Management and forecast of dynamic customer ...

Advanced Engineering Informatics 24 (2010) 96–106

Contents lists available at ScienceDirect

Advanced Engineering Informatics journal homepage: www.elsevier.com/locate/aei

Management and forecast of dynamic customer needs: An artificial immune and neural system approach Yih Tng Chong, Chun-Hsien Chen * School of Mechanical and Aerospace Engineering, Nanyang Technological University, 50 Nanyang Avenue, Singapore 639798, Singapore

a r t i c l e

i n f o

Article history: Received 28 October 2008 Received in revised form 4 June 2009 Accepted 15 June 2009 Available online 23 July 2009

a b s t r a c t The twenty-first century is marked by fast evolution of customer tastes and needs. Research has shown that customer requirements could vary in the temporal space between product conceptualization and market introduction. In such cases, the products generated might not fit the consumer needs as companies originally expected. This paper advocates the proactive management and forecast of the dynamic customer requirements in bid to lower the inherent risk in developing products for fast shifting markets. The research identified the principles of artificial immune and neural systems as a solution to the problem. A customer requirements analysis and forecast (CRAF) system is defined in this paper to address the issue. The system aims to support product development functions with quantitative and qualitative customer requirements information, in the pursuit of generating products for near future markets. A case study is presented in this article to illustrate the functions of the system. Ó 2009 Elsevier Ltd. All rights reserved.

1. Introduction While uncertainties in businesses can be managed to produce competitive advantages, they are also threats if left unaccounted. In the twenty-first century, many new product development (NPD1) businesses face the uncertainties of highly turbulent market. Blocker and Flint [1] found that the changing needs and rapidly evolving preferences of customers represent the key drivers of market turbulence. It is therefore sound to assert that customers are the key source of information in the bid to reduce uncertainties inherent in NPD projects [2]. Meeting or exceeding customer requirements is the ultimate target of total quality management [3], and this class of market information is an important capital for capturing product markets in competitive environments. The management of customer requirement information is generally concerned with requirement elicitation, analysis and specification [4]. While the processes of requirement management in product development have been relatively well-investigated, there have been only few studies that considered the temporal dimension of the critical information. The status is similar in the industry. Studies have shown that companies in general have neglected this aspect in product development [5,6]. In volatile markets, for instance the cellular phone market, it is imprudent for one to comment on the validity of customer requirement data without making any reference to the * Corresponding author. Address: School of Mechanical and Aerospace Engineering, Nanyang Technological University, North Spine (N3), Level 2, 50 Nanyang Avenue, Singapore 639798, Singapore. Tel.: +65 6790 4888; fax: +65 6792 4062. E-mail address: [email protected] (C.-H. Chen). 1 A list of acronyms is provided in the appendix (Table A). 1474-0346/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.aei.2009.06.003

frames of time. As the issue of dynamic customer requirements is increasingly valid due to such factors as consumer sophistication and competition, it is urgent and critical to recognise the issue in both practice and research. 2. The issue of dynamic customer requirements Requirement is the basis of product development, and it can vary with time. The variation of customer satisfaction attribute weights along the temporal dimension was elucidated in a longitudinal study performed by Mittal et al. [7]. Calantone and Sawyer [8] demonstrated the instability of market segments, which manifested the changing of consumer needs over time. Prior to ramping up production, design specifications are by necessity frozen. In cases where customer requirements shift substantially during the period between design freeze and market introduction, the final products may not satisfy the customers as intended. Factors that affect the variation of requirements include the length of production period, market volatility and competition intensity. Product developers, if unwary of the variable, may end up generating products not wanted by the customers. In another scenario where organizations react to changing requirements by modifying design specifications, undesirable delays in schedules can be resulted. The situation was identified in a study by Kärkkäinen and Elfvengren [9] as self-reinforcing ‘‘vicious” cycles in product development processes. 2.1. The significances of dynamic customer requirements There has been research that indicates the implications of shifting customer requirements on NPD successes. It was found in an

Y.T. Chong, C.-H. Chen / Advanced Engineering Informatics 24 (2010) 96–106

empirical analysis that a series of managerial problems could be resulted when the assessment of future customer needs has not been given adequate attention [9]. Design specifications based on customers’ future needs can reduce design iterations and reworks, as indicated by Kärkkäinen et al. [5]. The selection of research and development (R&D) projects requires foresight and good grasp of the market dynamics. When future customer needs could not be clearly identified, the return of investment of R&D projects would be of higher uncertainty, which might in turn undermine the overall NPD efforts [6]. In a similar vein, it is important for marketers to understand the evolving markets. Revision of marketing strategies, such as product repositioning and advertising appeals revamping [8] may be called for as the responses to the shifting market needs. In the manufacturing context where customer requirements are in the state of permanent flux [10], production planning is an extremely difficult task [11]. As shown by various studies, the dynamism of customer requirements bears implications on product development successes, ramified through such functions as design, marketing, production and R&D. Research in the area of dynamic customer needs will therefore be of considerable significances to the multi-disciplinary aspects of NPD. 2.2. A review of the current solutions Research of requirement management to date has preliminarily considered the temporal dimension of the information. Studies that have focused on or at least considered the time-based variation of customer requirements have offered two classes of solutions in general. The first approaches the problem by advocating sensitivity to the dynamic needs, while the second proposes methodologies that forecast future needs. 2.2.1. Sensitivity to the changing needs As customer requirements are in many cases dynamic, researchers deemed that customer satisfaction should be measured and assessed regularly. Reichwald et al. [12] proposed a distributed mini-factory organization that can improve firms’ sensitivity to requirement changes, essentially by the virtue of close proximity to customers. Apart from physically moving closer to the customers, Internet has been a common tool proposed for customer relationship management [13]. Modern technology makes it incredibly straightforward for companies to gather vast amount of data concerning individuals and their habits on daily basis [14,15]. Such data can be employed for pattern and trend tracking to gain competitive advantages. For instance, information on the usages of credit cards and supermarket loyalty cards is flowing in a continuous manner, and timely analysis would require a dynamic system [14]. Gunasekaran et al. [16] proposed a web-enabled quality function deployment (QFD) model that facilitates the continuous communication of customer needs information across geographical dispersed supply chain. A patent [17] describes the idea of having microprocessors embedded in products for bilateral communications between product developers and customers. The patent claims that changing customer requirements can be sensitively detected based on the method [18]. Wu and Shieh [19] proposed the application of Markov chain model on QFD to monitor the dynamism of customer requirements from probabilistic viewpoint. Similarly, a dynamic QFD (as opposed to the traditional static versions) was proposed to handle the constantly evolving customer needs [20]. The paradigm of agile manufacturing advocates continual assessments of customer requirements, as well as production flexibility that can response effectively to requirement changes [10]. 2.2.2. Forecasting of changing customer needs Reichwald et al. [12] commented that traditional market research methodologies focus only on current situation and often

97

do not contribute to the correct assessment of future customer requirements. Flores [21] and Shen et al. [22] separately proposed methods that involve seeking consumers’ opinions on their future requirements. While such approach is useful when historical data is unavailable for projections, the soundness of the results hinges on the sampled customers’ abilities to forecast the population’s future requirements. With historical data, time-series methods can be employed for forecasting. Xie et al. [23] employed the double exponential smoothing technique in projecting the importance levels of the requirements. The method is limited to forecasting quantitative data, and only of linear trend. Wu et al. [24] proposed the use of grey theory in forecasting, which requires only four past data points. Raharjo et al. [25] developed a method that prioritises quality characteristics that have greater confidence in meeting future customer requirements, in the context of the dynamic QFD. The method per se does not generate future customer requirement information, but requires it as input. Chen and Yan [26] analysed customer preferences using radial basis function neural network. While the method does not specifically ascertain future customer needs, it predicts customer preferences over a range of product options. Ha [27] proposed using knowledge engineering technique to analyse and predict the shifting of customers across market segments. 2.3. Moving forward In the studies of dynamic customer requirements, the intentions were in general to counter the uncertainties that the variable contributes to NPD. A study by Kärkkäinen and Elfvengren [9] recommended companies to develop better abilities to recognize customers’ future needs. Having noted that most approaches proposed have been reactive, Blocker and Flint [1] specifically recommended future research to develop tools that forecast the direction and rate of change of market segments. Kahn et al. [28] similarly suggested forecasting (an under-researched area) as a means of addressing uncertainties in product development. In this light, this article defines an approach that forecasts customer needs. Customer requirement information consists of both qualitative and quantitative aspects [23]. The former represents the types (or objects) of requirements, while the latter represents the degree of importance or preferences. The quantitative dimension has been the variable of interest focused in previous studies (e.g. Refs. [23,24]). Szakonyi [6] recommended that customer future needs should be identified in terms of new product characteristics (i.e. qualitative terms). As such, a solution that can handle both quantitative and qualitative variation will be valuable in addressing the issue. It is known that new customer requirements emerge while current ones obsolete over time. A common limitation of the current methodologies reviewed above is the necessity of continuous human interventions to update newly emerged needs and to strike out out-dated requirements. For instance, in the case of the dynamic QFD proposed by Adiano and Roth [18], manual reviews of customer surveys are required in the process of updating the matrix. Therefore, a dynamic system that autonomously learns both the quantitative and qualitative aspects of the data will be desirable for the benefits of NPD organizations. A review of literature of customer requirement management by Jiao and Chen [4] shows that the research of dynamic customer needs has not been addressed. Likewise in the industry, product developers generally paid little attention to identifying future customer needs [5]. Szakonyi [6] highlighted marketers’ lack of analytical skills as a plausible impedance to ascertaining future needs. It is clear that the research necessary to produce relevant theories, methodologies and tools has been modest in both areas of marketing [1,7] and engineering [23,26].

98


Given the increasingly fast moving product markets, research and development effort in the area is ever more valid. This work advocates product development organizations to sensitively stay in touch with the changing needs of customers, and to proactively identify future requirements. In view of the problem and the research boundaries described above, this article introduces an approach that autonomously manages and forecasts customer needs information. 3. The applicability of artificial immune and neural system Artificial immune system (AIS) is unique in a few aspects, which render it highly applicable to the management of dynamic customer requirements. Immune system is described as adaptive, self-organizing in nature, maintaining a memory of past encounters, and has the ability to continually learn via new encounters [29,30]. An intrinsic and unique ability of the AIS is its continuous capability to adapt to and to co-evolve with the environment [31]. This aspect, which is a requisite for addressing the issue of dynamic customer needs, is clearly distinct from other computational intelligence (CI) paradigms [14]. Specifically, the applications of ‘‘static” algorithm repeatedly, such as ANN (artificial neural network) and EA (evolutionary algorithm), may disregard previous data in time, thus missing out vital clues that are only factorable across time [15]. For instance, the clonal selection algorithm of the AIS paradigm was applied to deal with dynamic email classification [32,33]. Population-based CI techniques such as EA and swarm algorithm typically produce the fittest solution [34], while AIS models could treat the entire population as the solution [29]. The self-organized repertoire set in the AIS paradigm is therefore especially suited to represent an array of customer needs for a given product. Research has highlighted the abilities of AIS-based data analysis methods [14,35,36]. Several studies have been performed on AISbased learning systems that operate in continuously evolving environments, including for dynamic optimisation [37] and dynamic clustering of time-varying data [14,38–40]. Based on the established unsupervised machine learning algorithms, this work defines a domain specific dynamic population-based clonal selection algorithm in Section 4.2. While the AIS paradigm involves stochastic mechanisms such as somatic hypermutation, they may not be applicable to all problems. The considerations include the representation scheme and the role of the AIS with respect to the given problem (see Aickelin and Dasgupta [29]). In this context, where mutating the represented data is not applicable, the proposed system will be deterministic, apart from the random initial weights in the embedded ANN introduced below. Customer needs data can be non-linear, non-stationary, noisy and limited in quantity, all which pose challenge for the timebased learning machine – AIS. An artificial neural network (ANN), in particular the focused time delay neural network (FTDNN) is therefore embedded in the proposed AIS-based system. The FTDNN is described as focused since the short-termed memory is only at the filtering input layer. Static backpropagation is suited for network training in this case. Based on the input signal, the memory structure transforms samples into points within the input space. Time delay neural network, first applied by Waibel [41] for speech recognition, is a feedforward network with each synapse of the network designed as the finite impulse response filters. The network is advantageous when applied to characterize and forecast dynamic customer requirements, as discussed below:

which have limitations in modelling complex systems, ANNs are more general and flexible, therefore performing better. Traditional time-series models such as Box-Jenkins and ARIMA (autoregressive integrated moving average) methods assume that time-series under study are generated from linear processes [43]. As such, ANN models usually outperform these traditional techniques when time-series are non-linear and noisy [44]. As it is clearly neither reasonable nor practical to assume that customer needs variations are linear over time, direct applications of such traditional methods would be unsuitable. Non-linear models such as TAR (threshold autoregressive model) and state-space models require the model types and complexities predefined. Like many types of time-series in the real world, the a priori models of customer needs variations are not known at the outsets. The ANN approach does not require the prior assumption of function forms and is therefore more applicable in this study. Targeted training: Compared to unsupervised ANNs, FTDNN’s supervised backpropagation training is able to approximate the model in a targeted manner given the training input data. Time-based characterization: Neural network can be broadly classified into static and dynamic categories. Static networks per se, such as the multi-layer perceptron (MLP), suffer from limitation in modelling dynamic systems where time delays affect the dynamism [45]. FTDNN on the other hand bears topologies that explicitly address the dynamic input–output space. Specifically, FTDNN is a dynamic ANN that has short-term memory or delay elements built within the learning machine, differing from static networks that are at times coupled to windowed input signal. Due to this feature, FTDNN can be implemented to independently adjust the delay size and weights where applicable. Generally, when time imposes structures in the input space, dynamic FTDNN is a more rigorous approach as compared to static networks. Usability: Amongst the types of dynamic ANNs, FTDNN is reported as a balanced network for both processing power and simplicity [46]. In the light of the dynamic customer needs problem and the solutions (presented in Sections 2 and 3, respectively), Section 4 describes the engagement of the issue based on the principles of artificial immune and neural system. 4. A customer requirements analysis and forecast system A customer requirements analysis and forecast (CRAF) system is proposed to address the issue of dynamic customer requirement. Fig. 1 illustrates the system framework in the context of customer requirement management, in terms of three distinct phases – elicitation, analysis and application or specification as generalized by Jiao and Chen [4]. With regards to the elicitation phase, Chen et al. [47] proposed and investigated a customer-oriented approach to customer requirement elicitation. On the other end of the analysis phase, there has also been research that deals with problems of downstream processes such as design (see Ref. [48]). In the present study, the focus is on the analysis of the customer needs, which aims to produce vital information for downstream applications, as shown in Fig. 1. In Section 4.1, the data representation scheme for dynamic customer requirements is specified, and in Section 4.2 the algorithm of the CRAF system is presented. 4.1. Data representation of dynamic customer requirements

Function modelling: ANN has been shown to be universal approximators when adequate numbers of hidden neuron are used [42]. Compared to traditional statistical time-series models,

In the CRAF system, dynamic customer requirement information is modelled based on the design space framework (DSF)

99


Customer Requirement Elicitation

User Interface Layer

Design Application Manufacturing

Query Layer AIS

ANN

Application Marketing Application

Database

…

Elicitation

Applications

CRAF System Fig. 1. CRAF system framework in context.

postulated by Chong et al. [49]. DSF is simplified by considering only the customer requirement data objects in the postulated space of three axes – X, Y and Z. The result is a data structure referred to as the requirement space framework (RSF). An instance of the RSF is shown in Fig. 2. For the robustness of the data, customer requirements are represented in multiple levels of abstraction, as described along the Z axis of the RSF. Alternative values can exist for requirements, and details of customers requirements can be further represented by co-requirements. Correspondingly, requirement options are represented and associated along the Y axis while co-requirements are related along the X axis of the RSF (as shown in Fig. 2). Readers may refer to [49] for more details on the background of the data structure. Apart from the three dimensions of the information, the fourth facet of customer needs information, i.e. time, is prescribed to capture the dynamisms. Fig. 2 depicts an example of the RSF of a time instance. In this work, the learning and analysis of the dynamic customer needs data is primarily based on the AIS. As such, the data representation scheme of the CRAF system conforms to the basic conventions in the AIS research (see Refs. [50,51] for the concepts of AIS paradigm). The following describes the mapping of the RSF to the AIS-based data representation scheme. While Table 1 shows the correspondences between an immune system and the

Y

proposed system, Fig. 3 shows an abstraction of the data elements in the environment. As schematically depicted in the figure, customer requirements are represented by two types of data classes – antigens (Ag) and B-cells. In the AIS paradigm, antigens model the incoming data for analysis, while B-cells represent the composition of the system’s knowledge. Accordingly, dynamic customer requirement input data is modelled as Ag that ‘‘infiltrates” the CRAF system, while the system memory of customer needs is represented by B-cells (see Table 1 and Fig. 3). Artificial recognition ball (ARB) is a concept established in the AIS research (see Refs. [38,39]). It is applied in this work to represent groups of similar B-cells within the system memory. t Antigen set of I number at time t is represented as Ag ¼ t t t hAg 1 ; Ag i ; . . . ; Ag I i, and the representation of each antigen is in turn based on a chain of bio-molecules, i.e. Ag ti ¼ hAg ti;1 ; Ag ti;l ; . . . ; Ag ti;L i. Index l, where 1 < l 6 L, represents the specificity levels of the bio-molecules. In the studies of immunology [52,53], specificity index (SI) has been proposed to characterize the repertoire of cells’ ability to discern self from non-self. SI of zero denotes B-cells as universal glue, while SI of value one denotes infinite specificity. In this work, upper limit L denotes the lowest abstraction level in the system, while the highest abstraction level where bio-molecules can reside is denoted by l = 1. Note that Fig. 2 illustrates an

m0

m5

m4

X

m2

m1 m3

m6 m7 m8 Z

m

Bio-molecule

Abstraction plane

Option link

Examples of ARB

Co-requirement link

Fig. 2. An instance of the requirement space framework.

100


Table 1 Modelling based on immune system. Biological immune system

CRAF system

Antigens B-cells/ARBs Physical environment

Customer requirements (in product market) Customer requirements (in system memory) Product market

instance of the RSF with L = 2. Adopting the concept of specificity from immunological studies to model the generality of information – customer needs information in this case – is an originality of this work. A special case of a bio-molecule m0 (see Fig. 2), which exists on the theoretical specificity level of zero (i.e. l = 0), is pre-defined in each RSF. The node represents the root (i.e. the most abstract) requirement of a given product (see root node descriptions in [49]). Memory B-cells in the system at time t is categorically represented as ARBt ¼ hARB1 ; ARBk ; . . . ; ARBK i. As illustrated in Fig. 2, representation of each memory B-cell group is similarly based on a string of bio-molecules, each of which describes the data at distinct level of abstraction, i.e. ARBk ¼ hARBk;1 ; ARBk;l ; . . . ; ARBk;L i. States of ARBs are dichotomy, either active or inactive, i.e. Active(ARBk ) = 1 or 0. Bio-molecules, the building blocks of the system data, are modelled by symbolic attribute–value pairs, ARBk;l = hatt, vali, e.g. ARB3;2 = hcolour, blacki. Undefined (i.e. ‘‘don’t care”) bio-molecules of Ag and ARB are represented using the symbol #, e.g. Ag 53;2 = . The attribute–value pair of the root bio-molecule m0 serves as the tag of the given product or problem, which can be arbitrarily assigned. In this study, the concentration levels of B-cells found in the artificial recognition balls (ARBs) represent the importance levels of attributes and the preference levels of values, of the respective represented customer requirements. Function ConcðARBk ; tÞ outputs the (non-negative) number of B-cells in ARBk at time t. Fig. 4 illustrates a fundamental theory in immunological science [50]. It depicts the fluctuations of antibody concentration as the consequences of the infiltrating antigens. The interacting mechanisms within the immune system involve biological processes such as cloning, apoptosis, and cell activation. Fig. 4 also relates the types of responses, such as the primary and secondary responses [50]. As conveyed in this paragraph, the notion of temporal variation of B-cell population is metaphorically applied in this context to model the dynamism of customer needs. 4.2. The dynamic customer requirements analysis and forecast algorithm The proposed CRAF algorithm is illustrated in Fig. 5. Note that the input data is captured in Step 2, while Step 10 generates the results. The functions of the algorithm are specified as follows:

Antigens

ARB

CRAF System B-cells Environment

Fig. 3. A schematic diagram of the CRAF system.

1. Immune network initialisation: A set of active ARBs with the predetermined concentration level of ARBT (ARB Threshold) 0 at t = 1 is preinstalled in the system, i.e. ARB ¼ hARB1 ; ARBk ; . . . ; ARBK i, where ConcðARBk ; 0j8kÞ = ARBT and Active ðARBk j8kÞ ¼ 1. This step specifies the initial condition of the system (see Neal [39]). 2. Antigens introduction: Data Ag t ¼ hAg t1 ; Ag ti ; . . . ; Ag tI i is introduced to the system at time t. 3. Secondary immune response: Each introduced data packet Ag is presented to the active ARBs for stimulation. As depicted in Fig. 4, the biological immune system launches secondary responses when similar previously encountered antigens are detected. Likewise in the case of the CRAF system, the algorithm computes the levels of stimulation for all current encounters. As understood in theoretical immunology, pathogens can be recognized by more than one type of B-cells, and vice-versa. It follows that the stimulation levels Stk of the ARBs can be computed using Eq. (1).

Stk ¼

I X ½AðAg ti ; ARBk Þ ConcðAg ti Þ

ð1Þ

i¼1

where function A produces the affinity between the two operands, given by the affinity–distance relationship as shown in Fig. 6. Distance function D that quantifies the dissimilarity between an Ag and an ARB is given by Eq. (2).

DðAg ti ; ARBk Þ ¼

L X

al where al ¼ 0 if Ag ti;l ¼ ARBk;l or ARBk;l ¼ #

l¼1

al ¼ 1 if otherwise

ð2Þ

4. Primary immune response: In the stimulation of inactive ARBs in this step, cross-reactivity does not apply, i.e. stimulation in this step requires exact match, as expressed in Eq. (3).

Stk ¼

I X

Conc½Ag ti DðAg ti ; ARBk Þ ¼ 0

ð3Þ

i¼1

5. Apoptosis: The population of B-cells in the system is maintained by the process of apoptosis, i.e. programmed cell death. As in the biological immune system, this step keeps the size of the repertoire finite and tractable. The decay is computed using Eq. (4).

PK DecayðtÞ ¼ d

t k¼1 Sk

K

!

where d is the decay factor:

ð4Þ

6. Concentration update: The concentration level of each ARB is refreshed based on Eq. (5).

ConcðARBk ; tÞ ¼ ConcðARBk ; tÞ þ Stk DecayðtÞ

ð5Þ

7. Antigenic learning: Foreign antigenic datasets that are beyond the recognition of the existing ARBs are learnt by the system. Under the condition that a previously unencountered antigen is sufficiently different from the system’s ARBs to an extent dictated by parameter er (i.e. DðAg ti ; ARBk Þ P er ), a new inactive ARB of the same attributes and values will be registered in the system memory. 8. Activeness update: When the concentration level of an active ARB falls below the ARBT level for a specific period of time, it is deactivated in this phase. The period is termed as persistency period, or PP. On the other hand, inactive ARBs will be activated when the concentration level is maintained above the ARBT level over the specified PP. 9. Neural network training: The latest available set of B-cell concentration data is utilised to train the embedded focused time delay neural network, based on static backpropagation

101


Antibody Concentration

Cross-Reactive Response

Secondary Response

Primary Response

Lag Lag

Response to Ag1

Lag

Response to Ag1

... ...

... Response to Ag2

...

Antigens Ag1, Ag2

Antigen Ag1

Response to Ag1 + Ag3

Antigen Ag1 + Ag3

Time

Fig. 4. Profile of antibody concentration over time domain [50].

Affinity, A 1. Immune Network Initialisation

1.00 0.75

2. Antigens Introduction

A = 1 – (1/L) D

0.50 3. Secondary Immune Response

0.25 0.00

4. Primary Immune Response

0

Distance, D

L

Fig. 6. The affinity–distance relationship.

5. Apoptosis

6. Concentration Update

11. Repertoire Data Production

7. Antigenic Learning

10. Concentration Forecast

8. Activeness Update

9. Neural Network Training

Fig. 5. Flowchart representation of the CRAF algorithm.

algorithm. The numbers of delay, hidden layer and neuron are problem dependent. Experiments revealed that neural networks trained on inputs that are spaced by time interval s would have the capability to forecast s periods ahead, irrespective of the s length [54]. Longer-term forecast of multiple s can be implemented by the iterated multi/single-step approach [44,55]. Concentrations of B-cells (denoted as c) in the respective ARBs are passed to the ANN as vector [ct1 ; . . . ; ctK ]. The outputs of the network are in the format s ^tþs T [^ctþ 1 ; . . . cK ] . 10. Concentration forecast: The trained neural network is simulated to forecast future B-cell concentrations of the respective ARBs. 11. Repertoire data generation: The present and forecasted population datasets are produced in this step. The algorithm is looped (t = t + 1) by proceeding to Step 2. 5. A case study The development of personal computer is challenging due to the fast moving market. The requirements of customers evolve rap-

idly, relative to the span of the development cycle. It was estimated that the value of personal computers depreciate by an average of 50% for every year of use [56]. The trend can be associated with the high rate of new product introduction. Market foresight is in this case critical to the developers in acting and reacting in the industry. Especially in highly competitive market, robust customer requirements information plays key role in such processes as product planning, design, marketing, R&D and manufacturing. Customer requirement information can be derived from online or traditional market surveys, as the input to the CRAF system. Customer requirements may also be inferred from the sales of products, by regarding the acts of purchase as customer needs and preferences expressions. In this study, an appropriately scaled and disguised proprietary dataset from a personal computer manufacturer is employed [57]. The data relates the purchases of five types of personal computer (P1–P5) over 38 months, in a particular market segment (see Table 2). The personal computers have the assumed features as shown below. P1: Brand X processor 1.6 GHz, Brand A operating system (OS), tower chassis, 1 GB memory (RAM) (667 MHz), 80 GB hard disk (HDD) (5400 RPM), integrated graphic processor, CD-RW drive and diskette drive. P2: Brand Y processor 1.6 GHz, Brand A OS, tower chassis, 1 GB RAM (667 MHz), 160 GB HDD (5400 RPM), integrated graphic processor, DVD-ROM and diskette drive. P3: Brand X processor 2.0 GHz, Brand B OS, desktop chassis, 2 GB RAM (800 MHz), 320 GB HDD (5400 RPM), discrete graphic processor Type 1, DVD-ROM drive. P4: Brand Y processor 2.0 GHz, Brand A OS, desktop chassis, 2 GB RAM (800 MHz), 320 GB HDD (7200 RPM), discrete graphic processor Type 1, DVD recordable drive.

102


P5: Brand X processor 2.4 GHz, Brand B OS, compact chassis, 4 GB RAM (1067 MHz), 500 GB HDD (7200 RPM), discrete graphic processor Type 2, DVD recordable drive and media card reader. The CRAF algorithm was coded and implemented in the Matlab environment. Table 3 shows the parameters and values applied in the study. The case data was fed to the system, generating results as described in the rest of this section. Fig. 2 shows examples of the ARBs generated in the study. The depicted bio-molecules are m0 = hproduct, personal computeri, m1 = hchassis, toweri, m2 = hgraphic, integratedi, m3 = hchassis, compacti, m4 = hchassis, desktopi, m5 = hgraphic, discretei, m6 = htype, Xi, m7 = htype, Bi and m8 = htype, Ai. Note that bio-molecules m2 is an option of m5, and is less specific than m7. ARB, which is defined by a chain of bio-molecules, ranges in terms of specificity. For instance, in Fig. 2, the ARB that is defined by both m5 and m6 is more specific than the one that is only defined by m5. The simulated sequential processes of the algorithm resulted in a set of ARB objects that reflects the essential patterns in the data. The forecast of B-cells repertoire is made possible having modelled the dynamics of the evolution, in this case, based on the neural network technique. Fig. 7 depicts the qualitative learning of the system in terms of (1) total ARB – the aggregated objects learnt by the system over time, (2) inactive ARB – the premature and obsolete customer needs, and (3) active ARB – the relevant customer needs in points of time. Fig. 7 therefore at the same time reveals the dynamic processes that ascertain the validity of cus-

Table 2 Scaled personal computer sales data [57]. Month

P1

P2

P3

P4

P5

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38

3 24 49 58 67 71 85 92 155 80 145 198 195 237 309 302 205 413 314 378 452 334 239 250 139 255 69 20 30 18 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 28 206 253 352 290 362 317 224 204 339 175 190 196 177 272 216 203 137 80 69 149 111 133 92 73 60 66 56 1 0

0 0 0 0 0 0 0 0 0 0 0 0 42 75 112 121 144 166 135 134 180 119 160 230 276 468 323 173 206 259 218 192 201 151 156 183 123 68

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 66 108 127 496 751 507 407 465 308 389 256 298 285 277 117 43 89 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 11 128 353 518 386 312 319 378 420 298 291 335 178 231 215 226 77

Table 3 Applied parameters of the CRAF algorithm. Parameter

Value

ARBT PP d

5 3 0.3 1 2 8 8

er L Delay Hidden neurons

tomer needs over time. Customer requirements were shown to vary considerably over time. The relatively less stimulated ARBs (i.e. customer needs) faded and were deactivated while new ARBs were learnt and activated through the processes of stimulation. The sensitivity of the activation and deactivation is dependent on parameters ARBT and PP. They are empirically set in the experiment with reference to the noise characteristics of the input data. Researchers have proposed means to stabilise the size of total B-cells population to ensure that the resultant data is tractable and meaningful. Timmis and Neal [38] devised a resource limited scheme in the context of continuous learning. The population control mechanism in their work is based on the idea of competition amongst the ARBs for survival. In the present work, the total population and the number of ARBs, though should exhibit tractability, should not be capped. This is in consideration of the notion that new knowledge of customer needs (i.e. ARB) is continuously learnt. As such, the CRAF algorithm (in Step 5) maintains the overall population size by allowing only newly leant B-cells to add to the total population. The decay of B-cells through apoptosis process in the algorithm performs the necessary population regulation function. Fig. 8 shows the controlled increase of B-cells population in the system, with adherence to the tractability requirement. FTDNN is embedded in the system to sequentially learn the quantitative dynamism of the population. The ANN is trained by backpropagation algorithm. Fig. 9 shows an instance of the training result during the learning process, with mean squared error (MSE) as the performance index over the epochs. It is found that the MSE in general tends to zero during training, indicating positive progressions in the tasks of function approximation. The study has therefore demonstrated the function of the artificial neural network in approximating the variations of B-cells concentration levels over time domain.

Fig. 7. Total, active and inactive units of ARB.


Dynamic forecasts of B-cell concentration levels were found to be improving over the months. Fig. 10 charted the percentage error for the nine ARBs that had non-zero concentration levels though the forecasting period (note: forecast is not performed on ARBs with zero B-cells). The errors were generally maintained at low levels in the final months, as shown in Fig. 10. The mean percentage error of forecast dropped below 5% for the final 6 months, with each lower than the previous as time passes (see Fig. 11). The mean percentage error of forecast during the 38th month was noted to be at 0.79%. The low error indicates reasonable degree of accuracy for practical applications. Apart from generating qualitative forms of information (i.e. attributes, values and ‘‘active or inactive” statuses), the CRAF system at the same time produces past, current and forecasted concentration levels of the ARBs. The quantitative data represents the relative degree of preferences, at various levels of abstraction. Post-analysis of the output (qualitative and quantitative) data generated useful information in this case study. For instance, in Fig. 12, the system showed the rising preference for the tower type chassis, against the then common desktop type. The requirement for the compact type chassis though was premature in the market by the end of the 38th month, the forecast indicated signs of increasing preference during the 37th and 38th months. In another instance, the system reported the discrete type graphic card as being relevant (i.e. activated) around the 25th month, and thereafter progressively gained popularity against the integrated type (see Fig. 13). Meaningful graphical representations of B-cells concentration levels can be generated based on post-analysis functions (coded in Matlab in this case). This study demonstrated the dynamic functions of the CRAF system as an indicator of the current and future customer requirements. The generated sets of customer needs data were cross-referenced to the temporal space, making them exceptionally valuable in time-sensitive markets.

6. Conclusion Turbulent product markets render the treatment of customer requirements data as time-based variables relevant and critical. Being sensitive to the changes of customer requirements is argued here as a basic reactive strategy where volatile market is concerned. To better compete, the data should be applied to forecast future customer needs as an effective secondary strategy. Solutions proposed to date carry various facets of limitation, as reviewed in Section 2. The CRAF system is defined in this article to address the problem of fast shifting customer needs, and to overcome the limitations of previously proposed methodologies. In effect, the CRAF system studies the dynamics of pattern evolution to further per-

Fig. 8. Repertoire size in the CRAF system.

103

Fig. 9. An instance of FTDNN learning process.

form data forecast. The algorithm is specified to operate continuously so as to manage and forecast customer needs data dynamically. This characteristic is in contrast to traditional methods that treat temporal data in discrete approaches. Clearly, these traditional methodologies are less suited for application in fast changing product markets. The CRAF system proposed in this paper possesses the following unique functions that address the issue. The CRAF system operates on customer requirements information robustly modelled by the requirement space framework (RSF). The RSF is mapped to the ARB model founded in AIS research, which allows the application of AIS methodologies on customer needs data. The structure of RSF describes co-requirement and requirement option objects in multiple levels of abstraction. As a result, the CRAF system is capable of generating multi-level customer needs information in contrast to previous methods that operate on limited abstraction levels of information. QFD, as one of the conventional methods for capturing customer needs, does not represent the alternatives of requirements. Previous propositions (e.g. [19,24]) that applied QFD to manage and/or to forecast customer needs were therefore unable to perform the analysis of requirement options. In the case of the CRAF system, the RSF facilitates the computation, as illustrated in Section 5. The CRAF system is based on a dynamic and autonomous architecture inherent in the computational methodologies employed. By learning both the quantitative and qualitative aspects of the data, the system functions to identify relevant and inapplicable (premature or obsolete) customer needs. It therefore addresses the issue of managing dynamic customer needs, which has been involving human interventions in keeping the set of customer requirements up to date in real time. The CRAF system functions to approximate and forecast customer requirement data. Like many physical and social phenomena, the validity of customer requirement over time is likely to be non-linear (i.e. they may not be written as linear combination of independent components). The proposed system offers the approximations of non-linearity, without the need to assume specific models in prior. As a strategic move towards successful NPD, management may apply the derived qualitative and quantitative information to support decision-making processes. The information derived from the management and forecasting system may serve to reduce the uncertainties found inherently in product development

104


Fig. 10. Percentage error of concentration forecast.

Fig. 11. Mean percentage error of concentration forecast.

projects. We have discussed that the dynamism not only bears implications on product design processes, but also on the spectrum of NPD activities such as R&D, production and marketing. The outcomes of this class of research will therefore be of interest to various functions across NPD organizations and sup-

ply chains. In view of the increasingly fast changing markets, dynamic customer requirement management and forecast as well as the applications of the generated information on downstream processes are expected to be vital areas for future research.

Fig. 12. B-cell concentration of chassis designs.

Fig. 13. B-cell concentration of graphic card types.


Appendix A Table A List of acronyms. AIS ANN ARB ARBT ARIMA CI CRAF DSF EA FTDNN MLP MSE NPD PP QFD R&D RSF SI TAR

Artificial immune system Artificial neural network Artificial recognition ball Artificial recognition ball threshold Autoregressive integrated moving average Computational intelligence Customer requirements analysis and forecast Design space framework Evolutionary algorithm Focused time delay neural network Multi-layer perceptron Mean squared error New product development Persistency period Quality function deployment Research and development Requirement space framework Specificity index Threshold autoregressive model

References [1] C.P. Blocker, D.J. Flint, Customer segments as moving targets: integrating customer value dynamism into segment instability logic, Industrial Marketing Management 36 (2007) 810–822. [2] C. Herstatt, C. Stockstrom, B. Verworn, et al., ‘‘Fuzzy front end” practices in innovating Japanese companies in new product development, Journal of Product Innovation Management 3 (1) (2006) 43–60. [3] Y. Kondo, Customer satisfaction: how can I measure it?, Total Quality Management of Business Excellence 12 (7–8) (2001) 867–872. [4] J. Jiao, C.-H. Chen, Customer requirement management in product development: a review of research issues, Concurrent Engineering: Research and Applications 14 (3) (2006) 173–185. [5] H. Kärkkäinen, P. Piippo, K. Puumalainen, M. Tuominen, Assessment of hidden and future customer needs in Finnish business-to-business companies, R&D Management 31 (4) (2001) 391–407. [6] R. Szakonyi, Dealing with a nonobvious source of problems related to selecting R&D to meet customers’ future needs: weakness within an R&D organization’s and within a marketing organization’s individual operations, IEEE Transactions on Engineering Management 35 (1) (1988) 37–41. [7] V. Mittal, P. Kumar, M. Tsiros, Attribute-level performance, satisfaction, and behavioral intentions over time: a consumption-system approach, Journal of Marketing 63 (1999) 88–101. [8] R.J. Calantone, A.G. Sawyer, The stability of benefit segments, Journal of Marketing Research 15 (1978) 395–404. [9] H. Kärkkäinen, K. Elfvengren, Role of careful customer need assessment in product innovation management – empirical analysis, International Journal of Production Economics 80 (2002) 85–103. [10] I. Christian, H. Ismail, J. Mooney, S. Snowden, M. Toward, D. Zhang, Agile manufacturing transitional strategies, in: Proceedings of the Fourth SMESME International Conference, Aalborg, Denmark, 2001, pp. 69–77. [11] A. Persona, A. Regattieri, P. Romano, An integrated reference model for production planning and control in SMEs, Journal of Manufacturing Technology Management 15 (7) (2004) 626–640. [12] R. Reichwald, C.M. Stotko, F.T. Pillar, Distributed mini-factory networks as a form of real-time enterprise: concept, flexibility potential and case studies, in: B. Kuhlin, H. Thielmann (Eds.), The Practical Real-time Enterprise – Facts and Perspectives, Springer, Berlin, Heidelberg, 2005, pp. 403–434. [13] S. Maguire, S.C.L. Koh, C. Huang, Identifying the range of customer listening tools: a logical pre-cursor to CRM?, Industrial Management and Data Systems 107 (4) (2007) 567–586 [14] E. Hart, P. Ross, Exploiting the analogy between the immune system and sparse distributed memories, Genetic Programming and Evolvable Machine 4 (2003) 333–358. [15] E. Hart, J. Timmis, Application areas of AIS: the past, the present and the future, Applied Soft Computing 8 (2008) 191–201. [16] N. Gunasekaran, V.P. Arunachalam, A. Vinu Selva Kumar, Web-enabled integration of the voice-of-customer for continuous improvement and product development, International Journal of Service and Operations Management 2 (1) (2006) 78–94. [17] D.H. Abelow, US Patent US5999908, 1999. [18] C.-H. Chen, Y.T. Chong, W. Yan, Recent patents of product design technologies, Recent Patents on Engineering 1 (2) (2007) 195–202. [19] H.-H. Wu, J.-I. Shieh, Using a Markov chain model in quality function deployment to analyse customer requirements, International Journal of Advanced Manufacturing Technology 30 (2006) 141–146.

105

[20] C. Adiano, A.V. Roth, Beyond the house of quality: dynamic QFD, Benchmarking for Quality Management and Technology 1 (1994) 25–37. [21] F. Flores, Innovation by listening carefully to customers, Long Range Planning 26 (3) (1993) 95–102. [22] X.-X. Shen, M. Xie, K.-C. Tan, Listening to the future voice of the customer using fuzzy trend analysis in QFD, Quality Engineering 13 (3) (2001) 419–425. [23] M. Xie, K.C. Tan, T.N. Goh, Advanced QFD Applications, ASQ Quality Press, Wisconsin, 2003. [24] H.-H. Wu, A.Y.H. Liao, P.C. Wang, Using grey theory in quality function deployment to analyse dynamic customer requirements, International Journal of Advanced Manufacturing Technology 25 (2005) 1241–1247. [25] H. Raharjo, M. Xie, A.C. Brombacher, Prioritizing quality characteristics in dynamic quality function deployment, International Journal of Production Research 44 (23) (2006) 5005–5018. [26] C.-H. Chen, W. Yan, An in-process customer utility prediction system for product conceptualization, Expert Systems with Applications 34 (4) (2008) 2555–2567. [27] S.H. Ha, Applying knowledge engineering techniques to customer analysis in the service industry, Advanced Engineering Informatics 21 (2007) 293–301. [28] K.B. Kahn, F. Franzak, A. Griffin, S. Kohn, W.M. Christopher, Editorial: identification and consideration of emerging research questions, Journal of Product Innovation Management 20 (3) (2003) 193–201. [29] U. Aickelin, D. Dasgupta, Artificial immune systems, in: E.K. Burke, G. Kendall (Eds.), Search Methodologies: Introductory Tutorials in Optimization and Decision Support Techniques, Springer, New York, 2005, pp. 375– 399. [30] D. Dasgupta, Advances in artificial immune systems, IEEE Computational Intelligence Magazine Nov 2006 (2006) 40–48. [31] S. Stepney, J.A. Clark, C.G. Johnson, D. Partridge, R.E. Smith, Artificial immune systems and the grand challenge for non-classical computation, ICARIS 2003, LNCS 2787, 2003, pp. 204–216. [32] J. Kim, P.J. Bentley, Immune memory in the dynamic clonal selection algorithm, ICARIS 2002, 2002, pp. 59–67. [33] A. Secker, A.A. Freitas, J. Timmis, AISEC: an artificial immune system for email classification, in: R. Sarker, R. Reynolds, H. Abbass, T. Kay-Chen, R. McKay, D. Essam, T. Gedeon (Eds.), Proceedings of the Congress on Evolutionary Computation, Canberra, Australia, December 2003, IEEE, 2003, pp. 131–139. [34] S. Stepney, R.E. Smith, J. Timmis, A.M. Tyrell, Towards a conceptual framework for artificial immune systems, ICARIS 2004, LNCS 3239, 2004, pp. 53–64. [35] J. Timmis, M. Neal, J. Hunt, Data analysis using artificial immune systems, clustering analysis and Kohonen networks: some comparisons, in: Proceedings of the International Conference of Systems and Man and Cybernetics, Tokyo, Japan, IEEE, 1999, pp. 922–927. [36] J. Timmis, M. Neal, J. Hunt, An artificial immune system for data analysis, Biosystems 55 (2000) 143–150. [37] A. Gaspar, P. Collard, From GAs to artificial immune systems: improving adaptation in time dependent optimisation, in: Proceedings of the Congress on Evolutionary Computation, vol. 3, 1999, pp. 1859–1866. [38] J. Timmis, M. Neal, A resource limited artificial immune system for data analysis, Knowledge-based Systems 14 (2001) 121–130. [39] M. Neal, Meta-stable memory in an artificial immune network, ICARIS 2003, LNCS 2787, 2003, pp. 168–180. [40] E. Hart, P. Ross, Clustering moving data with a modified immune algorithm, EvoWorkshop 2001, LNCS 2037, 2001, pp. 394–430. [41] A. Waibel, Phoneme recognition using time delay neural networks, IEEE Proceeding ASSP-37 (1989) 328–339. [42] J. Park, I.W. Sandberg, Universal approximation using radial-basis-function networks, Neural Computation 3 (2) (1991) 246–257. [43] G. Zhang, B.E. Patuwo, M.Y. Hu, Forecasting with artificial neural networks: the state of the art, International Journal of Forecasting 14 (1998) 35–62. [44] J.M.P. Menezes Jr., A.B. Barreto, A new look at nonlinear time series prediction with NARX recurrent neural network, in: Proceedings of the Ninth Brazilian Symposium on Neural Networks (IEEE), 2006. [45] N.K. Sinha, M.M. Gupta, D.H. Rao, Dynamic neural network, Proceedings of the IEEE International Conference on Industrial Technology 1 (2000) 491–496. [46] J.C. Principe, N.R. Euliano, W.C. Lefebvre, Neural Adaptive Systems: Fundamentals Through Simulations, John Wiley and Sons, 2000. [47] C.-H. Chen, L.P. Khoo, W. Yan, A strategy for acquiring customer requirement patterns using laddering technique and ART2 neural network, Advanced Engineering Informatics 16 (2002) 229–240. [48] K.M. Saridakis, A.J. Dentsoras, Soft computing in engineering design – a review, Advanced Engineering Informatics 22 (2008) 202–221. [49] Y.T. Chong, C.-H. Chen, K.F. Leong, Human-centric product conceptualization using a design space framework, Advanced Engineering Informatics 23 (2009) 149–156. [50] L.N. de Castro, J. Timmis, Artificial Immune Systems: A New Computational Intelligence Approach, Springer-Verlag, Heidelberg, Germany, 2002. [51] L.N. de Castro, J. Timmis, Artificial immune systems as a novel soft computing paradigm, Soft Computing 7 (2003) 526–544. [52] M. Cohn, Degeneracy, mimicry and cross-reactivity in immune recognition, Molecular Immunology 42 (2005) 651–655.

106


[53] R.E. Langman, The specificity of immunological reactions, Molecular Immunology 37 (2000) 555–561. [54] L. Alan, F. Robert, Nonlinear signal processing using neural networks, prediction and system modeling, Los Alamos Report LA-UR-87-2662, Los Alamos National Laboratory, Los Alamos, NM, 1987. [55] E.M. Azoff, Neural Network Time Series Forecasting of Financial Markets, John Wiley & Sons, 1994.

[56] M.E. Doms, W.E. Dunn, S.D. Oliner, D.E. Sichel, How fast do personal computers depreciate? Concepts and new estimates, in: NBER Conference on Tax Policy and the Economy, Washington, DC, 2003. [57] A.A. Kurawarwala, H. Matsuo, Product growth models for medium-term forecasting of short life cycle products, Technological Forecasting and Social Change 57 (1998) 169–196.