A knowledge-based ingredient formulation system for chemical product development in the personal care industry

Computers and Chemical Engineering 65 (2014) 40–53

Contents lists available at ScienceDirect

Computers and Chemical Engineering journal homepage: www.elsevier.com/locate/compchemeng

A knowledge-based ingredient formulation system for chemical product development in the personal care industry C.K.H. Lee ∗ , K.L. Choy, Y.N. Chan Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hunghom, Hong Kong

a r t i c l e

i n f o

Article history: Received 18 July 2013 Received in revised form 17 February 2014 Accepted 10 March 2014 Available online 19 March 2014 Keywords: Knowledge management New product development Ingredient formulation Chemical product Personal care industry Case-based reasoning

a b s t r a c t The formulation of personal care products involves a trial-and-error approach to testing different combinations of chemicals. Specific knowledge plays an important role in creating the desired product properties. Without knowledge support tools, the formulation process becomes iterative. Furthermore, personal care products cannot be designed without analyzing market needs, and their development thus involves the collaboration of the formulators and the marketing teams. Miscommunication can reduce the efficiency and effectiveness of the process. This paper presents a knowledge-based ingredient formulation system for supporting chemical product development in the personal care industry. Case-based reasoning is used to solve ingredient formulation problems with reference to how similar past problems have been solved. The system also acts as a collaborative platform for sharing knowledge among the various stakeholders. A case study confirms the viability of the system, and the results show that the system provides formulators with key knowledge, enabling effective product formulation. © 2014 Elsevier Ltd. All rights reserved.

1. Introduction Spurred by the increasingly volatile environment, attention has shifted from process design to product development. As most chemical-based products, such as personal care products and cosmetics, are customer-oriented, new product development (NPD) cannot be isolated from marketing issues. In the conventional approach, chemical product development, especially in the personal care industry, is a cross-functional process involving two major units: (i) the Sales and Marketing Department and (ii) the Research and Development (R&D) Department, as shown in Fig. 1. Sales teams analyze the market trends and define the desired product attributes before creating a NPD enquiry for the formulators in the R&D Department. However, some product attributes are defined by the Sales and Marketing Department, such as the smoothness and softness of products, which are sensorial and thus subjective. Sales staff lack the technical knowledge to help them precisely express their preferences regarding those attributes. As a result, miscommunication between sales staff and formulators

∗ Corresponding author at: Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hunghom, Kowloon, Hong Kong. Tel.: +852 2766 6630. E-mail addresses: [email protected] (C.K.H. Lee), [email protected] (K.L. Choy), [email protected] (Y.N. Chan). http://dx.doi.org/10.1016/j.compchemeng.2014.03.004 0098-1354/© 2014 Elsevier Ltd. All rights reserved.

can easily occur as their perceptions of these subjective product attributes are different. If formulators interpret sales staff intentions or preferences for product attributes inaccurately, the ingredient formulation becomes iterative, lowing the efficiency and effectiveness of the entire NPD process. In addition, the ingredients involved in chemical products are often chosen based on previous experience with similar products (Cheng, Lam, Ng, Ko, & Wibowo, 2009). Without any reference to the previous similar NPD cases, formulators can make decisions based only on their own judgments. In this case, the decision quality involved in chemical product development cannot be guaranteed. A knowledge-based ingredient formulation system (KIFS) is presented in this study for addressing the needs in practice. There are two objectives to be achieved by the KIFS. First, the KIFS must act as a knowledge-based collaborative platform for translating the product requirements from sales into NPD knowledge, supporting formulators in formulating personal care products accordingly. This will eliminate any potential miscommunication problems and will improve the efficiency of chemical product development. Second, the KIFS provides decision support to formulators in their attempt to select the most appropriate ingredient formulae. To achieve this, a knowledge-based artificial intelligence (AI) technique, namely case-based reasoning (CBR), is embedded into the system. It is used to address the new chemical development problems by utilizing the knowledge gained in solving similar past cases.

C.K.H. Lee et al. / Computers and Chemical Engineering 65 (2014) 40–53

41

Fig. 1. Existing problems in chemical product development in the personal care industry.

In general, there are three types of NPD in the personal care industry. (i) The first type involves new products which are initialized by the R&D Department. Formulators and chemists initialize this type of NPD when they are inspired by new chemical ingredients, suppliers or information from trade exhibitions, trade conferences, market reports and published papers. After a prototype is created, the R&D Department will present it to the Sales and Marketing Department. Since marketing is not greatly involved, the functions provided by KIFS may be of less importance to this type of NPD. Therefore, it is not the main focus of this paper. (ii) The second type involves the development of new products which are targeted to a certain group of customers. Information for these kinds of new products is based on customer information. In usual practice, the Sales and Marketing Department creates product concepts such as “rejuvenating” and “refreshing”, and the R&D Department then formulates products based on the given concepts. Panel evaluation is then required by having both departments judge how well the prototype fits the concept. In the evaluation, some subjective product attributes are involved and without any knowledge support tools, miscommunication problems between different departments can occur easily. (iii) The third type is through reverse engineering in which some products from the markets or competitors are used as benchmarks. In this type of NPD, the Sales and Marketing Department will set some requirements on several subjective product attributes and the formulators and chemists have to interpret the meanings. For example, if the product used for benchmarking has a very smooth texture, the Sales and Marketing Department will require the prototype to have similar smoothness. If there are any misinterpretation, rework may be needed, lengthening the NPD cycle time. As the focus in this study is to develop a system which can improve the communication between the Sales and Marketing Department and the R&D Department, the effectiveness of the KIFS is confined to the second and the third types of NPD. This paper

is divided into six sections. Section 2 describes the past literature related to this study. Section 3 introduces the architecture of the KIFS. Section 4 contains a case study for demonstrating the feasibility of the system. Case study results and discussion of the system are presented in Section 5. Finally, a conclusion of this study is made in Section 6.

2. Literature review Faced with a dynamic and turbulent environment that requires flexibility for the changing business needs, organizations are seeking ways to improve their new product development (NPD) so as to maintain substantial growth for business survival (Chan & Ip, 2011; Jang, Dickerson, & Hawley, 2005; Zahay, Griffin, & Fredericks, 2004; Zapata, Varma, & Reklaitis, 2008). An NPD process is a sequence of steps or activities that an organization employs to conceive, design and commercialize a product (Aya˘g, 2005). Since most chemical-based products are formulated products containing different chemicals, active ingredients as well as additives (Conte & Gani, 2011; Mattei, Kontogeorgis, & Gani, 2012), its NPD process involves the generation and screening of a large number of chemical ingredients (Gani, 2004). The potential search space involved in the ingredient formulation process in chemical product development is large, constituting a significant portion of the entire NPD cycle time. A good formula is believed to be able to give a good compromise between different product properties (Claeys-Bruno, Lamant, Blasco, Phan-Tan-Luu, & Sergent, 2009). Ingredient formulation is usually undertaken through experimental trial-and-error techniques (Cheng et al., 2009; Conte, Gani, & Ng, 2011; Gani, 2004; Hill, 2009; Wibowo & Ng, 2004). This is a highly specialized task requiring a mix of skills from diverse disciplines to combine different chemical ingredients for creating the desired product properties (Conte, Gani, Cheng, & Ng, 2012). Without any knowledge support tools, chemical product development becomes iterative and time-consuming until an acceptable ingredient formula can be established. Since speeding up the product development is of paramount importance in enhancing competitiveness (Charpentier, 2009), the conventional approach is not generally appropriate for today’s time-sensitive markets.

42

C.K.H. Lee et al. / Computers and Chemical Engineering 65 (2014) 40–53

Fig. 2. CBR cycle.

In addition, a chemical product cannot be designed without considering the profitability of the product (Hill, 2009). Cussler and Moggridge (2001) suggested four steps for chemical product design: to define market needs, to generate ideas to meet the needs, to select among ideas, and to manufacture the product. In particular, the creation and initial screening of product ideas should be based not just on an analysis of what is possible to manufacture, but also on an analysis of what the market would like to purchase (Hill, 2009). From the perspective of consumers, the end-use properties of consumer-based products, such as personal care products, are often more important than the chemical composition (Charpentier & McKenna, 2004). This highlights the importance of marketing issues in chemical product development and of the collaboration between the technical and marketing teams. However, due to the cross-functional nature of NPD, there is a high chance of having miscommunication during collaboration (Cooper, 2003). According to Cheng et al. (2009), it is easier for chemical engineers to deal with marketing knowledge on the job, than for business personnel to acquire the technical skills. In view of this, it is useful to develop an intelligent system which can translate the marketing knowledge into technical specifications for the products (Hoegl & Schulze, 2005). In order to develop a knowledge-based system for assisting in the efficient formulation of chemical-based products, artificial intelligence (AI) techniques play an important role (Rowe & Roberts, 1998). Case-based reasoning (CBR) is a knowledge-based AI technique which is fundamentally different from other AI techniques. It solves a new problem case by reusing the knowledge and experience gained from solving a previous problem in a similar situation (Kolodner, 1991). Processes involved in a CBR cycle, as shown in Fig. 2, can be generally described by four “REs” (Aamodt & Plaza, 1994): -

Retrieve the most similar case or cases; Reuse the knowledge provided in the case for problem solving; Revise the proposed solution; Retain the new experience as a new case for future problem solving.

Because of the abovementioned processes, it is believed that CBR is a good candidate for solving problems in experience-rich domains such as NPD. Haque, Belecheanu, Barson, and Pawar (2000) developed a decision support system using CBR to provide relevant information and knowledge for concurrent product development. In their system, the highly structured nature of the cases combined with what-if analysis enabled engineers to investigate different design scenarios. Choy et al. (2009) and Moon and Ngai (2010) applied CBR to search for fabrics whose characteristics were similar to the input enquiry for fashion product development. Lee, Choy, Law, and Ho (2012) used CBR to provide users with the

operational guidelines and suggestions for developing garment samples. Tacit knowledge for sample development can thus be retained in the garment industry. It is found that CBR can manage useful knowledge in a structured way for NPD. Considering that the development of chemical products is often based on previous experience with similar products (Cheng et al., 2009), it is expected that CBR can provide important assistance to chemical engineers in their attempts to develop chemical products. Craw, Wiratunga, and Rowe (1998) presented a CBR approach for tablet formulation in pharmaceutical product development. The results showed that their system can propose useful chemical ingredients for tablets. Similar to pharmaceutical product development, personal care product development also requires the identification of ingredients. However, in the above work, no marketing knowledge was considered during NPD. Therefore, their approach cannot be fully applicable to the resolution of the NPD problems in the personal care industry in which end-use properties of products, based on marketing analysis, are crucial. Furthermore, Avramenko and Kraslawski (2006) used CBR to formulate fat and oil products. In their approach, some important functional attributes such as flavor have to be translated into analytical measurements like melting properties and oxidative stability for calculating any case similarity. It was assumed that users had the technical knowledge to translate the intended functionality into specific parameters. However, in a real industrial NPD environment, not every party, such as the marketing team, has the technical knowledge to do so. Thus, their approach may not be able to address the conventional communication problems between the marketing and design teams in the personal care industry. In summary, not much research work has focused on CBR for the development of personal care products. On the other hand, in order to develop an effective CBR system, the most critical issue is the design of a case retrieval mechanism (Qi, Hu, Peng, Wang, & Zhang, 2009). In general, CBR retrieval methods can be broadly categorized into two types, which are the inductive indexing and the nearest neighbor approaches (Chow, Choy, Lee, & Chan, 2005; Tsai, Chiu, & Chen, 2005). The inductive indexing approach determines which features best discriminate cases and generates a tree-like structure to organize cases. The nearest neighbor approach is an exhaustive searching method which evaluates the dissimilarity between all past cases and the new case. Though the inductive indexing approach is useful when a single case feature is dependent upon others, it can be impossible to retrieve cases when there is a case feature missing or unknown. On the other hand, the nearest neighbor approach is simple to use but time-consuming, as it makes comparisons with every case. To compose the niches of these two approaches, one of the most popular methods is to apply the inductive indexing approach to retrieve a set of matching cases, followed by the nearest-neighbor approach to rank the cases in the set according to the similarity to the new case (Choy, Lee, Lau, & Choy, 2005; Shin & Han, 2001; Wess, Althoff, & Derwand, 1994). In recent years, AI techniques such as fuzzy logic and the genetic algorithm (GA) have also been integrated with CBR to construct the retrieval mechanism. For instance, Wu, Lo, and Hsu (2008) developed a fuzzy CBR retrieval mechanism to retrieve product ideas that tend to enhance the functions of a given product. In their algorithm, linguistic variables in fuzzy theory are advocated to define the case attributes. Chiu (2002) integrated GA with CBR to determine a set of weighting values that can best formalize the match between the input case and the past cases. It is observed that there are many approaches for improving the performance of CBR systems in problem solving. However, CBR systems focusing on the ingredient formulation of personal care products have been scarce. As a pioneering study in this area, this paper presents a case-based system to provide knowledge support to the ingredient formulation process during chemical product development in the personal care industry. The incorporation of both inductive indexing and

C.K.H. Lee et al. / Computers and Chemical Engineering 65 (2014) 40–53

43

Fig. 3. The architecture of a knowledge-based ingredient formulation system.

the nearest neighbor approach is used in the system due to its well proven capabilities in effective case retrieval. The system is also designed in such a way that it can act as a collaborative platform between the marketing teams and R&D teams. The ingredient formulation of the new product at hand can be obtained by referring to similar past cases. 3. A knowledge-based ingredient formulation system In order to provide knowledge support for chemical product development, an intelligent system, the knowledge-based ingredient formulation system (KIFS), has been designed, as shown in Fig. 3. It acts as a knowledge-based platform to transform the NPD requirements provided by the Sales and Marketing Department into knowledge supporting R&D activities so that formulators can develop the desired chemical products in an effective manner. There are two modules in the KIFS responsible for the knowledge reformulation process: the Case Retrieval Module and the Case Adaptation Module. In the Case Retrieval Module, there is a case library for storing historical NPD cases, each of which records the details of the ingredient formulation of a particular product. A hybrid inductive indexing and nearest neighbor approach is used to develop a mechanism for retrieving suitable cases from the case library based on the input enquiries. In order to apply the inductive indexing approach, historical cases are stored in a tree-like structure in the case library. There are several levels in the induction tree, each of which contains different clusters. Cases are allocated in different clusters based on their product attributes such as product category and price range. For instance, cases describing the same product category, such as “Cream”, will be grouped in the same cluster of the tree at the “Product Texture” level. If the price ranges

of those products are the same, the corresponding cases will also be grouped together at the next level, which is the “Price Range” level. As a result, the number of cases in a cluster decreases when the price level increases. After the required product attributes of the new case are inputted by the users, a search path is identified to retrieve potentially useful cases in a particular cluster. This ensures that the potentially useful cases retrieved by the inductive indexing approach possess the product attributes defined by the users. A nearest neighbor approach is then used to prioritize the potentially useful cases based on their similarity values. It is expected that the case with the highest similarity value will be the most appropriate case for solving the ingredient formulation problem of the input case. Details of the ingredient formulation of the new case can be an adaptation of cases with highest similarity values. In the Case Adaptation Module, reuse and revision of a retrieved case are undertaken. Formulators are allowed to make reference to the retrieved case when they are formulating the ingredient formula for the new product at hand. After the validity of the ingredient formula has been determined, the details of ingredient formulations are recorded as parts of the content of a new case. The new case is then retained and stored in the case library for future use. Details of the two modules constituting the KIFS are discussed in the following sections. 3.1. Case Retrieval Module When a chemical product has to be launched in the market after market analysis, the Sales and Marketing Department inputs the relevant product attributes into the KIFS. These attributes contain the required data for determining both the physical and chemical

44

C.K.H. Lee et al. / Computers and Chemical Engineering 65 (2014) 40–53

properties of the active ingredients as well as the design of a base case formula. In general, the attributes can be categorized into two types: (a) objective product attributes, and (b) subjective product attributes. The objective product attributes can be regarded as the general description of the products, such as the product category, the color and functions of the product. On the other hand, the subjective product attributes can be regarded as personal preferences for the target product properties, such as the softness, stickiness and greasiness of the product. They are subjective as different people may have different perspectives, based on personal judgment and human sensitivity. Sales personnel in the Sales and Marketing Department express these attributes usually by verbal means but these attributes are hard to convert into exact quantitative values based on conceptual ideas. It is also difficult for sales personnel to do this as they lack specific domain knowledge such as chemical knowledge. Because of the subjective evaluation of some product properties, miscommunication problems usually occur between sales personnel and formulators when dealing with subjective product attributes. In the Case Retrieval Module of the KIFS, salespeople are required to input their expected degrees of the subjective product attributes. Similar past NPD cases possessing similar degrees of the subjective product attributes will be retrieved and sent to the formulators as knowledge support. This can thus eliminate the traditional miscommunication problems. To facilitate the case retrieval process, a number of historical NPD cases are initially stored in the case library in a tree-like structure. In each level of the tree, there are different clusters representing different objective product attributes. Based on their product attributes, cases are stored in different clusters of the tree. After all the required attributes are inputted into the KIFS, the Case Retrieval Module begins to browse in the case library in order to retrieve potentially useful cases using the inductive indexing approach, with a search path. The search path is a path connecting every cluster containing the defined objective product attributes. Along it, cases stored in the cluster at the last level are retrieved as potentially useful cases. Using this inductive indexing approach can guarantee that the potentially useful cases retrieved must contain the objective product attributes defined by the sales personnel. To further resolve the suitability of cases, the nearest neighbor approach is applied to calculate the similarity values of each potentially useful case. This approach is used to compare the subjective product attributes between potentially useful cases and the new input case. The similarity value (S) is calculated according to Eq. (1):

n

S=

i=1

wi × sim(f )

n

i=1

wi

(1)

where wi is the weighting of the subjective product attribute i, and sim(f) is the similarity function between the potentially useful case and the input case. After calculation, cases are ranked according to their similarity values in descending order. Cases with high similar values are regarded as significant cases for the generation of the final solution. Thus, they will be transferred to the next module where case adaptation takes place. 3.2. Case Adaptation Module The Case Adaptation Module is responsible for the processes of new case creation case reuse, case revision and case retention. Since it is expected that cases with high similarity value are capable of solving the ingredient formulation problems of new cases, the details of ingredient formulation of new cases can be derived through an adaptation of cases with high similarity values. As the

KIFS only supports the decision-making process involved in chemical product development, users, mainly the formulators in the R&D Department, still have the authority to decide if the solution stated in the retrieved case is applicable or not to solving the problems of new cases. In most situations, the retrieved case only matches the new case to a certain degree. Thus, the case has to be revised according to the actual situation and the revised solution still has to be verified by means of experiments and testing. If the retrieved case is applied directly to solve the new problem without any modification, the new product at hand will simply be based on unmodified similar past products. Therefore, in chemical product development, it is always essential to revise the case content of any new products. For example, the ingredient formulae may be revised by adding new, refined concentrations of active ingredients. After the details of the ingredient formulae are recorded as a new case, the new case is retained in the case library for future use. Because of the case revision and retention process in the Case Adaptation Module, the KIFS is given a learning capability. The quality of the retrieved case can thus be enhanced as the number of cases stored in the case library increases over time. The tacit knowledge for chemical product development is also retained in the form of cases.

4. Case study The viability of the proposed KIFS is verified by means of a case study in which a pilot run of the system is undertaken in a Hong Kong-based manufacturing company in the personal care industry. The case company, MFI, is an international enterprise dedicated to providing customers with quality chemical-based products, such as body and skin care products, and carries a comprehensive lineup of internationally renowned brands. It has been a listed company on the HK stock market since 2007, and was the first company to attain ISO certification in Hong Kong. In recent years, in order to deal with the challenges brought by globalization and to maintain a substantial growth of business, MFI has made two strategic changes to entering the retailing market. Firstly, MFI has developed bath and body care products as well as skin care products, aimed at customers who are looking for natural beauty care products in Hong Kong and China. Secondly, it has entered the cosmetic and fashion accessories retailing market in China through the acquisition of one of the famous retailers which has about 2000 outlets. With such changes in organizational business strategies, more chemical products have to be developed in order to respond to the market demand. Given the complexity of the ingredient formulation process, a decision support system is needed to provide knowledge support for assisting chemical product development. Currently, there are two existing problems in the case company: (i) lack of guidelines for decision making, and (ii) miscommunication between the sales teams and R&D formulators.

(i) Lack of guidelines for decision making In current practice, there are no formal documents or guidelines provided for formulators when they are developing new chemical products. In fact, some products having the same functional requirements or properties may require similar combinations of ingredients and experimental NPD procedures. However, without any decision support tools, no references can be made to NPD projects of a similar nature. This may result in repeated efforts in dealing with similar ingredient formulation problems. Furthermore, when some experienced formulators resign or retire, the product quality is lowered during the period of training of new formulators who receive insufficient knowledge support from the company. This will significantly affect

C.K.H. Lee et al. / Computers and Chemical Engineering 65 (2014) 40–53

the competitive advantage of the company in today’s competitive business environment. (ii) Miscommunication between sales teams and R&D formulators In the company, miscommunication between sales teams and formulators always happens when their perceptions of some sensorial properties of products, such as softness and greasiness of the product, are different. Formulators have to guess the preferences of sales personnel in order to transform the conceptual product ideas into technical specifications. If their perceptions are different, the product formula developed has to be adjusted, lengthening the NPD cycle time. To tackle the above problems, the company trial launched the KIFS to handle their new NPD projects. The details of the system’s construction are presented in the following sections. 4.1. Construction of the KIFS Four phases are involved in the implementation of the KIFS in the case company, they are: (i) Design of the content of the case, (ii) Data collection and storage, (iii) Construction of an induction tree for case allocation, and (iv) Definition of subjective product attributes and weightings for case ranking. 4.1.1. Design of the content of the case In this phase, a survey is conducted within the R&D Department of the case company. The aim is to identify important information for developing a chemical product so as to design the content. To ensure the effectiveness of the KIFS, the content of a case displayed to users should include the important information so that users are able to make use of it for NPD. Based on the survey results, it is found that critical information supporting the ingredient formulation can be divided into four aspects, which are (a) ingredients, (b) procedures, (c) parameter settings, and (d) equipment. Firstly, ingredients refer to any chemicals involved in the product such as active ingredients, solvents and additives. Secondly, the procedures refer to the workflow of how the chemical ingredients are mixed together so as to achieve the desired specification of the products. Thirdly, parameter settings mean the parameters involved in the entire ingredient formulation process, including the concentration of ingredients, time, speed and temperature used in experiments and performance tests. Fourthly, equipment means different kinds of equipment and tools which are required when developing the chemical product. These four aspects are essential for ingredient formulation and are thus included in the content of cases as an output of the system. 4.1.2. Data collection and storage Data collection is conducted to build the case library. Since the data important for ingredient formulation come from diverse sources, a centralized database is used for data consolidation and storage. For instance, data related to chemical ingredients are initially stored in the internal database in the R&D Department while the customer data and product data are stored in the Customer Relationship Management section (CRM) and SAP, respectively. As data from different sources may be stored in different formats, Extensible Markup Language (XML) is used as the standard format for data exchange. Data in the database are stored in relational tables as shown in Fig. 4.They are used for ease of construction of the case library and for providing extra information for any retrieved cases. 4.1.3. Construction of an induction tree for case allocation In this phase, an induction tree is constructed to allocate cases into different clusters for inductive indexing. It is used to retrieve cases which potentially contain the objective product attributes

45

defined by users. According to domain experts, there are six objective product attributes which have to be defined before ingredient formulation. They are product application, product category, product texture, product function, regulation compliance and price range. The hierarchy of the induction tree is designed according to these six objective product attributes. Therefore, it consists of the six levels shown in Fig. 5, ranked according to the importance of the attributes. The most important attribute will be put in the first level while the least important one will be put in the last level. The six levels are: -

Level 1: Product Application Level 2: Product Category Level 3: Product Texture Level 4: Product Function Level 5: Regulation Compliance Level 6: Price Range

In the KIFS, historical cases are allocated in different clusters of the tree based on their product attributes. The “Product Application” of cases is firstly matched in Level 1, followed by “Product Category” and then “Product Texture”, and so on. The number of cases decreases when the number of levels increases. As a result, the number of cases is narrowed down by the induction indexing approach before the application of the nearest neighbor approach commences. The overall case retrieval speed can thus be increased. 4.1.4. Definition of subjective product attributes and weightings for case ranking To execute the nearest neighbor approach, different subjective product attributes have to be defined to describe the sensorial properties of the product. There are six subjective product attributes identified as important for ingredient formulation, which are: -

Greasiness Smoothness Softness Spreadability Stickiness Viscosity

The above six attributes are commonly used by the Sales and Marketing Department to describe some sensorial product properties. In addition, it is more convenient for sales personnel to describe these properties by verbal means rather than by expressing them in quantitative values. In view of that, a 100-point scale is used for salesmen to show their preference for each attribute. Domain experts are invited to give scores to each of the above attributes for historical cases. Besides the six subjective product attributes, other non-rhelogical properties are also important. However, it could be difficult to quantify non-rhelogical properties by requiring users to input a score ranging from 1 to 100. Therefore, users are not required to input them for ease of the use of the system. Instead, non-rhelogical properties such as color will be displayed to users directly after they retrieve a potentially useful case. The calculation of similarity values is used to compare the score of each attribute between the new case and the historical cases. Furthermore, these attributes may have different priorities for different product types. Therefore, weightings have to be assigned to the attributes based on their importance. If the attribute is relatively more important, its weighting will be relatively higher. To facilitate the case ranking in the KIFS, there are three levels of importance: (i) less important, (ii) important, and (iii) very important, with a weighting value of 1, 3, and 5, respectively. Users are allowed to

46

C.K.H. Lee et al. / Computers and Chemical Engineering 65 (2014) 40–53

Fig. 4. Data stored in the database in relational tables.

give the importance level of each attribute and the corresponding weighting will be considered in the calculation of similarity values.

4.2. Operation flow of the KIFS Fig. 6 shows a comparison between the conventional approach and the proposed approach with the use of KIFS for chemical product development. In the conventional approach, the Sales and Marketing Department defines the NPD requirements and sends NPD inquiries to the R&D Department. Formulators in the R&D Department have to interpret the preferences of the salesmen regarding subjective product attributes when they design a product formula. If the product formula is not accepted by the sales people, they have to design it again, thus lengthening the entire NPD cycle time.

Usually, subjective evaluation is one of the main causes of conflict between salesmen and formulators. For example, salesmen may require the product to be softer, less greasy or sticky. However, it is difficult for them to express their preferences precisely. Without the use of any knowledge support tools, the ingredient formulation becomes repetitive as the formulators have to develop several product formulae before there is one which can fulfill the expectations of the sales personnel. On the other hand, the KIFS provides a platform to convey the NPD requirements from sales to relevant NPD knowledge supporting R&D activities, so that formulators can have reference to similar products while solving the NPD problem at hand. With the knowledge supported by the KIFS, formulators are not required to interpret sales staff preferences subjectively but are provided with clearer product specifications for NPD. There are seven main steps involved in operating the KIFS, which are discussed in the following sections.

C.K.H. Lee et al. / Computers and Chemical Engineering 65 (2014) 40–53

Fig. 5. Induction tree.

Fig. 6. A comparison between (a) conventional approach and (b) proposed approach (KIFS) for chemical product development.

47

48

C.K.H. Lee et al. / Computers and Chemical Engineering 65 (2014) 40–53

Fig. 7. User interface for data input into the KIFS.

4.2.1. Step 1: Defining/refining the requirements of the new product The first step is performed by the sales personnel who are required to define or refine their requirements of the new product. Fig. 7 shows the user interface for data input. Users are allowed to input both objective and subjective product attributes into the KIFS. There are different objective product attributes, describing physical product properties, for salesmen to select and input into the system. For each subjective product attribute describing sensorial product properties, users are required to give a score, representing their preference, ranging from 1 to 100. For instance, when the salesmen define their preference on the stickiness of the product, they have to input the degree from “non-sticky” to “very-sticky” on a 100-point scale. In addition, users have to define the weightings of these attributes to indicate the importance of the attributes. The weightings and scores are used in the nearest neighbor approach for calculating the similarity values of cases in Step 3.

4.2.2. Step 2: Retrieving potentially useful cases After the data are inputted in the KIFS, the KIFS retrieves potentially useful cases by using the inductive indexing approach. Based on the objective product attributes input in Step 1, a search path connecting clusters with the defined objective product attributes is formed. For example, if the Sales and Marketing Department would like to develop a leave-on product which is a face cream with a moisturizing function, with a selling price within USD101 to USD250, and has to fulfill the IFFR Compliance, clusters representing the corresponding attributes are all connected to form a search path for case retrieval, as shown in Fig. 8. Along the search path, cases stored in the identified cluster at the last level are retrieved as potentially useful cases.

4.2.3. Step 3: Ranking the potentially useful cases In this step, the nearest neighbor approach is applied to calculate the similarity value of each potentially useful case based on the score for each subjective product attribute input in Step 1. Potentially useful cases are ranked according to their similarity values in descending order. By using the formula (100 − |pR − pI |)/100 × 100% where pI is the score of preference, input by users, and pR is the score of preference stored in the potentially useful case, the value of the similarity function of each attribute is calculated. Substituting the similarity function into Eq. (1), the similarity value of the case is obtained. As illustrated in Fig. 9, a potentially useful case “A” is chosen for demonstrating how the similarity value is generated. After calculating the similarity value of every potentially useful case retrieved in Step 2, the KIFS will display the top five cases with high similarity values in descending order, as shown in Fig. 10. The case with the highest similarity values is expected to be the case most capable of solving the problems in developing the new product, due to its high degree of similarity. However, the KIFS solely supports the decision-making process without making any decisions and the similarity values of each case act only as a reference for users. Thus, users have to decide if the cases displayed by the KIFS are capable of solving the new problem, which will be performed in Steps 4–6.

4.2.4. Step 4: Selecting a case After cases are retrieved by the KIFS, sales personnel are allowed to view the details of each case so as to select an appropriate case that they think is similar to the target new product. By referring to the product ID stored in a case, they can look for the corresponding product and check if its subjective product attributes can fulfill their sensorial expectations. If so, the corresponding case can be selected and passed to R&D Department for the design of a new product formula using the selected case as a reference. Otherwise,

C.K.H. Lee et al. / Computers and Chemical Engineering 65 (2014) 40–53

49

Fig. 8. Potentially useful case retrieval process by the inductive indexing approach.

sales personnel repeat Step 1 to refine their NPD requirements until a similar case is retrieved. In this step, sales staff only select a case possessing the product attributes which are similar to their expectations. Due to their lack of domain knowledge, such as chemical knowledge, the case they select will be sent to R&D Department for further analysis. 4.2.5. Step 5: Reusing the knowledge of the selected case The retrieved case contains information and knowledge related to the ingredients used in the product, the working procedures involved for product development, the parameters such as time and temperature set during the procedures, and the equipment required. Fig. 11 shows the content of a retrieved case. It tells the formulators how a similar product was formulated in the past. Thereby, it acts as a direct guideline for formulators to develop new products possessing similar product properties. As mentioned in Section 4.1.4, some non-rhelogical properties, such as the color of the product, are not defined by users. However, these properties are important during NPD. In the KIFS, the product color of a retrieved case is shown in the system output interface as shown in Fig. 11. This reminds users to revise the case, especially when they prefer another color for the new product. On the other hand, properties such as odor is not included in the system as it is believed that requirements on the odor of a product can be easily catered for by fine-tuning the ingredient formulae at a later stage. Therefore, odor is relatively of less importance when developing an initial formula.

In fact, the subjective product attributes defined in this paper are case-sensitive. It is recommended that individual companies should determine their own attributes for their specific uses. 4.2.6. Step 6: Revising the case Revision of the case is essential to differentiate the new product from other similar products. As product safety is important in personal care products, different performance and stability tests are needed to verify the revision. When a final product formula is obtained, the details of the revision are recorded as a new case. 4.2.7. Step 7: Retaining the case After the content of the new case is revised, the case is retained and stored in the case library for future use. As a result, the number of cases increases the longer the KIFS operates. As new cases are revised in Step 6 before being stored in the system, relevant NPD knowledge can be recursively updated and retained in the system. With a larger case base, the chance of having retrieved cases with high similarity values is increased. This provides the KIFS with a learning capability which can enhance the quality of decisions in the long run. 5. Results and discussion In this section, three key performance indicators (KPIs) are selected for measuring the effectiveness of the KIFS. They are (i) average ingredient formulation time per NPD project, (ii) average

50

C.K.H. Lee et al. / Computers and Chemical Engineering 65 (2014) 40–53

Fig. 9. Calculation of the similarity value of a potentially useful case by the nearest neighbor approach.

5.1. Shortened NPD cycle time with a faster track of ingredient formulae

Fig. 10. Interface showing a list of ranked cases.

hit rate, and (iii) average number of trials per NPD project. During a nine-month period in a pilot run, the measurement of KPIs was conducted every three months. The results are shown in Table 1. The results show that the KIFS outperforms traditional approaches in ingredient formulation in chemical product development by offering a series of benefits which include:

The traditional human-based decision making process in chemical product development without the use of KIFS takes a lot of time in searching for an appropriate chemical ingredient formula. This involves a series of trial and error procedures, causing a long NPD cycle time. The ingredient formulation time refers to the time starting from the generation of a new product idea to the establishment of an initial ingredient formula. Such an initial formula indicates which ingredients and in which ratios can be considered in the formulation. Without KIFS, the company takes around 4 h on average to obtain an initial formula that still requires further finetuning. The company produces around 1200 new products per year. Among them, some require the establishment of new ingredient formulae, while some require the modification or fine-tuning based on ingredients of developed products. The former one involves a longer product development time while the latter one requires a shorter product development time. In the company, a relatively large portion of new products belong to the latter ones. As a result, it is found that the average ingredient formulation time per NPD project without KIFS, 4 h, could be relatively short if one compares

Table 1 Comparison of results with and without the use of KIFS. KPI

Without KIFS

Average ingredient formulation time per NPD Project Average hit rate Average number of trials per NPD project

4h 82% 5

With KIFS 3 months

6 months

9 months

3.2 h 87% 3.5

2.6 h 91% 2.2

2h 93% 2

C.K.H. Lee et al. / Computers and Chemical Engineering 65 (2014) 40–53

51

Fig. 11. A retrieved case.

it with that of a company that is a specialist in developing new formulae without benchmarking. With the use of KIFS, the average ingredient formulation time per project is shortened from 4 h to 2 h. Such an improvement is due to a faster tracking of ingredient formulae with the use of KIFS. Users are required to input only the requirements of the new product and they will then be provided with ingredient formulation solutions based on similar past cases. Having the retrieved solutions as a reference, users can have knowledge support for their decisions, increasing the efficiency of NPD.

5.2. Improved decision quality and customer satisfaction The average hit rate is used to describe the success rate of developed product formulae accepted by customers. After receiving a request from the Sales and Marketing Department, the R&D Department will search for an appropriate ingredient formula which can create the desired product properties to satisfy the customers. Without the use of KIFS, miscommunication between sales teams and formulators always occurs, causing the formulators to fail to create the desired product properties. With the use of KIFS, the average hit rate is improved from 82% to 93%. This shows that the NPD decisions supported by the KIFS are of better quality and can meet customers’ expectations. The KIFS acts as a common platform for sharing knowledge between the Sales and Marketing Department and the R&D Department. The NPD requirements defined by the sales staff are successfully conveyed to the R&D Department and are transferred into NPD knowledge supporting the ingredient formulation process through the use of KIFS. The subjective product attributes are now expressed in terms of a score ranging from 1 to 100. This helps eliminate the traditional miscommunication problems. The R&D Department can thus make better decisions which, in turn, can enhance customer satisfaction.

5.3. Increased effectiveness in NPD with better knowledge management Before the pilot run of the system, formulators have to test various formulae by means of experiments, tests and by trial-and-error, to verify their formulae. Stability tests, patch tests and compatibility tests are done to verify the formulae. In most cases, stability tests and compatibility tests can be performed concurrently during a three-month period of time. During the period, patch tests, each of which last for around 2 days, are also performed. These three main experiments are sequence-independent and the formulae are only accepted if they pass all the tests. As all these are highly specialized tasks, less experienced formulators may lack the capability to formulate chemical formulae effectively. For instance, they may lack experience and knowledge related to ingredients used in the products and the required procedures, making the trial process more iterative. However, after the use of KIFS, the average number of trials per project is significantly reduced from 5 to 2 as the expert knowledge is initially stored in the system as cases. When formulators are developing chemical products, the knowledge is delivered to them as a guideline to support their decision making processes. With such a knowledge support tool, the tacit knowledge for ingredient formulation is retained and the number of trials per project is reduced, increasing the overall effectiveness of NPD.

5.4. Continuous NPD improvement From Table 1, it can be seen that the three KPIs are improved over time. Because of the case retention process in the KIFS, more cases can be stored in the system when the system operates for a longer period of time, allowing the case base to expand continuously. When there is new case input to the system in the future, more potentially useful cases can be retrieved by the inductive indexing approach and the chance of obtaining a higher similarity

52

C.K.H. Lee et al. / Computers and Chemical Engineering 65 (2014) 40–53

for manufacturers in their attempt to set the right prices based on the value of a product to the end customers. In ideal cases, a product should be positioned on the value equivalent line (Bagajewicz, 2007) where products at the value equivalent line are sold at the right price from the perspective of consumers. However, in practice, manufacturers always face a tradeoff between price and product attributes. With the use of the KIFS, the economic factor is considered during case retrieval. Thereby, the potentially useful cases suggested are historical NPD cases in which the products are near the value equivalent line. In addition to price range of the product, there are still some other economic factors such as the effects of competition affecting the profitability of a product. In Fig. 12, different aspects are suggested for consideration for shifting products to the value equivalent line. They include product presentation, packaging, advertising, composition, production costs and selling price as mentioned by Bagajewicz et al. (2011). Hence, it is suggested that these factors can be included in the KIFS on demand if one aims to capture the role of microeconomics more comprehensively. Fig. 12. A framework for capturing the role of microeconomics.

6. Conclusions value of a case by the nearest neighbor approach can be increased. As a result, the probability of having a solution from a retrieved case which is capable of solving the new problem is increased and continuous NPD improvement can thus be achieved. 5.5. Alleviated the effect brought by subjective judgments to improve result sensitivity Since different people have different perceptions, a certain level of discrepancies between them is unavoidable when they define subjective properties based on their own preference. To alleviate the effect brought by subjective judgments, the KIFS is equipped with two features to increase the chance of obtaining the same results when different users work on a same product concept. Firstly, the searching mechanism of the KIFS can filter these mis-match cases from the target cases through the tree-like structure of the case library. In the induction tree, cases are stored in different clusters based on objective product attributes. Since the attributes are objective data, different users are able to perform the same search when they are referring to the same product concept. Only cases stored in the selected cluster are retrieved as potentially useful cases for further selection by users. Secondly, the KIFS provides users with a list of five potentially useful cases after case retrieval. Therefore, when there are some differences in the users’ input, the same results can still be generated as long as the differences between the similarity values of the cases are not large. Users are also allowed to view details of each case in order to guide them to find the most suitable one. 5.6. Captured the role of microeconomics in product design In the induction tree embedded in the KIFS, the price range of a product is included to capture the role of microeconomics in product design. This is an important feature of the KIFS for guiding the company to advance the formulation not only in an effective way, but also in a profitable way. Though different case companies have to define their own attributes based on their specific needs while adopting the KIFS, it is recommended that the price range has to be viewed as an important attribute and to be included in the induction tree. While consumers prefer buying products with superior attributes at a lower price, manufacturers prefer selling products at a higher price even if the product attributes are inferior. Hence, a framework for capturing the role of microeconomics in product design, as shown in Fig. 12, is developed so as to provide guidelines

This paper presents a KIFS using CBR techniques applied to the ingredient formulation process. As there are a large number of chemicals available for chemical product development, the potential search space for obtaining an appropriate formulation for creating certain product properties is very large. The use of CBR in the KIFS identifies a smaller number of previous design cases which possess the desirable product properties, thereby greatly reducing the search space that has to be traditionally investigated by a series of trial and error procedures. Considering that tacit knowledge and experience serve as a good starting point in chemical product development, the case retrieved by the KIFS facilitates the retention of tacit knowledge and experience gained from the development of similar products in the past. It provides direct guidelines for formulators to develop products based on previously developed products with similar product characteristics. From the case study, it is found that the use of KIFS can increase both the efficiency and effectiveness of chemical product development, with better knowledge management in the personal care industry. In addition, it allows sales personnel, who lack technical skills, to input new product requirements based on their marketing analysis in linguistic terms, such as “not soft” and “very soft”, when describing product properties. Their preferences for properties are expressed in terms of scores on a standardized scale. Therefore, the R&D Department can understand their preferences more accurately and the marketing knowledge can be concurrently considered in chemical product development. This narrows the gap between marketing and R&D as the marketing information is inputted into the KIFS and transformed into knowledge for supporting the NPD activities. Suggested future work includes the integration of fuzzy logic into the KIFS so as to help formulators determine the quantitative values of the parameters, such as time and temperature, for developing chemical products. Acknowledgement The authors would like to thank the Research Office of the Hong Kong Polytechnic University for supporting the current project (Project Code: RPXV). References Aamodt, A., & Plaza, E. (1994). Case-based reasoning: Foundational issues, methodological variations, and system approaches. AI Communications. IOS Press, 7(1), 39–59.

C.K.H. Lee et al. / Computers and Chemical Engineering 65 (2014) 40–53 Avramenko, Y., & Kraslawski, A. (2006). Similarity concept for case-based design in process engineering. Computers and Chemical Engineering, 30(3), 548–557. Aya˘g, Z. (2005). A fuzzy AHP-based simulation approach to concept evaluation in a NPD environment. IIE Transactions, 37(9), 827–842. Bagajewicz, M. J. (2007). On the role of microeconomics, planning, and finances in product design. AIChE Journal, 53(12), 3155–3170. Bagajewicz, M., Hill, S., Robben, A., Lopez, H., Sanders, M., Sposato, E., et al. (2011). Product design in price-competitive markets: A case study of a skin moisturizing lotion. AIChE Journal, 57(1), 160–177. Chan, S. L., & Ip, W. H. (2011). A dynamic decision support system to predict the value of customer for new product development. Decision Support Systems, 52(1), 178–188. Charpentier, J. C. (2009). Perspective on multiscale methodology for product design and engineering. Computers and Chemical Engineering, 33(5), 936–946. Charpentier, J. C., & McKenna, T. F. (2004). Managing complex systems: Some trends for the future of chemical and process engineering. Chemical Engineering Science, 59(8–9), 1617–1640. Cheng, Y. S., Lam, K. W., Ng, K. M., Ko, R. K. M., & Wibowo, C. (2009). An integrative approach to product development – A skin-care cream. Computers and Chemical Engineering, 33(5), 1097–1113. Chiu, C. (2002). A case-based customer classification approach for direct marketing. Expert Systems with Applications, 22(2), 163–168. Chow, H. K. H., Choy, K. L., Lee, W. B., & Chan, F. T. S. (2005). Design of a knowledgebased logistics strategy system. Expert Systems with Applications, 29(2), 272–290. Choy, K. L., Chow, K. H., Moon, K. L., Zeng, X., Lau, H. C. W., Chan, F. T. S., et al. (2009). A RFID-case-based sample management system for fashion product development. Engineering Applications of Artificial Intelligence, 22(6), 882–896. Choy, K. L., Lee, W. B., Lau, H. C. W., & Choy, L. C. (2005). A knowledge-based supplier intelligence retrieval system for outsource manufacturing. Knowledge-Based Systems, 18(1), 1–17. Claeys-Bruno, M. C., Lamant, J., Blasco, L., Phan-Tan-Luu, R., & Sergent, M. (2009). Development of a skin care formulation using experimental design. Chemometrics and Intelligent Laboratory Systems, 96(2), 101–107. Conte, E., & Gani, R. (2011). Chemicals-based formulation design: Virtual experimentations. Computer Aided Chemical Engineering, 29, 1588–1592. Conte, E., Gani, R., & Ng, K. M. (2011). Design of formulated products: A systematic methodology. AIChE Journal, 57(9), 2431–2449. Conte, E., Gani, R., Cheng, Y. S., & Ng, K. M. (2012). Design of formulated products: Experimental component. AIChE Journal, 58(1), 173–189. Cooper, L. P. (2003). A research agenda to reduce risk in new product development through knowledge management: A practitioner perspective. Journal of Engineering and Technology Management, 20(1), 117–140. Craw, S., Wiratunga, N., & Rowe, R. (1998). Case-based design for tablet formulation. In Advances in case-based reasoning. Berlin, Heidelberg: Springer. Cussler, E. L., & Moggridge, G. D. (2001). Chemical product design. USA: Cambridge University Press. Gani, R. (2004). Computer-aided methods and tools for chemical product design. Chemical Engineering Research and Design, 82(11), 1494–1504.

53

Haque, B. U., Belecheanu, R. A., Barson, R. J., & Pawar, K. S. (2000). Towards the application of case based reasoning to decision-making in concurrent product development (concurrent engineering). Knowledge-Based Systems, 13(2), 101–112. Hill, M. (2009). Chemical product engineering – The third paradigm. Computers and Chemical Engineering, 33(5), 947–953. Hoegl, M., & Schulze, A. (2005). How to support knowledge creation in new product development: An investigation of knowledge management methods. European Management Journal, 23(3), 263–273. Jang, N., Dickerson, K. G., & Hawley, J. M. (2005). Apparel product development: Measures of apparel product success and failure. Journal of Fashion Marketing and Management, 9(2), 195–206. Kolodner, J. L. (1991). Improving human decision making through case-based decision aiding. AI Magazine, 12(2), 52–68. Lee, C. K. H., Choy, K. L., Law, K. M. Y., & Ho, G. T. S. (2012). Decision support system for sample development in the Hong Kong garment industry. In Technology Management for Emerging Technologies (PICMET), 2012 Proceedings of PICMET’12 (pp. 754–761). IEEE. Mattei, M., Kontogeorgis, G. M., & Gani, R. (2012). A systematic methodology for design of emulsion based chemical products. Computer Aided Chemical Engineering, 31, 220–224. Moon, K. L., & Ngai, E. W. T. (2010). R&D framework for an intelligent fabric sample management system: A design science approach. International Journal of Operations & Production Management, 30(7), 721–743. Qi, J., Hu, J., Peng, Y. H., Wang, W., & Zhang, Z. (2009). A case retrieval method combined with similarity measurement and multi-criteria decision making for concurrent design. Expert Systems with Applications, 36(7), 10357–10366. Rowe, R. C., & Roberts, R. J. (1998). Artificial intelligence in pharmaceutical product formulation: Knowledge-based and expert systems. Pharmaceutical Science & Technology Today, 1(4), 153–159. Shin, K. S., & Han, I. (2001). A case-based approach using inductive indexing for corporate bond rating. Decision Support Systems, 32(1), 41–52. Tsai, C. Y., Chiu, C. C., & Chen, J. S. (2005). A case-based reasoning system for PCB defect prediction. Expert Systems with Applications, 28(4), 813–822. Wess, S., Althoff, K. D., & Derwand, G. (1994). Using k-d trees to improve the retrieval step in case-based reasoning. In Topics in case-based reasoning. Berlin, Heidelberg: Springer. Wibowo, C., & Ng, K. M. (2004). Product-centered processing: Manufacture of chemical-based consumer products. AIChE Journal, 48(6), 1212–1230. Wu, M. C., Lo, Y. F., & Hsu, S. H. (2008). A fuzzy CBR technique for generating product ideas. Expert Systems with Applications, 34(1), 530–540. Zahay, D., Griffin, A., & Fredericks, E. (2004). Sources, uses, and forms of data in the new product development process. Industrial Marketing Management, 33(7), 657–666. Zapata, J. C., Varma, V. A., & Reklaitis, G. V. (2008). Impact of tactical and operational policies in the selection of a new product portfolio. Computers and Chemical Engineering, 32(1–2), 307–319.