider spectrum of real world problems
in various fields like decision support systems, control systems,
medical diagnoses systems, effect-based planning, and prediction
systems.
Keywords— Fuzzy cognitive map, Dynamic causal networks,
Knowledge representation, Dynamic inference, Fuzzy
systems, Complex systems modeling.
I. INTRODUCTION
Fuzzy cognitive map is a fuzzy signed directed graph that
incorporates a set of concepts and a set of causal relationships
between such concepts [1, 2 and 6]. FCM represents a
modeling technique for complex systems originating from a
combination of fuzzy logic and neural networks. Modeling
dynamic systems using fuzzy cognitive maps (FCM) has
several advantages [2, 5, 7 and 10]: Firstly, it is very simple
and intuitive to understand in terms of the underlying formal
model and its execution. Secondly, FCM is characterized by
flexibility in system analysis and design, ability to deal with
fuzzy information, and adaptability to any given domain.
Finally, inferences are carried out by numerical calculations
instead of symbolic deductions. The potential of this work
relies on enhancing the classical fuzzy cognitive map
approach to make it able to model and reason under dynamism
that represents the natural metaphor of many real world
problems. The main contribution of this work is to make FCM
approach more flexible, reliable and dynamic. This paper aims
to MAKE FCM approach able to take into considerations the
prevailing conditions in the environment under which the
FCM will: firstly, represent the knowledge that may be
sensitive to changes in its surrounding environment and
secondly, make the inferences required to take decisions under
recent prevailing conditions. Also, we note that in a wide
spectrum of real world problems, there may be a set of special
type of concepts which: firstly, does not affect directly any
FCM base concepts but affects one or more of the defined
FCM causal relationships; secondly, may affect some
concepts and its effect should be passed in just some
subsequent paths of the effected concepts. Such prevailing
environment conditions and such special type of concepts will
be called the environment dominant variables set (V).
Accordingly, the aim is to make DFCM approach more
generic by taking into consideration the effect of V on both of
knowledge representation and inference process. This paper is
organized as follows: section 2 gives an overview of FCM.
Previous and related works were introduced in section 3.
Section 4 presents the extended DFCM. An illustrative
example was introduced in section 5. Section 6 addresses the
conclusion.
II. FUZZY COGNITIVE MAPS
Fuzzy cognitive maps are fuzzy signed directed graphs with
feed back [9]. FCM represents a modeling technique for
complex systems originating from a combination of fuzzy
logic and neural networks [6]. FCM describes the structure of
a system in a set of concepts and a set of causal relationship
between these concepts. Inference or reason in FCM is
achieved via numerical calculations through a product
operation between the concept vector and the causalrelationship matrix. Fig. 1 shows a simple FCM which
consists of a set of concepts represented by nodes and a set of
weights represented by directed arrows. Each concept Ci has a
state value that measures its strength. The state value can be a
crisp value in {0, 1} or a fuzzy value within [0,1]. The weight
Wij represents the causal relationship or the influence degree
between two concepts: a causal concept Cj, and an effect
concept Ci. The direction of a causal relationship directed
arrow specifies which concept affects the other one. While,
the value of the weight Wij specifies how strongly the causal
Faculty of Computers and Information - Cairo University
AOM -99
The 7th International Conference on INFOrmatics and Systems (INFOS 2010) – 28-30 March
Applied Optimization and Metaheuristics Track
concept Cj affects the effect concept Ci. While a positive value
of Wij represents a proportional effect, a negative value
represents an inversely proportional effect and zero Wij
represents the absence of an effect. [1-5].
Fig. 1 A simple fuzzy cognitive map
In a simple FCM that has n concepts, the knowledge
representation model has the following n_dimension concepts
vector (C) and n*n weight matrix (W).
C = [c1, c2, c3,.…….cn]
W =
(1)
w 11
w 21
w 12
w 22
w 1n
w 2n
w n1
wn2
w nn
(2)
The inference process in a FCM can be described as follows:
Multiply the weights matrix (W) by the concept vector
(C). The result of the multiplication operation is a ndimensional vector Y as in equation (3) and (4).
Apply the augmented effect yi as an input of concept ci
decision making function to compute its new value as
depicted in equation (5).
Accordingly, each step of the inference becomes a matrix
multiplication followed by applying the accumulated results to
the concept decision making function to deduce the new value
of the concept as a result of changes in its causal concepts.
Y = W×C =
w11
w21
w12
w22
wn1 wn 2
= [y1,y2,y3………….yn]
w1n
w2 n
c2
wnn
cn
c1
(3)
n
yi =
wij c j ,
i = 1,2,
n
(4)
j =1, j ≠ i
where ci represents the value of concept ci, yi represents
the sum of the products of each causal concept value cj
and the weight value between the two concept nodes wij.
cinew = f i ( y i ) + ciold
(5)
III. PREVIOUS AND RELATED WORK
Many studies of FCM have proved that it has many
advantages when modeling complex systems. Applications of
FCM approach have been growing rapidly in many real world
domains which almost are complex. The authors in [5] used
FCM approach to design intelligent software agents. The
developed intelligent agent became able to represent the
domain knowledge and carry out the inferences depending on
numerical representation and calculations instead of symbolic
representation and inference. The used FCM model is a 2element tuple: {C, W} as depicted in equation (1) and (2). A
weight may be dynamic when it depends on its causal
concept. Reasoning about a new state of concept ci is
computed as in equation (5). According to this model, there
are a set of fuzzification mapping functions that map the
interface concepts real values into fuzzy values within the
range [-1, 1]. Also, the model used the dynamic weight idea in
which a weight may be a function in its causal concept value.
Yet, the used FCM model ignores the effects of the
surrounding atmosphere on the causal relationships
representation between concepts.
In process control problems [9], FCM approach was applied
to represent the accumulated knowledge from domain experts
in order to simulate and control the system. In each step of
cycling FCM, the concepts will freely interacted with each
other in respect to the predefined causal relationship between
them. The new value for each concept is calculated by
applying the augmented effect from the causal concepts to that
concept decision making function as depicted in equation (5).
Yet in real world, the surrounding atmosphere may affect the
causal relationship between the interacted concepts in the
designed FCM for process control problems. The used FCM
model ignores a natural fact that, the causal relationship may
be a function of the environment prevailing conditions. For
example, naturally, the higher the surrounding atmosphere
temperature the higher the steaming rate and the lower the
temperature lose rate and vice versa. Also, the wetter the
atmosphere the lower the steaming rate and vice versa. The
used model assumes that the environment has no effect on the
designed FCM causal relationships.
In the work of [6], a fuzzy cognitive agent has been
developed.
This
agent
provides
personalized
recommendations to on-line customers in e-commerce sites.
Fuzzy cognitive agents are designed to give personalized
suggestions based on the on-line customer preferences and
experts' domain knowledge. The agent knowledge model of
FCM is defined as a 2-element tuple: FCM = {C, W} as
depicted in equations (1) and (2). The inference will take place
as depicted in equations (3) and (5). An algorithm based on
the self organized map learning technique is used to learn the
membership function for transferring the base case into fuzzy
case base. The learned membership function for each concept
represents its fuzzification mapping function. In this model of
FCM, the fuzzification mapping function were learned to map
where fi is the decision making function of the concept ci.
Faculty of Computers and Information - Cairo University
AOM -100
The 7th International Conference on INFOrmatics and Systems (INFOS 2010) – 28-30 March
Applied Optimization and Metaheuristics Track
the real concept value into a fuzzy value within [-1, 1]. Also,
the weights of a FCM can be learned from a fuzzy training set
for each concept. The main problem here concerns with the
natural dependency that may exist of both of the fuzzification
mapping functions and the causal relationships between
concepts on the environment prevailing conditions. Also, the
learned weights became fixed values and don't have the
flexibility to respond to changes in the environment prevailing
conditions.
Dynamic cognitive networks (DCN) proposed in [3] have
their roots in the domain of cognitive mapping. DCN aimed to
overcome some shortages of FCM when dealing with many
real-world problems: firstly, FCM is based on static causeeffect schemes and their inference mechanism is able to
provide just the potential of causal relationship between their
concepts, secondly, FCM does not provide a robust and
dynamic mechanism to represent the strength of the cause and
the degree of the effect and finally, FCM lacks the temporal
concept that is crucial in many real world applications. The
authors proposed equation (6) for compute concepts new
values trying to overcome the above shortages. Such equation
provides non-linear causal functions, decaying mechanism,
and temporal effect mechanism. Yet, such causal functions are
mainly function of only its causal concept value.
c
t +1
i
t
c + s (1 − c )
c + s (1 + c )
c +s
t
= F(ci , si ) =
t
t
t
i
i
i
t
t
t
i
i
i
t
t
i
i
t
t
i
i
t
t
i
i
(7)
c ≥ 0, s ≥ 0
c < 0, s < 0
otherwise
where
n
t
wij c j , j ≠ i and c i is the value of concept Ci at time t.
Sit =
j =1
This mechanism has the advantage of keeping the same
relative strength among weights because of dividing each of
them by same saturating value
s
t
max
using equations (8) and
(9) [Yaman et al., 2009], [Tsadiras et al., 1999]. Yet, as the
number of concepts increases, the absolute values of the
weights decrease as a result of dividing it by a bigger value.
Accordingly, concepts can't effectively affect each other
[Tsadiras et al., 1999].
n
n
Sit,max =
| wij | | c j | =
| wij | i = 1,2,....n
(8)
(s1t,max, s2t ,max,.....sit,max)
(9)
j =1
j =1
t
S max
= max
S
t
i
On the other hand, it lacks keeping the relative effect of
over concept ci. That is: the first two cases in equation
(7); that apply the whole value of the external influence
c
t +1
j
=
(
(
t
min 1, max − 1, (1 − d j ). c j + R
))
(6)
where
R=
c
t +1
j
+
w.
ij
i =1,
t
i ≠ j , i >0
c
t
i
to the concept not-activated degree; have different relative
effect of third case. The average effect of S t over ci in first
i
n
and
S
t
max
−t t
. +
t max ci i =1,
n
−
w.
ij
t
max
−t t
.
t max ci
t
c
i ≠ j , i <0
two cases in equation (7) can be computed respectively as
follows:
AVG
case 1
c
represents the new value of concept cj at time t+1, dj
is a decaying factor, wij+:[0, 1] to [-1, 1] represents the
dynamic weights of the positive domain of the causal concept,
wij-:[0, -1] to [-1, 1] represents the dynamic weights of the
negative domain of the causal concept, tmax is the time
horizon until the deactivation of the existing causal
relationship.
FCM model used in [10] was developed to model and
handle effect-based operations aiming to determine or predict
alternative courses of actions to realize the aims of a military
effect-based operation. In this model of FCM a new set of
attributes were added to the concept value calculation
algorithm. These attributes include influence possibility,
influence duration, influence permanence and dynamic
influence. At each step of cycling the FCM, the value of
concept ci will be obtained using equation (7). The dynamic
unit influence degree through the influence duration is
modeled by reducing the links influence duration by one unit
and recalculating the unit influence degree for every step.
( s it ) =
AVG
t
≥ 0 , s it ≥ 0
i
t
case 2
s it (1 − 0 ) + s it ( 1 − 1 )
1 t
=
si
2
2
( S i) =
S
t
i
and,
(1 + 0 ) +
S
t
i
(10)
( 1 + ( − 1 ))
2
c
t
i
≤ 0, s
t
i
< 0
=
1
2
S
t
i
(11)
Reasonably, in order to keep the relative effect of s it over ci
in the third case, the decision making function of concepts
may be replaced by c it + 1 S it .
2
As a result of studying previous models of FCM, a need has
been appeared for a more flexible, reliable and dynamic FCM
for more effective knowledge representation and inference
models to cope efficiently with a wider spectrum of real world
problems.
IV. THE PROPOSED EXTENDED DYNAMIC FUZZY
COGNITIVE MAP DFCM
The definition of FCM in previous works [1-13] deal with
FCM as a 2-element tuple that includes a set of nodes and a
set of directed arrows. A node represents a concept that
indicates an entity, a state or a characteristic of the system. On
the other hand, a directed arrow represents the weighted
causal relationship between the causal and the effect concepts.
Faculty of Computers and Information - Cairo University
AOM -101
The 7th International Conference on INFOrmatics and Systems (INFOS 2010) – 28-30 March
Applied Optimization and Metaheuristics Track
Such weight may be static crisp value or fuzzy value and may
depend on the value of its causal concept. Yet, none of these
previous works take into considerations the prevailing
conditions in the environment under which the FCM will:
firstly, represent the knowledge that may be sensitive to
changes in its surrounding environment and secondly, make
the inferences required to take decisions under recent
prevailing conditions. Also, it ignores the existence of special
type of concepts which: firstly, does not affect directly any
FCM base concepts but affects one or more of the defined
FCM causal relationships; secondly,
may affect some
concepts and its effect should be pass in just some subsequent
paths. Such prevailing environment conditions and such
special type of concepts we will call the set of environment
dominant variables V. Accordingly, we propose to identify a
set of environment dominant variables V which is
comprehensive to represent the current environment state.
Actually, this assertion aims to make the FCM approach more
flexible and reliable in representing a wider range of real
world problems through satisfying the following needs:
As a natural metaphor, there may be some special set of
concepts that have no direct causal relationship with the
base concepts of the targeted FCM although they may
affect the strength of some of its causal relationships. If
such special concepts were added to the designed FCM it
would be isolated-nodes forever which may duplicate the
complexity and complicate both of knowledge
representation and inference.
There may be some concepts that affect other ones but its
effect shouldn't be passed forward in subsequent paths.
Hence, they should be represented in the designed FCM
in a flexible and reliable manner that insures their natural
effect behavior.
The dynamic weights proposed in [6] make the weight
value depends on its causal concepts value. But in real
world, the strength of the causal relationship between two
concepts may depend not only on the strength of its
causal concept but on the environment prevailing
conditions which almost are dynamic.
The proposed DFCM approach aims to fulfill the above cases
through a set of enhancements on both of knowledge
representation and inference phases.
Assuming a DFCM that has n concepts and m
environment dominant variables, the model has a concepts
vector, weight matrix and variable vector as depicted in
equation (1), (2) and (13) respectively.
V = [v1, v2, v3,.…….vm]
The state value of any concept Ci can be within [-1, 1]
that represents the strength of the concept. There is a
fuzzification mapping function for each interface concept that
maps its real world value into a fuzzy value within [-1, 1].
There are many types of functions [14, 15] that can be used as
fuzzification mapping functions like logistic, sigmoid (logistic
or hyperbolic), Gaussian and S membership functions.
Occasionally, the prevailing environment conditions may
affect the shape, the slops and the control points of such
functions. Hence, such membership functions may be
sensitive to the ongoing changes in the environment. Hence,
the same actual value for a concept Ci will be mapped to
different fuzzy values with different environment states.
Dynamic effects of V over the DFCM causal relationships
may be represented using one or more of equations (14), (15),
(16) and (17). Such equations are subjective and depend on
the nature of the problem and the expert point of view.
fw∞− (wij , vk , α ) =
where
f
∞
w−
w −αv (1+ w ) v ≥ 0
w +αv w
v 0
k
ij
k
ij
k
ij
(14)
k
relationship strength wij in response to a positive value of an
environment variable vk and vice versa, and is an effect
factor.
f w∞+ (wij , vk ,α ) =
where
f
∞
w+
w + αv (1 − w ) v ≥ 0
w + αv w
v 0
ij
k
ij
k
ij
k
ij
(15)
k
indicates an increase in the positive causal
relationship strength wij in response to a positive value of an
environment variable vk and vice versa, and is an effect
factor.
A. Knowledge Representation Phase
where
f
x
w−
w − αv w
v ≥0
w + αv (1 + w ) v 0
ij
k
ij
k
ij
k
ij
(16)
k
indicates a decrease in the negative causal
relationship strength wij in response to a positive value of an
environment variable vk and vice versa, and is an effect
factor.
f wx+ (wij , vk ,α ) =
where
The value of any concept ci, causal relationship wij or
environment dominant variable Vk can be within the bipolar
interval [-1, 1].
ij
indicates an increase in the negative causal
f wx− ( wij , vk , α ) =
The proposed model of DFCM is defined as a 3-element
tuple as follows:
DFCM = {C, W, V}
(12)
where,
C = { ci | ci ∈ [-1, 1], i = 1, 2, …n} is a set of concepts.
W={wij|wij ∈ [-1, 1], i, j = 1,2,…n } is a weight matrix.
V = {vk | vk ∈ [-1, 1], k = 1,2,…..m } is a set of
environment dominant variables.
(13)
f
x
w+
w − αv w
v ≥0
w + αv (1 − w ) v 0
ij
k
ij
k
ij
k
ij
(17)
k
indicates a decrease in the positive causal
relationship strength wij in response to a positive value of an
environment variable vk and vice versa, and is an effect
factor.
Faculty of Computers and Information - Cairo University
AOM -102
The 7th International Conference on INFOrmatics and Systems (INFOS 2010) – 28-30 March
Applied Optimization and Metaheuristics Track
B. Reasoning Phase
The inference process in the enhanced DFCM is carried out
by multiplying the adjusted weights values matrix (Wd); as a
response to the effect of its causal concept and the
environment dominant variables at time t (Vt); by the concept
matrix at time t (Ct). The result of the multiplication operation
will be a new 1 x n vector Y as shown in equation (18).
w11d w12d
w1dn
c1t
d
w d w22
w2dn c 2t
(18)
Y = W d × C t = 21
wnd1 wnd2
= [y1,y2,y3………….yn]
wijd c tj ,
yi =
j = 1,2,
d
wnn
variables. Fig. 3 shows the fuzzification mapping function for
car price attribute C1. It is a hyperbolic tangent function whose
output is in the range [-1, 1].
c nt
n
(19)
Fig. 2 The Knowledge model of car_evaluation problem
j ≠i
where yi represents the augmented influence from all
causal concepts over concept xi.
Price strength
Accordingly, the effects between the concepts of the designed
DFCM will reflect any changes in the prevailing environment
conditions as depicted in equation (19). Hence, applying yi as
an input of the decision making function of the concept x;, that
takes into consideration the values of V using equation (20),
will result in the new value of the concept ci.
t +1
c
i
(
= f i yi ,
c )=
t
i
c + y (1− c )
c + y (1+ c )
t
t
t
i
i
i
t
t
t
i
i
i
1 t
ci + 2 yi
t
t
t
i
i
t
t
c ≥ 0, y ≥ 0
c < 0, y < 0
i
Price real value
(20)
i
otherwise
where fi is the decision making function of the concept ci that
takes into consideration the current values of the environment
dominant variables V.
C. AN ILLUSTRATIVE EXAMPLE
The following example illustrates the ability of applying
the proposed DFCM to an intelligent software agent to make it
able to reason in a dynamic environment. The case study
assumes a car-purchasing domain. An intelligent software
agent that represents a customer is in charge of taking a
purchasing decision about a specific class of cars produced at
same year. The agent is concerned with three attributes about
the exhibited cars: price C1, warranty C2, and fuel consuming
rate C3. According such attributes values the agent should
decide how much a specific car satisfies the buyer in
accordance to the prevailing environment conditions. The set
of environment dominant variables V includes: competition
degree V1, fuel price V2 and buyer financial state V3.
According to the proposed model, the representation of carevaluation problem is depicted in Fig. 2.
According to the proposed model any fuzzification
mapping function may be uncertain, i.e., its shape may depend
on the values of one or more of the environment dominant
Fig. 3 Different possible shapes of price fuzzification mapping function at
different times depending on competition degree and customer financial state.
Intuitively, we can note that both of the competition
degree and the financial state of the buyer affects sharply the
shape and the steepness of the price fuzzification mapping
function. Hence, the fuzzification mapping function should be
flexible to respond dynamically to changes in both of them.
That is, the same real price will be mapped to different fuzzy
values in different competition degrees and different buyer
financial states. As a result the price mapping function can be
defined as depicted in equation (21). Accordingly, the user
financial state and the competition degree will affect the
behavior of measuring the strength of a given price i.e. how
much a given price is high or strong.
2
(21)
−1
1+ e
Where θ = pavg + v3 ( pmax − pmin ) / 2, λ = 0.25v1, x is the price
real value, controls the steepness of the function, and is the
function threshold value, pmax, pmin and pavg represent highest,
lowest and average prices.
FC1 (x) =
− λ ( x −θ )
Also, the fuzzification mapping functions for warranty FC2 can
be defined as follows:
Faculty of Computers and Information - Cairo University
AOM -103
The 7th International Conference on INFOrmatics and Systems (INFOS 2010) – 28-30 March
Applied Optimization and Metaheuristics Track
2
FC2 (x) =
1 + e − 0.225 ( x − 6 )
−1
(22)
Assuming that, the initial causal relationship matrix W0 is set
as shown in equation (24).
where Fc2 represents the fuzzification mapping function of
warranty concept.
W
0
C1
= C2
Fuel consuming rate fuzzy value
C1
C2
C3
C4
0
− 0 .1
0 .3
0
0
0
0
0
C3
0
0
0
0
C4
− 0 .6
0 .4
− 0 .5
0
(24)
Assuming that, the following notes represent an expert's point
of view indicating which variables in V affect which weights
in W and how much an effect will take place.
• An increase in competition degree V1 will increase the
negative influences of both of price W41 and fuel consuming
rate W43 and the positive influence of warranty W42 on
satisfaction degree; and vise versa. Such effects can be
Fuel consuming
rate real value
represented
by
f w∞− (w41, v1 ,0.5) ,
f w∞− (w43 , v1 ,1)
and
∞
w+
f (w42, v1,0.5) respectively using equations (14) and (15)
Fig. 4 Fuel consuming rate fuzzification mapping function with threshold
depends on fuel price.
Fig. 4 and equation (23) represent the fuzzification mapping
function for fuel consuming rate C3 as a hyperbolic tangent
function with a threshold depends on fuel price V3. That is, the
higher the fuel price, the lower the threshold value and vice
versa.
FC3 (x) =
2
1+ e
(23)
−1
− λ ( x −θ )
where θ = f avg − v2 ( f max − f min ) / 2, λ = 0.25 , Fc3 represents
the fuzzification mapping function of fuel consuming rate,
fmax , fmin , favg represent highest, lowest and average fuel
consuming rates for 20 Km, V3 denotes the strength of current
fuel price.
TABLE I
• An increase in fuel price V2 will increase the negative
influence of fuel consuming rate W43 on satisfaction degree;
and vice versa. Such effect can be represented as
f w∞− (w43 , v2 ,1) using equation (14)
• An increase in the buyer financial state V3 will decrease the
negative influences of both of price W41 and fuel consuming
rate W43; and vice versa. Such effects can be represented
as
f w−x (w41, v3 ,0.6) and f w−x (w43, v3 ,1) using equation (16).
According to defining such dynamic causal relationship
functions will reflex the environment prevailing condition as
shown in TABLE .
TABLE II
DYNAMIC EFFECTS OF V ON W
Wij0
Wij
THE CONCEPTS MAPPED FUZZY VALUES WITH RESPECT TO V
AT TIMES T1 AND T2.
Mapping the concepts real values into
fuzzy values
V
At time
C
V
t1
Real
Value
t2
V1
0.3
0.4
V2
-0.5
0.3
V3
0.6
-0.2
C1
C2
C3
90,000
$
9
month
3 L/
20km
Fuzzy
value at
t2
W42
-0.149
0.889
W43
0.325
0.325
-0.245
-0.075
Accordingly for specific car, the fuzzification mapping
process for interface concepts real values will be sensitive to
the environment dominant variables V. Assuming that
pmax=110, pmin=30, fmax =6 and fmin=2. The result of the
fuzzification mapping process of the interface concepts C1, C2,
and C3 at times t1 and t2 is shown in TABLE I.
Wijd (t1)
Wijd (t2)
f w∞− (w41, v1 ,0.5)
-0.66
-0.68
f wx− (w41, v3 ,1)
-0.264
-0.744
fw∞+ (w42, v1,0.5)
0.49
0.52
f w∞− (w43 , v1 ,1)
-0.65
-0.7
f w∞− (w43 , v2 ,1)
-0.325
-0.79
f wx− (w43, v3 ,1)
-0.130
-0.83
-0.6
W41
Fuzzy
value
at t1
The effect of V on
Wij
0.4
-0.5
Accordingly, the causal relationship matrices after the
influences of V at times t1 and t2 become as shown in
equations (25) and (26) respectively
W
Faculty of Computers and Information - Cairo University
t1
C1
= C2
C3
C4
C1
0
0
0
- 0.264
C2
− 0 .1
0
0
0 . 49
C3
0 .3
0
0
- 0.130
AOM -104
C4
0
0
0
0
(25)
The 7th International Conference on INFOrmatics and Systems (INFOS 2010) – 28-30 March
Applied Optimization and Metaheuristics Track
C1
0
C2
− 0 .1
C3
0 .3
C4
0
2
0
0
0
0
3
0
0
0
0
− 0 . 744
0 . 52
− 0 . 83
0
C1
W
t2
= C
C
C
4
(26)
Accordingly, the inference is carried out at times t1 and t2 as
follows:
w1t1j C tj1 = w12t1 * C 2t1 + w13t1 * C 3t1
y1t1 =
j
= -0.1 * 0.325 + 0.3 * -0.245= - 0.106
(
C1t1 +1 = f C C1t1 , y1t1
) =c
t1
1
t
+ y1t1 (1− c11)
= -0.149 - 0.106 * 0.851= - 0.239
t
t
t
t
t
t
w1t1j C tj1 = w411 * C11 + w421 * C 21 + w431 * C 31
t1
4
y =
j
= -0.264 * -0.239 + 0.49 * 0.325 +
- 0.130 * -0.245 = 0.254
(
C4t1 +1 = f C C4t1 , y4t1
)
=
in some cases, the designer may not able to add such concepts
because of its partial effect in some subsequent paths of the
effected concept. The proposed model enables the designer of
a DFCM to make any dynamic causal relationship function in
not only its causal concept but also in the environment
prevailing conditions if needed. Also, the proposed model
allows dynamic fuzzification mapping for interface concepts.
This assertion aims to reflex the environment prevailing
conditions on the behavior of fuzzification process.
Accordingly, same interface concept real value will be
mapped to different fuzzy values in response to changes in the
environment. Such contribution makes the inference process
environment sensitive. Accordingly, the result will be more
reliable and robust. An illustrative example shows that: when
evaluating specific car, although it may be satisfying at time
t1 it may become not satisfying at time t2. Hence, the same
fuzzy cognitive map knowledge and concept values infer
deferent outcomes in response to changes in the environment
prevailing conditions and the result became more reliable.
f C (0,0.254 ) = 0.254
REFERENCES
[1]
In the same way, the satisfaction degree at time t2 will be
computed as follows:
w1t 2j C tj2 = w12t2 * C 2t2 + w13t2 * C 3t2
y1t2 =
[2]
[3]
j
= -0.1 * 0.325 + 0.3 * -0.075
= -0.0325 - 0.0225= - 0.055
(
C1t2 +1 = f C C1t 2 , y1t 2
) =c
t2
1
[4]
t
+ y1t2 (1− c12 )
= 0.889 – 0.5 * 0.055 = 0.861
t2
y4 =
w1t 2j C tj2 =
[5]
[6]
j
t2
41
t2
1
t2
t2
w * C + w42
* C 2t2 + w43
* C 3t2
= -0.744 * 0.861 + 0.52 * 0.325 +
- 0.83 * -0.075 = -0.409
(
)
C 4t2 +1 = f C C 4t2 , y 4t2 = f C (0,0.254 ) =-0.409
D. CONCLUSION
In this paper, we introduced an enhanced version of fuzzy
cognitive map approach called "DFCM". The proposed
enhancements aim to make FCM approach able to represent
the domain knowledge and carry out the inference process in
accordance to the dynamism that exist in most of real world
problems. In the proposed DFCM, a new vector of special set
of concepts or variables V was added classical FCM model.
Such concepts represent the environment prevailing
conditions that may have direct effects on some causal
relationships of the designed DFCM although it don't have
any direct causal relationship with any of the designed DFCM
base-concepts. If it were valid to add such concepts to the
designed FCM base-concepts, it will be isolated nodes and
will duplicate the complexity of the solution in average. Also
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
S. Bueno, Jose L. Salmeron, "Benchmarking main activation functions
in fuzzy cognitive maps2", Expert Systems with Applications 36 : 5221–
5229, 2009.
P.R. Innocent, R.I. John, "Computer aided fuzzy medical diagnosis ",
Information Sciences, 162(2004)81-104.2004.
D.E. Koulouriotis, I.E. Diakoulakis, D.M. Emiris, C.D.
Zopounidis,"development of dynamic cognitive networks as complex
systems approximators: validation in financial time series", Applied Soft
Computing 5 : 157–179, 2005.
S. Lee, B. G. Kim, and K. Lee, "Fuzzy cognitive map-based approach to
evaluate EDI performance: a test of causal model", ELSEVIER, Expert
Systems with Applications 27: 287–299, 2004.
C.Y. Miao, A. Goh, Y. Miao, Z.H. Yang, "Agent that model, reason and
makes decisions", Knowledge-Based Systems, 15: 203-211, 2002.
C. Miao, Q. Yang, H. Fang, and A. Goh,, "A cognitive approach for
agent-based personalized recommendation", Knowledge-Based Systems
20: 397-405, 2007.
C. D. Stylios, Voula C. Georgopoulos, and P. P.Groumpos, "the use of
fuzzy cognitive maps in modeling systems", 2000.
A. K. Tsadiras, K. G. Margaritis, "An experimental study of the
dynamics of the certainty neuron fuzzy cognitive maps",
Neurocomputing, 24:95-116, 1999.
A. Tsadiras and K. G. Margaritis, “Cognitive mapping and
certaintyneuron fuzzy congnitive maps,” Inform. Sci., vol. 101, pp. 109–
130,1997.
A. K. Tsadiras, "Comparing the inference capabilities of binary,
trivalent and sigmoid fuzzy cognitive maps", Information Sciences 178
(2008) 3880–3894, 2008.
D. Yaman, S. Polat, "A fuzzy cognitive map approach for effect-based
operations", Information Sciences, 179:382-403, 2009.
S. Zhou and Z. Q. Liu, "Fuzzy Causal Networks: General Model,
Inference, and Convergence", IEEE Tansactions on fuzzy systems, vol.
14, no. 3, June 2006.
B. Kosko , " Fuzzy Engineering", Prentice-Hall, New Jersey, 1997.
M. Negnevitsky, “Artificial Intelligence: A guide to intelligent
Systems”, Biddles Ltd, King’s Lynn, Great Britain, 2005.
O. Castillo and P. Melin, "Type-2 Fuzzy Logic: Theory and
Applications", Spring-Verlag, Berlin, Heidelbrerg, 2008.
Faculty of Computers and Information - Cairo University
AOM -105