An agent-based cognitive mapping system for sales opportunity analysis

An agent-based cognitive mapping system for sales opportunity analysis

Expert Systems with Applications 38 (2011) 7016–7028 Contents lists available at ScienceDirect Expert Systems with Applications journal homepage: ww...

1MB Sizes 0 Downloads 26 Views

Expert Systems with Applications 38 (2011) 7016–7028

Contents lists available at ScienceDirect

Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa

An agent-based cognitive mapping system for sales opportunity analysis Namho Lee a, Jae Kwon Bae b,⇑, Chulmo Koo c a

Consulting Group, SAP Korea, Dogok 2-dong, Kangnam, Seoul 135-700, Republic of Korea Department of Railroad Management Information, Dongyang University, #1 Gyochon-dong, Punggi, Yeongju, Gyeongbuk 750-711, Republic of Korea c College of Business, Chosun University, #375 Seosuk-dong, Dong-gu, Gwangju 501-759, Republic of Korea b

a r t i c l e

i n f o

Keywords: Cognitive map (CM) Multi-agent system (MAS) Particle swarm optimization (PSO) Agent-based cognitive mapping system (ABCMS)

a b s t r a c t The cognitive map (CM) is a representation of the causal relationships existing among the decision elements of a given object and/or problem, and also describes experts’ tacit knowledge. The CM has proven particularly useful in solving unstructured problems with many variables and causal relationships. Taking into consideration the amount of CM application studies thus far conducted in various fields, there has been relatively little research focused on the process of developing a CM. There have been some studies concerning the CM design process, most notably those conducted by Nelson, Nadkarni, Narayanan, and Ghods (2000) and Annibal, Tatiana, Susan, Julie, and Arthur (2006); however, these have failed to come up with a systematic approach in terms of the essential CM elements, which include: (1) the number of nodes, (2) the cause-and-effect relationships (arrows) among nodes, and (3) the strength of the cause-and-effect relationships (causality coefficients). The principal objective of this study, then, was to, first, determine the number of nodes that constitute a CM; second, to extract the cause-and-effect relationships among the nodes; and third, to devise an objective and systematic approach by which the causality coefficient within a single framework can be obtained. To accomplish this objective, our study adopts a CM-based mechanism referred to as the agent-based cognitive mapping system (ABCMS). Moreover, in order to extract effectively the three key elements of a CM, this study introduces the concepts of the multi-agent system (MAS) and particle swarm optimization (PSO), which allow for the construction of an ABCMS with a flexible and dynamic mechanism. Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction The cognitive map (CM) is a representation of the causal relationships existing among the decision elements of a given object and/or problem, and also describes experts’ tacit knowledge. The CM has proven to be of particular use in attempts to solve unstructured problems with many variables and causal relationships. Examples include information system requirements analysis (Montazemi & Conrath, 1986), distributed decision process modeling on networks (Zhang, Wang, & King, 1994), geographical information systems (Liu & Satur, 1999), electronic commerce web site design (Lee & Lee, 2003), knowledge management (Noh, Lee, Kim, Lee, & Kim, 2000), decision analysis (Zhang, Chen, & Bezdek, 1989), business process redesign (Kwahk & Kim, 1999), complex war games (Klein & Cooper, 1982), and strategic planning problems (Ramaprasad & Poon, 1985). Although CM has been applied in various fields of the social and natural sciences (Lee, Han, Song, & Lee, 1998; Liu & Satur, 1999; Styblinski & Meyer, 1991), the majority of these applications have involved using the CM to solve

⇑ Corresponding author. Tel.: +82 54 630 1298; fax: +82 54 630 1275. E-mail address: [email protected] (J.K. Bae). 0957-4174/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2010.12.013

specific problems and evaluate cause-and-effect relationships among elements of a problem, using the CM in the decision-making process of nonstructural problems via the CM inference function (Kardaras & Karakostas, 1999), or applying the CM as a single mechanism of artificial intelligence (Miao, Liu, Siew, & Miao, 2001). Taking into consideration the number of CM application studies that have been conducted in a variety of fields, it appears that relatively little research has been focused specifically on the process by which a CM is constructed. There have been a few studies concerning the CM design process, including those conducted by Nelson et al. (2000) and Annibal et al. (2006); however, these studies have failed to come up with a systematic approach to CM design in terms of the essential elements of the CM, which are: (1) the number of nodes, (2) the cause-and-effect relationships (arrows) among nodes, and (3) the strength of the cause-and-effect relationships (causality coefficients). In order not only to assess simple cause-and-effect relationships among the nodes but also to generate a CM that can utilize inference, which is one of the primary benefits of the CM, a more detailed, precise and objective technique for determining the three aforementioned essential elements of CM design is required. Thus, the principal objective of this study is to, first, determine the number of nodes that constitute a CM; second, to determine

N. Lee et al. / Expert Systems with Applications 38 (2011) 7016–7028

the cause-and-effect relationships among the nodes; and third, to provide a more objective and systematic approach to obtaining the causality coefficient within a single framework. In order to extract effectively the three key elements of a CM, flexible and organic coordination among the elements is of critical importance. However, the three elements of a traditional CM are fixed, which makes it difficult to modify them dynamically according to the relevant circumstances, and thus limits the capacity of the CM to solve real world problems. Based on the findings of a previous study that addressed the concept of an agent to a conventional CM (Miao, Goh, Miao, & Yang, 2001), this study proposed the concept of a multi-agent system (MAS), allowing for the construction of a CM with a flexible and dynamic mechanism, and thus overcome the aforementioned disadvantages. Viewing each node of a CM based on the MAS concept from the perspective of an intelligent agent, the three key CM elements are as follows: (1) the number of agents, (2) relationships among agents, and (3) strength of relationship. In other words, from the perspective of the MAS, the process of CM design involves determining the most appropriate number and relationships of the agents for the resolution of a particular problem from a number of possible relationships among agents; this can be regarded as a variety of emergent behavior. To accomplish this research objective, our study adopts a CMbased mechanism referred to as the agent-based cognitive mapping system (ABCMS). In order to extract effectively the three key elements of an ABCMS, this study applied the concepts of MAS and particle swarm optimization (PSO) to construct an ABCMS with a flexible and dynamic mechanism. Furthermore, the application was developed using the NetLogo simulation tool in order to implement the proposed framework and mechanism, and apply them to real world problems. The efficacy of the developed application was evaluated by analyzing the sales opportunities of the consultation business unit of a multinational software company. The remainder of this paper is structured as follows. Section 2 presents a research background review focused on the cognitive map and particle swarm optimization concepts. Section 3 describes the proposed ABCMS using an illustrative example, and elucidates its architecture and workflow. Some experimental results are presented and analyzed in Section 4, and finally our concluding remarks are provided in Section 5.

2. Literature review 2.1. Cognitive map (CM) The CM, first introduced by Axelrod (1976), has been utilized to represent knowledge in the political and social sciences, describing the cause-effect relationships that are perceived to exist among the elements of a given environment. The principal concern of a CM is to assess whether the state of one element can be perceived to have an influence on the state of the other. As one example, it might be proposed that if a city’s sanitation facilities were improved, the incidence of disease would decrease. However, better sanitation facilities, which would initially precipitate a decrease in disease incidence, might then lead indirectly to an increase in the city’s population. This growth might consecutively trigger an up shift in garbage, more bacteria, and increased disease as a result of the positive chain of influence between related elements, thereby ultimately nullifying or overshadowing the initial advantage gained by the effected improvements in sanitation (Montazemi & Conrath, 1986). From this example, from the standpoint of interpreting the operation of a CM, it can be readily grasped that positive causal links (denoted as ‘+’ in CM) should be regarded as excitatory relationships, whereas negative causal links (denoted

7017

as ‘’ in CM) can be viewed as inhibitory relationships between elements. How can a CM with the above characteristics be used in our decision making process? To answer this question, we will address three possible uses of a CM in decision making. As for the first, irrelevant data can be defined as those data pertaining to factors that have no effect on the target decision outcome. By analyzing the connectivity between the factors depicted in the CM, one can easily determine whether or not data regarding a particular factor are relevant to a given decision. To illustrate this, consider Fig. 1, wherein causal links between six factors are depicted. Suppose that one is concerned with the state of F, because it will affect a target decision to be made. Fig. 1 shows us that both A and D directly influence F, and C has an indirect influence on F through D. Conversely, neither B nor E influences F indirectly or directly. Those two factors (B and E) are therefore ‘‘irrelevant’’ with regard to outcome F. Determinations such as this are frequently difficult to make by direct inquiry or observation, but can be made much clearer via the appropriate use of a CM. As for the second kind of CM usage, one may pose the question, ‘‘Does CM-based causal knowledge as a whole work as well as we expect?’’ To answer this, let us consider a case in which one finds the value of F to be relatively high, while all the values of A and D are relatively low. According to the CM, low A and low D values should result in a low F value. Thus, the results are counter to the prediction of the CM. This may occur as the result of other as-yet-unknown factors affecting F in this particular situation, possibly factors that were identified but not originally regarded as relevant. In such a case, efforts should be re-focused on identifying those factors relevant to F, thereby preventing the performance of CM-based knowledge against abrupt degradation under turbulent decision making conditions. The third type of CM usage appears to hold great promise in the field of expert systems. The CM must be sufficiently improved to deal with the uncertainty and vagueness frequently associated with information in decision environments, allowing it to be used as a knowledge engineering tool for extracting causal knowledge from environments (Kosko, 1986; Taber, 1994). A schema map can be organized into a matrix capable not only of evolving over time but also of generating outputs (or consequence vectors) for some specific inputs (or stimulus vectors), allowing for ‘‘what-if’’ analyses. A CM, then, can be developed in accordance with the following three steps: Step 1: The first important step is to clarify the purpose for which the CM is being built. If the purpose is ill-defined, the search for relevant factors is likely to lack direction, and the CM might rapidly grow to an unmanageable size.

Fig. 1. An example of a CM.

7018

N. Lee et al. / Expert Systems with Applications 38 (2011) 7016–7028

Step 2: The next step is to identify the relevant factors – those which may influence a decision. Step 3: The final step is to find causal relationships among the factors identified in Step 2. These can be determined via several methods: they can be adapted from a decision maker’s statements (Axelrod, 1976), derived from questionnaires prepared specifically for this purpose, or generated via neural networkbased learning. 2.2. CM research As shown in Fig. 2, CM studies are classified into two categories: those in the social sciences, and those applied to the natural sciences and engineering. Studies in both categories can also be classified into those dealing with methodology, and application studies in which problems are solved by the application of the CM under various circumstances. The CM has been actively utilized in both the social and natural sciences. As shown in Table 1, examples of social science CM applications include political issues (Hart, 1977; Robert, 1976), marketing (Warren, 1995), organizational issues (Kwahk & Kim, 1999; Wright, 2004), knowledge management (Boegl, Adlassnig, Hayashi, Rothenfluh, & Leitich, 2004), and strategic planning (Lee et al., 1998). As shown in Table 2, natural science and engineering CM applications include e-business (Lee, Kim, Chung, & Kwon, 2002; Lee & Kwon, 2006; Lee & Lee, 2003), technical engineering (Ndousse & Okuda, 1996; Styblinski & Meyer, 1991), geographic information system (Liu & Satur, 1999; Satur & Liu, 1999), and intelligent agent (Brahim, 2002; Miao, Goh, Miao, & Yang, 2001; Rai & Kim, 2002) applications. The breadth of these examples illustrates the utility of the CM in diverse fields of application. Most social science studies dealing with methodology tend to address cognitive mapping itself, such as Self-Q (Bougon, 1983), repertory grip techniques (Reger, 1990), mental imagery (Zmud, Anthony, & Stair, 1993), and revealed causal mapping (Nelson et al., 2000) research. These studies have all generally entailed the use of CM within a psychological framework, to express implicit knowledge possessed by human beings. However, these studies were, from a technical point of view, somewhat limited in terms of utilizing the devised CM. On the other hand, methodology studies conducted in the natural sciences include CM structures based on mathematical approaches (Schneider, Shnaider, Kandel, & Chew,

1998; Silva, 1995), and extensions of CM mechanisms (Chunyan, Angela, Yuan, & Zhonghua, 2001; Hagiwara, 1992; Pal & Konar, 1996; Park & Kim, 1995; Satur & Liu, 1999). Since the early 1990s, nearly constant efforts have been exerted to identify and improve the limitations of the traditional CM mechanism (Hagiwara, 1992; Park & Kim, 1995; Schneider et al., 1998). Typically, the limitations of traditional CM include: (1) fixed relation among nodes, (2) linearity of relations, and (3) the lack of a time concept. Thanks to assiduous research activities undertaken to overcome these disadvantages, the intelligent agent concept has emerged (Chunyan et al., 2001; Miao, Goh, Miao, & Yang, 2001). Although a great deal of CM research in the natural science and engineering fields has been focused on the CM inference mechanism, there has been a relative dearth of research focused on the process of obtaining or developing a CM. Upon review of the CM studies thus far conducted in the social and natural sciences, it can be fairly confidently asserted that no study has yet systematically dealt with the decision making process of the three key elements of a CM: nodes, relations among nodes, and relation strength. 2.3. Particle swarm optimization (PSO) The PSO concept is a swarm intelligence method that differs from well-known evolutionary computation algorithms, such as the genetic algorithms (GAs), in that the population is not manipulated through operators inspired by human DNA procedures. Rather, in PSO, the population dynamics simulate the behavior of a ‘‘flock of birds’’, wherein social sharing of information occurs and individuals profit from the discoveries and previous experience of all other companions during the ‘‘search for food’’. Thus, each member referred to as a ‘‘particle’’ of the population (which is called the ‘‘swarm’’), is assumed to ‘‘fly’’ over the search space, seeking promising regions on the landscape. For instance, in the single-objective minimization case, regions previously visited by other particles in the swarm are ascribed lower function values. In this context, each particle is treated as a point into the search space, which adjusts its own ‘‘flight’’ according to its flying experience, as well as the flight experiences of the other particles. First, let us define the notation adopted in this paper: assuming that the search space is D-dimensional, the ith particle of the swarm is represented by the D-dimensional vector Xi = (xi1, xi2, . . ., xiD) and the best particle in the swarm, i.e. the particle with the

Fig. 2. Taxonomy of CM studies.

7019

N. Lee et al. / Expert Systems with Applications 38 (2011) 7016–7028 Table 1 CM research in social science. Category 1

Category 2

Contents

Researcher

Methodology

Mapping technique

Self-Q Repertory grid techniques Mental imagery Revealed causal mapping

Bougon (1983) Reger (1990) Zmud et al. (1993) Nelson et al. (2000)

Application

Political issues

Influence diagram Examining the beliefs of foreign policy elites Marketing strategy Executive perceptions and decision making Organizational conflict Constructs of how real respondents perceive appraisal systems Method to elicit tacit knowledge Strategic planning simulation Information system planning

Ross and Hall (1980) Hart (1977), Robert (1976) Warren (1995) Clarke and Mackaness (2001) Kwahk and Kim (1999) Wright (2004) Boegl et al. (2004) Lee et al. (1998) Kwahk and Kim (1999)

Marketing Organizational issues Knowledge management Strategic planning

Table 2 CM research in engineering science. Category 1

Category 2

Contents

Researcher

Methodology

Construction

Method for automatically constructing fuzzy cognitive map Combined matrices of FCM Extended fuzzy cognitive maps Fuzzy time cognitive maps Cognitive reasoning using fuzzy neural nets Contextual fuzzy cognitive map Agent inference model

Silva (1995) Schneider et al. (1998) Hagiwara (1992) Park and Kim (1995) Pal and Konar (1996) Satur and Liu (1999) Chunyan et al. (2001)

Web-mining Knowledge based news miner Web design EDI control B2B negotiation Electronic circuit analysis Network management Contextual fuzzy cognitive map for decision support in GIS Contextual fuzzy cognitive map framework for GIS Dynamic inference model for intelligent agent CM-review tool used to test multi agent environment Control agent by CM

Lee et al. (2002) Hong and Han (2002) Lee and Lee (2003) Lee and Lee (2007) Lee and Kwon (2006) Styblinski and Meyer (1991) Ndousse and Okuda (1996) Liu and Satur (1999) Satur and Liu (1999) Miao, Goh, Miao, and Yang (2001) Brahim (2002) Rai and Kim (2002)

Extension

Application

e-Business

Technical Engineering GIS Intelligent Agent

smallest function value, is denoted by the index g. The best previous position (i.e. the position giving the best function value) of the ith particle is recorded and represented as Pi = (pi1, pi2, . . ., piD), while the position change (velocity) of the ith particle is represented as Vi = (vi1, vi2, . . ., viD). Following this notation, the particles are manipulated according to the following Eqs. (1) and (2):

v id ¼ wv id þ c1 s1 ðpid  xid Þ þ c2 s2 ðpgd  xid Þ xid ¼ xid þ vv id

ð1Þ ð2Þ

where d = 1, 2, . . ., D; N is the size of the population; i = 1, 2, . . ., N; v is a constriction factor that controls and constricts the magnitude of the velocity; w is the inertia weight; c1 and c2 are two positive constants; s1 and s2 are two random numbers within the range [0, 1]. Eq. (1) determines the ith particle’s new velocity as a function of three terms: the particle’s previous velocity; the distance between the best previous position of the particle and its current position, and finally; the distance between the swarm’s best experience (the position of the best particle in the swarm) and the 4th particle’s current position. Then, according to Eq. (2), the ith particle ‘‘flies’’ towards a new position. In general, the performance of each particle is measured in accordance with a fitness function, which is problem-dependent. In optimization problems, the fitness function is usually the objective function under consideration. The role of the inertia weight, w, is considered to be crucial for the PSO’s convergence. The inertia weight is employed to control the impact of the history of previous velocities on the current velocity of each particle. Thus, the parameter w regulates the trade-off between the glo-

bal and local exploration abilities of the swarm. A large inertia weight facilitates global exploration (searching new areas), whereas a small one tends to facilitate local exploration, i.e. fine-tuning the current search area. A suitable value for the inertia weight w balances the global and local exploration ability, and consequently reduces the number of iterations required to locate the optimal solution. A general rule of thumb suggests that it is better to initially set the inertia to a large value, in order to facilitate a global exploration of the search space, and then gradually decrease it to acquire more refined solutions. Thus, a time-decreasing inertia weight value is used. The initial swarm can be generated either randomly or using a sobol sequence generator (Press, Vetterling, Teukolsky, & Flannery, 1992), which ensures that the particles will be distributed uniformly throughout the search space. From the above discussion, it is obvious that PSO resembles, to some extent, the ‘‘mutation’’ operator of GA. Note, however, that in PSO the ‘‘mutation’’ operator is guided by the particle’s own ‘‘flight’’ experience and also benefits from the swarm’s flight experience. In other words, the PSO can be considered a performing mutation with a ‘‘conscience’’, as described by Eberhart and Shi (2000). 3. The agent-based cognitive mapping system (ABCMS) 3.1. Background Traditionally, many subjective factors were reflected in the design of a CM. The general process of drawing a CM as in the exam-

7020

N. Lee et al. / Expert Systems with Applications 38 (2011) 7016–7028

ple displayed on the right in Fig. 3 involves individuals or experts participating in a discussion to (1) extract nodes related to problem solving, (2) display cause-and-effect relationships with arrows, and (3) indicate the strength of cause-and-effect relations. For this study, however, in order to minimize the subjective factors in the CM design process, insufficient cognitive information was not propagated based on subjective judgment. Rather, basic information regarding the nodes and relations was left intact within the realm of the obtained information, and the insufficient CM was trained using actual data to complete the CM, as shown in Fig. 3. If the process of the CM completing the node relations on its own based on self-learning and insufficient information is analyzed with an eye toward the multi-agent concept, it becomes an intelligent agent with live nodes, and the relations among the agents become the causal relations expressed with arrows in a traditional CM as shown in Fig. 4. Furthermore, the condition of one agent having a cause-and-effect relationship with another agent can be expressed as an event. There have been earlier studies that attempted to introduce the concept of an intelligent agent in a CM in order to render it more flexible and dynamic (Chunyan et al., 2001). However, these studies focused on the CM inference methodology and developed CM to offer more flexible and dynamic inference functions by incorporating the agent concept into the CM on the basis of agent oriented modeling (AOM) methodology (Yim, Cho, Kim, & Park, 2000), but failed to improve the CM design process. The ABCMS proposed in this study not only employed the agent concept explained earlier for CM inference, but also maximized the advantages of the MAS by applying the swarm simulation technique with CM’s learning mechanism to derive the optimal CM from the obtained information. The MAS can use the swarm simulation concept to induce emergent behavior from the interactions of multi-agents (Rouff, Vanderbilt, Hinchey, Truszkowski, & Rash, 2004). In swarm simulations, each agent interacts with other agents to satisfy its own goals, but multiple agents interact based on simple action strategies, leading to logical and stable behavioral outcomes referred to as emergent behaviors (Bonabeau, Dorigo, & Theraulaz, 1999), which convey meaningful information to decision makers. 3.2. PSO in ABCMS The determination of causal relationships and causality coefficients are of major significance in CM, because the ABCMS encompasses tacit knowledge through supervised learning. For example, decision makers may want some factors to lie within certain prespecified boundaries. The ABCMS learns these boundaries through PSO.

In our case, the PSO is constructed to find the causal relationships and causality coefficients that will minimize the error function. Let us say that Ci denotes ith concept node of the ABCMS, and Ai represents the actual value of the ith concept node Ci, 1 6 i 6 N. The ABCMS restricts the values of the concept nodes in bounds such as Amin  Ai  Amax ; i ¼ 1; . . . ; N, where the upper i i max bound Ai and lower bound Amin are pre-determined by the infori mation system (IS) project experts. In the ABCMS, a concept node, Ci, is represented by an agent, and thus a complete CM can be regarded as a set of multi-agents that interrelate dynamically with one another, affecting other agents through changing causality coefficients over time. Thus, the principal objective of the ABCMS is to find a causality coefficient matrix, E = [Eij], i, j = 1, . . ., N, which leads CM to a steady state at which the output concept nodes lie in their pre-specified boundaries, while the causality coefficients retain their physical meaning. The ABCMS finds the causality coefficient matrix E = [Eij], i, j = 1, . . ., N by imposing constraints given by past experience in IS Pm 2 b projects. To do this, the objective function i¼1 ðAouti  A outi Þ is considered, provided that the number of output concept nodes m, Aout ¼ ½Aouti ; i ¼ 1; . . . ; m represent the actual value vector of b out ¼ ½ A b out ; i ¼ 1; . . . ; m the comthe output concept nodes, and A i ^ out is obtained puted value vector of the output concept nodes. A by multiplying the output concept node vector Cout with E, and organizing the corresponding computed values into a vector consisting of m output concept nodes, where vector   0; i–j C out ¼ , i = 1, . . ., N and j = 1, . . ., m. Aouti is obtained C outi ; i ¼ j via the application of numerous causality coefficients [Eij] until Pm 2 b the value of the objective function i¼1 ðAouti  A outi Þ is reduced in comparison with the previous one. Eij is the causality coefficient of the causal relationship connecting concept node Ci to concept node Cj, and is randomized following a uniform distribution [1, 1] when there are no constraints on it. In the swarm optimization, therefore, the causality coefficient matrix E is computed by minimizing the objective function mentioned earlier under the constraints on the node value such as Amin  Ai  Amax ; i ¼ 1; . . . ; N i i and/or the constraints on the causality coefficient such as 1 6 Eij 6 0.5 or 0.5 6 Eij 6 1. Therefore, the PSO adopted in the ABCMS can be summarized as follows: Stage 1: Search local optimal. Step 1a: Select random vector Einitial in Rn space. Step 1b: Search Ek that minimize objective function [1, 1] near Einital vector. Step 1c: If Aouti ðk þ 1Þ  Aouti ðkÞ Then, step Stage 1, otherwise repeat 1b.

Fig. 3. Traditional CM vs. ABCMS.

N. Lee et al. / Expert Systems with Applications 38 (2011) 7016–7028

7021

Fig. 4. Constructing CM by ABCMS.

Stage 2: Evolution Step 2a: Select Einitial that satisfy Aouti ðk þ 1Þ  Aouti ðkÞ in Rn space. Step 2b: Repeat Step 1b. Step 2c: Repeat Step 1c. Termination rule: After N number of evolution, If Aouti ðk þ 1Þ  Aouti ðkÞ  d. Then stop. 3.3. Architecture of the ABCMS As explained earlier, the ABCMS proposed in this study is based heavily on MAS and PSO. Besides, its primary purpose is to induce a final CM based on insufficient node information and to allow the final CM to resolve problems. The multi-agent environment adopted in the ABCMS is based on the NetLogo environment. As depicted in Fig. 5, the ABCMS consists principally of three databases, two engines, and the user interface. Database 1 (DB1) contains the basic information about the CM nodes. Database 2 (DB2) stores the information about the final CM, and Database 3 (DB3) harbors the case data of the nodes for training purposes. Using the CM information stored in DB2, the inference engine makes inferences based on the input node values to provide the inference processes and results to the user, via the user interface. The swarm optimization

S Swarm Simulation Si l ti E Engine i

DB1

DB2

DB3 3

Node Pool & Draft Relation

g Knowledge Base (Complete CM)

T i i Training Case Data

Inference Engine

User Interface

Fig. 5. Architecture of the ABCMS.

engine updates the CM in DB2 using the basic information stored in DB1 as well as the training data in DB2 to complete the CM. From the entity and service aspects, an ABCMS consists primarily of two entities and three services as shown in Fig. 6. Entities include node agents with node information as their attributes, and events that involve a status change in a particular node affecting another node as a result of the cause-and-effect relationships among node agents. Services include the inference engine responsible for performing CM inference, the swarm optimization engine that trains the CM and finds CM that satisfy the given conditions, and the user interface that provides the user with information.

3.4. Workflow of the ABCMS The process of constructing a CM using the ABCMS is displayed in Fig. 5. The basic information about the nodes that constitute the CM must first be prepared. Basic information refers to the fundamental information regarding the cause-and-effect relationship among the nodes of the problem. Rather than complete information about the nodes that constitute the CM of the corresponding problem and the cause-and-effect relationships among the nodes, the basic information signifies restricted information about the nodes held by the user who is attempting to design the CM. In other words, basic information provides a general idea about the factors (nodes) expected in resolving the problem, as well as information regarding the relations among the nodes. Such basic information is concerned with cause-and-effect relationships between nodes ranging from 1.0 to +1.0; however, since the users do not possess exact information about the relationships, the PSO of the ABCMS supports eight kinds of linguistic alternatives: (a) Strong Positive: a relation with a potential strength of +0.5 or greater; (b) Strong Negative: a relation with a potential strength of 0.5 or less; (c) Weak Positive: a relation with a potential positive strength of +0.5 or less; (d) Weak Negative: a relation with a potential negative strength of 0.5 or greater; (e) Positive: a relation with a potential strength of 0 or greater; (f) Negative: a relation with a potential negative strength of 0 or less; (g) Some Relation: a relation with a possible cause-and-effect relationship, but difficult to judge between positive and negative; and (h) Unknown: unable to assess the node relationship. Fig. 7 shows a workflow of a CM constructed by the ABCMS. The nodes in the node pool are provided as agents (Task s), and these

7022

N. Lee et al. / Expert Systems with Applications 38 (2011) 7016–7028

Fig. 6. Entity and service of the ABCMS.

node agents recognize their draft-relation to establish the draftrelation information with other node agents according to the eight types described above. Tasks t through x are the CM training process using the PSO technique. Task t randomly configures a single draft-relation set within the known information scope of Rn space, which functions as the start point of the PSO process. The training data inference result values based on Tasks u and v are compared with the actual training data result values to obtain the error value. If the error does not decrease below the threshold value d, the PSO process stops. The user reviews the CM drawn by the PSO process in Task u, and if the CM recommended by the ABCMS is satisfactory, the entire process is complete. Otherwise, some nodes added to the node pool are removed, the draft-relation among specific nodes is modified as 1 , and the PSO process is again executed. in Task 1 3.5. CM main issue addressing by the ABCMS As mentioned in the introduction, despite a number of previous studies some questions about the CM remain to be adequately

addressed, namely: (1) how to determine the adequate number of nodes, (2) how to establish the cause-and-effect relationships among the nodes, and (3) how to determine the strengths of the cause-and-effect relationships. The objective of this study was to allow the ABCMS to solve these three problems within a single frame. The three decision making points are not sequentially processed through the ABCMS, but rather in parallel, as is described in Fig. 7. (1) Number of nodes determination When attempting to solve a particular problem using a CM, the nodes that constitute the CM are usually considered first. The CM is designed such that it contains no redundant nodes, and the node pool is created with a set of nodes to construct the CM on the basis of expert opinions. However, the process of selecting specific nodes from the node pool to construct the CM was different for each individual. To resolve such problems, the threshold v was established for the cause-and-effect relation between nodes, and only those nodes with one or more relations with a relation strength of v or greater in the PSO process as explained in Section 2.3 were included in the CM. Consequently, among the nodes

Fig. 7. Workflow of the ABCMS.

7023

N. Lee et al. / Expert Systems with Applications 38 (2011) 7016–7028

constructed only with the six draft-relation types (the eight draftrelation types described in Section 3.4 minus strong positive and strong negative), those with strengths less than the threshold value v in the PSO process are removed from the node pool for CM construction. Although uncommon, there are some cases in which a node is included in the node pool, but its relations with other nodes are unclear. This is regarded as one of the factors critical to solving a problem using the CM, but there is no information regarding relations with other factors. In such cases, the ABCMS seeks an adequate relation by searching the range of combinations

of all cause-and-effect relationships that the node can have in the PSO process. In other words, the ABCMS establishes the (h) Unknown direction status explained in Section 3.4 for all such nodes, with every node in the node pool except itself. If the PSO process reveals a relation greater than the threshold value v, the corresponding node is selected for inclusion in the CM. According to these procedures, the nodes that constitute the CM can be selected from the node pool. After reviewing the nodes recommended by the ABCMS, the end user can confirm the candidate 10 , nodes for the CM from the node pool as shown in Fig. 7. Task s

Table 3 Nodes and fuzzy conversion value. No.

Node name (factor)

Description

Node status (attributes)

Fuzzy value

1

Project risk (PR)

Overall project risk

2

Contract type (CT)

Contract type: time and material, fixed price

3

Difficulty (DI)

Difficulty of implementation

Very high High Medium Low Very low Subcontractor and time and material Prime contractor and time and material Subcontractor and fixed price Prime contractor and fixed Price Stable product and industry reference and experienced resource (Low) Industry reference and no experienced resource (Medium) New product and no reference and no experienced resource (High) Clear Medium Unclear Need to make reference

1.0 0.5 0.0 0.5 1.0 1.0 0.5 0.5 1.0 1.0

Market has much potential No special impact Available

0.5 0.0 1.0

Not available Strategic account First deal Normal Bad site (Bad customer) No issue expected Non-standard collection condition Historical credit issue A: more than 30% B: 20–30% C: 10–20% D: less than 10% A: more than 1M USD B: 500K  1M USD C: 100–500K USD D: less than 100K USD A: No issue B: Execution is possible but caution required C: Difficult execution D: Expected to be a very difficult project A: Strategically very important deal

1.0 1.0 0.8 0.0 1.0 1.0 0.5 1.0 1.0 0.7 0.3 1.0 1.0 0.7 0.0 0.5 1.0 0.5 0.5 1.0 1.0

B: Average deal C: Strategically unimportant deal A: Deal with good financial conditions B: Small revenue but no problems in terms of margin and collection C: Margin less than 20% or collection issues expected D: Not a good deal in terms of revenue, margin and collection Green

0.0 1.0 1.0 0.5

Yellow Amber Red

0.5 0.0 0.5

4

Scope (SC)

Clarity of project scope definition

5

Impact on next deal (IN)

Degree of impact the outcome of this deal will have on subsequent deals

6

Resource availability (RA)

Available personnel for presales activities and project implementation for this deal

7

Relationship (RL)

Relationship with the customer

8

Collection (CO)

Potential problems with collection

9

Margin (MA)

Expected margin

10

Revenue (RE)

Revenue scale of the deal

11

Delivery (DE)

Evaluation from the delivery of project execution

12

Strategic (ST)

Evaluation from strategic aspect rather than financial or project delivery perspective

13

14

Financial (FI)

Sales opportunity index (SOI)

Financial evaluation

Evaluation of overall sales opportunity

0.5 1.0 1.0 0.0 1.0 1.0

0.5 1.0 1.0

7024

N. Lee et al. / Expert Systems with Applications 38 (2011) 7016–7028

Fig. 8. DRI result for target problem.

Table 4 Notation of draft-relation. Notation

Description

Notation

Description

SP SN WP WN

Strong positive Strong negative Weak positive Weak negative

P N S U

Positive Negative Some Relation Unknown

or specific nodes can be added or removed prior to carrying out another process. (2) Relation determination The process of designing a CM frequently involves cases in which there is no certainty regarding the causal relations and/or experts offering different opinions. In such cases, the CM is usually completed on the basis of subjective judgment. In order to improve this point, this study proposes an approach involving the determi-

nation of an uncertain relation using a method similar to that used for the selection of the CM nodes from the node pool. For relations with an absolute strength value of greater than 0.5-(a) Strong Positive and (b) Strong Negative-the relation is certain. Therefore, among draft-relations with the remaining six types, only those with a value higher than the threshold v are selected via the PSO process. Cause-and-effect relationships and directions are, in turn, assigned to uncertain relations with types such as (g) Some Relation and (h) Unknown Direction. (3) Strength determination Finally, the relation strengths between the nodes have generally been displayed with a value between 1.0 and +1.0 when a CM was designed, and an average value was obtained on the basis of expert opinions. However, this approach also reflected many subjective factors, and it was not easy for someone dealing with a CM for the first time to express the strength using a value between 1.0 and +1.0. In an effort to address these problems, this study classified the initial draft-relations into eight values for providing the basic information. Then, a specific value between 1.0 and +1.0 was sought via the PSO process. For example, if there was no conviction regarding the causal relation between Node A and Node B, but some positive or negative relation was anticipated, ‘‘Some Relation’’ would be assigned to the draft-relation. Then, the ABCMS establishes the relation strength between Nodes A and B rAB as [0.1, +0.1], which is the section value between 0.1 and +1.0, and finds a particular value R within the range rAB via the PSO process, as explained in Section 3.2. 4. Experiments 4.1. Target problem The methodology proposed in this paper was applied to sales opportunity analysis for the consultation projects of a multinational corporate software company’s consulting business unit.

Table 5 Recommended relation (Iteration 1). From node

Relation strength

To node

Accept (Y/N)

[1] Project_Risk (PR)

1.0 0.5 0.5 0.5 0.5 1.0 0.5 0.5 0.5 0.5 0.5 1.0 1.0 1.0 1.0 0.5 0.5 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 0.5 0.5 1.0 1.0

[3] Difficulty (DI) [8] Collection (CO) [11] Delivery (DE)

Y Y Y N Y N N Y Y N N Y Y Y Y Y N N N Y Y Y Y Y Y Y Y Y Y Y

[2] Contract_Type (CT)

[3] Difficulty (DI)

[4] Scope (SC) [5] Impact_on_next deal (IN) [6] Resource_Availability (RA)

[7] Relationship (RL) [8] Collection (CO) [9] Margin (MA) [10] Revenue (RE) [11] Delivery (DE) [12] Strategic (ST) [13] Financial (FI)

[9] Margin (MA)

[13] Financial (FI) [1] Project_Risk (PR)

[6] Resource_Availability (RA) [11] Delivery (DE) [3] Difficulty (DI) [12] Strategic (ST) [14] Sales_Opportunity (SOI)

[10] Revenue (RE) [11] Delivery (DE) [3] Difficulty (DI) [12] Strategic (ST) [7] Relationship (RL) [13] Financial (FI) [13] Financial (FI) [13] Financial (FI) [14] Sales_Opportunity (SOI) [14] Sales_Opportunity (SOI) [14] Sales_Opportunity (SOI)

N. Lee et al. / Expert Systems with Applications 38 (2011) 7016–7028

7025

Fig. 9. The ABCMS application developed by NetLogo.

We gathered data from a multinational IT consulting firm located in Seoul, South Korea. This organization is involved in corporate software implementation consultation, including enterprise resource planning (ERP), customer relationship management (CRM), and package products. There are approximately 80 consultants, and additional work staff can be outsourced or recruited from overseas, depending on the circumstances. There can be a lack of a work force in a particular area at a certain time point.

However, as the target utilization for an individual consultant is 70%, new consultants cannot be hired at every instant at which an additional work force is required. In order to lead a potential sales opportunity to an actual contract, fierce competition occurs among the competitors, and additional people are required at the pre-sales stage. Although every sales representative claims that his account is important and requests priority in personnel assignment, a work force cannot be supported at every opportunity from

Fig. 10. The Final ABCMS.

7026

N. Lee et al. / Expert Systems with Applications 38 (2011) 7016–7028

the perspective of the organization managing the consultants. There is always the problem of selecting the most important opportunities for the allocation of additional staffing. This organization holds weekly meetings attended by a consulting project manager, four resource managers, and four sales managers. The weekly meetings help the organization to determine which sales opportunities should receive priority for additional staffing. Major decision criteria include financial revenue and margin, stability in project delivery, and the strategic aspects with regard to the market situation and the client. However, the process by which priority is determined at a meeting attended by people with differing interests is subject to debate, and there is a definite need for a more systematic method of assigning priority.

4.2. CM construction using the ABCMS Step 1: Prepare node pool The sales opportunity index (SOI) indicating the importance of a sales opportunity was established as the final node and 21 major factors related to SOI were determined by three experts, each with 10 or more years of experience. From the 21 nodes drawn, the experts conducted discussions to eliminate similar factors, ultimately leaving 14 in the final node pool as shown in Table 3. Step 2: Draw draft relationship In order to survey the draft-relations for this study, the experts extracted 7 pieces of relational information for the 14 nodes using the draft-relation inquiry (DRI) form shown in Fig. 8. The experts were instructed to indicate the eight draft-relation types mentioned earlier according to Table 4. From the results, the seven nodes on which all experts agreed were assigned as draft-relations, and the relations of the remaining nodes were tagged as ‘‘Unknown’’; ABCMS was to identify the relations in those cases. The draft-relation information surveyed with

the DRI form was converted into section fuzzy values according to the causal relation coding rules (See Table 5). Step 3: Preparing training case data As explained in Section 3.3, an ABCMS completes a CM from the draft-relations based on the actual training data. 100 actual sales opportunities were used for this study. Three consulting industry managers first evaluated the node statuses for the 100 sales opportunities according to the criteria in Table 3, and the fuzzy value of each node was determined by referencing the node fuzzy value. Among the 100, data from 50 were utilized for CM training and the remaining fifty were employed for CM testing. Step 4: CM training Based on the draft-relation information and the training case data prepared in Steps 2 and 3, the CM training is conducted using the ABCMS application developed under the NetLogo environment as shown in Fig. 9. For this experiment, the relation strength threshold v was set to 0.1. The ABCMS application trains the CM through the PSO process of Section 3.2. There were three iterations until the final CM was extracted as shown in Fig. 10 (Here, iteration 1 in Fig. 7). refers to a single execution of Task 1 After the first iteration, the ABCMS recommended 30 relations (See Table 5), eight relations that the experts deemed completely insignificant were removed, and a condition was set such that those eight relations did not generate relations in the next iteration. The second iteration resulted in the ABCMS recommending 22 relations (See Table 6). Three relations were again eliminated by the experts. The third iteration yielded the final CM shown in Fig. 10, after the experts agreed that there were no more relations to remove.

4.3. Evaluation In order to evaluate the effectiveness of an ABCMS in designing a CM, this study used inferences for test data made on the basis of a

Table 6 Recommended relation (Iteration 2). From node

Relation Strength

[1] Project_Risk (PR)

0.5 1.0 1.0 0.5 0.5 1.0 1.0 0.5 0.5 0.5 0.5 0.5 0.5 1.0 1.0 0.5 0.5 0.5 1.0 0.5 0.5 1.0

[2] Contract_Type (CT) [3] Difficulty (DI)

[4] Scope (SC) [5] Impact_on_next deal (IN) [6] Resource_Availability (RA) [7] Relationship (RL) [8] Collection (CO) [9] Margin (MA) [10] Revenue (RE) [11] Delivery (DE) [12] Strategic (ST) [13] Financial (FI)

To node

Accept (Y/N)

[8] Collection (CO) [11] Delivery (DE) [9] Margin (MA) [1] Project_Risk (PR) [6] Resource_Availability (RA) [11] Delivery (DE) [3] Difficulty (DI) [12] Strategic (ST) [14] Sales_Opportunity (SOI) [10] Revenue (RE) [11] Delivery (DE) [12] Strategic (ST) [7] Relationship (RL) [13] Financial (FI) [13] Financial (FI) [13] Financial (FI) [14] Sales_Opportunity (SOI) [14] Sales_Opportunity (SOI) [14] Sales_Opportunity (SOI)

N Y Y Y N Y Y Y Y Y Y Y Y N Y Y Y Y Y Y Y Y

Table 7 Adjusted fuzzy value for output node. No.

Node

Description

Node status

Old fuzzy value

New fuzzy value

14

Sales opportunity index (SOI)

Assessment for overall sales opportunity

Green Yellow Amber Red

1.0 0.5 0.0 0.5

0.6 6 X 0.4 6 X < 0.5 0.2 6 X < 0.4 X < 0.2

N. Lee et al. / Expert Systems with Applications 38 (2011) 7016–7028

7027

Fig. 11. Experiment results.

CM obtained from an ABCMS, and the results of expert evaluations were compared with those of the ABCMS inferences. Prior to conducting inferences for the 50 test data, calibration was required for the fuzzy conversion value of node 14 (See Table 7 and Fig. 11), which was the final node. This was because the fuzzy value of node 14 obtained from CM inference was bound to be different from the fuzzy conversion value prepared based on the evaluation guide for each node status prior to CM construction. Therefore, after completing training through the ABCMS, the trained results were reviewed and the inference results of node 14 were calibrated for conversion into Green, Yellow, Amber, and Red. When the inferences for 50 test data based on the CM devised by the ABCMS were evaluated, 45 were found to correspond with the experts’ evaluations. With a 90% degree of correspondence, the CM completed with ABCMS support proved significant for the determination of sales opportunities.

5. Conclusions CM has been studied in various fields of the social and natural sciences, and has proven to be an effective decision making tool, as well as a useful means of displaying knowledge. However, the process of obtaining a CM involves subjective judgment to determine (1) the number of nodes, (2) relations among the nodes, and (3) relation strengths. This presents a barrier to the CM becoming a more objective and systematic decision making tool. The ABCMS proposed in this study improved most of the above limitations. The ABCMS is composed of multiple agents; interactions among multi-agents in the ABCMS allow for more effective sales opportunity analyses. One advantage of this approach is that the proposed ABCMS can behave more intelligently using a PSO, whereas multi-agents working together in an iterate manner dis-

play emergent behaviors derived from the complicated interactions among multiple agents over time. The emergent behaviors deduced by the ABCMS reveal some important implications for decision makers who are trying to handle sales opportunity analysis. In particular, the ABCMS recommended possible relations for decision makers based on limited initial data. Further, the decision maker reviews these recommended relations and provides the system with feedback. This type of interaction is a typical function of decision support systems (DSS). This process in ABCMS minimizes the subjective judgment of the decision maker in the construction of CM. The results of testing the CM created by the ABCMS for actual data turned out to be significant, and comparison with a conventional CM approach for the users also proved statistically significant. This study focused on the process of CM extraction with an ABCMS. However, if a CM becomes an agent like the ABCMS proposed in this study and is integrated with swarm optimization, the limitations of conventional CM can be mitigated, thus rendering the CM approach more readily expandable into other areas of application. The significance of this study lies in its provision of a foundation for future studies. The fields with room for additional improvements for the traditional CM properties based on ABCMS to offer more dynamic CM include: CM with a time gap consideration among nodes, CM subject to modification according to relations between nodes and inference functions, and CM with variable node direction according to changes in the status of particular nodes. Such refinements shall, however, be left for future studies. References Annibal, J. S., Tatiana, B. C., Susan, M. G., Julie, M. H., & Arthur, V. H. (2006). Methodological node: A methodology for constructing collective causal maps. Decision Sciences, 37(2), 263–283.

7028

N. Lee et al. / Expert Systems with Applications 38 (2011) 7016–7028

Axelrod, R. (1976). Structure of decision: The cognitive maps of political elites. Princeton, NJ: Princeton University Press. Boegl, K., Adlassnig, K. P., Hayashi, Y., Rothenfluh, T. E., & Leitich, H. (2004). Knowledge acquisition in the fuzzy knowledge representation framework of a medical consultation system. Artificial Intelligence in Medicine, 30, 1–26. Bonabeau, E., Dorigo, M., & Theraulaz, G. (1999). Swarm intelligence. From natural to artificial systems. New York: Oxford University Press. Bougon, M. G. (1983). Uncovering cognitive maps- the Self-Q technique. In G. Morgan (Ed.), Beyond method: Strategies for social research (pp. 173–188). Beverly Hills, CA: Sage. Brahim, C. D. (2002). Causal maps: Theory, implementation, and practical applications in multiagent environments. IEEE Transactions on Knowledge and Data Engineering, 14(6), 1201–1217. Chunyan, M., Angela, G., Yuan, M., & Zhonghua, Y. (2001). A dynamic inference model for intelligent agents. International Journal of Software Engineering & Knowledge Engineering, 11(5), 509–528. Clarke, I., & Mackaness, W. (2001). Management ‘Intuition’: An interpretative account of structure and content of decision schemas using cognitive maps. Journal of Management Studies, 38(2), 147–172. Eberhart, R. C., & Shi, Y. (2000). Comparing inertia weights and constriction factors in particle swarm optimization. Congress on Evolutionary Computing, 1, 84–88. Hagiwara, M. (1992). Extended fuzzy cognitive maps. In Proceedings of the First IEEE International Conference on Fuzzy Systems, New York, p. 795–780. Hart, J. A. (1977). Cognitive maps of three Latin American policy makers. World Politics, 30(1), 115–140. Hong, T., & Han, I. (2002). Knowledge-based data mining of news information on the internet using cognitive maps and neural networks. Expert Systems with Applications, 23(1), 1–8. Kardaras, D., & Karakostas, B. (1999). The use of fuzzy cognitive maps to simulate the information systems strategic planning process. Information and Software Technology, 41, 197–210. Klein, J. H., & Cooper, D. F. (1982). Cognitive maps of decision-makers in a complex game. Journal of the Operational Research Society, 33(1), 63–71. Kosko, B. (1986). Fuzzy cognitive maps. International Journal of Man-Machine Studies, 24, 65–75. Kwahk, K. Y., & Kim, Y. G. (1999). Supporting business process redesign using cognitive maps. Decision Support Systems, 25(2), 155–178. Lee, K. C., Han, J. H., Song, Y. U., & Lee, W. J. (1998). A fuzzy logic-driven multiple knowledge integration framework for improving the performance of expert systems. International Journal of Intelligent System in Accounting, Finance, and Management, 7, 213–222. Lee, K. C., Kim, J. S., Chung, N. H., & Kwon, S. J. (2002). Fuzzy cognitive map approach to web-mining inference amplification. Expert Systems with Applications, 22(3), 197–211. Lee, K. C., & Kwon, S. J. (2006). The use of cognitive maps and case-based reasoning for B2B negotiation. Journal of Management Information Systems, 22(4), 337–376. Lee, K. C., & Lee, S. (2003). A cognitive map simulation approach to adjusting the design factors of the electronic commerce web sites. Expert Systems with Applications, 24(1), 1–11. Lee, K. C., & Lee, S. J. (2007). Causal knowledge-based design of EDI controls: An explorative study. Computers in Human Behavior, 23(1), 628–663. Liu, Z. Q., & Satur, R. (1999). Contextual fuzzy cognitive map for decision support in geographic information systems. IEEE Transactions on Fuzzy Systems, 7(5), 495–507. Miao, C., Goh, A., Miao, Y., & Yang, Z. (2001). A dynamic inference model for intelligent agents. International Journal of Software Engineering and Knowledge Engineering, 11(5), 509–528. Miao, Y., Liu, Z. Q., Siew, C. K., & Miao, C. Y. (2001). Dynamical cognitive network: An extension of fuzzy cognitive map. IEEE Transactions on Fuzzy Systems, 7(5), 760–770. Montazemi, A. R., & Conrath, D. W. (1986). The use of cognitive mapping for information requirements analysis. MIS Quarterly, 10(1), 45–56.

Ndousse, T.D., & Okuda, T. (1996). Computational intelligence for distributed fault management in networks using fuzzy cognitive maps. In Proceedings of the 1996 IEEE International Conference on Communications. New York, p. 1558–1562. Nelson, K. M., Nadkarni, S., Narayanan, V. K., & Ghods, M. (2000). Understanding software operations support expertise: A revealed causal mapping approach. MIS Quarterly, 24(3), 475–507. Noh, J. B., Lee, K. C., Kim, J. K., Lee, J. K., & Kim, S. H. (2000). A case-based reasoning approach to cognitive map-driven tacit knowledge management. Expert Systems with Applications, 19(4), 249–259. Pal, S., & Konar, A. (1996). Cognitive reasoning using fuzzy neural nets. IEEE Transactions on Systems, Man. And Cybernetics, 26(4), 616–619. Park, K. S., & Kim, S. H. (1995). Fuzzy cognitive maps considering time relationships. International Journal of Human-Computer Studies, 42(2), 157–168. Press, W. H., Vetterling, W. T., Teukolsky, S. A., & Flannery, B. P. (1992). Numerical recipes in FORTRAN 77: The art of scientific computing (second ed.). Cambridge University Press. Rai, V. K., & Kim, D. H. (2002). Principal–agent problem: A cognitive map approach. Electronic Commerce Research and Applications, 1(2), 174–192. Ramaprasad, A., & Poon, E. (1985). A computerized interactive technique for mapping influence diagrams (MIND). Strategic Management Journal, 6(4), 377–392. Reger, R. K. (1990). Managerial thought structures and competitive positioning. In A. Huff (Ed.), Mapping Strategic Thought (pp. 71–88). NY: J. Wiley. Robert, F. S. (1976). Strategy for the energy crisis: The case of commuter transportation policy. In R. Axelrod (Ed.), Structure of decision: The cognitive maps of political elites. NJ: Princeton University Press. Ross, L. L., & Hall, R. I. (1980). Influence diagrams and organizational power. Administrative Science Quarterly, 25(1), 57–71. Rouff, C., Vanderbilt, A., Hinchey, M., Truszkowski, W., & Rash, J. (2004). Properties of a formal method for prediction of emergent behaviors in swarm-based systems. In Proceedings of the Software Engineering and Formal Methods (SEFM 2004), IEEE Computer Society, Washington, DC, USA, p. 24–33. Satur, R., & Liu, Z. Q. (1999). A contextual fuzzy cognitive map framework for geographic information systems. IEEE Transactions on Fuzzy Systems, 7(5), 481–494. Schneider, M., Shnaider, E., Kandel, A., & Chew, G. (1998). Automatic construction of FCMs. Fuzzy Sets and Systems, 93(2), 161–172. Silva, P. C. (1995). New forms of combined matrices of fuzzy cognitive maps. In Proceedings of the IEEE International Conference on Neural Networks, New York, p. 771–76. Styblinski, M. A., & Meyer, B. D. (1991). Signal flow graphs vs. Fuzzy cognitive maps in application to qualitative circuit analysis. International Journal of ManMachine Studies, 35(2), 175–186. Taber, R. (1994). Fuzzy cognitive maps model social systems. AI Expert, 9(7), 19–23. Warren, K. (1995). Exploring competitive futures using cognitive mapping. Long Range Planning, 28(5), 10–21. Wright, R. P. (2004). Mapping cognitions to better understand attitudinal and behavioral responses in appraisal research. Journal of Organizational Behavior, 25(3), 339–374. Yim, H., Cho, K., Kim, J. & Park, S. (2000). Architecture-Centric object-oriented design method for multi-agent systems. In Proceedings of the Fourth International Conference on Multi-Agent Systems (ICMAS 2000), Boston, Massachusetts, USA, p. 469–70. Zhang, W. R., Chen, S. S., & Bezdek, J. C. (1989). Pool2: A generic system for cognitive map development and decision analysis. IEEE Transactions on System, Man and Cybernetics, 19(1), 31–39. Zhang, W. R., Wang, W., & King, R. S. (1994). A-Pool: An agent-oriented open system shell for distributed decision process modeling. Journal of Organizational Computing, 4(2), 127–154. Zmud, R. W., Anthony, W. P., & Stair, R. M. (1993). The use of mental imagery to facilitate information identification in requirements analysis. Journal of Management Information Systems, 9(4), 175–191.