doi:10.1016/j.eswa.2006.09.018 www.elsevier.com/locate/eswa Expert Systems with Applications 34 (2008) 530–540 Expert Systems with Applications A fuzzy CBR technique for generating product ideas Muh-Cherng Wu *, Ying-Fu Lo, Shang-Hwa Hsu Department of Industrial Engineering and Management, National Chiao Tung University, Hsin-Chu, Taiwan, ROC Abstract This paper presents a fuzzy CBR (case-based reasoning) technique for generating new product ideas from a product database for enhancing the functions of a given product (called the baseline product). In the database, a product is modeled by a 100-attribute vector, 87 of which are used to model the use-scenario and 13 are used to describe the manufacturing/recycling features. Based on the use- scenario attributes and their relative weights – determined by a fuzzy AHP technique, a fuzzy CBR retrieving mechanism is developed to retrieve product-ideas that tend to enhance the functions of the baseline product. Based on the manufacturing/recycling features, a fuzzy CBR mechanism is developed to screen the retrieved product ideas in order to obtain a higher ratio of valuable product ideas. Experiments indicate that the retrieving-and-filtering mechanism outperforms the prior retrieving-only mechanism in terms of generating a higher ratio of valuable product ideas. � 2006 Elsevier Ltd. All rights reserved. Keywords: New product development; Case-based reasoning; Fuzzy CBR; Fuzzy AHP 1. Introduction Product life cycle is getting shorter at this age. How to enhance the productivity of new product development is very important. A typical process of new product devel- opment includes product idea generation, conceptual design, detailed design, and economic justification of design. Among these procedures, generation of product ideas may be the most important because the other proce- dures are intended to realize and justify the generated ideas. Various methods for creating product ideas have been published in the literature. According to the degree of computerization, the previous methods can be grouped into three categories: (1) manual-based approach, (2) computer- aided approach, and (3) computer-generated approach. The manual-based approach is to generate product ideas by asking an individual or a group of people to think freely or think under a guided process. Examples of this approach 0957-4174/$ - see front matter � 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.09.018 * Corresponding author. Tel.: +886 35 731 913; fax: +886 35 720 610. E-mail address: mcwu@cc.nctu.edu.tw (M.-C. Wu). include brainstorming method (Higgins, 1994; Nijssen & Lieshout, 1995; Osborn, 1963), forced relationships method (Higgins, 1994; Hisrich, Ingram, & Peters, 1991; Kotler, 1994), focus groups method (Higgins, 1994; Hisrich et al., 1991; Kotler, 1994; Nijssen & Lieshout, 1995), attribute listing method (Higgins, 1994; Kotler, 1994; Linda, 1991; Nijssen & Lieshout, 1995), check-list method (Higgins, 1994; Kotler, 1994), morphological analysis (Higgins, 1994; Kotler, 1994; Linda, 1991; Nijssen & Lieshout, 1995), and synectics method (Higgins, 1994; Kotler, 1994). Techniques of this approach are carried out solely through human, without using any computing facilities in the creation of new product ideas. The computer-aided approach intends to use computer to guide a person’s thinking process for creating or realiz- ing product ideas. A typical technique of this approach encodes the innovative rules listed by TRIZ (Mann, 2003; Rantanen & Domb, 2002) in a computer program. Through a series of human–computer interaction activities, users of the computer program can find a number of design tem- plates to realize a user-desired product function. Example software of the technique includes Goldfire Innovator (2006), Trisolver (2006), and Creax.com (2006). mailto:mcwu@cc.nctu.edu.tw M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 531 The computer-generated approach is intended to create ideas for enhancing the function of a baseline product by retrieving ‘‘scenario-compatible’’ products from database (Wu, Lo, & Hsu, 2006). The retrieved products are similar enough to the baseline product in the scenario where the baseline product is used. Product functions of the retrieved ones are called product ideas. This approach would gener- ate a large amount of product ideas in a very short time. However, one weakness is that many less-valued products ideas may be generated, which would consequently require a large amount of human efforts to screen them. To alleviate the weakness, this paper presents a CBR (case-based reasoning) technique combined with a fuzzy AHP method for retrieving product ideas that tend to be more-valued, and from the retrieved ones screening out those ideas that tend to be less-valued. As shown in Fig. 1, the research framework involves three modules. The first module is to establish a product database in which a product is encoded by a vector involv- ing 100 attributes. Of these attributes, 87 ones represent the scenario of using a product and the other 13 represent the scenario of manufacturing and recycling the product. Each attribute is defined by a linguistic variable of fuzzy theory (Zadeh, 1975). The second module is to characterize the scenario of use for target customers. The 87 attributes for modeling the scenario of use are classified into five categories (also called dimensions). The technique of fuzzy AHP is used to deter- mine the relative weight for each of the five dimensions in order to understand the preferences of target customers. The weighting of product attributes is intended to help retrieving more-valued products ideas; that is, less-valued ideas may not be generated in the retrieval stage. The third module firstly retrieves product ideas whose functions tend to be attachable to the baseline product, and then filters out those retrieved ideas that tend to be less-valued. The retrieving mechanism is by using the 87 Establishing product database Weighting the use-scenario of target customers Retrieve and screen Product ideas Fig. 1. Research framework. product attributes that have been weighted to characterize the scenario of use for target customers. The screening mechanism is by using the 13 manufacturing and recycling attributes. The research framework is developed by examining the three main stages of a product life cycle – manufacturing, using, and recycling. The 100-attribute product representa- tion for generating new product ideas are developed based on the various costs/benefits concerned in a product life cycle. We retrieve product ideas from the perspective of enhancing the usability in order to increase product value; and filter out product ideas from the perspectives of reduc- ing product cost – reducing the manufacturing/recycling costs. The remainder of this paper is organized as follows. Sec- tion 2 reviews the literature on case-based reasoning (CBR) technique. Section 3 describes the method for representing a product by a vector of 100 attributes. Section 4 presents the fuzzy AHP method for determining the relative weights to characterize target customers’ scenario of use. Section 5 describes the CBR method for retrieving and screening product ideas. Experiment results are presented in Section 6 and concluding remarks are in Section 7. 2. Case-based reasoning Case-based reasoning (CBR), a well-known artificial intelligence technique, is a process for solving a new prob- lem case by referring to the solutions of similar past cases (Aamodt & Plaza, 1994; Kolodner & Leake, 1996; Marling, Sqalli, Rissland, Munoz–Avila, & Aha, 2002). In a CBR system, a database for storing the past cases has to be avail- able. To solve a new problem case by CBR, similar past cases are first retrieved and their associated solutions are then used to aid users to develop solutions for the new case. Two survey papers of CBR can be referred to Watson and Marir (1994), de Mantaras and Plaza (1997). A CBR system involves three modules: (1) a case repre- sentation scheme, (2) a similarity metric, and (3) a case- retrieval mechanism. A case representation scheme is to model a case by a set of attributes for characterizing the case at a particular application. A similarity metric is for measuring the similarity between any two cases. A case- retrieval mechanism is designed to retrieve the past cases that are similar enough to the new case. To make a CBR system more user-friendly, some studies proposed a fuzzy-CBR approach. This approach advocates using linguistics variables in fuzzy theory to valuate the case attributes. A linguistic variable is represented by a nat- ural language form as well as by a fuzzy number. The text description is intended to help users resolve the uncertainty issues while they valuate the case attributes. The fuzzy number representation and the associated fuzzy operators are used to calculate the similarity metrics for implement- ing the case-retrieving mechanism. Much literature based on such a fuzzy-CBR approach has been published. 532 M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 Examples include Hirota et al. (1997), de Mantaras and Plaza (1997), Ruet and Geneste (2002), and Chan (2005). The CBR paradigm has been applied in a wide variety of design problems. The applications include architecture design (Trousse & Visser, 1993), mechanical design (Maher & Garza, 1997), and some other design problems. In the CBR applications for product design, existing design archi- tectures/parameters are retrieved to aid the realization of a product function (Bilgic & Fox, 1996; Maher, Balachan- dran, & Zhang, 1995) or to help engineers develop a new design that could fulfill the downstream requirements (Belecheanu, Pawar, Barson, Bredehorst, & Weber, 2003). These previous CBR studies for product design focus on the engineering aspect – realizing a design for a given func- tional requirement. Yet, this paper’s focus – how to apply CBR to create new and marketable functional require- ments for a given product is rarely studied. 3. Product database In the product database for generating and screening ideas, a product is encoded by a vector consisting of 100 attributes, where 87 ones are used to characterize the sce- nario of using the product, and 13 attributes are used to describe the manufacturing and recycling characteristics. For a product, each attribute is described by a linguistic variable in fuzzy theory. 3.1. Product attributes for creating ideas This research generates new product ideas based on the following hypothetical assertion – two products that are similar in their scenario of use have chance to be combined into a new product. That is, the new product would involve the main functions of the two products. To generate new product ideas for a given product (called the baseline prod- uct), we aim to identify some other products that are sim- ilar to the baseline product in its scenario of use. This research models a product use-scenario from five dimensions, which are developed based on the notion of UCD (user centered design) – a design paradigm originally proposed by Norman and Draper (1986). The UCD notion advocates that a product should be designed based on the Fig. 2. The proposed product representation is based on a UCD paradigm. user’s needs and the scenario where they use the product. As shown in Fig. 2, the five dimensions involve: interface attributes, task type, physical feature, environment, and user characteristic characteristics; and they are deployed into 20 sub-dimensions (also called groups) that ultimately yield 87 attributes (Table 1). The first dimension – interface attributes are to identify the medium through which the product interacts with the user. Here, the medium denotes a particular portion of the user’s body. This research classifies the interface attri- butes into three groups: sensory modality, response modal- ity, and interface point. The interface attributes are composed of 18 attributes in total. The second dimension – task type is to model the tasks to be performed by the user through using the product. According to users’ needs, the tasks are categorized into seven groups: eating, clothing, living, transportation, educa- tion/entertainment, working, and health care. These seven groups are further characterized by 30 attributes based on the execution process in each group. The third dimension – physical feature is intended to describe a product from the aspects of physical size, mobil- ity, and scaleability. Ten attributes are used to model the dimension of physical feature. The fourth dimension – environment is intended to char- acterize the environment where the product is used. The characterization involves three groups: sociality, physical place, and the harshness of environment. Ten attributes are used to model the environment dimension. The fifth dimension – user characteristics are intended to describe which groups of users tend to use the product. The dimension is characterized by the following demographic features: gender, age, profession, and job title. These five features are further described by 19 attributes. 3.2. Product attributes for screening ideas Based on the aforementioned product modeling method, new product ideas of the baseline product may be retrieved by applying the fuzzy CBR technique. Surely, a retrieved product idea and the baseline product to a certain extent are ‘‘compatible’’, from the perspective of using the two products. However, from the perspectives of manufactur- ing/recycling, these two products may not be economically justifiable to combine them into one. That is, we assert that a product idea will be discarded, if it has few commonality with the baseline product in terms of manufacturing/recyc- ling attributes. As shown in Table 1, 13 attributes are used to model the features of manufacturing and recycling, which are grouped into four groups. The first group describes the materials of a product, which involves the following four types: (a) metal, (b) non- metal, (c) animals, and (d) plants. A product may be com- posed of several types of materials. The percentage of a particular type of material used in a product is described by an attribute. Table 1 Product representation for cell phones and ball pens M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 533 534 M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 The second group describes the processing mechanisms, which involves three types: (a) physical processes, (b) chemical processes, and (c) biological processes. A product may be manufactured by more than one type of processes. The percentage of a particular type of process used to man- ufacture a product is described by an attribute. The third group describes the energy resources used in a product, which involves four types: (a) electricity, (b) bat- teries, (c) solar energy, and (d) oil/gas. A product may use more than one type of energy resources. The percentage of energy resources used by a product is described by an attribute. The fourth group uses two attributes model the pro- cesses and results of recycling a product, which involves (a) the easiness in recycling a product, and (b) the usability of a recycled product. The easier is a recycling process, the higher is its attribute value; the higher is the usability of a recycled product, the higher is its attribute value. 3.3. Linguistic variables for describing product attributes To facilitate users to characterize a product, each of the 100 product attributes is described by a linguistic variable in fuzzy theory. A linguistic variable, appearing in a natu- ral language form, represents a human’s judgment, which can be further modeled by a triangular fuzzy number (Zadeh, 1975). Of the 100 product attributes, the first 98 attributes are described by five linguistic variables: extreme relevance, high relevance, relevance, low relevance, and no relevance. The associated fuzzy number of a linguistic variable, ~A ¼ðl1; m1; r1Þ, is shown in Fig. 3, where l1 denotes the 0.0 0.25 0.5 0.75 1.0 No relevance Low relevance Relevance High relevance Extreme relevance xA ~ X μ ( ) Fig. 3. Linguistic variables and the associated fuzzy numbers. M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 535 leftmost coordinate, m1 is the central coordinate, and r1 the rightmost coordinate on the x-axis (Moon & Kang, 2001). The last two attributes (attribute 99 and 100), which char- acterize the recycling characteristics, are described by another five linguistic variables: very high, high, normal, low, and very low. The fuzzy numbers in Fig. 3 are also used to represent these linguistic variables. 4. Weighting each dimension of target customers’ use-scenario In retrieving product ideas, we have to determine the rel- ative importance of the five dimensions of target custom- ers’ use-scenario. To resolve the vagueness caused by human judgment, we use a widely used methodology – fuzzy AHP (analytical hierarchy process) to determine the weighting for each of the five dimensions. The computational procedure of the fuzzy AHP meth- odology (Zadeh, 1975) is summarized below, where the arithmetical operators of fuzzy numbers are defined in Appendix 1. Step 1: Define linguistic variables for pair-wise compari- son. We use linguistic variables to compare the relative importance between any two dimensions. These linguistic variables include ‘‘absolutely impor- tant’’, ‘‘very strongly important’’, ‘‘ essentially important’’, ‘‘weakly important’’ and ‘‘equally important’’ on a five level scales, where between any two consecutive scales an intermediate scale Table 2 Linguistic variables used in the fuzzy AHP Fuzzy number ~1 ¼ð1; 1; 1Þ ~3 ¼ð2; 3; 4Þ ~5 ¼ð4; 5; 6Þ ~7 ¼ð6; 7; 8Þ ~9 ¼ð8; 9; 10Þ ~2 ¼ð1; 2; 3Þ; ~4 ¼ð3; 4; 5Þ; ~6 ¼ð5; 6; 7Þ; ~8 ¼ð7; 8; 9Þ is additionally defined so that 9 scales are created. Table 2 shows the resulting 9 scales that are repre- sented by 9 triangular fuzzy numbers (Chiou, Tzeng, & Cheng, 2005): ~1; ~2; . . . ; ~9. Step 2: Establish the comparison matrix ~A. By performing a pair-wise comparison for any two of the five concerned dimensions, a fuzzy matrix ~A is constructed. Linguis Equally Weakly Essenti Very st Absolu Interme ~A ¼ 1 ~a12 � � � ~a1n ~a21 1 � � � ~a2n .. . .. . . . . .. . ~an1 ~an2 � � � 1 2 6664 3 7775 where tic va imp imp ally im rongl tely i diate if i ¼ j; ~aij ¼ 1, if i 6¼ j; ~aij ¼ ~a�1ji and ~aij ¼ð~a1ij � a2ij ��� � �~aNijÞ=N ~akij: customer k’s judgment on the relative importance between dimension i and j N: total number of target customers in- terviewed. Step 3: Calculate the fuzzy weight of each row in ~A (Buck- ley, 1985). ~Z i ¼ð~ai1 � ~ai2 ��� �� ~ainÞ 1 n 8i ¼ 1; 2; . . . ; n ~W i ¼ ~Z i � ~Z 1 � ~Z 2 ����� ~Z n � ��1 8i ¼ 1; . . . ; n Step 4: Defuzzication of ~W i and ~A (Teng & Tzeng, 1993). aij ¼ Defuzzyð~aijÞ W i ¼ Defuzzyð ~W iÞ where the function Defuzzy is stated in Appendix 1. riables ortant ortant portant y important mportant values between two adjacent judgments Table 3 Values of RI n 1 2 3 4 5 6 7 8 9 10 11 RI 0.00 0.00 0.58 0.90 1.12 1.24 1.32 1.41 1.45 1.49 1.51 536 M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 Step 5: Normalization of Wi Ŵ i ¼ W iPn i¼1W i : Step 6: Consistency check. (1) Compute W �i as follows:2 3 �2 3 A � ~W 1 ~W 2 .. . ~W n 66664 77775 ¼ W 1 W �2 .. . W �n 66664 77775 where A = [aij]. (2) Compute kmax ¼ 1 n W �1 Ŵ 1 � � þ W �2 Ŵ 2 � � þ�� �þ W �n Ŵ n � �� � where kmax is called maximum eigenvalue. (3) Compute CI ¼ kmax � n n � 1 where CI is called consistency index. (4) Compute CR = CI/RI, where CR is called consistency ratio, and RI is called the average random consistency index of randomly generated matrices of size n · n. The values of RI have been provided by Saaty (1980) as shown in Table 3. (5) Consistency check. If CR 6 0.1, then the pair-wise comparison matrix is reasonably consistent and W _ i; 1 6 i 6 n; is the result- ing weighting of dimension i. If CR > 0.1, then the pair-wise comparison results are inconsistent and the pair-comparison procedure has to be updated. Notice that W _ i; 1 6 i 6 5; is the weighting factor of dimension i in the product representation. The five dimen- sions include 87 use-scenario attributes. The attributes in each dimension are of the same weighting – the weight of their parent dimension. Let wj represent the weighting fac- tor of attribute j. Then, wj ¼ W _ i if attribute j belongs to dimension i. In summary, the results of the fuzzy AHP yield the weighting of each use-scenario attribute (wj, 1 6 j 6 87). 5. Retrieving and filtering product ideas Given a product database, this research uses two mech- anisms to propose new product ideas for enhancing the functions of a baseline product. Firstly, from the database, we retrieve products that tend to be compatible to the baseline product – from the perspective of product usabil- ity. Secondly, we filter out the retrieved products whose combinations with the baseline product tend to become costly – from the perspective of manufacturing and recycling. The retrieving and filtering mechanisms are explained by referring to a scenario stated below. A product database P = {Pi, i = 1, . . . , K} has been established, where K is a huge positive integer number and P i ¼ ½~pij�; 1 6 j 6 100; denotes the vector representation of product i. Let B ¼ ½~bj�; 1 6 j 6 100; represents the baseline product. Notice that ~pij and ~bj are fuzzy numbers that indicate the attri- butes of a product. The purpose is to retrieve some prod- ucts from database P, whose combinations with baseline product B may enhance the resulting product usability in a low-cost manner. 5.1. Retrieving mechanism The procedure of the retrieving mechanism is described below. Step 1: Define a retrieving threshold Hr 2 [0,1]. Step 2: Identify the important attributes of the baseline product B ¼ ½~bj�; 1 6 j 6 100. S ¼fjj~bj is of high relevance or extreme relevanceg (Refer to Fig. 3). Let N(S) denote the number of attributes in set S. Step 3: Form retrieval key sets. Ask users to randomly cluster the attributes in S into several subsets Tk, 1 6 k 6 q = dN(S)/me, where each subset involves m or (m � 1) attri- butes. That is, S ¼ Sq k¼1T k , where subset Tk is called a retrieval key set. Step 4: With respect to the retrieval key set Tk, compute product relevance metric between products P i ¼ ½~pij� and baseline product B ¼ ½~bj�. ~RT ki ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP j�T k½ ~bj � Rð~bj; ~pijÞ� wj� 2 P j�T k½bj � wj� 2 vuut for k ¼ 1; 2; . . . q where wj is the weighting factor of attribute j as de- rived in Section 4, Rð~bj � ~pijÞ as defined below rep- resents the relevance metric of jth attribute between baseline product B and product Pi. Rð~bj; ~pijÞ¼ 1 �j~bj � ~pijj M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 537 Step 5: Defuzzication of ~RT ki RT ki ¼ DefuzzyðR T k i Þ for k ¼ 1; . . . ; q Step 6: Retrieve products compatible tobaseline product B from database P. Qk ¼fP ijR T k i P H rg Q ¼ [q k¼1 Qk where Q represents the set of retrieved products. Several distinct points in the retrieving mechanism are explained further. First, only important attributes in base- line product B are included in a retrieval key set. Experi- ments indicate that the inclusion of less important attributes would lead to the retrieval of a huge number of products that are irrelevant to baseline product B. Sec- ond, a retrieval key set involves only a few number of important attributes. With the inclusion of all important attributes in a retrieval key set, the number of retrieved products tends to be very small; and their functions tend to be too close to the baseline product B and cannot be seen as a good product idea. Third, in the computation of ~RT ki , ~bj denotes the value of jth attribute and wj denotes its relative importance from the perspective target customer. That is, ~bj is independent of target customers while wj is dependent on target customers. 5.2. Filtering mechanism The retrieved products in set Q are relevant to the base- line product B, from the perspective of usability. However, some of these products may not be compatible to product B from the perspectives of manufacturing and recycling. We use a filtering mechanism to filter out these incompat- ible products in Q. The procedure of the filtering mecha- nism is described below. Step 1: Define a filtering threshold, Hf 2 [0,1]. Step 2: Define the filtering key set Tm, m = 1, . . . , 4. As stated in Section 3, we use 13 attributes (attri- butes 88–100) to model the features of manufac- turing and recycling, which are grouped into four categories. The attributes in mth category forms a filtering key set, denoted by Tm. Step 3: Compute compatible metric between product Pi and B, with respect to Tm. Table 4 Detailed data in the fuzzy AHP for characterizing target customers, where CI = 0.08199 and CR = 0.07321 Dimension ~W i Wi Ŵ i Interface attributes (0.10499, 0.15366, 0.22404) 0.160896 0.1540 Task type (0.25824, 0.36120, 0.50480) 0.374747 0.3588 ~CT mi ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP j�T m½~bj � Rðbj; pijÞ� 2 P j�T m½~bj� 2 vuut for m ¼ 1; 2; . . . 4 where Rð~bj; ~pijÞ¼ 1 �j~bj � ~pijj. Physical feature (0.10252, 0.14837, 0.21435) 0.155079 0.1485 Step 4: Defuzzication of ~CT mi Environment (0.10462, 0.15165, 0.21995) 0.158739 0.1520 User characteristics (0.12496, 0.18512, 0.27504) 0.195040 0.1867 CT mi ¼ Defuzzyð~C T m i Þ; for m ¼ 1; . . . ; 4 Step 5: Filter out the incompatible products from set Q F m ¼fP ijCT mi < H fg F f ¼[4m¼1F m where Ff represents the set of incompatible prod- ucts in set Q. Step 6: Determine X, the set of products that may enhance the function of product B effectively X ¼ Q � F f : 6. Experiments An empirical study is carried out to compare the effi- ciency and effectiveness between the proposed retrieving- and-filtering mechanism and the retrieving-only mechanism published in Wu et al. (2006) that has been justified to be better than the traditional brainstorming approach. Two experiments for the comparison are performed; one exper- iment uses cell phone and the other uses ball pen as the baseline products. A prototype product database is established for the experiments, which involves 1600 products and is coded by Microsoft Access. The retrieving/filtering mechanism is coded by ASP.NET (Active Server Page.NET), with its interface developed by Macromedia Dreamwaver, and Microsoft Internet Explorer is used as the vehicle for web browsing. To determine the dimensional weighting of target cus- tomers’ use-scenarios, 30 female subjects, aged 19–30, are invited to perform the pair-wise comparison required by the fuzzy AHP method. Results and the associated data of the AHP process are listed in Table 4. 6.1. Generating product ideas Five senior undergraduate students are invited as exper- iment subjects. Using cell phone as the baseline product, each step in executing the retrieving-and-filtering mecha- nism is explained below. Step 1: Set the retrieving threshold, Hr = 0.5. Step 2: Of the 87 product attributes of a cell phone, 10 attributes are automatically identified as impor- tant attributes (highlighted in Table 1). 538 M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 Step 3: From the 10 important attributes, each subject is asked to freely form four retrieval key sets; each set involves either 3 or 2 attributes. Step 4: Based on the retrieval key sets, the retrieving mechanism will generate a set Q – the set of retrieved products. Step 5: Set the filtering threshold, Hf = 0.5. Step 6: Based on the four filtering key sets, the filtering mechanism automatically identifies a set Ff that represents the incompatible products in set Q. Step 7: The system computes the set X = Q � Ff. In the aforementioned procedure, set X represents the product ideas proposed by the retrieving-and-filtering mechanism and set Q represents the product ideas pro- posed by the retrieving-only mechanism. 6.2. Comparison of product-idea-generating mechanisms The performance metric for comparing the two product- idea-generating mechanisms is called creative ratio, as defined below. CZ ¼ NðZ gÞ NðZÞ where Z represents a set of product ideas, Zg is a subset of Z that includes only good product ideas, and N(Z) repre- sents the number of product ideas in set Z. The objectives of the experiments are to compare the value of CX and CQ (Table 5). As stated, X is a subset of Q. A random filtering mechanism would filter out good product-ideas at a probability of CQ; this on average yields that CX = CQ. By contrast, a less-effective filtering mecha- nism would yield that CX < CQ, while an effective filtering mechanism would yield that CX > CQ. That is, CX > CQ indicates that the average time required to identify one good product-idea is less. The method for justifying whether a product-idea is good is through expert’s evaluation. Three experts familiar with new product developments are invited to evaluate the generated product-ideas based on three criteria – original- ity, valuableness, and usefulness (Besemer & O’Quin, 1986). Each criterion is rated in a five-point scale – the higher the better. A product-idea is justified by averaging Table 5 Comparing the performance of the retrieving-and-filtering and the retrieving-o Cell phone Retrieving Retrieving + filtering N(Q) N(Qg) CQ (%) N(X) N(Xg) CX Sub_1 302 30 9.93 134 23 17.1 Sub_2 179 21 11.73 82 17 20.7 Sub_3 254 24 9.45 119 16 13.4 Sub_4 254 26 10.23 98 18 18.3 Sub_5 207 26 12.56 104 19 18.2 Mean 239 25.4 11.02 107 18.6 17.7 the points of the three criteria; an average point greater than 4.0 is regarded as a good idea. The results of the two experiments are shown in Table 5. For each baseline product, the mean of CX is greater than that of CQ. A t-test for cell phone (a = 0.05, t-value = �7.641, and P-value = 0.002) indicates CX > CQ is statisti- cally significant. Another t-test for ball pen (a = 0.05, t-value = �5.287, and P-value = 0.006) also supports the finding CX > CQ. These two findings conclude that the pro- posed filtering mechanism is effective. That is, the average time required to manually identify one good product-idea from the retrieved products is reduced if we enhance the retrieving mechanism by a filtering mechanism. However, the advantage of CX > CQ is offset by a draw- back – N(Qg) > N(Xg). That is, some good product-ideas are filtered out in the filtering mechanism. One may ques- tion that what is the ‘‘net benefit’’ of developing the filtering mechanism. Is the ‘‘net benefit’’ positive or negative? Con- sider the comparison of the retrieving mechanism and an exhaustively-listing mechanism – taking all products in the database as generated product-ideas. Surely, the exhaustively-listing mechanism can always generate more numbers of good ideas than the retrieving mechanism, at the expense of paying more expert time to identify good product ideas. This analogy illustration may explain the need for developing an effective filtering mechanism. Moreover, the performance of the filtering mechanism can be regulated by giving different values to the filtering threshold (Hf). In the case of Hf = 0, the filtering mecha- nism is halted; that is, only the retrieving-mechanism works. Increasing the value of Hf tends to pay more atten- tion to the filtering mechanism. Users of the product-idea- generating systems may iteratively give different Hf value, depending upon the number of retrieved ideas and how much expert time is available to evaluate these ideas. Sup- pose an initial assignment of Hf = 0.3 yields 1 million product-ideas. This would lead the users to reassign a higher Hf value to reduce the numbers of the retrieved ideas. Table 6 gives 23 product-ideas for enhancing the func- tion of cell phone. Table 7 describes 18 product-ideas for enhancing the function of ball pen. To our knowledge, some of these ideas for enhancing the baseline products are currently not available in the market. nly mechanisms Ball pen Retrieving Retrieving + filtering (%) N(Q) N(Qg) CQ (%) N(X) N(Xg) CX (%) 6 225 22 9.78 108 17 15.74 3 219 18 8.22 114 12 10.53 4 247 23 9.31 139 18 12.95 7 211 18 8.53 94 11 11.70 7 198 15 7.58 78 11 14.10 6 220 19.2 8.73 106.6 13.8 12.95 Table 7 Generated product ideas for ball pen Lipstick Compass Clinical thermometer Laser pointer Baton Electric torch Eyebrow pencil Mp3 Ultraviolet rays tester Massage stick GPRS USB Flash memory drive Voice recorder Pill box Pregnancy test pen Swatch (Timer) LED Language translator Table 6 Generated product ideas for cell phone Alcohol tester MP3 Language learning machine Pulse detector e-Map PDA (personal digital assistant) Radio and Walkman RFID Mosquito prevention set Electronic pet (Game) e-Book USB Flash memory drive Stun gun for security Cosmetic box Language translator/Dictionary Anti-camera detector Pedometer Remote controller for door/car Clinical thermometer Digital wallet OBU (car navigation system) Barcode scanner Strobe light M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 539 7. Concluding remarks This research presents a case-based reasoning (CBR) approach combined with the fuzzy AHP method to generate new product ideas that tend to be valuable for enhancing the baseline product. The generation of new product ideas is through a retrieving-and-filtering mecha- nism operated on a product database, where a product is modeled by 100 attributes—87 ones model the use-scenario and the other 13 model the manufacturing-and-recycling features. The retrieving mechanism is a fuzzy CBR technique that utilizes the use-scenario attributes for retrieving prod- uct-ideas. The retrieved product-ideas are subsequently screened by a filtering mechanism – a CBR technique that utilizes the manufacturing-and-recycling attributes as the fil- tering criteria. The filtering mechanism is proposed to filter out less valuable product ideas in order to save time for subsequent product-idea evaluation by experts. A prototype system has been implemented for justifying the contribution of the filtering mechanism. Experiments show that the retrieving-and-filtering mechanism outper- forms the prior retrieving-only mechanism in terms of creative ratio. That is, the retrieving-and-filtering mecha- nism generates higher percentage of ‘‘good product ideas’’ as opposed to that of the retrieving-only mechanism. Appendix 1. Arithmetical operators for fuzzy numbers The fuzzy arithmetic operators used in this research are defined below (Laarhoven & Pedrycz, 1983; Zadeh, 1975), by referring two fuzzy numbers ~a1 ¼ðl1; m1; r1Þ and ~a2 ¼ðl2; m2; r2Þ. (1) Addition operator: � ~a1 � ~a2 ¼ðl1 þ l2; m1 þ m2; r1 þ r2Þ (2) Subtraction operator: � ~a1 � ~a2 ¼ðl1 � l2; m1 � m2; r1 � r2Þ (3) Multiplication operator: � ~a1 � ~a2 ¼ðl1 l2; m1 m2; r1 r2Þ (4) Division operator: / ~a1=~a2 ¼ l1 r2 ; m1 m2 ; r1 l2 � � (5) Inverse power operators ~a�1=n1 ¼ðr �1=n 1 ; m �1=n 1 ; l �1=n 1 Þ (6) Distance of two fuzzy numbers (Chen, 2000) Dð~a1;~a2Þ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1=3½ðl1 �l2Þ 2 �ðm1 �m2Þ 2 �ðr1 �r2Þ 2� q (7) Defuzzication operator (Teng & Tzeng, 1993) Defuzzyð~a1Þ¼ jðr1 � l1Þþðm1 � l1Þj=3 þ l1 References Aamodt, A., & Plaza, E. (1994). Case-based reasoning: foundational issues, methodological variations, and system approaches. AI Com- munications Journal, 7(1), 39–59. Belecheanu, R., Pawar, K. S., Barson, R. J., Bredehorst, B., & Weber, F. (2003). The application of case-based reasoning to decision support in new product development. Integrated Manufacturing System, 14(1), 36–45. Besemer, S., & O’Quin, K. (1986). Analysis of creative products: refinement and test of a judging instrument. Journal of Creative Behavior, 20(2), 115–126. Bilgic, T., & Fox, M. S. (1996). Constraint-based retrieval of engineering design cases: context as constraints. In J. Gero & F. Sudweeks (Eds.), Artificial intelligence in design ’96 (pp. 269–288). Dordrecht: Kluwer Academic Publishers. Buckley, J. J. (1985). Ranking alternatives using fuzzy numbers. Fuzzy Sets and Systems, 15(1), 21–31. Chan, F. T. S. (2005). Application of a hybrid case-base reasoning approach in electroplating industry. Expert Systems with Application, 29, 121–130. Chen, C. T. (2000). Extensions of the TOPSIS for group decision-making under fuzzy environment. Fuzzy Sets and Systems, 114, 1–9. Chiou, H. K., Tzeng, G. H., & Cheng, D. C. (2005). Evaluating sustainable fishing development strategies using fuzzy MCDM approach. Omega, 33, 223–234. Creax.com. (2006). Available from http://www.creax.com/tools.htm. de Mantaras, R. L., & Plaza, E. (1997). Case base reasoning: an overview, IIA research report 96. AI Communications Journal, 10, 21–29. Goldfire Innovator: Invention-machine.com. (2006). Available from http://www.invention-machine.com/. Higgins, J. M. (1994). 101 Creative problem solving techniques: The handbook of new ideas for business. New Management Pub. Co. Hirota, K., Yoshino, H., Xu, M. Q., Zhu, Y., Li, X. Y., & Horie, D. (1997). An application of fuzzy theory to the case-based reasoning of the CISG. Journal of Advanced Computational Intelligence and Intel- ligent Informatics, 1(2), 86–93. Hisrich, R. D., Ingram, T. N., & Peters, M. P. (1991). Marketing decisions for new and mature products, 2/e. Maxwell Macmillam (pp. 172–174). Kolodner, J. L., & Leake, D. B. (1996). A tutorial introduction to case- based reasoning. In D. Leake (Ed.), Experiences lessons and future directions (pp. 31–66). Menlo Park, CA: AAAI Press. http://www.creax.com/tools.htm. http://www.invention-machine.com/ 540 M.-C. Wu et al. / Expert Systems with Applications 34 (2008) 530–540 Kotler, P. (1994). Marketing management: Analysis, planning, and control, 8/e. Englewood Cliffs, NJ: Prentice-Hall (pp. 319–328). Laarhoven, P. J. M., & Pedrycz, W. (1983). A fuzzy extension of Saaty’s priority theory. Fuzzy Sets and Systems, 11(3), 229–241. Linda, R. (1991). Generating and screening new products ideals. Industrial Marketing Management, 20, 287–296. Maher, M. L., Balachandran, M. B., & Zhang, D. M. (1995). Case-based reasoning in design. Hillsdale, NJ: Lawrence Erlbaum. Maher, M. L., & Garza, A. G. S. (1997). Case-based reasoning in design. IEEE Expert(March–April), 34–41. Mann, D. (2003). Hand-on systematic innovation. CREAX Press. Marling, C., Sqalli, M., Rissland, E., Munoz–Avila, H., & Aha, D. (2002). Case-based reasoning integrations. AI Magazine, 23(1), 69–86. Moon, J. H., & Kang, C. S. (2001). Application of fuzzy decision making method to the evaluation of spent fuel storage options. Progress in Nuclear Energy, 39(3–4), 345–351. Nijssen, E. J., & Lieshout, K. F. M. (1995). Awareness, use, and effectiveness of models and methods for new product development. European Journal of Marketing, 29(10), 27–43. Norman, D. A., & Draper, S. W. (1986). User-centered system design: New perspectives on human–computer interaction. Hillsdale, NJ: Lawrence Earlbaum Associates. Osborn, A. (1963). Applied imagination. New York: Charle Scribners Son (pp. 151–165). Rantanen, K., & Domb, E. (2002). Simplified TRIZ: New product-solving applications for engineers and manufacturing professionals. ST. Lucie Press. Ruet, M., & Geneste, L. (2002). Search and adaptation in fuzzy object case-base. Advances in case-based reasoning. In Proceedings of the sixth European conference, ECCBR 2002, Aberdeen, Scotland, United Kingdom, September 4–7, 2002 (pp. 350–364). Saaty, T. L. (1980). The analytic hierarchy process: Planning, priority setting, resource allocation. New York: McGraw-Hill, p. 20. Teng, J. Y., & Tzeng, G. H. (1993). Transportation investment project selection with fuzzy multi-objective. Transportation Planning and Technology, 17, 91–112. Trisolver.com. (2006). Available from http://www.trisolver.com/. Trousse, B., & Visser, W. (1993). Use of case-based reasoning techniques for intelligent computer-aided-design systems. In Proceedings of the IEEE international conference on systems, man and cybernetics (Vol. 3, pp. 513–517). Watson, I., & Marir, F. (1994). Case-based reasoning: a review. The Knowledge Engineering Review, 9(4), 327–354. Wu, M. C., Lo, Y. F., & Hsu, S. H. (2006). A case-based reasoning approach to generating new product ideas. Journal of Advance Manufacturing and Technology, 30, 166–173. Zadeh, L. A. (1975). The concept of a linguistic variable and its application to approximate reasoning. Information Sciences, part1:8, 199–249, part2:8, pp. 301–357, part3: pp. 43–80. http://www.trisolver.com/ A fuzzy CBR technique for generating product ideas Introduction Case-based reasoning Product database Product attributes for creating ideas Product attributes for screening ideas Linguistic variables for describing product attributes Weighting each dimension of target customers ' use-scenario Retrieving and filtering product ideas Retrieving mechanism Filtering mechanism Experiments Generating product ideas Comparison of product-idea-generating mechanisms Concluding remarks Arithmetical operators for fuzzy numbers References