You are on page 1of 7

Expert Systems with Applications 37 (2010) 72887294

Contents lists available at ScienceDirect

Expert Systems with Applications


journal homepage: www.elsevier.com/locate/eswa

Short communication

The design of cognitive maps: A study in synergy of granular computing and evolutionary optimization
Witold Pedrycz *
Department of Electrical and Computer Engineering, University of Alberta, Edmonton, Canada T6R 2G7 Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland

a r t i c l e

i n f o

a b s t r a c t
Cognitive maps and fuzzy cognitive maps offer interesting and transparent modeling capabilities by functioning at a level of conceptual entities (nodes) and their relationships expressed either at the qualitative level of excitatory/inhibitory relationships or being further numerically quantied as encountered in fuzzy cognitive maps. While there has been a vast array of conceptual enhancements, a relatively less attention has been paid to the design of the maps especially when dealing with an algorithmic way of forming the map. The objective of this study is to offer a design strategy in which starting with experimental evidence in the form of numeric data, those data are transformed into a nite and small number of concepts (nodes) of the map and afterwards the connections of the map are estimated. We show that techniques of Granular Computing, especially fuzzy clustering are effectively used to form concepts (nodes) of well-articulated semantics. In the sequel, we show the use of global optimization in the form of Particle Swarm Optimization (PSO) to carry out calibration of the connections of the interrelationships between the nodes of the map. Numeric examples are concerned with the representation of time series and their visualization in the form of fuzzy cognitive maps. Further interpretation of the maps is also discussed. 2010 Elsevier Ltd. All rights reserved.

Keywords: Cognitive maps Design Granular computing Concept development Particle Swarm Optimization (PSO) Parametric learning Fuzzy clustering Fuzzy C-means Time series representation

1. Introduction In spite of their visible architectural and functional differences, cognitive maps, automata, fuzzy automata, Petri nets, fuzzy Petri nets share one visible commonality: all of them operate on symbols, concepts, notions, abstractions but not plain numbers. The assumed level of abstraction is required to develop the foundations of these constructs and the constructs themselves are aimed at capturing essential dependencies between the abstract entities. Information granulation plays a pivotal role in the buildup of conceptual constructs. As such, Granular Computing is a contributing technology positioned at the conceptual as well as algorithmic level of these constructs. Cognitive maps and fuzzy cognitive maps have been around for several decades starting from some early works of Kelly (1955) and Axelrod (1976). Even a brief retrospective summary of the area shows very apparently that the main emphasis is on the concept, its formal analysis (Chaib-Draa & Desharnais, 1998) and its renement and generalizations as well as some studies devoted to the analysis of such maps (Konar & Chakraborty, 2005; Kosko, 1986;
* Address: Department of Electrical and Computer Engineering, University of Alberta, Edmonton, Canada T6R 2G7. E-mail address: pedrycz@ee.ualberta.ca 0957-4174/$ - see front matter 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2010.03.006

Salmeron, 2009; Styblinski & Meyer, 1991; Stylios & Groumpos, 1999). There has been a wealth of conceptual developments in the area with a remarkable plethora of generalizations of the generic concept in terms of nodes and connections with an obvious intent to enhance the knowledge representation capabilities. Relatively far less attention has been paid to the comprehensive and practically sound ways of design of the cognitive maps. Interestingly, from the very inception of maps, one can see a predominant trend of accepting a position of a manual, user-driven development of the maps. Handcrafting the structure via a manual development of a collection of concepts (nodes) and then forming a suitable set of connections is a common practice. The level of difculties with this approach increases quite fast and is exacerbated by two important practical factors. First, the maps tend to be larger comprising a signicant number of nodes. Second, while in the generic form of the maps, we admitted only the +1 and 1 connections, in fuzzy cognitive maps the values of the connections being anywhere in-between 1 and +1. It becomes difcult, if not impossible, to estimate the numeric values of such connections in a manual (expert-type) of mode. We badly need some algorithmic support with this regard. This has been a signicant reason to envisage some design schemes including various optimization techniques such as e.g., Evolutionary optimization and simulated

W. Pedrycz / Expert Systems with Applications 37 (2010) 72887294

7289

annealing, see (Alizadeh and Ghazanfari, 2009; Ghazanfari, Alizadeh, Fathian, & Koulouriotis, 2007; Kim, Kim, Hong, & Kwon, 2008; Papageorgiou & Groumpos, 2005; Papageorgiou, Stylios, & Groumpos, 2006; Stach, Kurgan, Pedrycz, & Reformat, 2005). In spite of these pursuits, one can conclude that those developments are quite introductory in their nature. The other fundamental design issue is about the formation of the concepts. While it would be difcult to come up with a fully automated scheme of eliciting the concepts from domain knowledge, more thorough investigation could be focused on the work on concept formation in presence of numeric data. The objective of this study is twofold. First, we intend to show a way of forming abstract constructs concepts (nodes) on a basis of experimental evidence thus demonstrating a way in which we transform numeric data into a collection of semantically meaningful concepts (nodes) in this way forming a blueprint of the structure of the cognitive map. Second, we develop a scheme of parametric optimization of cognitive maps through the evolutionary optimization realized by means of Particle Swarm Optimization (PSO). We start with a brief recollection of the underlying ideas of cognitive maps and processing supported therein (Section 2). In the sequel, in Section 3, we elaborate on the formation of concepts (nodes) of the map showing that they can be formed on the basis of numeric experimental evidence through fuzzy clustering. The learning of the maps via Evolutionary Optimization is discussed in Section 4 while in Section 5 we provide some numeric experimentation. 2. The architecture Cognitive map is composed of a collection of nodes and directed links (edges) between the nodes. The nodes represent concepts, say A1, A2, . . . , Ac envisioned in the description of the problem/process while the edges represent relationships between the concepts or more precisely a way in which one concept affects another one. The edges are directed and the connections between the nodes could be asymmetric. In the original concept of cognitive maps, there are two numeric values associated with the edges, that is +1 (which indicates positive impact exerted on the node) and 1 (which stands for an inhibitory nature of inuence occurring between the nodes). Traditionally, in fuzzy cognitive maps, the strength of connections is quantied to be in-between 1 and 1 even though the membership degrees vary in the unit interval. As an example, we are interested in building a cognitive map for a time series or a multivariable time series (viz. a collection of scalar time series; such as those encountered in stock market, multilead EEG or ECG signals). The concepts of interest here could be formed on a basis of descriptors of amplitude of the signal (or signals) and the changes in their amplitudes. More formally, we could envision them as elements of the Cartesian product of information granules in the space of amplitude and change of amplitude. For instance, considering that we distinguish between L (low), M (medium), H (high) amplitude and likewise L, M, and H change of amplitude, we can have a collection of nodes A where each of them comes with a well-dened semantics

ules, say through the concepts of sets or rough sets. One can also envision the use of probability in the quantication of the strengths of the relationships. In all these cases, however, as we encounter the range of [0, 1], some provisions need to be made with regard to the inhibitory nature of the relationships. More formally, an activation level of the ith node in discrete time moment k is governed by the expression

Ai k 1 f

c X j1

! wij Aj k w0i ; 1

where wij e [0, 1] are the strengths of the edges and w0i e [0, 1] is the bias associated with the ith node. Ai(k) is an activation level of the ith concept in the kth time instance. We can also introduce a more concise vector notation. Denote the vector of the connections of ith expression in (1) by wi, wi = [wi1, wi2, . . . , wic]T. Likewise the vector of activation levels is described as A(k) = [A1(k) A2(k) Ac(k)]T. Overall we have Ai k 1 f wT Ak w0i . i The monotonically nondecreasing nonlinear function f is used to transform the weighted combination of the activation levels coming from the remaining nodes of the network. Typically, sigmoid function is sought, say f(u) = 1/(1 + exp(u)) or its parametric version is considered where f(u) = 1/(1 + exp(ru)), u e R and r > 0. Admitting this enhancement, we rewrite (1) as follows

Ai k 1 fi

c X j

! wij Aj k w0i fi

c X j

! wij Aj k w0i ; ri ; 2

where the steepness parameters ri is associated with the individual node of the map. The role of this parameter is to provide some additional calibration of the activation level of the node. The steepness inuences the sensitivity of the node as becomes clearly visible in Fig. 1; higher values of r increase the steepness of the function and make it more sensitive to the changes of u. More specically, the derivative df/du becomes higher for increasing values of u. An illustration of the details of the cognitive map along with the notation used with each node of the map is presented in Fig. 2. Interestingly, the computing faculties associated with each node could be made more advanced to the point of endowing each node with some local neural networks or other locally available processing scheme (which might be a certain polynomial, neural network or local rule-based architecture, etc.) We can also note that when the processing capabilities are reduced then cognitive maps and fuzzy cognitive maps start resembling automata and fuzzy automata, respectively.

0.5

A fL L; M M; S S; . . . ; L Sg:
For instance, the node associated with the Cartesian product (L L) describes a concept of high amplitudes and high changes of amplitude. The symbols L, M, H are examples of information granules. Here we are concerned with them being represented as fuzzy sets (more specically, Cartesian products of fuzzy sets) however one could envision any other formal representation of the information gran-

u
Fig. 1. A family of sigmoid functions indexed by the values of r; thick solid line relates to the function with the highest value of this parameter (r = 4.0).

7290

W. Pedrycz / Expert Systems with Applications 37 (2010) 72887294

A(k + 1)s being computed on a basis of A(k)s available in the kth time moment and the transformed vector zk+1 which with the use of fuzzy granulation is transformed into the activation levels of the fuzzy sets of nodes, namely B1(k + 1), B2(k + 1), . . . ,Bc(k + 1)
node-i wij

! 2 N c c X X X 1 wij Ai k w0i ; ri : Bi k 1 fi N 1c k i j

wik
node-j node-k

Fig. 2. The computational details of the fuzzy cognitive map; shown are connections between the nodes and a way of processing realized at each node. Dotted connection (link) indicates an inhibitory nature of interaction.

3. Fuzzy clustering: from data to information granules Fuzzy clustering is an effective way of forming information granules on a basis of numeric (experimental) data. Fuzzy C-Means (FCM) Bezdek (1981) is one of the commonly used clustering techniques. As the method itself is well-documented in the literature, the reader may refer e.g., to Bargiela and Pedrycz (2003) for further details regarding optimization, algorithmic enhancements and experimentation. In the context of this study, we note that the objective function Q is regarded as a sum of weighted distances between the data points zk, k = 1, 2, . . ., N and the prototypes (vi) where the weights are the corresponding entries of the partition PP matrix, viz. the membership degrees that is Q c N um i k ik 2 kzk v i k . The data are viewed as vectors of amplitude and changes of amplitude of the time series reported at the consecutive time moments, say zk = [xk, Dxk]T in case of multivariable time series. The fuzzication coefcient (factor) m plays an important role impacting the shape of membership functions and in this sense affecting the nature of the concept. There are three interesting cases with this regard: (a) if m tends to 1, say 1.05 or less, the membership functions become very close to the characteristic functions so most of the elements assume values 0 or 1 with a very narrow range of arguments for which this membership functions assumes values in-between 0 and 1, (b) quite commonly, the value of :m is equal to 2 and in this case we encounter a shape of the membership functions that resemble Gaussian functions (which is smooth with a large number of entries characterized by intermediate membership values), (c) high values of m, say 4 or 5, lead to the spiky shape of the membership function where now there is a single element (prototype) with the membership degree of 1 and the neighboring elements assume far lower values. In this case, on average, the membership grade is equal to 1/c.

The normalization used in V is completed over the number of nodes (c) and number of data (N 1) which makes the value of this index independent of the characteristics of the map and the amount of data used for the training process. This allows for a comparative analysis of results even though they could have been realized in different learning conditions. As far as the parametric exibility is concerned, the adjustable parameters of the map include the matrix of weights, vector w0, biases and steepness parameters of the nonlinear functions. In many cases, the steepness parameters are not adjustable and set to 1. The gradient-based learning is one of the available learning alternatives. Here one could be faced with the high dimensionality of the problem which has direct implications on the effectiveness of this learning method causing its inability to support global exploration of the search space and a signicant exposure to numerous traps of local minima. Evolutionary optimization is an interesting avenue to follow. While genetic algorithms have been studied in the past, the associated computing burden is high. Particle Swarm Optimization (PSO) (Kennedy & Eberhart, 1995) arises here as a viable alternative which offers possibilities of global, population-based optimization yet not imposing a very heavy computational overload. In a nutshell, PSO involves a population of particles whose dynamics is guided by the mechanisms of social interactions and personal experience. The details of the PSO as well as their numerous variants are documented in the literature, cf. (Chatterjee & Siarry, 2006; Coelho, de Moura Oliveira, & Boaventura Cunha, 2005; Paterlini & Krink, 2006; van den Bergh & Engelbrecht, 2006). The search space involves the weights, biases and a vector of steepness parameters. All of them are arranged in a single vector (particle) which is then subject to the operations used in the PSO. Each particle explores the search space and during this search it adheres to some quite intuitively appealing guidelines navigating the search process: (a) to some extent the particle tries to follow its previous direction, and (b) it looks back at the best performance both at the level of the individual particle and the entire population and tries to duplicate these successive moves. In this sense the algorithm exhibits some societal facets as there is some collective search of the problem space along with some component of individual memory (local behavior) incorporated as an integral part of the search mechanism. More formally, these components of the search strategy can be described as follows

4. Learning The process of learning is crucial to the design of the cognitive map and to its further performance. We discuss a supervised version of the learning process. Let us remark that majority of the existing learning schemes applied here is of unsupervised nature, e.g., Hebbian learning, however as the training data are available, this fully justies the use of the supervised mode of learning. The data available for learning come in the form of the sequence of vectors zk, k = 1, 2, . . . , N. More precisely, these vectors are transformed into the levels of activation of the information granules A1(k), A2(k), . . . , Ac(k) and as such used in the learning procedure. The performance index is dened as a normalized sum of distances |||| between predicted levels of activation of the nodes,

zt 1 zt v t 1 ==update of position of the particle; v t 1 nvt /1 p xt /2 ptotal xt ==update of speed of the particle; 4
where p denotes the best position (the lowest performance index) reported so far for this particle while ptotal is the best position overall developed so far across the whole population. The two other parameters of the PSO, that is /1 and /2 are random numbers drawn from the uniform distribution U[0, 2] dened over the [0, 2] interval that help build a proper mix of the components of the speed; different random numbers affect the individual coordinates of the speed. As seen in (4), the current speed v(t) is scaled by the inertial weight (n) smaller than 1 whose role is to articulate some factor of resistance to change the current speed (the values of the inertia weight

W. Pedrycz / Expert Systems with Applications 37 (2010) 72887294

7291

are kept below 1). Second, we relate to the memory of this particle. Thirdly, there is some societal aspect of the optimization scheme expressed by its reliance on the best performance reported across the whole population. 5. Experimental studies The experiments reported in this section illustrate the performance of the learning method as well as look carefully at the cognitive maps constructed there. Synthetic time series. This one-dimensional time series comes from the website of time series, http://robjhyndman.com/TSDL/ and is presented in Fig. 3 in which we represent it by showing the values of xk, changes in the signal Dxk and illustrate the distribution of data in the x Dx coordinates as well. The last graph tells us about the structure in data and a location of possible clusters. The FCM algorithm was run to reveal the structure in the twodimensional data (xk, Dxk). The clustering were carried out for a varying number of clusters ranging from c = 210 clusters. In the algorithm, the fuzzication coefcient was equal to 2.0, the distance was expressed as a weighted Euclidean distance and the method itself was run for 70 iterations (which was more than sufcient to reach convergence). The results obtained in terms of the minimized objective function are shown in Fig. 4. As expected, the values of Q get lower with the increasing number of clusters however the reduction of Q becomes less visible when going beyond a certain number of clusters, say 5 or 6. The optimization of the cognitive map was realized as described in Section 4. We started with c = 2 nodes of the map and moved up to 10 nodes. The minimized performance index, Fig. 5, exhibits an interesting tendency: for the xed value of the slope of the nonlinear function, its values decrease but then move up slightly. In case of the optimized slopes of the functions, there is a continuous decrease in the values as the number of nodes (concepts) gets higher. While there is no benecial effect of bringing the optimized slope in case of lower number of clusters, the improvement is visible when the values of c are higher (which is again illustrated in Fig. 5).

350 300 250 200 150 100 50 0 2 4 6 8 10

Fig. 4. Objective function of the FCM versus varying number of clusters, c = 2, . . . , 10.

0.14 0.12 0.1 0.08 0.06 0.04 0.02 0 2 3 4 5 6 7 8 9 1 0

Fig. 5. Performance index versus the number of nodes of the map; the optimization of the slopes of the nonlinear functions becomes more advantageous with the increase of the number of the nodes: black line optimization of steepness coefcients, grey line- xed value of steepness equal to 1.0.

Let us look in more detail into one case where the number of the nodes was set to 5. The choice of this particular number stems from the fact that at this value the changes in the values of Q become less intensive. The prototypes produced by the FCM

5 4 3 2 1 0 -1 0 -2 -3 -4 -5

50

100

150

200

250

300

350

400

5 4 3 2 1 0 -1 0 -2 -3 -4 -5

50

100

150

200

250

300

350

400

5 4 3 2 1 0

-5

-3

-1 -1
-2 -3 -4 -5

Fig. 3. Representation of synthetic time series: (a) sequence of samples xk, (b) sequences of differences Dxk, and (c) x Dxk plot.

7292

W. Pedrycz / Expert Systems with Applications 37 (2010) 72887294

5 4 3 2 1 -5 -3 -1 0 -1 -2 -3 -4 -5

xk

tative of the structure in the data. Moreover with each of them we can associate some clearly dened semantics, say

L LS ML SS SL L;
where L, M, and S stand for the symbols (fuzzy sets) of Large, Medium, and Small. For instance (L, L) captures the concept where we have signal and its change that are negative () of high amplitude. Running the PSO learning process, we end up with the following optimized values of the connections:

xk

w1 0:359 0:079 w3 0:676

0:090

0:066

0:632;

w2 0:377 0:727 1:000 0:595 0:685;


Fig. 6. Prototypes (concepts) superimposed on the original data.

0:427 0:401

0:526 0:833;

(-L, -L)

(-S, -M)

w4 0:477 0:393 w5 0:619 0:582

0:234 0:425 0:016; 0:471 1:000 0:641;

(L, L)

(L, S)

(-S , S)
Fig. 7. The resulting cognitive map shown are the connections whose values exceed the threshold value of 0.30.

v 1 2:03 1:86; v 2 0:20 0:89; v 3 1:48 0:25; v 4 0:27 0:39; v5 1:38

1:93

are superimposed on the data as illustrated in Fig. 6. They are located at the regions of the highest density of data and are represen250 200 150 100 50 0

while the bias values associated with the corresponding nodes of the map are given as 0.98, 1.00, 0.70, 0.54, 1.00. The slopes of the nonlinear functions are equal to 1.83, 3.51, 1.86, 4.17, and 3.74, respectively. They are clearly indicative of the diversied sensitivity of the individual nodes where the fourth node having the highest value of the sensitivity. The lowest sensitivity comes with the rst node. If we consider only the most signicant connections, say all of those whose absolute values exceed the threshold of 0.30, the graphic visualization of the map is shown in Fig. 7. Notably the third and the fth node exhibit the highest level of connectivity by being affected by a signicant number of other nodes in the map. Sunspot data. This data set comes from http://robjhyndman.com/TSDL/ and is concerned with the number of sunspots recorded in consecutive years. The data are displayed in the same manner as in the previous experiment viz. through successive values of xk, Dxk and the pairs (xk, Dxk), refer to Fig. 8. The FCM was run for several values of c ranging from c = 39. The results of the PSO optimization reported in terms of the values of the performance index are shown in Fig. 9.
100 80 60 40 20 0 -20 -40 -60 -80
0 500 1000 1500 2000 2500 3000

500

1000

1500

2000 100 80 60 40 20 0

2500

3000

-30

-20 -40 -60 -80

20

70

120

170

220

270

Fig. 8. Plot of sunspot data: (a) sequence of samples xk, (b) sequences of differences Dxk, and (c) x Dxk plot.

W. Pedrycz / Expert Systems with Applications 37 (2010) 72887294

7293

We look at the details for case of 6 clusters (nodes). Here the prototypes are given as follows, see also Fig. 10:
(H, -S))

(M, M)

v 1 109:03 v 3 55:72 v 5 115:66

3:51; 26:16; 29:14;

v2 52:89 v4 45:11 v6 9:81

12:89; 3:18; 1:23:


(-M, -L)

Their semantics associated with the nodes is concerned with the quantication of the dynamics of the time series. More specically, the meaning of the nodes one can consider reads in the following manner

(S, -S)

H SM MM LM SL LS S:
The results are also provided in terms of the detailed parameters of the nodes of the map: connections
(L, L) (M,-S)

w1 0:733 0:088 0:246 0:041 0:663

0:241;

Fig. 11. Main links in the fuzzy cognitive map; visualized are the connections with the absolute values higher than 0.4.

w2 0:104 0:328 0:461 0:395 0:032 0:145; w3 0:554 0:518 0:306 0:358 0:661 0:090; w4 0:614 0:782 0:516 0:890 0:145 0:224; w5 0:824 0:577 1:000 0:680 1:000 w6 0:748
bias vector

1:000; 0:524 0:703:

0:162

0:247

0:343

w0 0:625
slope vector

0:503

0:783

0:813

1:000

0:210

The node representing the concept of the lowest amplitude and the lowest dynamics (L, L) is the most connected with other nodes (concepts) however itself is not sustainable as the local feedback is strong and highly suppressive (the connection assumes the value of 1). The other nodes are less connected yet most of them have a positive internal feedback look. The exceptions are nodes (M, M) and (M, L). The node (S, S) has a sustainable internal feedback however interaction with the two other nodes of the map is inhibitory. Furthermore we note that there is some clear difference in the sensitivity of the nodes with the two of them characterized by high values of r (over 6) and the fourth one with r of 2.6. 6. Conclusions Being cognizant of the main conceptual pursuits developments of cognitive maps and a lions share of interest being put into their analysis, this study has raised a need for their design aspects which are equally important especially in cases when manual development is not necessarily a viable alternative. We have raised and addressed two key design phases that is the formation of information granules forming the nodes (concepts) of the map and the parametric buildup of the map coming with the estimation of the connections between the nodes. We stressed the hybrid nature of the development framework of fuzzy cognitive maps which exploits the settings of Granular and Evolutionary Computing. In particular, we demonstrated how information granules are formed on a basis of available numeric evidence and exemplied the design process in the context of time series analysis. While the study raises the point of importance of the solid design practices of cognitive maps and underlines the need for the use of automated tools, there are a number of unexplored topics which deserve further attention. First, structural optimization of the maps still remains an open issue. The design presented here was predominantly aimed at the parametric optimization with some activities along the line of the formation of information granules. Yet the problem of the construction of concepts in a general way is still open. Some innovative ways of taking advantage of lattices of concepts might be sought as a viable alternative yet this aspect still remains to be thoroughly explored. Second, Granular Computing brings a wealth of concepts and techniques of hierarchical information granulation which could effectively support the design of hierarchical cognitive maps. References
Alizadeh, S., & Ghazanfari, M. (2009). Learning FCM by chaotic simulated annealing. Chaos, Solitons and Fractals, 41(3), 11821190. Axelrod, R. (1976). Structure of decision: The cognitive maps of political elites. Princeton, NJ: Princeton University Press.

r 4:972 6:150 4:828 2:573 6:150 4:076:


Admitting a threshold of 0.50 (treated here as an example), the plot of the map showing the connections exceeding the threshold is given in Fig. 11.
0.06 0.05 0.04 0.03 0.02 0.01 0 1 2 3 4 5 6 7

Fig. 9. Values of performance index for the number of nodes ranging from 3 to 9; black bars xed slope (of the value of 1.0) and grey bars optimized values of the slopes.

100 80 60 40 20 0 -30 -20 -40 -60 -80


Fig. 10. The distribution of prototypes of the clusters vis--vis original numeric data.

xk xk
20 70 120 170 220 270

7294

W. Pedrycz / Expert Systems with Applications 37 (2010) 72887294 Kosko, B. (1986). Fuzzy cognitive maps. International Journal of ManMachine Studies, 24, 6575. Papageorgiou, E. I., & Groumpos, P. P. (2005). A new hybrid method using evolutionary algorithms to train fuzzy cognitive maps. Applied Soft Computing, 5(4), 409431. Papageorgiou, E., Stylios, C., & Groumpos, P. (2006). Unsupervised learning techniques for ne-tuning fuzzy cognitive map causal links. International Journal of HumanComputer Studies, 64, 727743. Paterlini, S., & Krink, T. (2006). Differential evolution and particle swarm optimisation in partitional clustering. Computational Statistics and Data Analysis, 50(5), 12201247. Salmeron, J. L. (2009). Augmented fuzzy cognitive maps for modelling LMS critical success factors. Knowledge-Based Systems, 22(4), 275278. Stach, W., Kurgan, L., Pedrycz, W., & Reformat, M. (2005). Genetic learning of fuzzy cognitive maps. Fuzzy Sets Systems, 153, 371401. Styblinski, M. A., & Meyer, B. D. (1991). Signal ow graphs versus fuzzy cognitive maps in application to qualitative circuit analysis. International Journal of Man Machine Studies, 35, 175186. Stylios, C. D., & Groumpos, P. P. (1999). Fuzzy cognitive maps: A model for intelligent supervisory control systems. Computers in Industry C, 39(3), 229238. van den Bergh, F., & Engelbrecht, A. P. (2006). A study of particle swarm optimization particle trajectories. Information Sciences, 176(8), 937971.

Bargiela, A., & Pedrycz, W. (2003). Granular computing: An introduction. Dordrecht: Kluwer Academic Publishers. Bezdek, J. C. (1981). Pattern recognition with fuzzy objective function algorithms. NY: Plenum Press. Chaib-Draa, B., & Desharnais, J. (1998). A relational model of cognitive maps. International Journal of HumanComputer Studies, 49, 181200. Chatterjee, A., & Siarry, P. (2006). Nonlinear inertia weight variation for dynamic adaptation in particle swarm optimization. Computers and Operations Research, 33(3), 859871. Coelho, J. P., de Moura Oliveira, P. B., & Boaventura Cunha, J. (2005). Greenhouse air temperature predictive control using the particle swarm optimisation algorithm. Computers and Electronics in Agriculture, 49(3), 330344. Ghazanfari, M., Alizadeh, S., Fathian, M., & Koulouriotis, D. E. (2007). Comparing simulated annealing and genetic algorithm in learning FCM. Applied Mathematics and Computation, 192(1), 5668. Kelly, G. A. (1955). The psychology of personal constructs. Norton: New York. Kennedy, J., & Eberhart, R. C. (1995). Particle swarm optimization. Proceedings of IEEE international conference on neural networks, (Vol. 4, pp. 19421948). NJ: IEEE Press. Kim, M. C., Kim, C. O., Hong, S. R., & Kwon, I. H. (2008). Forwardbackward analysis of RFID-enabled supply chain using fuzzy cognitive map and genetic algorithm. Expert Systems with Applications, 35(3), 11661176. Konar, A., & Chakraborty, U. K. (2005). Reasoning and unsupervised learning in a fuzzy cognitive map. Information Sciences, 170(24), 419441.

You might also like