PDF (2242 KB) - World Scientific

July 27, 2011 15:23 S0129065711002833

International Journal of Neural Systems, Vol. 21, No. 4 (2011) 277–296 c World Scientific Publishing Company
DOI: 10.1142/S0129065711002833

HYBRID NEURAL INTELLIGENT SYSTEM TO PREDICT BUSINESS FAILURE IN SMALL-TO-MEDIUM-SIZE ENTERPRISES M. LOURDES BORRAJO Department of Inform´ atica, University of Vigo Edificio Polit´ecnico, Campus Universitario As Lagoas s/n Ourense, 32004, Spain [email protected] BRUNO BARUQUE Department of de Ingenier´ıa Civil, University of Burgos Esc. Polit´ecnica Superior, Edificio C C/Francisco de Vitoria, 09006, Burgos, Spain [email protected] EMILIO CORCHADO∗ , JAVIER BAJO† and JUAN M. CORCHADO‡ Departamento de Inform´ atica y Autom´ atica University of Salamanca, Plaza de la Merced s/n 37008, Salamanca, Spain ∗ [email protected] † [email protected] ‡ [email protected] During the last years there has been a growing need of developing innovative tools that can help small to medium sized enterprises to predict business failure as well as financial crisis. In this study we present a novel hybrid intelligent system aimed at monitoring the modus operandi of the companies and predicting possible failures. This system is implemented by means of a neural-based multi-agent system that models the different actors of the companies as agents. The core of the multi-agent system is a type of agent that incorporates a case-based reasoning system and automates the business control process and failure prediction. The stages of the case-based reasoning system are implemented by means of web services: the retrieval stage uses an innovative weighted voting summarization of self-organizing maps ensemblesbased method and the reuse stage is implemented by means of a radial basis function neural network. An initial prototype was developed and the results obtained related to small and medium enterprises in a real scenario are presented. Keywords: Hybrid neural intelligent system; CBR; MAS; Business failure.

they are achieving their goals.6–8 Such mechanisms are constructed around a series of organizational policies and specific procedures referred to as “controls”, and conform to the structure of the business control of the company. As a consequence of this, the need has arisen for periodic internal audits. However, evaluating and predicting the evolution of these

1. Introduction During the last years there has been a growing need of developing innovative tools that can help SMEs (small to medium sized enterprises) to predict business failure as well as financial crisis.1–5 SMEs require a business control mechanism in order to monitor their modus operandi and analyze whether

277

July 27, 2011 15:23 S0129065711002833

278

M. L. Borrajo et al.

types of business entities, which are characterized by their great dynamism, tends to be a complicated process. It is necessary to construct models that facilitate the analysis of the work carried out in changing environments such as finance. The processes carried out inside a company are grouped into functional areas17 denominated “Functions”. A Function is a group of coordinated and related activities that are systematically and iteratively carried out during the process of reaching the company’s objectives.53 The functions that are usually carried out within a company, as studied within the framework of this research, are: Purchases, Cash Management, Sales, Information Technology, Fixed Assets Management, Compliance to Legal Norms and Human Resources. Each one of these functions is broken down into a series of activities. For example, the Information Technology function is divided into the following activities: Computer Plan Development, Study of Systems, Installation of Systems, Treatment of Information Flows, and Security Management. Agents and multi-agent systems (MAS) have become increasingly relevant for developing distributed and dynamic intelligent environments.69,71,72,74,75 Agents are computational entities that can be characterized through their capacities in areas such as autonomy, reactivity, proactivity, social abilities, reasoning, learning and mobility.11,68,70 These capacities make the multiagent systems very appropriate for constructing intelligent environments.76–81 An agent can act as an interface between the user and the rest of the elements of the intelligent environment.82,83 This paper presents an innovative solution to model distributed adaptive systems in business environments. It is based on a multi-agent architecture11 that can integrate Web services, and incorporates a novel planning mechanism that makes it possible to determine potential risky situations based on existing activities and previous results. The core of the multi-agent system is a hybrid neural intelligent mechanism aimed at detecting and predicting business failures. There are various artificial intelligence techniques such as artificial neural networks,112–114,101,109–111,99 Bayesian networks,99 and fuzzy logic100,102,103,105–107 which have been applied to business failure prediction. While these techniques can be applied to failure detection and

prediction, the knowledge obtained cannot be incorporated into successive tests and included in subsequent analyses. This paper presents a hybrid neural intelligent system based on CBR (Case-Based Reasoning) which uses past experiences to solve new problems.12 CBR facilitates the creation of strategies similar to the processes followed in SMEs. The recovery of information from previous experiences simplifies the prediction process by detecting and eliminating relevant and irrelevant patterns detected in previous analyses. The retrieve phase of the hybrid neural intelligent system incorporates a novel version of the SOM13 (Self-Organizing Map) based on the application of ensembles techniques and called WeVoS-SOM.14 The reuse stage incorporates a radial basis function neural network.104 The revise and retain stages implement a decision support system for experts. In Ref. 53 we presented a system composed of two case-based reasoning systems to detect the associate risk in the activities of SMEs in the textile sector and generate recommendations to improve the erroneous processes. In Ref. 54 we presented a decision support tool based on a case-based reasoning system that automates the internal control processes of a SME. The approach presented in this article is an evolution of our previous works that proposes an innovative perspective53,54 where a multi-agent system that integrates within web services is used to model the organizational structure of a SME. This perspective allows taking into account the advantages of virtual organizations to model the SME functions. Moreover, the core of the system is a CBR-BDI agent with the ability to distribute its internal functionalities in web services which notably improves the efficiency in the distribution of tasks. The new approach proposes a new method for the retrieval stage of the CBR system, the WeVoS-SOM, that notably improves the case’s recovery reducing the final quantity of cases stored and making it easier to recover the most similar cases to the problem introduced. The article is structured as follows: the next section briefly introduces the problem that motivates this research and revises the state of the art. Section 3 presents the multi-agent system for failure prediction in small and medium enterprises. Section 4 describes the novel strategies incorporated in the stages of the CBR cycle of the CBR-BDI evaluator agent. Section 5 describes a case study

July 27, 2011 15:23 S0129065711002833

Hybrid Neural Intelligent System to Predict Business Failure

specifically developed to evaluate the hybrid neural intelligent system presented within this study. Finally, Sec. 6 presents the results obtained after testing the model and Sec. 7 present the conclusions of this study. 2. The Crisis Prediction Problem Many things in our business and auditing environment are changing at an increasing rate. Increased competition and the need for faster and better information for decisions mark today’s business environment. In addition, systems are complex and many times provide on-line access facilities. This complexity means that auditors have more and different kinds of work to do than they had earlier.15 SMEs require an internal control mechanism in order to monitor their modus operandi and to analyze whether they are achieving their goals constructed around a series of organizational policies and specific procedures. The establishment of objectives (which is not a function of the Internal Control) is a previous condition for control risk evaluation, which is the main goal of the Internal Control. In general terms, the administration of an enterprise has three large categories of objectives when designing an internal control structure16 : Effectiveness and efficiency of operations; reliability of financial reporting and compliance with applicable laws and regulations. The current technological advances have involved considerable modifications in the area of auditing, basically characterized by the following features17 : • Progressive increase in the number and level of complexity of audit rules and procedures. • Changes in the norms of professional ethics, which demand greater control and quality in auditing. • Greater competitiveness between auditing firms. • Development of new types of auditing. The internal auditor must monitor the internal controls directly and recommend improvements on them. Therefore, all the activities carried out inside the organization can be included, potentially, within the internal auditors’ remit. Establishing and maintaining proper internal controls is an essential, ongoing, and complex process for today’s organizations. Internal auditors utilize risk measures to allocate critical audit resources to compliance, operational, or financial activities within the organization.18,19 Risk

279

assessment involves recognizing patterns in the data such as complex data anomalies and discrepancies, that perhaps conceal one or more error or hazard conditions.20–22 Despite the significance of performing risk analysis, there are several factors that cause this process to be viewed negatively. Most risk analysis processes use questionnaires for gathering asset, threat, and vulnerability data. These questionnaires are often very long and difficult to answer, making risk analysis time consuming and very costly. Moreover, the traditional process of risk analysis is not able to handle the contingency of organization, such as organizational culture, characteristics of IS, and characteristics of organization. Generally, the tasks of risk analysis for an information system require ‘deep knowledge’ and include various difficult problems. Together, these circumstances have made the audit profession increasingly competitive. Consequently, the need has arisen for new techniques and tools, which can be provided by information technology and artificial intelligence. Risk analyzers have been found to reason by analogy using prior experiences rather than by IF–THEN rules. The CBR system12,23 is useful for tasks using predicates that are ill defined, that lead to inconsistent outcomes, and which have incomplete rules to apply.24 Moreover, CBR is good for the tasks that are experience intensive and hard to acquire domain experience such as risk analysis.10,25,26 Different alternatives have been checked for CBR,30,32 as the use of the Electre method,28 the use of hybrid strategies,29 SVM108,97 or Majority voting combination.27,31 The aim is to achieve more relevant, more suitable information, in order to help auditors make decisions faster and thereby increase the efficiency and quality of auditing. The following section presents a system developed to facilitate the auditing process. 3.

Multi-Agent System for Preventing Crisis Failure in Medium-Sized Enterprises

Agent and multi-agent systems11 combine classical and modern functional architecture aspects. The integration and interoperability of agents and multi-agent systems with SOA and Web Services approaches has been recently explored.33 Bonino da Silva et al. in Ref. 34 propose merging multi-agent

July 27, 2011 15:23 S0129065711002833

280

M. L. Borrajo et al.

techniques with semantic web services to enable dynamic, context-aware service composition. Ricci et al. in Ref 35 have developed a java-based framework to create SOA and Web Services compliant applications, which are modeled as agents. Liu in Ref 36 proposes a multi-agent architecture to develop inter-enterprise cooperation systems using SOA and Web Services components and communication protocols. Our proposal not only provides communication and integration between distributed agents, services and applications; it also proposes a new method to facilitate the development of distributed multi-agent systems by means of modeling the functionalities of the agents and the systems as services. More concretely, the paper presents a multi-agent enterprise control system aimed at developing a generic model useful in any type of SME. The enterprises that participated in the experiment were very interested in a tool such as the one developed within the framework of this investigation, where the constructed multiagent system is able to determine the state of each of the activities and calculate the associated risk. It also detects any inefficient processes and predicts crisis situations. Figure 1 shows the basic schema of the proposed architecture, where all requests and responses are

Fig. 1.

handled by the agents in the platform. The system is modeled as a modular multi-agent architecture, where services and applications are managed and controlled by deliberative BDI (Belief, Desire, Intention) agents.37–39 There are different kinds of agents in the architecture, each one with specific roles, capabilities and characteristics: — Business Agent. This agent was assigned for each SME in order to collect new data and allow consultations. The SME can interact with the system by means of this agent, introducing information and receiving predictions. — Evaluator Agent. It is responsible for the evaluation and predictions of potential risky situations. Every time that it is necessary to obtain a new estimate of the state of an activity, the agent evolves through several phases. On the one hand, this evolution allows the multi-agent system, to identify the latest situations most similar to the current situation in the retrieval stage, and to adapt the current knowledge in the reuse stage in order to generate an initial estimate of the state of the activity being analysed. The activity selected will then serve as a guide for establishing a risk level for the activity, its function, and

Crisis multi-agent architecture basic schema.

July 27, 2011 15:23 S0129065711002833

Hybrid Neural Intelligent System to Predict Business Failure

the company itself, to develop in a more positive way. The retention phase guarantees that the system evolves in parallel with the firm, basing the corrective actions on the calculation of the error previously made. — Expert Agent. This agent helps the auditors and enterprise control experts that collaborate in the project to provide information and feedback to the multi-agent system. These experts generate prototypical cases from their experience and they receive assistance in developing the Store agent case-base. — Store Agent. This agent has a memory that has been fed with cases constructed with information provided by the enterprise (through its agent) and with prototypical cases identified by 34 enterprises control experts, using personal agents who have collaborated and supervised the developed model. The Evaluator agent is a CBR-BDI agent with advanced reasoning abilities that provided great adaptation and learning capacities. This agent type is the core of the multi-agent system and is presented in detail in Sec. 4. 4. CBR-BDI Agents for Crisis Classification and Prediction The application of agents and multi-agent systems facilitates taking advantage of the inherent capabilities of the agents. Nevertheless, it is possible to increase the reasoning and learning capabilities by incorporating a case-based reasoning (CBR)12 mechanism into the agents. In the case at hand, a CBR classifier agent (CBR-BDI)11,40 is responsible for classifying the enterprise situation and predict possible crisis. In the BDI (Beliefs Desires Intentions) model, the internal structure of an agent and its capacity to choose is based on mental aptitudes: agent behaviour is composed of beliefs, desires, and intentions.41 A BDI architecture has the advantage of being intuitive and capable of rather simply identifying the process of decision-making and how to perform it. Case-based Reasoning is a type of reasoning based on the use of past experiences.12,23 The purpose of case-based reasoning systems is to solve new problems by adapting solutions that have been used to solve similar problems in the past. The

281

fundamental concept when working with case-based reasoning is the concept of case. A case can be defined as a past experience, and is composed of three elements: A problem description which describes the initial problem, a solution which provides the sequence of actions carried out in order to solve the problem, and the final state which describes the state achieved once the solution was applied. The way in which cases are managed is known as the case-based reasoning cycle. These systems manage cases (past experiences) to solve new problems. The way in which cases are managed is known as the CBR cycle. The CBR cycle shown in Fig. 2 consists of four sequential phases: retrieve, reuse, revise and retain. The retrieve phase starts when a new problem description is received. Similarity algorithms are applied in order to retrieve from the cases memory the cases with a problem description more similar to the current one. Once the most similar cases have been retrieved, the reuse phase begins. In this phase the solutions of the cases retrieved are adapted to obtain the best solution for the current case. The revise phase consists of an expert revision of the solution proposed. Finally, the retain phase allows the system to learn from the experiences obtained in the three previous phases and updates the memory case in consequence.

Fig. 2. cycle.

Diagram including a CBR-BDI agent reasoning

July 27, 2011 15:23 S0129065711002833

282

M. L. Borrajo et al.

Fig. 3.

Internal structure of the CBR-BDI agent.

The following sub-sections present the internal structure of the CBR-BDI Evaluator agents used to predict and prevent crisis in medium-sized enterprises. This agent collaborates with the Expert Agent to provide a decision support tool for the internal auditor of the SME, who is the main user of the proposed model. As can be seen in Fig. 3, the retrieval stage consists of a novel version of the SOM13,42 based on the application of ensembles techniques and called WeVoS-SOM.14 The reuse stage incorporates a radial basis function neural network.87–95 The revise and retain stages implement a decision support system for experts. 4.1. Retrieval stage This stage is aimed at retrieving the most similar cases to the current problem. To do so, the approach used is to organize the case base in a way that cases with similar characteristics are stored “close” to each other, by means of topology preservation

models. The learning process used in this kind of artificial neural networks make only the most similar unit of the map (called Best Matching Unit or BMU) react when a new data entry is presented. By keeping track of which units react to which cases, is easy to keep also record of the similarity between cases, as it is determined by their proximity on the map obtained. That way, the algorithm can not only recover similar cases, but also order them according to their degree of similarity.65,66 This enables also the reduction of the case base, as is easy to find if the case to be inserted is similar to a high degree to already existent cases and therefore can be discarded.59 The specific technique used is the training of several Self-Organizing Maps using the classical bagging algorithm61,62 and a final step fusing all together to obtain a resultant map reflecting the best characteristics of each of its composing maps. The novelty of this approach lies in the algorithm used for the fusion of maps.

July 27, 2011 15:23 S0129065711002833

Hybrid Neural Intelligent System to Predict Business Failure

The SOM family of algorithms is one of the most common and well known among Prototype Learning Algorithms, included in the family of cluster based learning.67 The advantage that offers against other of the most commonly used models of that family (such as K-Nearest Neighbours) is its topological ordering.58,60 This allows the user to order the recovered cases according to the similarity to a given case, rather than obtaining an unordered set of examples. The authors of the paper have compared the WeVoSSOM algorithm in other works with similar simpler model such as the SOM and ViSOM.14 As studies have proved, this algorithm is the best at ordering its composing units, and therefore it improves the results of the model including the simple SOM, regarding the reduction of the size of the case base. This can also be seen in the results section. The bagging meta-algorithm consists on repeating several overlapping re-samplings over the dataset. That way, it simulates the obtaining of several different datasets from the same source as the original one. Each of the re-sampled datasets is used for the training of a single classifier. That way each is trained with very similar, but not exactly the same characteristics. Doing so, the errors that each of them will commit will not necessarily overlap, having therefore room for general improvement within the ensemble. The idea behind the novel fusion method presented in this study, Weighted Voting Superposition-SOM (WeVoS-SOM),14 is to obtain the best position for each neuron and its neighbours, by choosing among several possibilities that

Fig. 4. Adaptation of the units of the final map according to the composing units of different maps of the ensemble.

283

were explored by generating different, but similar, composing maps. As a consequence, the final map is expected to be better than its composing maps, but keeping one of the most important features of this type of algorithms: its topological ordering. WeVoS is an improved version of an algorithm presented in several previous works on superposition.14,43–45 In this study, it is applied to the recovery of cases with a description similar to the current situation of a given enterprise. The final map is obtained again on a neuron-byneuron basis. First, the neurons of the final map are initialized by calculating the centroids of the neurons in the same position of the map grid in each of the trained maps. This is calculated as in Eq. (1).
wi (1) wneuAvg = 1/|Wk | · wi ∈Wk

being wi the weights of an individual neuron and |Wk | denoting the number of neurons that are participating in the calculation of the centroid. In this case, it coincides with the number of maps trained in the ensemble. Then, a recalculation of the final position of that neuron uses the information associated with the neurons in that same position in each map of the ensemble. For each neuron, a sort of voting process is performed, as in Eq. (2):   bp,m qp,m · M (2) Vp,m = M i=1 bp,i i=1 qp,i where Vp,m is the weight of the vote for the neuron included in map m of the ensemble, in position p. N is the total number of maps in the ensemble, bp,m is the binary vector used for marking the data set entries recognized by neuron in position p of map m, and qp,m is the value of the desired quality measure for a neuron in position p of map m. Associated with each map the algorithm has binary matrix (b) with as many rows as neurons in the map and as many columns as number of entries. Initially all values are initialized to 0 and when a neuron is selected as the BMU for an entry, the bit for that neuron and that entry is switched to 1. The algorithm can make use of any of the quality/error measures in topology preservation literature as long as it can be calculated for each neuron individually, as opposed to a general map measure. In the case of the experiments performed in the paper

July 27, 2011 15:23 S0129065711002833

284

M. L. Borrajo et al.

the Goodness of Adaptation proposed in Ref. 63 and expressed in Eq. (3). |kv
|−1

d(xi ) =
xi − vi
+ min

i


k=0

×
wIi(k) − wIi(k+1)


(3)

where vi and vi represent the weights of the first BMU and the second BMU respectively, corresponding to data entry xi . Ii(k) and Ii(k+1) represent indexes of the kth and the kth + 1 neurons in the minimum path from vi to vi , both neurons being direct neighbours in the map grid. According to that definition, that is, the first neuron in the path is the first BMU for data entry xi and; that is, the last neuron in the path, corresponds to the second BMU for data entry xi . In summary, the algorithm takes into account both the recognition of units in a certain position, as expressed in Eq. (4); and the quality of adaptation of those units, as expressed in Eq. (5). B(p) =

N
E


BMUp (xj )

(4)

i=1 j=1

being BMUp (xj ) a function that outputs a 1 if the neuron in position p is the BMU for data entry xj and 0 otherwise. Q(p) =

N


qp,i

(5)

i=1

being qp,i the quality measure calculated for position p in map i. The weights of the neurons are fed into the final network as with the data inputs during the training phase of a SOM,13 considering the “homologous” neuron in the final map as the BMU (Best Matching Unit). The weight of the vote (Vp,m ) is used as the learning rate (α) for the update of the final neuron of the fused map in the characteristic weight update of the SOM (Eq. (6)): wk (t + 1) = wk (t) + α(t)η(v, k, t) × (x(t) − wk (t))

(6)

where x is the input to the network, wk is the weight vector associated with neuron k, while wv is the weight vector associated to the winning unit in the

lattice, also called best matching unit (BMU). α(t) is the learning rate of the algorithm, η(v, k, t) is the neighbourhood function (usually a gaussian function) where v represents the position of the BMU for the particular x of time t and k the positions of the units in the neighbourhood of this one. The characteristics vector of the final neuron will be updated towards the characteristics vector of the composing maps correspondent neurons. The difference of the updating performed for each “homologous” neuron in the composing maps depends on the quality measure calculated for each neuron. The higher the quality (or the lower the error) of the neuron of the composing map; the stronger the neuron of the fused map updated towards the weights of that neuron. Either a single or a linear combination of several quality measures can be used to determine the quality of a unit. The number of data inputs recognized by each neuron is also taken into account in this quantization of the “best suitability” of one neuron or another for the same position in the final map. In short, the fusion algorithm will consider “more suitable” weights of a composing neuron to be the weights of the neuron in the final map according to both the number of inputs recognized and the adaptation quality of the neuron. The model, called WeVoS is described in detail in Table 1. This approach takes into account not only the characteristics of a single neuron, but also the topographic ordering of its neighbourhood. This new approach14 obtains more maps that are more faithful to the inner structure of the data set from a visualization point of view.

4.2. Reuse stage This phase aims to obtain an initial estimation of the state of the activity analysed. In order to obtain this estimation, RBF networks are used.9,11,45–47,55–57 As in the previous phase, the number of attributes of the problem case depends on the activity analyzed. Therefore, it is necessary to establish a RBF network system, one for each of the activities to be analysed. The k cases retrieved in the previous phase are used by the RBF network as a training group that allows it to adapt its configuration to the new problem encountered before generating the initial estimation. The system presented in this article has

July 27, 2011 15:23 S0129065711002833

Hybrid Neural Intelligent System to Predict Business Failure

Table 1.

285

Algorithm 1. Weighted voting summarization algorithm.

Input: Set of trained topology-preserving maps: M =
M1 . . . MN , training data set: S Output: A final fused map: Mfus 1: Select a training set S =
(x1 , y1 ) . . . (xE , yE ) 2: train several networks by using the bagging meta-algorithm: M1 . . . MN WeVoS (M1 . . . MN ) 3: For each map Mi ∈ M 4: calculate the quality/error measure chosen for ALL neurons in the map. Calculate which is the neuron recognizing each entry of the dataset (sj ) in the map (BM U (sj )). These two values are used in Eq. (2). 5: end 6: calculate recognition rate for each position (Eq. (4).) 7: calculate an accumulated total of the quality/error for each position (Eq. (5)) 8: For each unit position p in Mi 9: initialize the fused map (Mfus ) by calculating the centroid (wc ) of the neurons of all maps in that position (p) as in Eq. (1) 10: end 11: For each map Mi ∈ M 12: For each unit position p in Mi 13:

calculate the vote weight (Vp,Mi ) using Eq. (2).

14: Feed the weights vector of neuron wp into the fused map (Mfus ) as if it was an input to the network. The weight of the vote (Vp,Mi ) is used as the learning rate (α) for the update of the final neuron of the fused map in the characteristic weight update of the SOM. Eq. (6) 15: The position of that neuron (p) is considered as the position of the BMU (v). This causes the neuron of the fused map (wp∗ ) to approximate the neuron of the composing ensemble (wp,m ) according to the quality of its adaptation. 16:

end

17: end

a RBF network for each of the activities that will be evaluated for a SME. Each of the RBF networks has as inputs as tasks are evaluated for the activity. The topology of each of the RBF networks used in this task consists of an input layer with as many neurons as attributes possessed by the input vector which constitutes the problem descriptor, a hidden layer with 14 centres, and an output layer with a single neuron corresponding to the variable to be estimated (correction level or state of activity analyzed in percentage). The number of centers (14) was empirically obtained after various tests. The centers are the neurons in the hidden layer of the RBF neural network. The RBF network is characterized by its ability to adapt, to learn rapidly, and to generalize.

Specifically, within this system the network acts as a mechanism capable of absorbing knowledge about a certain number of cases and generalizing from them. During this process, the RBF network, interpolates and carries out estimations without forgetting part of those already carried out. The Store agent acts as a permanent memory capable of maintaining many cases or experiences by consulting the WeVoS-SOM model of the case base; while the RBF network, in the Evaluator agent, acts as a short term memory, able to recognize recently learnt patterns and to generalize from them. 4.3. Revise stage The objective of this phase is to guarantee or refute the initial solution proposed by the RBF network,

July 27, 2011 15:23 S0129065711002833

286

M. L. Borrajo et al.

obtaining a final solution and working out, apart from this, the control risk. Faced with the initial estimation generated by the RBF network, the internal auditor will be responsible of deciding whether to accept the aforementioned solution. It will be based on the specific knowledge he possesses about the company in which he is working. If the estimation provided by the system is considered to be valid, the system will take this solution as the final solution and, in the next phase of the CBR cycle, will store a new case in the case base, formed from the problem case and the final solution. The system will assign this new case with an initial reliability of 100%. If, on the contrary, the internal auditor considers that the solution provided by the system is not valid, they will provide their own solution that the system will take as the final solution. This, along with the problem case, will be stored as a new case in the case memory in the following phase. This new case will be assigned a reliability of 30%. This value has been determined bearing in mind the opinion of various auditors as to the importance that should be assigned to the personal opinion of the internal auditor. As of the final solution: the state of the activity, the system must work out the control risk associated with the activity. All developed activity in the enterprise sector has an associated risk that will indicate the negative influence that it exerts over the proper running of the enterprise. That is to say, the control risk of an activity measures the impact, on the whole on all of the managerial processes that the current state of the aforementioned activity represents. In this investigation it is considered that the risk level can assume one of these three possible values: low, medium and high. The calculation of the control risk level associated with an activity is based on the current state of the activity and on its level of importance. This last value is obtained after analysing the information obtained from a series of surveys (98 in total) answered by auditors all over Spain. In the previously mentioned surveys the auditors were asked to rank from 1 to 10 the importance that they gave the above-mentioned activity together with the function to which it belongs. The greater the importance is; the greater weighting the activity will have within the internal control system. Then, as of the level of importance of the activity, and bearing in mind the final solution obtained

after the revision phase, the control risk level can be worked out. The calculation of the control risk level is carried out using if-then rules48–50 in which the importance the auditors have assigned to activity is compared to the final solution or state of the activity. 4.4. Retain stage The retain phase is aimed at learning from the new past experience. As stated in the previous subsection, a case together with its corresponding solution is stored in the memory of cases if it is considered as reliable, that is, a confidence level of 100%. Moreover, the system also stores those past experiences that are considered as not reliable by the expert. These past experiences are considered of interest by the CBR because they can be used as an indicator for invalid solutions, although this learning ability is not taken into consideration in this study. In this phase, the risk level associated with the case is also stored in the memory of cases, that is, the if-then rules that contain the know-how of the auditors.

5.

Case Study: Castilla and Le´ on Business Scenario

The case study was aimed to detect and predict possible failures in SMEs in Castilla y Le´ on. The experiment consisted on the construction of the initial prototype of memory of cases and then in predicting potential risky situations for the enterprises taken into considerations. The data used to construct the model were obtained by surveys conducted with enterprise experts in the different functional areas of various enterprises, using the Expert agents. This type of survey attempts to reflect the experience of the experts in their different fields. For each activity, the survey presents two possible situations: the first one tries to reflect the situation of an activity with an incorrect activity state, and the second one tries to reflect the situation of an activity with a satisfactory activity state. Both situations will be valued by a human expert using a percentage. Figure 5 shows a generic survey relative to any activity.51 Each activity is composed of tasks, and each task has an importance rate, and values of realization for both incorrect and satisfactory activity state. These parameters are explained below in the analysis of the

July 27, 2011 15:23 S0129065711002833

Hybrid Neural Intelligent System to Predict Business Failure

Fig. 5. Table 2.

Generic experts’ survey.

Case structure.

Problem

287

Solution

Case Input Function Activity Reliability Activity number vector number number state

case structure. The data acquired by means of surveys were used to build the prototype cases for the initial Store agent case base. Table 2 shows the case structure that constitutes the case base. Each case is composed of the following attributes, that can be observed in Fig. 5: • Case number: Unique identification: positive integer number. • Input vector: Information about the tasks (n subvectors) that constitute an industrial activity: ((IR1, V1), (IR2, V2), . . . , (IRn, Vn)) for n tasks. Each task sub-vector has the following structure (IRi,Vi): — IRi: importance rate for this task within the activity. It can only take one of the following values: VHI (Very high importance), HI (High Importance), AI (Average Importance), LI

Table 3. Assessment of the different levels of importance of the tasks. Importance rate VHI (Very High Importance) HI (High Importance) AI (Average Importance) LI (Low Importance) VLI (Very Low Importance)

Numeric value 5 4 3 2 1

(Low Importance), VLI (Very low importance). To transform this attribute in a numeric value, the conversion process, shown in Table 3, was applied. — Vi: Value of the realization state of a given task: a positive integer number (between 1 and 10). • Function number: Unique identification number for each function • Activity number: Unique identification number for each activity • Reliability: Percentage of probability of success. It represents the percentage of success obtained using the case as a reference to generate predictions.

July 27, 2011 15:23 S0129065711002833

288

M. L. Borrajo et al.

• Activity State: degree of perfection for the development of the activity, expressed by percentage. This is the solution of a problem case. The case study presented in this work has been implemented under the framework of the Crisis project. This project was oriented to detect possible risky situations in SMEs, taken into account the crisis that affects the market. A multi-agent system was implemented and 18 SMEs participated in the experiment and were assigned a personal business agent. The enterprises were situated in different sectors and located in the Spanish region of Castilla y Le´ on. The economic context is the same for all the SMEs. The system was tested during 24 months, from January 2008 to December 2009, tuned and improved taking into account the experience acquired using a total of 196 cases. Next section presents the results obtained after testing the novel hybrid neural intelligent system in the real scenario.

6.

Results

The approach has been applied to a real case scenario in the Spanish region of Castilla y Le´ on, for enterprises from different sectors. The human experts (the auditors) have remarked on the advantages of using the presented multi-agent system approach as a

Fig. 6.

decision support system for detecting potential risky situations in SMEs, and have especially noted the facility in acquiring knowledge and explanations. The data obtained demonstrates that the application of this hybrid neural intelligent system caused a positive evolution in all enterprises. This evolution was reflected in the reduction of inefficient processes. The indicator used to determine the positive evolution of the companies was the state of each of the activities analysed. After analysing one of the company’s activities, it was necessary to prove that the state of the activity (valued between 1 and 100) had increased beyond the state obtained in the previous three month period. The system considers small changes in the tasks performed in the SMEs, and all the experts that participated in the experiments considered 3 months as a significant time to evaluate the evolution of a SME related to these changes. Thus, it was possible to conclude that the inefficient processes had been reduced within the same activity. When there was measurable improvement in the majority of activities (above all those most relevant to the enterprise), it could be affirmed that the enterprise had improved its situation. Figure 6 shows two screenshots of examples of activities for an expert agent corresponding to activities for the same enterprise, Information Technologies and Inmobilized. Figure 6 shows a positive evolution of the

Example of monitoring and prediction for two activities: Information technology and capital assets.

July 27, 2011 15:23 S0129065711002833

Hybrid Neural Intelligent System to Predict Business Failure

activities. Taking into account the global results, in 86% of these companies, the number of inconsistent processes was reduced improving the state of activities by an average of 18%. In general, it could be said that these results demonstrate the suitability of the techniques used for the integration of the developed hybrid neural intelligent control system. The best results occurred in the smaller sized enterprises. This is due to the fact that these businesses have a greater facility to adopt and adapt to the changes suggested by the system’s predictions. In order to have an objective evaluation of the prediction abilities of the hybrid neural intelligent system presented in this study, we have evaluated the system with different configurations and compared the results obtained: (i) Using a RBF network45 ; (ii) Using a CBR mechanism12 with a RBF network in the Reuse phase; (iii) Using a CBR with a SOM13 in the Retrieve phase and a RBF in the Reuse phase; (iv) Using a CBR with WeVoS-SOM14 in the Retrieve phase and a RBF in the Reuse phase. Table 4 shows a summary of the results obtained. The four different techniques are compared, using an incremental case base containing from 10 to 190 cases. There is a new prediction for each of the new cases. Table 4 shows the evolution of the results along with the increase in the number of cases stored in the memory of cases. The numerical results shown the average result of a series of tests performed with the available information. The number of tests carried out in each iteration (every time the case base grows) is ten percent of the size of the case base (for instance, when the case base contained 100 elements, 10 tests were performed, and so on). This percentage

Table 4. Percentage of good predictions obtained with four different techniques and for a different amount of cases. Number RBF (%) CBR + CBR + SOM + CBR + of cases RBF (%) RBF (%) WeVoSSOM + RBF (%) 10 50 100 150 190

46 48 53 56 61

38 43 47 58 64

42 47 58 64 72

43 48 63 74 87

289

of the data used to test the correction of the system was not previously used in its training. Those cases are also part of the historical information used, so they can be used to check the correction of the predictions generated by the system. The results for each of the techniques being analyzed improved when the number of cases stored was increased. The incorporation of the WeVoS-SOM system in the Retrieve phase provides more accurate predictions. The size of the maps was calculated as suggested in Ref. 64, while the size of the ensemble was also increased form 2 maps up to 5, depending on the data managed. The improvement over the CBR-SOM consists of using the WeVoS-SOM algorithm to structure the case base, enhancing the organizational characteristics of the case base. More visual results of those experiments can be found on Fig. 7(a). In Fig. 7(a) the X axis represent the number of cases used in each of the tests and the Y axis represents the percentage of successful predictions. Figure 7(b) provides a comparison between the sizes of the memory of cases of the different methods used for comparison. X axis presents each of the methods taken into consideration and the Y axis represents the size of the memory of cases, in number of cases. In the last two methods the data are structured and analyzed before being stored. The size of the case base in the CBR + WeVoS + RBF system is smaller mainly from using an organising algorithm like WeVoS-SOM that structures the information depending on its values, and not only on the different elements available. With the same amount of initial information, the system does not store redundant information, thus reducing the final quantity of cases stored and making it easier to recover the most similar cases to the problem introduced. The reason for the decrease in size when using the SOM-based algorithm is that, once a new case is needed to be added, it can easily be compared with the most similar ones to it. If the new case is very similar to one (or more) of the cases already stored in the case base, it is not necessary to store it again, as we already have a representative of a very similar situation stored. The WeVoS-SOM takes these organization capabilities a step even further. Figure 7(c) shows how the performance of the CBR system improves when complex calculation processes are externalized. The X axis represents the two systems and the Y axis represents the efficiency of the system

July 27, 2011 15:23 S0129065711002833

290

M. L. Borrajo et al.

(a)

Fig. 7.

(b)

(a) % Prediction results. (b) Size of the memory of cases. (c) Efficiency of the CBR system.

(in %). In previous studies,11 the reasoning (CBR) and planning (CBP) mechanisms are integrated into the agents. These mechanisms are busy almost all the time, overloading the respective agents. Because CBR mechanism is the core of the system, it must be available at all times. The system depends on this mechanism to generate all decisions, so it is essential that it has all processing power available in order to increase overall performance. In addition, the use of CBR and CBP mechanisms into deliberative BDI agents makes these agents complex and unable to be executed on mobile devices. In this study these mechanisms have been modelled as services, so any agent can make use of them. Figure 7(c) shows the average number of free computational resources in the computer used for the experiments (Intel Core2 DUO 2,8 GHz, 8 GB RAM) for the CBR + Web Services configuration and for the CBR configuration. As can be seen in Figure 6(c), the use of web services increases the processing power of the CBRBDI agent. Moreover, we observed a decrease in the number of crashed agents during tests when web services are used, which notable improves the hability to recover from errors.

Table 5. Multiple comparison procedure among different techniques. RBF

RBF CBR + RBF * CBR + SOM + * RBF CBR + WeVoS + * RBF

CBR + RBF

CBR + CBR + SOM + WeVoS + RBF RBF

= *

*

Table 5 shows a multiple comparison procedure (Mann-Whitney test52 ) used to determine which models are significantly different from the others. The asterisk indicates that these pairs show statistically significant differences at the 99.0% confidence level. Table 5 shows that the CBR + WeVoS + RBF system presents statistically significant differences compared to the other models. We had certain problems implementing the system, partly because the management and experts were not familiar with the use of computational

July 27, 2011 15:23 S0129065711002833

Hybrid Neural Intelligent System to Predict Business Failure

291

architecture should incorporate a shared memory of cases to compare data from different firms, but with the guarantee of data privacy.

7.

Fig. 8. % Average satisfaction degree related to the average number of retrieved cases.

devices and multi-agent systems, so some courses were given to introduce them to these technologies and teach them how to use the system interface. The proposed novel hybrid neural intelligent system is a unique system useful for dynamic environments, and open enough to be used in other enterprise environments. The experts noted that the behavior of the system improves as the number of cases in the memory of cases grows. Figure 8 shows a representation of the average satisfaction degree of the experts related to the growth of the number of cases. The expert satisfaction is obtained by means of surveys that were completed for different sizes of the memory of cases and taking into account the effectiveness of the system. 24 experts participated in the experiments and provided a satisfaction degree that takes values in the interval [0, 100]. We found some difficulties with this system, because some of the SMEs were reticent to complete surveys. Management are the most reticent about trusting the system for several reasons: (i) they do not trust the partiality of the system, and are reluctant to facilitate their internal data (ii) updating the information about the enterprise requires specialized human resources and time. However, the auditors and experts believe that the CBR-BDI agents may favour their work and provide a highly appreciated decision support tool. They believe that this hybrid neural intelligent architecture has more advantages than disadvantages and that the system helped them to detect inconsistent processes in the enterprises. They tend to argue that the hybrid neural intelligent

Conclusions

This study has presented a hybrid neural intelligent system to predict enterprise failure in SMEs. It applies a hybrid reasoning system specifically designed to analyze data from enterprises and predict potential failures, in order to provide an innovative hybrid neural method for exploring the failure prediction process and extracting knowledge in the form of rules which help the human experts to understand the prediction process and to obtain conclusions about the relevance of the situation of the enterprise. If the system does not use information from banks or policy makers, is because the model proposed in this article is based on a previous study that does not take these considerations into account.51 However, we are working in an improvement of the system that takes into consideration information about financial crisis and external factors. Nowadays, it is highly recommended for the SMEs to have available tools for predicting potential risky situations and have a better understanding of their internal functioning. The approach presented in this article provides the SMEs with a decision support tool that can contribute to predict failures and identify malfunctioning, as demonstrated in the experimental results. Management staff are, in general, reticent about trusting Artificial Intelligencebased systems, but the incorporation of the CBR paradigm helps to solve this situation by using past experiences. The proposed approach was applied to different SMEs (taking into account the particular problem/casuistic of each SME), and we have shown a high percentage of success to improve the behavior of the SMEs that participated in the experiments. However, it is not possible to extract general conclusions (i.e. All the SMEs should revise task X in Activity X), because this is not the aim of the proposed model Although the defined model has not been tested in big enterprises, we believe that it could work adequately, although changes would take place more slowly than in SMEs. We are moving in this direction and expect that an evaluation of the system will soon be possible in a major international company from the textile sector. Moreover, additional

July 27, 2011 15:23 S0129065711002833

292

M. L. Borrajo et al.

work is still required to improve the adaptation of the CBR-BDI agents when the size of the memory of cases is very high and to learn from the cases with invalid solutions. Besides, it is necessary to increase the sample size by incorporating more SMEs to the experiments as well as explore different sets of samples. These are our next challenges.

Acknowledgments This research has been partially supported through the MICINN under project IDI-20110343; project of the Spanish Ministry of Science and Innovation TIN2010-21272-C02-01 (funded by the European Regional Development Fund). The authors would also like to thank the vehicle interior manufacturer, Grupo Antolin Ingenieria S.A., under project MAGNO2008 — 1028. — CENIT Project funded by the Spanish Government.

References 1. A. Khashman, A neural network model for credit risk evaluation, International Journal of Neural Systems 19(4) (2009) 285–294. 2. J. Sun and H. Li, Financial distress early warning based on group decision making, Computers & Operations Research 36(3) (2009) 885–906. 3. H. Li and J. Sun, On performance of case-based reasoning in Chinese business failure prediction from sensitivity, specificity, positive and negative values, Applied Soft Computing, in press. 4. H. Li, H. B. Huang, J. Sun et al., On sensitivity of case-based reasoning to optimal feature subsets in business failure prediction, Expert Systems with Applications, in press. 5. J. Sun and H. Li, Financial distress prediction based on serial combination of multiple classifiers, Expert Systems with Applications 36(4) (2009) 8659–8666. 6. J. Sun and H. Li, Listed companies’ financial distress prediction based on weighted majority voting combination of multiple classifiers, Expert Systems with Applications 35(3) (2008) 818–827. 7. H. Li and J. Sun, Ranking-order case-based reasoning for financial distress prediction, KnowledgeBased Systems 21(8) (2008) 868–878. 8. J. Sun and H. Li, Data mining method for listed companies’ financial distress prediction, Knowledge-Based Systems 21(1) (2008) 1–5. 9. J. M. Corchado, F. D´ıaz, L. Borrajo and F. Fdez-Riverola, Redes Neuronales Artificiales: Un Enfoque Pr´ actico, Departamento de publicaciones de la Universidad de Vigo, Spain (2000).

10. H. Li and J. Sun, Gaussian case-based reasoning for business failure prediction with empirical data in China, Information Sciences 179(1–2) (2009) 89–108. 11. J. Bajo, Y. De Paz, J. F. De Paz and J. M. Corchado, Integrating case planning and RPTW neuronal networks to construct an intelligent environment for health care, Expert Systems with Applications 36(3–2) (2009) 5844–5858. 12. J. Kolodner, Case-Based Reasoning Morgan Kaufmann, San Francisco (1993). 13. T. Kohonen, Self-Organization and Associative Memory, Springer, Berlin (1984). 14. B. Baruque and E. Corchado, A weighted voting summarization of SOM ensembles, Data Mining and Knowledge Discovery 21(3) (2010) 398–426. 15. E. Koskivaara, Artificial Neural Networks in Auditing: State of the Art. TUCS Technical Report No 509 (2003). 16. AICPA, The AICPA Audit Committee Toolkit. c 2005 by the American Institute of Copyright  Certified Public Accountants, Inc., New York, New York (2005). 17. A. S´ anchez, Los Sistemas Expertos en la Contabilidad, Biblioteca Electr´ onica de Contabilidad 1(2) (1995). 18. J. L. Colbert, Risk, Internal Auditor 3 (1995) 19. S. Ramamoorti, A. D. Bailey and R. O. Traver, Risk assessment in internal auditing: A neural network approach, International Journal of Intelligent Systems in Accounting, Finance & Management 8(3) (1999) 159–180. 20. J. R. Coakley and C. E. Brown, Artificial neural networks applied to ratio analysis in the analytical review process, in Neural Networks in Finance and Investing: Using Artificial Intelligence to Improve Real-World Performance, R. R. Trippi and E. Turban (eds), Irwin, Chicago, IL (1996) 123–162. 21. J. Bedard and S. F. Biggs, Pattern recognition, hypothesis generation, and auditor performance in an analytic task, The Accounting Review 66(3) (1991) 622–643. 22. R. Libby, Availability and the generation of hypotheses in analytic review, Journal of Accounting Research 23 (1985) 648–667. 23. A. Aamodt and E. Plaza, Case-based reasoning: Foundational issues, methodological variations, and system approaches, AI Commun. 7(1) (1994) 39–59. 24. B. W. Morris, SCAN: A case-based reasoning model for generating information system control recommendations, Intelligent Systems in Accounting, Finance and Management 3 (1994) 47–63. 25. C. Jung, I. Han and B. Suh, Risk analysis for electronic commerce using case-based reasoning, Int. J. Intell. Sys. Acc. Fin. Mgmt. 8 (1999) 61–73.

July 27, 2011 15:23 S0129065711002833

Hybrid Neural Intelligent System to Predict Business Failure

26. A. M. Salem, Case based reasoning technology for medical diagnosis, World Academy of Science, Engineering and Technology 31 (2007). 27. H. Li and J. Sun, Forecasting business failure in China using hybrid case-based reasoning, Journal of Forecasting, in press. 28. H. Li and J. Sun, Hybridizing principles of the Electre method with case-based reasoning for data mining, European Journal of Operational Research 197(1) (2009) 214–224. 29. H. Li and J. Sun, Business failure prediction using hybrid case-based reasoning, Computers & Operations Research 37(1) (2010) 137–151. 30. H. Li and J. Sun, Predicting business failure using multiple case-based reasoning combined with support vector machine, Expert Systems with Applications 36(6) (2009) 10085–10096. 31. H. Li and J. Sun, Majority voting combination of multiple case-based reasoning for financial distress prediction, Expert Systems with Applications 36(3) (2009) 4363–4373. 32. H. Li, J. Sun and B. Sun, Financial distress prediction based on OR-CBR in the principle of k-nearest neighbors, Expert Systems with Applications 36(1) (2009) 643–659. 33. L. Ardissono, G. Petrone and M. Segnan, A conversational approach to the interaction with web services, Computational Intelligence 20 (2004) 693–709. 34. L. O. Bonino da Silva, F. Ramparany, P. Dockhorn, P. Vink, R. Etter and T. Broens, A service architecture for context awareness and reaction provisioning, in Proc. IEEE Congress on Services (2007) 25–32 35. A. Ricci, C. Buda and N. Zaghini, An agentoriented programming model for SOA & web services, in Proc. 5th IEEE Int. Conf. Industrial Informatics (Vienna, Austria, 2007), pp. 1059– 1064. 36. X. Liu, A multi-agent-based service-oriented architecture for inter-enterprise cooperation system, in Proc. 2nd Int. Conf. on Digital Telecommunications (IEEE Computer Society, Washington, DC, 2007). 37. N. R. Jennings and M. Wooldridge, Applying agent technology, Applied Artificial Intelligence 9(4) (1995) 351–361. 38. M. E. Bratman, D. Israel and M. E. Pollack, Plans and resource-bounded practical reasoning, Computational Intelligence 4 (1988) 349–355. 39. A. Pokahr, L. Braubach and W. Lamersdorf, Jadex: Implementing a BDI-Infrastructure for JADE Agents, in EXP — In Search of Innovation (Special Issue on JADE ), Department of Informatics, University of Hamburg, Germany (2003), 76–85. 40. R. Laza, R. Pav´ on and J. M. Corchado, A Reasoning Model for CBR BDI Agents Using an Adaptable Fuzzy Inference System, in R. Conejo,

41.

42.

43.

44.

45.

46.

47.

48.

49.

50.

51.

52.

53.

293

M. Urretavizcaya and J.-L. P. De-La Cruz (eds.). Springer (2003). J. M. Corchado and R. Laza, Constructing deliberative agents with case-based reasoning technology, International Journal of Intelligent Systems 18 (12) (2003) 1227–1241. Y. Y. Chen and K. Y. Young, An SOM-based algorithm for optimization with dynamic weight updating, Int. Journal of Neural Systems 17(3) (2007) 171–181. E. Corchado, B Baruque and H. Yin, Boosting unsupervised competitive learning ensembles, Proceedings ICANN, Part I. Lecture Notes in Computer Science, Vol. 4668, Springer (2007), pp. 339–348. B. Baruque, E. Corchado, A. Mata and J. M. Corchado, Ensemble methods for boosting visualization models, Proceedings IWANN, Part I. Lecture Notes in Computer Science, Vol. 5517, Springer (2009), pp. 165–173 B. Baruque, E. Corchado, A. Mata and J. M. Corchado, A forecasting solution to the oil spill problem based on a hybrid intelligent system information sciences, 180(10) (2010) 2029–2043; B. Fritzke, Fast Learning with Incremental RBF Networks, Neural Processing Letters 1(1) (1994) 5–2. R. Savitha, S. Suresh and N. Sundararajan, A fully complex-valued radial basis function network and its learning algorithm, Int. Journal of Neural Systems 19(4) (2009) 253–267. J. F. Qiao and H. G. Han, A repair algorithm for radial basis function neural networks and its applicatoin to chemical oxygen demand modeling, Int. Journal of Neural Systems 20(1) (2010) 63–74. G. H. Lee, Rule-based and case-based reasoning approach for internal audit of bank, KnowledgeBased Systems 21(2) (2008) 140–147. N. Cercone, A. Aijun and C. Chan, Rule-induction and case-based reasoning: Hybrid architectures appear advantageous, IEEE Transactions on Knowledge and Data Engineering 11(1) (1999) 166–174. C. Marling, E. Rissland and A. Aamodt, Integrations with case-based reasoning, The Knowledge Engineering Review 20(3) (2006) 241–245. J. C. Ya˜ nez, Importancia del sistema de Control Interno en la Auditor´ıa Legal, Contrastes Emp´ıricos, PhD Thesis, Universidade de Vigo, Spain (2003). H. B. Mann and D. R. Whitney, On a test of whether one of two random variables is stochastically larger than the other, The Annals of Mathematical Statistics 18 (1947) 50–60. M. L. Borrajo, J. M. Corchado, C. Yanez, C., F. Fdez-Riverola and F. Diaz, Autonomous internal control system for small to medium firms, in Proc.

July 27, 2011 15:23 S0129065711002833

294

54.

55.

56.

57.

58.

59.

60.

61. 62.

63.

64. 65.

66.

M. L. Borrajo et al.

ICCBR 2005, Lecture Notes in Computer Science, Vol. 3620, Springer (2005), pp. 106–121. J. M. Corchado, M. L. Borrajo, M. A. Pellicer and C. Y. Yanez, Neuro-symbolic system for business internal control, in Proc. ICDM 2004, Lecture Notes in Computer Science Vol. 3275, Springer (2005), pp. 1–10. R. V. Mayorga and J. Carrera, A radial basis function network approach for the computational of inverse continuous time variant functions, International Journal of Neural Systems 17(3) (2007) 149–160. W. Pedrycz, R. Rai and J. Zurada, Experienceconsistent modeling for radial basis function neural networks, International Journal of Neural Systems 18(4) (2008) 279–292. Q. Junfei and H. Honggui, A repair algorithm for radial basis function neural network with application to chemical oxygen demand modeling, International Journal of Neural Systems 20(1) (2010) 63–74. A. Fatehi and K. Abe, Flexible structure multiple modeling using irregular self-organizing maps neural network, International Journal of Neural Systems 18(4) (2008) 347–370. M. Oja, G. O. Sperber, J. Blomberg and S. Kaski, Self-organizing map-based discovery and visualization of human endogenous retroviral sequence groups, International Journal of Neural Systems 15(3) (2005) 163–179. N. Keeratipranon and F. D. Maire and H. Huang, Manifold learning for robot navigation, International Journal of Neural Systems 16(5) (2006) 383–392. L. Breiman, Bagging predictors, Machine Learning 24 (1996) 123–140. D. Ruta and B. Gabrys, Classifier selection for majority voting, Information Fusion 6 (2005) 63–81. S. Kaski and K. Lagus, Comparing self-organizing maps, in International Conference on Artificial Neural Networks (ICAAN’96), Springer, Lecture Notes in Computer Science, Springer, Bochum, Germany (1996) 809–814 T. Kohonen, Self-Organizing Maps, Springer Series in Information Sciences, Springer (1995), 30. T. Kohonen, S. Kaski, K. Lagus, J. Saloj¨ arvi, J. Honkela, V. Paatero, E. Oja and S. Kaski (eds.), Self organization of a massive text document collection Kohonen Maps, Elsevier (1999), 171–182. K. Lagus, Text Mining with the WEBSOM, in Acta Polytechnica Scandinavica, Mathematics and Computing Series no. 110, Espoo 2000, 54 pp. Published by the Finnish Academy of Technology. ISBN 951-666-556-X. ISSN 1456–9418. UDC 004.032.26:025.4.03:004.5.

67. F. Chang, C. C. Lin and C. J. Lu, Adaptive prototype learning algorithms: Theoretical and experimental studies, J. Mach. Learn. Res 7 (2006) 2125–2148. 68. J. Gero and U. Kannengiesser, Agent-based interoperability without product model standards, Computer-Aided Civil and Infrastructure Engineering 22(3) (2007) 80–97. 69. D. Monticolov, V. Hilaire, S. Gomes and A. Koukam, A multi-agent system for building project memories to facilitate design process, Integrated Computer-Aided Engineering 15(1) (2008) 3–20. 70. H. C. Chen, M. Goldberg, M. Magdon-Ismail and W. A. Wallace, Reverse engineering a social agentbased hidden Markov model, International Journal of Neural Systems 18(6) (2008) 491–526. 71. D. L´ opez-Par´ıs and A. Braz´ alez-Guerra, A new autonomous agent approach for the simulation of pedestrians in urban environments, Integrated Computer-Aided Engineering 16(4) (2009) 283–297. 72. J. Bajo, J. F. De Paz, S. Rodr´ıguez and A. Gonz´ alez, Multi-agent system to monitor oceanic environments, Integrated Computer-Aided Engineering 17(2) (2010) 131–144. 73. I. Marti, V. R. Tomas, L. A. Garcia and J. J. Martinez, A Multi-agent system for managing adverse weather situations on the road network, Integrated Computer-Aided Engineering 17(2) (2010) 145–155. 74. J. M. Fern´ andez de Alba and J. Pav´ on, Talking agents: A distributed architecture for interactive installation-artworks, Integrated Computer-Aided Engineering 17(3) (2010) 243–259. 75. B. Chen and W. Liu, Mobile agent computing paradigm for building a flexible structural health monitoring sensor network, Computer-Aided Civil and Infrastructure Engineering 25(7) (2010) 504–516. 76. S. Ceppi and N. Gatti, An algorithmic game theory framework for the study of wholesale electricity markets, Integrated Computer-Aided Engineering 17(4) (2010) 273–290. 77. M. Prymek and A. Horak, Multi-agent approach to power distribution network modelling, Integrated Computer-Aided Engineering 17(3) (2010) 291–303. 78. M. J. Santofimia, X. del Toro, P. Roncero, F. Moya, M. A. Martinez and J. C. Lopez, A qualitative agent-based approach to power quality monitoring and diagnosis, Integrated Computer-Aided Engineering 17(4) (2010) 305–319. 79. S. Mullen and G. Onsongo, Decentralized agentbased underfrequency load shedding, Integrated Computer-Aided Engineering 17(4) (2010) 321–329.

July 27, 2011 15:23 S0129065711002833

Hybrid Neural Intelligent System to Predict Business Failure

80. J. Solanki, S. K. Solanki and N. Schulz, Multi-agent-based reconfiguration for restoration of distribution systems with distributed, Integrated Computer-Aided Engineering 17(4) (2010) 331–346. 81. Q. Yu, N. Schulz and A. K. Srivastava, Multi-agent based reconfiguration of AC-DC shipboard distribution power system, Integrated Computer-Aided Engineering 17(4) (2010) 347–357. 82. L. Cristaldi, F. Ponci, M. Riva and M. Faifer, Multi-agent systems: An example of power system dynamic reconfiguration, Integrated ComputerAided Engineering 17(4) (2010) 359–372. 83. R. Badaway, B. Hirsch and S. Albayrak, Agent-based coordination techniques for matching supply and demand in energy networks, Integrated Computer-Aided Engineering 17(4) (2010) 373–382. 84. A. Karim and H. Adeli, CBR model for freeway work zone traffic management, Journal of Transportation Engineering 129(2) (2003) 134–145. 85. A. Waheed and H. Adeli, Case-based reasoning in steel bridge engineering, Knowledge–Based Systems 18(1) (2005) 37–46. 86. G. F. Jr. Sirca and H. Adeli, Case-based reasoning for converting working stress design-based bridge ratings to load factor design-based ratings, Journal of Bridge Engineering 10(4) (2005) 450–459. 87. Q. Junfei and H. Honggui, A repair algorithm for radial basis function Neural network with application to chemical oxygen demand modeling, International Journal of Neural Systems 20(1) (2010) 63–74. 88. P. Patrinos, A. Alexandridis, K. Ninos and H. Sarimveis, Variable selection in nonlinear modeling based on RBF networks and evolutionary computation, International Journal of Neural Systems 20(5) (2010) 365–379. 89. D. Wu, K. Warwick, Z. Ma, M. N. Gasson, J. G. Burgess, S. Pan and T. Z. Aziz, Prediction of Parkinson’s disease tremor onset using radial basis function neural network based on particle swarm optimization, International Journal of Neural Systems 20(2) (2010) 109–116. 90. P. Anand, B. V. N. Siva Prasad and C. Venkateswarlu, Modeling and optimization of a pharmaceutical formulation system using radial basis function network, International Journal of Neural Systems 19(2) (2009) 127–136. 91. R. Savitha, S. Suresh and N. Sundararajan, A fully complex-valued radial basis function network and its learning algorithm, International Journal of Neural Systems 19(4) (2009) 253–267. 92. H. Adeli and A. Karim, Fuzzy-wavelet RBFNN model for freeway incident detection, Journal of Transportation Engineering 126(6) (2000) 464–471.

295

93. A. Karim and H. Adeli, Comparison of the fuzzy — wavelet RBFNN freeway incident detection model with the California algorithm, Journal of Transportation Engineering 128(1) (2002) 21–30. 94. A. Karim and H. Adeli, Radial Basis function neural network for work zone capacity and queue estimation, Journal of Transportation Engineering 129(5) (2003) 494–503. 95. S. Ghosh-Dastidar, H. Adeli and N. Dadmehr, Principal component analysis-enhanced cosine radial basis function neural network for robust epilepsy and seizure detection, IEEE Transactions on Biomedical Engineering 55(2) (2008) 512–518. 96. H. C. Lian, No-reference video quality measurement with support vector regression, International Journal of Neural Systems 19(6) (2009) 457–464. 97. Q. Cong, W. Yu and T. Chai, Cascade process modeling with mechanism-based hierarchical neural networks, International Journal of Neural Systems 20(1) (2010) 1–11. 98. Y. Yang and B. L. Lu, Protein subcellular multilocalization prediction using a min-max modular support vector machine, International Journal of Neural Systems 20(1) (2010) 13–28. 99. X. Jiang, S. Mahadevan and H. Adeli, Bayesian wavelet packet denoising for structural system identification, Structural Control and Health Monitoring 14(2) (2007) 333–356. 100. H. Adeli and X. Jiang, Neuro-fuzzy logic model for freeway work zone capacity estimation, Journal of Transportation Engineering 129(5) (2003) 484–493. 101. A. B. Senouci and H. Adeli, Resource scheduling using neural dynamics model of adeli and park, Journal of Construction Engineering and Management 127(1) (2001) 28–34. 102. F. Wu, F. Sun and H. Liu, A dual-model jumping fuzzy system approach to networked control systems design, International Journal of Neural Systems 20(1) (2010) 51–62. 103. J. A. Iglesias, P. Angelov, A. Ledezma and A. Sanchis, Human activity recognition based on evolving fuzzy systems, International Journal of Neural Systems 20(5) (2010) 355–364. 104. P. Patrinos, A. Alexandridis, K. Ninos and H. Sarimveis, Variable selection in nonlinear modeling based on RBF networks and evolutionary computation, International Journal of Neural Systems 20(5) (2010) 365–379. 105. R. Scherer, Designing boosting ensemble of relational fuzzy systems, International Journal of Neural Systems 20(5) (2010) 381–388. 106. N. Wang, M. Joo, X. Y. Meng and X. Li, An online self-organizing scheme for parsimonious and accurate fuzzy neural networks, International Journal of Neural Systems 20(5) (2010) 389–403.

July 27, 2011 15:23 S0129065711002833

296

M. L. Borrajo et al.

107. X. Li, M. J. Er, B. S. Lim, J. H. Zhou, O. P. Gan and L. Rutkowski, Fuzzy regression modelling for tool performance prediction and degradation detection, International Journal of Neural Systems 20(5) (2010) 405–419. 108. J. Liang, Z. Wang and X. Liu, Global synchronization in an array of discrete-time neural networks with mixed coupling and time-varying delays, International Journal of Neural Systems 19(1) (2009) 57–63. 109. H. Adeli and A. Karim, Neural network model for optimization of cold-formed steel beams, Journal of Structural Engineering 123(11) (1997) 1535–1543. 110. H. Adeli and A. Karim, Scheduling/cost optimization and neural dynamics model for construction, Journal of Construction Management and Engineering 123(4) (1997) 450–458.

111. A. Karim and H. Adeli, Fast automatic incident detection on urban and rural freeways using wavelet energy algorithm, Journal of Transportation Engineering 129(1) (2003) 57–68. 112. H. Adeli and H. S. Park, Counter propagation neural network in structural engineering, Journal of Structural Engineering 121(8) (1995) 1205–1212. 113. H. Adeli and H. S. Park, A neural dynamics model for structural optimization — theory, Computers and Structures 57(3) (1995) 383–390. 114. H. Adeli and H. S. Park, Optimization of space structures by neural dynamics, Neural Networks 8(5) (1995) 769–781.