Beruflich Dokumente
Kultur Dokumente
6, DECEMBER 2011
AbstractThe theory of fuzzy cognitive maps (FCMs) is a pow- as a collection of discrete-time recurrent irregular neural net-
erful approach to modeling human knowledge that is based on works [5]. Moreover, different from neural networks, FCMs are
causal reasoning. Taking advantage of fuzzy logic and cognitive relatively flexible and simple to use to model human knowl-
map theories, FCMs enable system designers to model complex
frameworks by defining degrees of causality between causal ob- edge: The important concepts of systems are modeled as graph
jects. They can be used to model and represent the behavior of vertices and the causal relationships between them are repre-
simple and complex systems by capturing and emulating the hu- sented as arcs; the weight of an edge indicates the relationship
man being to describe and present systems in terms of tolerance, strength between the related concepts. Unlike typical expert
imprecision, and granulation of information. However, FCMs lack systems that explicitly implement IFTHEN rules, FCMs model
the temporal concept that is crucial in many real-world appli-
cations, and they do not offer formal mechanisms to verify the rules by means of a simple adjacency matrix, and the FCMs
behavior of systems being represented, which limit conventional inference engine is based on an easy, iterative, and numeric pro-
FCMs in knowledge representation. In this paper, we present an cess. Consequently, FCMs are able to represent knowledge and
extension to FCMs by exploiting a theory from formal languages, implement inference operators with greater flexibility, as com-
namely, the timed automata, which bridges the aforementioned in- pared with many other available methods. Nevertheless, in spite
adequacies. Indeed, the theory of timed automata enables FCMs
to effectively deal with a double-layered temporal granularity, ex- of these benefits, conventional FCMs are not capable to deal with
tending the standard idea of B-time that characterizes the iterative the concept of time in a direct way. In consequence, real (social,
nature of a cognitive inference engine and offering model checking political, mechanical, electronic, etc.) systems ineluctably have
techniques to test the cognitive and dynamic comportment of the dynamic behaviors that FCMs are unable to model.
framework being designed. In this discuss about the proposal of a formal approach to
Index TermsFuzzy cognitive maps (FCMs), fuzzy systems, support the definition of a new timed cognitive inference en-
systems dynamic, timed automata. gine, which is named timed-automata-based fuzzy cognitive map
(TAFCM), that is able to improve FCMs in terms of represen-
I. INTRODUCTION tation, simulation, and validation. In order to achieve this aim,
the proposal relies on a double-layered temporal granularity,
UZZY cognitive maps (FCMs) represent a well-defined
F approach to modeling knowledge and inference that are
the most important components of any intelligent system. They
where the low-level layer is based on well-known FCMs timing
mechanisms, such as the base time (B-time) that is introduced
by Carvalho [6], whereas the high-level layer proposes a novel
were initially introduced by Kosko [1] to model systems into the
timing mechanism that is named top time (T-time). T-time acts
area of knowledge domains, for instance, political science, mil-
as a supervisor mechanism that exploits low-level concepts in
itary science, history, international relations, and organization
order to provide FCMs with new timing mechanisms, i.e., the
theory. They have recently emerged as a powerful modeling and
cognitive era and cognitive configuration, that further improve
simulation technique that is applicable to numerous areas. For
the time representation in qualitative system dynamics. In par-
instance, FCMs have been exploited to model the behavior of
ticular, this synergic approach enables the representation of a
complex applications as control plant systems, electrical circuit
generic system as a biological entity that lives its existence by
analysis, cooperative manmachines environments, distributed
crossing a sequence of time periods (cognitive eras), each one
group-decision support, adaptation and learning, etc. [2][4].
representing the longest interval time in which the system does
FCMs are the inference networks that use cyclic, directed
not change its cognitive configuration, i.e., its concepts, causal
graphs to represent the fuzzy causal relationships between fuzzy
relationships, and the B-time value. The double-layered tempo-
concepts, but unlike feedforward neural networks, FCMs feed-
ral granularity and the T-time idea will be integrated with FCMs
back cycles enable a temporal dependence that classifies FCMs
by means of timed automata theory [7], which is largely ex-
ploited to represent complex systems by means of a collection
of states and transitions depending upon temporal events.
Manuscript received June 18, 2010; revised January 8, 2011; accepted April Our method is innovative in that for the first time in the liter-
14, 2011. Date of publication June 16, 2011; date of current version December ature, we model cognitive systems with a well-defined dynamic
5, 2011.
The authors are with the Department of Computer Science, University representation, which is based on a solid mathematical founda-
of Salerno, Fisciano, Salerno 84084, Italy (e-mail: gacampora@unisa.it; tion. In consequence, other than purely temporal benefits, this
loia@unisa.it). choice provides system verification techniques able to be used
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. early in the system design cycle to detect logical bugs before
Digital Object Identifier 10.1109/TFUZZ.2011.2159799 it is implemented. Moreover, because the proposed approach
clearly separates temporal details from cognitive features of a to be described in the network. The value set can be a binary set,
given system, it strongly simplifies the system design stage, as a fuzzy set, or a continuous interval. In addition, the procedure
will be shown in Section VI. of how the causes take effect is modeled by a dynamic system.
The paper is organized as follows. In Section II, related works
are presented. Sections III and IV are, respectively, devoted to
introduce the basic concepts about the timed automata and FCM. B. Fuzzy Cognitive Maps and Temporal Issues
In Section V, which represents the paper core, we present the In spite of their wide applicability, FCMs lack the temporal
formal definition of TAFCMs. Finally, Section VI is devoted concept to support long-term, dynamic problems. For this rea-
to highlight the benefits that are derived from the TAFCMs son, several extended approaches have been proposed in order
modeling through a set of experiments together with the related to try to solve this important issue.
performances analysis. The extended FCM (EFCM) and the fuzzy time cognitive map
(FTCM) [25] were proposed in 1992 and 1995, respectively, to
II. RELATED WORKS extend FCMs with a discrete temporal domain in which time lags
were inserted between the evolving iterations, and the delaying
The first part of this section is devoted to introduce some
of the effect was studied. Later, decaying was introduced in
real-world scenarios that are modeled through FCMs, whereas
1997 as an independent attribute of concepts by Tsadiras et al.
the second part presents some of the most important temporal
[26]. Unfortunately, these approaches naively use deterministic
extensions of FCMs.
values to represent, in an ad hoc fashion, the delay present in
a causal relationship; this strategy is not at all applicable to
A. Fuzzy Cognitive Maps and Real-World Applications
modeling the dynamics of a system. Rather, it is important to
Initially, FCMs have been used in different research [8] and take a holistic view of the system for which FCMs should be
application fields to plan and make decisions in the field of in- designed in accordance with its dynamic behaviors.
ternational relations and political developments [9], [10] and to Another important result is given from fuzzy cognitive net-
analyze graph theoretic behavior [11]. Successively, they have works (FCNs) [27]: An alternative operational extension of
been proposed as a generic system for decision analysis [12] and FCMs to support the close interaction with the system they
for distributed cooperative agents [13]. In the control-related describe and, consequently, become appropriate for control ap-
themes, FCMs have been used to model and support enhanced plications and adaptive decision making. This improves standard
control environments [14], [15] to represent failure models and FCMs by introducing 1) the representation level (the cognitive
effects analysis for a system model [16] and to model the super- graph); 2) the updating mechanism that receives feedback from
visor of control systems [17]. the real system; and 3) the storage of the acquired knowledge
Stylios and Groumpos [18] developed a novel soft computing throughout the operation. FCNs have been successfully applied
approach to model the supervisor of manufacturing systems by to model a maximum power point tracker in order to improve
the usage of FCMs. energy conversion efficiency [28].
FCMs were used as a computationally inexpensive technique More recently, a significant approach, i.e., the dynamical cog-
to develop actors intelligence in a virtual world [19]. Simu- nitive network (DCN), was proposed by Miao and Liu [29].
lations that involve human actors might combine FCMs with DCNs are extensions of FCMs in a serious attempt to model
expert systems in order to model the soft, emotional aspect of dynamics in a continuous time domain. Based on the theory
human decision making, as well as the formal, logical side. An of dynamic systems, Miao and Liu built a DCN that supports
FCM links causal events, actors, values, goals, and trends in a FCMs, CMs, and nonlinear systems. Theoretically, DCNs can
fuzzy feedback dynamical system. support a full set of time-related features. It is also the first
FCM-based models [20] were proposed as a methodology paper in the literature to study the two aspects of the causality
to effectively understand how neophyte web users form the separately, i.e., causes and effects. However, the DCN relies on
cognitive neurological networks that result in a mental pathway, the Laplacian framework to describe the causal relationships.
or cognitive map, that makes the route more navigable to further The transformation between fuzzy knowledge and Laplacian
information, as well as the information they set out to find. functions imposes more modeling efforts to system designers.
In [21], the use of an FCM to analyze the diffusion process Over the past few years, temporalized FCM [30] has been
of a data warehouse in a bank was described. proposed to define a complete discrete temporal extension of
In [22], a formal methodology that is based on FCMs to rep- the FCM. To reduce the complexities that are brought by the
resent different types of knowledge in a collection of intelligent temporalization, a design approach was developed to construct
agents was introduced. the map, by the use of simplified patterns and fuzzy logic-
FCMs and fuzzy rule bases are used to support the causal based effect functions to capture fuzzy knowledge from domain
knowledge acquisition and reasoning processes. In [23], an in- experts.
tegrated process is presented for rating of the intensity of causal However, one of the most significant brands of research in
relationship generating mean FCMs, assessing group consen- the field of FCMs applied to systems dynamics is that related
sus, and supporting the building of group consensus. In [24], an to rule-based FCMs (RB-FCMs) [31], [32]. RB-FCMs are cog-
extension of the FCM was proposed, where each concept can nitive maps that, unlike FCMs, use fuzzy variables that are de-
have its own value set, which depends on how precisely it needs fined through fuzzy membership functions to deal with systems
1042 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 19, NO. 6, DECEMBER 2011
TABLE I
FUZZY COGNITIVE MAP BEHAVIOR
value to each clock, i.e., it is a mapping from X to R. A clock with si S and i [C R], for all i 0, satisfying the fol-
interpretation for X satisfies a clock constraint over X if lowing requirements.
and only if evaluates to true by the use of the values given 1) Initiation: s0 S0 and 0 (x) = 0 for all x C.
by . 2) Consecution: For all i 1, there is an edge in E of the form
Now, a precise definition of a timed transition table, which si1 , si , i , i , i , such that (i1 + i i1 ) satisfies
define the timed automaton behavior, is given. i , and i equals [i 0](i1 + i i1 ).
Definition 3: A timed transition table A is a tuple As an example, consider the timed automata that are shown in
, S, S0 , C, E, where we have the following. Fig. 3. It has two clocks x and y, two control locations l0 and l1 ,
1) is a finite alphabet. and an edge from the control locations l0 to l1 labeled with the
2) S is a finite set of states. guard y 1 x 2, the action name a, and the reset set {x}.
3) S0 S is a set of start states. By the assumption that all clock variables are initially set to zero
4) C is finite set of clocks. and the initial control location is l0 , the graph starts in the state
5) E S S 2C (C) is the collection of (l0 , {x = y = 0}). As the clocks increase synchronously with
transitions. time, it may evolve to all states of the form (l0 , {x = y = t}),
An edge s, s
, a, , represents a transition from state s to where t is a nonnegative real number. At any state with t [1, 2],
state s
on the input symbol a. The set C represents the it may change to the state (l1 , {x = 0, y = t}) by following
collection of clocks to be reset with this transition, and is a the edge from l0 to l1 that resets x. However, it can also be
clock constraint over C. idle forever in the location l0 since there is no way of forcing
If (, ) is a timed word that is viewed as an input to an au- progress in the semantics of timed automata.
tomaton, it presents the symbol i at time i . If each symbol i is The timed transition table, together with the run concept, is
interpreted to denote an event occurrence, then the correspond- the main notion used in our approach to embedding dynamism
ing component i is interpreted as the time of occurrence of i . in the standard FCM definition.
Given a timed word (, ), the timed transition table A starts in
one of its start states at time 0 with all clocks initialized to 0.
As time advances, the values of all clocks change, reflecting the V. MERGING TIMED AUTOMATA AND FUZZY COGNITIVE MAPS
elapsed time. On the time i , A transits from s to s
by the use In an attempt to solve the temporal drawbacks of the FCM that
of some transition of the form s, s
, i , , reading the input is discussed in Section III, we propose TAFCMs that add tempo-
i , if the current values of clocks satisfy . With this transition, ral concepts to the standard FCM by the use of a timed automaton
the clocks in are reset to 0 and, thus, start continuing time with whose possible behaviors define all the potential sequences of
respect to the time of occurrence of this transition. Formally, cognitive eras (and the related cognitive configurations) that the
this timed behavior is captured by introducing runs of timed system could cover during its life cycle. In particular, TAFCMs
transition tables. improve FCMs by associating each state in a timed automaton
ACAMPORA AND LOIA: ON THE TEMPORAL GRANULARITY IN FUZZY COGNITIVE MAPS 1045
These properties will allow us to formally introduce the w((c3 , c5 )) = 0.9, w((c4 , c7 )) = 0.9
TAFCMs.
w((c5 , c6 )) = 0.9, w((c5 , c7 )) = 0.9
A. Timed-Automata-Based Fuzzy Cognitive Maps w((c6 , c1 )) = 0.3, w((c7 , c6 )) = 0.8. (4)
The first step toward the definition of TAFCMs is to redefine
the standard FCMs by means of the graph theory. (A similar Example 2: The following definitions are related to the FCM
representation approach may also be used for the RBFCM.) that is shown in Fig. 1 at iteration 4:
From this point of view, an FCM F can be defined through
a graph, i.e., M = (V, E), that models the map structure, a
function a that associate a real value with each map concept, a M = (V, E)
function w that associate a real value with each map relationship,
and a real value btim e R that is used to model the iteration V = {c1 , c2 , c3 , c4 , c5 , c6 , c7 }
length in cognitive inference. Formally, an FCM F is a 4-tuple a(c1 ) = 0.395
(M, a, w, btim e ), where
a(c2 ) = 1.000
M = (V, E)
a(c3 ) = 0.927
V = {ci |i = 1, . . . , n}
a(c4 ) = 1.000
E = {(ci , cj )|ci , cj V }
a(c5 ) = 1.000
a : V [0, 1]
a(c6 ) = 1.000
w : E [1, 1]
a(c7 ) = 0.085
btim e R. (3)
E = {(c1 , c4 ), (c1 , c3 ), (c2 , c1 ), (c3 , c2 ), (c3 , c5 ),
Here, V is the set of nodes that compose the FCM; a is a function (c4 , c7 ), (c5 , c6 ), (c5 , c7 ), (c6 , c1 ), (c7 , c6 )}
that associates an activation value with each concept in V ; E is
the set of causal relationships between concepts; w is a function w((c1 , c4 )) = 0.9, w((c1 , c3 )) = 0.6
that associates an edge in E with a real weight in [1, 1]; btim e w((c2 , c1 )) = 0.1, w((c3 , c2 )) = 0.7
is a real variable modeling the length of a cognitive iteration;
in the context of TAFCM, the B-time is used to determine how w((c3 , c5 )) = 0.9, w((c4 , c7 )) = 0.9
many cognitive iterations have to be computed during a given w((c5 , c6 )) = 0.9, w((c5 , c7 )) = 0.9
cognitive era.
Examples 1 and 2 show two samples of an FCM modeled w((c6 , c1 )) = 0.3, w((c7 , c6 )) = 0.8. (5)
through (3). In particular, these examples show the FCM of
Fig. 1 at, respectively, iterations 0 and 4.
Example 1: The following definitions are related to the FCM In other words, (4) and (5) represent the cognitive configura-
that is shown in Fig. 1 at iteration 0: tion of the FCM that is shown in Fig. 1 during its first and fifth
cognitive eras.
M = (V, E) The formal graph view of an FCM represents only a static vi-
sion of our cognitive system. The successive step is to introduce
V = {c1 , c2 , c3 , c4 , c5 , c6 , c7 }
a collection of operators able to transform the cognitive structure
a(c1 ) = 0.300 that is defined in (3). These operators represent the fundamental
operations with which the proposed cognitive/dynamic models
a(c2 ) = 0.300
are constructed. They will change the cognitive configuration of
a(c3 ) = 0.300 a given FCM (M, a, w, btim e ) with M = (V, E) by following
the rules:
a(c4 ) = 0.300
1) to add concepts;
a(c5 ) = 0.300 2) to add causal relationships;
3) to remove concepts;
a(c6 ) = 0.300
4) to remove causal relationships;
a(c7 ) = 0.300 5) to magnify/reduce the strength of a causal relationships;
6) to magnify/reduce the level of system concept;
E = {(c1 , c4 ), (c1 , c3 ), (c2 , c1 ), (c3 , c2 ), (c3 , c5 ),
7) to modify the btim e value.
(c4 , c7 ), (c5 , c6 ), (c5 , c7 ), (c6 , c1 ), (c7 , c6 )} Definition 5 (Adding a Novel Cognitive Concept ): This
operator modifies in a direct way the set V that contains the
w((c1 , c4 )) = 0.9, w((c1 , c3 )) = 0.6
fuzzy cognitive concepts. Let ct be the concept to add to VC
w((c2 , c1 )) = 0.1, w((c3 , c2 )) = 0.7 and v [0, 1] the real value associated with the new concept ct .
ACAMPORA AND LOIA: ON THE TEMPORAL GRANULARITY IN FUZZY COGNITIVE MAPS 1047
Then, the behavior of is defined as follows: where the behavior of function, named w
, is the same as that
of w because E
E. Then, M
= (V, E
) is the graph that
VC
= VC {ct }
result from the application, and w
is the weight function
a
: VC
[0, 1] with a
/VC = a and a
(ct ) = v. (6) of F
.
Definition 11 (Additive Modification of a Causal Relation-
Then, M
= (VC
, E) is the new graph that results from ap-
ship ): The operator named modifies the value of a given
plication and a
is the new function that associates the concepts
causal relationship by adding it to a real value . This task is
with the corresponding values.
accomplished by opportunely redefining the weight-assignment
Definition 6 (Removing a Cognitive Concept ): The operator
function w. Let (ci , cj ) be the causal relationship to modify in
removes a cognitive concept, named ct , from the concepts set
additive way; then, w
: E [1, 1] is the modified function
V . Then
VC
= VC {ct } that is defined as follows:
a
: VC
[0, 1] (7) w
((cl , cm ))
where the behavior of the function a
is the same as that of a w((cl , cm )), if (cl , cm ) = (ci , cj )
because VC
VC . Then, M
= (VC
, E) is the new graph that
w((cl , cm )) + , if (cl , cm ) = (ci , cj )
results from application.
and 1 w((ci , cj )) + 1
Definition 7 (Additive Modification of a Cognitive Concept = 1, if (cl , cm ) = (ci , cj )
): This operator modifies the behavior of the function a by
and w((cl , cm )) + < 1
changing the value that is related to a concept ct VC in additive
way. Let r R be a real number; then
1, if (cl , cm ) = (ci , cj )
and w((cl , cm )) + > 1
a(cl ), if cl = ct
(12)
a(cl ) + r, if cl = ct and 0 a(ct ) + r 1
a
(cl ) = Then, M
= (V, E) is the new graph, and w
is its weight
1, if cl = ct and a(ct ) + r 1
function.
0, if cl = ct and a(ct ) + r 0. Definition 12 (Multiplicative Modification of a Causal Rela-
(8) tionship ): This operator changes the value of a given causal
Definition 8 (Multiplicative Modification of a Cognitive Con- relationship by multiplying it for a real value . Analogously to
cept ): This operator changes the behavior of the function a by the previous operator, this task is performed by redefining the
the modification of the value that is related to a concept ct VC weight-assignment function w. Let (ci , cj ) be the causal rela-
through a multiplicative constant. Let r R+ be a real number; tionship to modify in multiplicative way; then, w
: E [1, 1]
then is the modified function that is defined as follows:
a(cl ), if cl = ct
a
(cl ) = a(cl ) r, if cl = ct and 0 a(ct ) r 1 (9) w
((cl , cm ))
1, if cl = ct and a(ct ) r 1 w((cl , cm )), if (cl , cm ) = (ci , cj )
w((cl , cm )) , if (cl , cm ) = (ci , cj )
Definition 9 (Adding a Novel Cognitive Causal Relationship
and 1 w((ci , cj )) 1
): This operator is used to add a novel causal relationship to an
FCM F . Let (ci , cj ) be the causal relationship with ci , cj V = 1, if (cl , cm ) = (ci , cj )
and (ci , cj ) E, and let wij [1, 1] be the weight of rela-
and w((cl , cm )) < 1
tionship (ci , cj ); then
1, if (cl , cm ) = (ci , cj )
E
= E {(ci , cj )} and w((cl , cm )) > 1.
w
: E
[1, 1] with w
/E = w and w
((ci , cj )) = wij . (13)
(10) Then, M
= (V, E) is the new graph, and w
is its weight
function.
M
= (VC , E
) and w
are, respectively, the graph and weight Operators and , which change the interconnection values
function that result from the application of the operator. of a given FCM, take inspiration from FCNs (see Section II-B).
Definition 10 (Removing a Cognitive Causal Relationship Definition 13 (B-time Updating ): As we previously stated,
): This operator removes a given causal relationship from each cognitive era can be characterized by a different value of the
the collection of causal relationships E. Let (ci , cj ) be the re- B-time parameter. Because each cognitive era is determined by
lationship to be removed, with ci , cj V , (ci , cj ) E, and a given state of timed automaton modeling, the T-time concept
w((ci , cj )) = wij [1, 1]; then of our system (as will be shown in the following section), it is
E
= E {(ci , cj )} possible to define an operator that assigns each automaton
state a real number (expressing the B-time value):
V
= V {ci , cj }
w
: E
[1, 1] (11) fBTim e : S R, btim e = fBTim e (s), and s S. (14)
1048 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 19, NO. 6, DECEMBER 2011
Definition 14 (Cognitive Identity ): The operator rep- S with a one or more cognitive configurations that are obtained
resents a simple function identity that transforms an FCM by sequential application over F 0 of the cognitive operators in
(M, a, w, btim e ) in itself. This operator does not have a prac- S S 2C (C) Cop . Then, we have the following,
tical usefulness, but it is necessary for the introduction of the Definition 17 (Cognitive evolution): Let TA = (F 0 , TM ) be
TAFCMs definition in a simpler way. a TAFCM that is defined over a timed cognitive transition table
Once the operators are defined, changing the FCM configu- , S, S0 , C, EC with S = {s0 , s1 , . . . , s|S |1 } being the finite
ration, it is possible to define the cognitive operator set set of automaton states; let F be the collection of all FCMs that
are defined by means of (3) over a set {cl R : l = 1, . . . , k} of
Cop = {, , , , , , , , , }
initial cognitive concepts; let = {1 , 2 , . . . , || } be a subset
which allows us to redefine the timed automata concept in or- of operators in Cop used to define the edges in EC . Then, the
der to introduce a novel kind of transition edges to change the cognitive evolution over a state s0 S0 is
cognitive configuration of the modeled system. In particular, the
: N S F
following edges set is replaced to the standard transitions set
of timed automata defined, inductively, as follows.
1) Base case (i = 0). Let s0 S0 be an initial state of
EC S S 2 (C) Cop
C
(15)
, S, S0 , C, EC ; then
and the novel definition of timed automata that are based on
(0) = (s0 , F 0 ).
cognitive edges is as follows.
Definition 15: A timed cognitive transition table At is a tuple 2) Inductive step. Let (i 1), with i > 1, be the cognitive
, S, S0 , C, EC where we have the following. pair that is defined as
1) is a finite alphabet.
(i 1) = (si1 , F i1 )
2) S is a finite set of states.
3) S0 S is a set of start states. where si1 S and F i1 F ; then
4) C is finite set of clocks.
(i) = (si , F i )
5) EC S S 2C (C) Cop is the collection
of cognitive transitions. An edge s, s
, a, , , , with with si S, F i = i (F i1 ), i and si1 , si , a,
Cop , produces the same effect of a standard transition , , i EC .
s, s
, a, , , but it individuates the task that is defined The image of the function , i.e., I , can be finite or
by the operator Cop . The set C represents the infinite. This depends upon the topology of graph modeling the
collection of clocks to be reset with this transition, and component TM of the TAFCM. Indeed, if the topology of TM
is a clock constraint over C. contains cycles, then the edge si1 , si , a, , , i EC can
We are now in a position to present a formal definition be crossed more times and, consequently, various FCM can be
of a TAFCM and the properties that characterize its dynamic associated with the same state sh S, i.e., sh st st+c 1
behavior. st+c 2 st+c k and {(st+c 1 , F t+c 1 ), (st+c 2 ,
Definition 16: A TAFCM T A is an ordered pair composed F t+c 2 ), . . . , (st+c k , F t+c k ), . . .} I .
by an FCM, i.e., F 0 = (M, a, w, btim e ) that represents an initial More intuitively, (17) shows the sequence of pairs composing
cognitive configuration of a system, together with a timed cogni- a cognitive evolution over s0 S0 together with the fuzzy cog-
tive transition table TM that represents the mathematical entity nitive transformations obtained by exploiting the i operators:
acting as the joining point between cognitivism and dynamism (0) : s0 S0 F0
in system modeling. Formally
1
TA = (F 0 , TM ) (16)
(1) : s1 S F1
The TAFCM properties that define the dynamic behavior of 2
a system are cognitive evolution and cognitive run.
In the following definitions, we will use the subsequent nota- (2) : s2 S F2
tion: the symbols si and i (i N) represent, respectively, the 3
(17)
states of an automaton that are defined by means of Definition .. ..
15 and the cognitive operators that are defined in the collection . .
Cop ; at the same time, the symbols sj and j (j N) represent j
an infinite collection of variables that assume, respectively, the
(j) : sj S Fj
value of states si and operators i .
The cognitive evolution is a mapping among the states S j +1
contained in TM and the collection of possible cognitive con- .. ..
figurations that are obtained starting from F 0 . Specifically, the . .
cognitive evolution is a mathematical succession, which is gen- Obviously, the cognitive evolution represents only a mapping
erated in an inductive way (by the usage of an approach similar to between the states of the timed automaton TM and the col-
the regular expression definition [40]), which maps each state in lection of cognitive configurations that are computable starting
ACAMPORA AND LOIA: ON THE TEMPORAL GRANULARITY IN FUZZY COGNITIVE MAPS 1049
i.e., k Q. Specifically, k is the time amount between the Then, the nth-order cognitive run rcn is the sequence of first n
clock updating. elements of rc :
If TA = (F 0 , TM ) is a TAFCM that models a given system, 1 ,1 2 ,2 n 1 n 1
then the set of cognitive run rc that is defined over the timed rcn : s0 , 0 s1 , 1 . . . , sn 1 , n 1
1 2 n 1
language L, generated by TM , completely describes the collec-
tion dynamic behaviors of the system, whereas the cognitive run where the mapping between the automaton states si and the
rc defined over a single word wi L defines a precise dynamic FCM F i is computed by the nth-order cognitive evolution that
behavior of the system, i.e., wi defines the T-time. is related to .
Definition 19 (T-time): If TA = (F 0 , TM ) is a TAFCM, TM At this point, a simple example of a TAFCM is given. In this
is a timed automaton recognizing the timed language, i.e., sample, a timed cognitive system that describes the behavior of
L = {w1 , w2 , w3 , . . . , wi , . . .}, wi is a timed word, and rc is an ecosystem, which is formed by three initial elements preda-
a cognitive run that is defined over wi , then wi is a T-time of the tors (P2 ), preys (P1 ), and grass for preys (G), is given. Unlike
system. the standard FCM, our approach allows us to dynamically mod-
Starting from the T-time definition, a formal description of ify the cognitive concepts of the system. For instance, in our
cognitive era and cognitive configuration is given. case, the concept hunter (H) can be added in a precise time
Definition 20 (Cognitive Era and Cognitive Configuration): during the system life cycle in order to modify the cognitive
If rc is a cognitive run that is defined over the T-time wi = concepts, G, P1 , and P2 in a novel and unpredicted way. Figs.
(, ) L: 6 and 7 show, respectively, the initial cognitive configuration
F 0 of the proposed system and the modified timed automaton
1 ,1 2 ,2 3 ,3
rc : s0 , 0 s1 , 1 s2 , 2 TM that defines the dynamic behavior, whereas Fig. 8 shows a
1 2 3
finite cognitive evolution 8 of the proposed TAFCM. In de-
then the time interval between the instants i and i+1 is the tail, 8 associates each state in TM with eight FCMs that are
ith cognitive era of system and the FCM F i that depicts the obtained from F 0 by the sequential application of operators in
system during the same interval is defined as the ith cognitive = {, , , , , , , }. It is important to note that each
configuration. state in S may correspond to one or more FCMs, because the
Both the cognitive evolution and the cognitive run are poten- same state can occur in different times during a cognitive run to
tially based on the infinite concept in the system. In fact, the map various cognitive eras.
cognitive evolution can exploit an infinite application of cogni- However, as mentioned in the previous section, one of the
tive operators in to compute the mappings between the state main objectives of proposed paper is to provide a cognitive sim-
si and the FCM F i , whereas the cognitive run uses a timed ulation that is capable of answering questions as whatifwhen.
word, which is defined as a infinite sequence of ordered pairs, In order to compute this simulation over the given TAFCM, it is
1050 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 19, NO. 6, DECEMBER 2011
Fig. 7. Cognitive timed automaton that describes the dynamic behavior of the
ecosystem sample.
Fig. 11. Cognitive configurations that are related to the fourth-order cognitive evolution of w 1 .
1054 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 19, NO. 6, DECEMBER 2011
Fig. 12. Cognitive configurations that are related to the fourth-order cognitive evolution of w 2 .
ACAMPORA AND LOIA: ON THE TEMPORAL GRANULARITY IN FUZZY COGNITIVE MAPS 1055
TABLE II
TABULAR VIEW OF FOURTH-ORDER COGNITIVE EVOLUTIONS THAT ARE RELATED TO T-TIMES w 1 AND w 2
Fig. 13. Concept values (c4 , c8 , c7 ) evolution that is related to the fourth-order cognitive run of w 1 .
In this paper, we have presented a novel fuzzy inference en- shown by the experiment, a TAFCM is a promising framework
gine, i.e., the TAFCM, to support the modeling of dynamic sys- for dynamic, decision support systems. However, TAFCMs that
tems through a cognitive approach. A TAFCM is a synergetic are introduced in this paper exploits expert knowledge to design
approach to solving the drawbacks of the standard FCM. We pro- the collection of cognitive eras and configurations representing
posed a double-layered temporal granularity, based on which, the temporal behavior of the system under design. In the future,
we introduced the concepts of cognitive configuration, cogni- the proposed approach may be embedded to some machine-
tive era, and cognitive runs. It is the TAFCMs ability to express learning approaches, such as evolutionary computation, where
temporal uncertainty that has made it possible for us to use timed population-based search algorithms may automatically find suit-
automata, which play a crucial role in its implementation. As able T-times or cognitives era and configurations to improve the
1056 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 19, NO. 6, DECEMBER 2011
Fig. 14. Concept values (c4 , c8 , c7 ) evolution that is related to the fourth-order cognitive run of w 2 .
system behavior. In particular, we will investigate how genetic [11] Z. Wen-Ran and C. Su-Shing, A logical architecture for cognitive maps,
and memetic algorithms could be used to evolve T-time chro- in Proc. IEEE Int. Conf. Neural Netw., Jul. 1988, pp. 231238.
[12] W. Zhang, S. Chen, and J. Bezdek, Pool2: A generic sys-
mosomes toward a suboptimal systems behavior [43], [44]. tem for cognitive map development and decision analysis, IEEE
Trans. Syst., Man, Cybern., vol. 19, no. 1, pp. 3139, Jan./Feb.
1989.
ACKNOWLEDGMENT [13] W.-R. Zhang, S.-S. Chen, W. Wang, and R. King, A cognitive-map-
based approach to the coordination of distributed cooperative agents,
The authors would like to thank the anonymous reviewers IEEE Trans. Syst., Man Cybern., vol. 22, no. 1, pp. 103114, Jan./Feb.
for their comments, which were very helpful in improving the 1992.
quality and presentation of the paper. They would also like to [14] M. A. Styblinski and B. D. Meyer, Signal flow graphs vs fuzzy cognitive
maps in application to qualitative circuit analysis, Int. J. Man-Mach.
thank Prof. Z.-Q. Liu for fruitful and precious discussions on Stud., vol. 35, no. 2, pp. 175186, 1991.
FCMs. [15] S. Lee and I. Han. (2000). Fuzzy cognitive map for the design of EDI
controls. Inf. Manag. [Online]. vol. 37, no. 1, pp. 3750. Available: http://
REFERENCES www.sciencedirect.com/science/article/B6VD0-3XVPG51-4/2/e0a7e8aa
275c397a2d43d39eff4efcac
[1] B. Kosko, Fuzzy cognitive maps, Int. J. Man-Mach. Stud., vol. 24, [16] C. E. Pelaez, and J. B. Bowles, Using fuzzy cognitive maps as a system
pp. 6575, 1986. model for failure modes and effects analysis, Inf. Sci., vol. 88, no. 14,
[2] C. D. Stylios and P. P. Groumpos, Fuzzy cognitive maps in modeling pp. 177199, 1996.
supervisory control systems, J. Intell. Fuzzy Syst., vol. 8, pp. 8398, [17] C. D. Stylios and P. P. Groumpos, The challenge of modeling supervisory
2000. systems using fuzzy cognitive maps, J. Intell. Manuf., vol. 9, pp. 339
[3] C. D. Stylios, V. C. Georgopoulos, G. A. Malandraki, and S. Chouliara, 345, 1998.
Fuzzy cognitive map architectures for medical decision support systems, [18] C. D. Stylios and P. P. Groumpos, A soft computing approach for mod-
Appl. Soft Comput., vol. 8, no. 3, pp. 12431251, 2008. elling the supervisor of manufacturing systems, J. Intell. Robot. Syst.,
[4] W. Stach, L. Kurgan, and W. Pedrycz, Parallel learning of large fuzzy cog- vol. 26, no. 3-4, pp. 389403, 1999.
nitive maps, in Proc. Int. Joint Conf. Neural Netw., Aug. 2007, pp. 1584 [19] J. Dickerson and B. Kosko, Virtual worlds as fuzzy cognitive maps, in
1589. Proc. IEEE Annu. Int. Symp. Virtual Reality, Sep. 1993, pp. 471477.
[5] J. Carvalho, On the semantics and the use of fuzzy cognitive maps in [20] G. Meghabghab, Mining users web searching skills through fuzzy cog-
social sciences, in Proc. IEEE Int. Conf. Fuzzy Systems, Jul. 2010, pp. 1 nitive state map, in Proc. Joint 9th IFSA World Congr. 20th NAFIPS Int.
6. Conf., Jul. 2001, pp. 429434.
[6] J. Carvalho, Mapas cognitivos baseados em regras difusas: Modelacao e [21] M. S. Khan, M. Quaddus, A. Intrapairot, and A. Chong, Modelling data
simulacao da dinamica de sistemas qualitativos, Ph.D. dissertation, Dept. warehouse diffusion using fuzzy cognitive mapsA comparison with the
Electr. Comput. Eng., Inst. Superior Tecnico, Univ. Tecnica de Lisboa, system dynamics approach, in Proc. 3rd Western Australian Workshop
Lisbon, Portugal, 2002. Inf. Syst. Res., 2000, pp. 17.
[7] R. Alur, A theory of timed automata, Theor. Comput. Sci., vol. 126, [22] C. Stylios, V. Georgopoulos, and P. Groumpos, Introducing the theory
pp. 183235, 1994. of fuzzy cognitive maps in distributed systems, in Proc. IEEE Int. Symp.
[8] Y. Miao and Z.-Q. Liu, On causal inference in fuzzy cognitive maps, Intell. Control, Jul. 1997, pp. 5560.
IEEE Trans. Fuzzy Syst., vol. 8, no. 1, pp. 107119, Feb. 2000. [23] A. Mobolurin. (1997). Generating consensus fuzzy cognitive maps,
[9] R. Taber. (1991). Knowledge processing with fuzzy cognitive maps. in Proc. IASTED Int. Conf. Intell. Inf. Syst., Washington, DC:
Expert Syst. Appl. [Online]. vol. 2, no. 1, pp. 8387. Available: IEEE Comput. Soc. [Online]. p. 231. Available: http://portal.acm.org/
http://www.sciencedirect.com/science/article/B6V03-47XNRK2-1B/2/ citation.cfm?id=846229.848793
1bc65261711e31d73e0478f919bcdc91 [24] B. Kosko, Neural Networks and Fuzzy Systems: A Dynamical Systems
[10] W. Pedrycz. (2010). The design of cognitive maps: A study in synergy of Approach to Machine Intelligence. Upper Saddle River, NJ: Prentice-
granular computing and evolutionary optimization. Expert Syst. Appl. [On- Hall, 1991.
line]. vol. 37, no. 10, pp. 72887294. Available: http://www.sciencedirect. [25] K. S. Park and S. H. Kim, Fuzzy cognitive maps considering time re-
com/science/article/B6V03-4YNT455-4/2/dd3afbb18ed0eade5383c8285 lationships, Int. J. Hum.-Comput. Stud., vol. 42, no. 2, pp. 157168,
0a6cf35 1995.
ACAMPORA AND LOIA: ON THE TEMPORAL GRANULARITY IN FUZZY COGNITIVE MAPS 1057
[26] A. Tsadiras and K. Margaritis, Cognitive mapping and certainty neu- Giovanni Acampora (M07) received the Laurea
ron fuzzy cognitive maps, Inf. Sci., vol. 101, no. 12, pp. 109130, (cum laude) and Ph.D. degrees both in computer sci-
1997. ence, from the University of Salerno, Salerno, Italy,
[27] T. L. Kottas, Y. S. Boutalis, and M. A. Christodoulou, Fuzzy cognitive in 2003 and 2007, respectively.
network: A general framework, Intell. Decision Technol., pp. 183196, Since 2007, he has been a Research Associate with
2007. the Department of Mathematics and Computer Sci-
[28] T. Kottas, Y. Boutalis, and A. Karlis, New maximum power point tracker ence, University of Salerno. He is also a member of
for PV arrays using fuzzy controller in close cooperation with fuzzy the Multi-Agent Laboratory, University of Salerno,
cognitive networks, IEEE Trans. Energy Convers., vol. 21, no. 3, pp. 793 and is scientifically coresponsible for the COnsorzio
803, Sep. 2006. di RIcerca Sistemi ad Agenti (CORISA) Research
[29] Y. Miao, Z.-Q. Liu, C. K. Siew, and C. Y. Miao, Dynamical cognitive Centre. From September 2003 to June 2007, he was
networkAn extension of fuzzy cognitive map, IEEE Trans. Fuzzy also involved with the Centro Regionale di Competenza - Information and Com-
Syst., vol. 9, no. 5, pp. 760770, Oct. 2001. munication Technology (CRDC-ICT) Domotic project, where he was engaged
[30] H. Zhong, C. Miao, Z. Shen, and Y. Feng, Temporal fuzzy cognitive in research on multiagent systems and artificial intelligence applied to ambient
maps, in Proc. IEEE Int. Conf. Fuzzy Syst., (IEEE World Congr. Comput. intelligence environments. He designed the Fuzzy Markup Language, which is
Intell.), Jun. 2008, pp. 18311840. an XML-based environment for modeling transparent fuzzy systems. His current
[31] J. Carvalho and J. Tome, Rule based fuzzy cognitive maps-qualitative research interests include novel algorithms design approaches that are inspired
systems dynamics, in Proc. 19th Int. Conf. North Amer. Fuzzy Inf. Pro- by natural systems such as swarm intelligence; evolutionary and memetic strate-
cess. Soc., 2000, pp. 407411. gies; investigating the designing of novel humancomputer interaction systems
[32] J. Carvalho and J. Tome, Rule based fuzzy cognitive mapsExpressing based on integration among haptic hardware; virtual reality and augmented re-
time in qualitative system dynamics, in Proc. 10th IEEE Int. Conf. Fuzzy ality technologies; formal methods from the language theory area, and the study
Syst., 2001, pp. 280283. of temporal effects on the behavior of fuzzy systems modeled through fuzzy
[33] J. Carvalho, L. Wise, and A. Murta, Fuzzy modeling and simulation controllers and fuzzy cognitive maps. He also serves as a reviewer for several
of purse-seine fishing skippers behavior, in Proc. Annu. Meeting North international journals and conferences.
Amer. Fuzzy Inf. Process. Soc., May 2008, pp. 16. Dr. Acampora is a member of the IEEE Computational Intelligence Society
[34] L. B. Torres Simoes Wise, J. Carvalho, A. Murta, P. Abreu, A. C. Technical Committee on Standards. In this context, he also served as the Chair
Fernandes, A. L. Ferreira, E. Henriques, D. Silva, G. Pestana, and of the Task Force on Taxonomy and Terminology and as the Vice Chair of
J. A. B. Tome, Cognitive maps from onboard observers as a tool for the Task Force on New Standard Proposals. Since 2010, he has beeen serving
fishery management, in Proc. ICES Annu. Sci. Conf., 2009, pp. 18. as Secretary and Treasurer of the Italian Chapter of the IEEE Computational
[35] J. Carvalho, M. Carola, and J. Tome, Using rule-based fuzzy cognitive Intelligence Society.
maps to model dynamic cell behavior in Voronoi based cellular automata,
in Proc. IEEE Int. Conf. Fuzzy Syst., 2006, pp. 16871694.
[36] J. Aguilar, A survey about fuzzy cognitive maps papers, Int. J. Comput.
Cogn., vol. 3, no. 2, pp. 2733, 2005.
[37] A. K. Tsadiras. (Oct. 2008). Comparing the inference capabilities of
binary, trivalent and sigmoid fuzzy cognitive maps. Inf. Sci. [On-
line]. 178, pp. 38803894. Available: http://portal.acm.org/citation.cfm?
id=1401271.1401533
[38] Y. Boutalis, T. Kottas, and M. Christodoulou, Adaptive estimation of
fuzzy cognitive maps with proven stability and parameter convergence,
IEEE Trans. Fuzzy Syst., vol. 17, no. 4, pp. 874889, Aug. 2009.
[39] C. Norstrom, A. Wall, and W. Yi. (1999). Timed automata
as task models for event-driven systems, in Proc. Sixth
Int. Conf. Real-Time Comput. Syst. Applications. Washington, Vincenzo Loia (SM08) received the Bachelors de-
DC: IEEE Comput. Soc. [Online]. pp. 182204. Available: http:// gree in computer science from the University of
portal.acm.org/citation.cfm?id=519167.828781 Salerno, Fisciano, Italy, in 1984 and the Ph.D. de-
[40] J. E. Hopcroft, R. Motwani, and J. D. Ullman, Introduction to Automata gree in computer science from the University of Paris
Theory, Languages, and Computation, 3rd ed. Boston, MA: Addison- VI, Paris, France, in 1989.
Wesley, 2006. Since 1989, he has been a Faculty member with
[41] J. Bengtsson, K. Larsen, F. Larsson, P. Pettersson, and W. Yi. the University of Salerno, where he teaches Operating
(1996). UPPAAL: A tool suite for automatic verification of real- Systems, Semantic Web, and Multi-Agents Systems.
time systems, in Proceedings DIMACS/SYCON Workshop Hybrid Syst. He is currently a Full Professor of computer science
III. Secaucus, NJ: Springer-Verlag. [Online]. pp. 232243. Available: with the Department of Mathematics and Computer
http://portal.acm.org/citation.cfm?id=239587.239611 Science. He is the author of more than 190 original
[42] K. Larsen, P. Pettersson, and W. Yi, Model-checking for real-time sys- research papers in international journals, e-book chapters, and international
tems, in Fundamentals of Computation Theory (Lecture Notes in Com- conference proceedings. His current research interests include merging soft
puter Science Series) vol. 965, H. Reichel, Ed. Berlin, Heidelberg, computing and agent technology to design technologically complex environ-
Germany: Springer-Verlag, 1995, pp. 6288. ments, with particular interest in web intelligence applications.
[43] Z. Zhou, Y. S. Ong, M. H. Lim, and B. S. Lee, Memetic algorithm using Dr. Loia is the Co-Editor-in-Chief of Soft Computing and the Editor-in-Chief
multi-surrogates for computationally expensive optimization problems, of Ambient Intelligence and Humanized Computing. He serves as an Editor for
Soft Comput., vol. 11, no. 10, pp. 957971, 2007. 14 other international journals. He has been the Chair of the Emergent Tech-
[44] J. Tang, M. H. Lim, and Y. S. Ong, Diversity-adaptive parallel memetic nologies Technical Committee of the IEEE Computational Intelligence Society,
algorithm for solving large scale combinatorial optimization problems, where he is currently the Chair the of Task Force Intelligent Agents.
Soft Comput., vol. 11, no. 9, pp. 873888, 2007.