中華管理學報 第二卷 第一期 第75-87頁 民國九十年 An Application of Hierarchy-Oriented Case-Based Reasoning in the Pilot Production Stage Tzu-Fu Chiu Department of Industrial Management, Aletheia University 32, Jenli St., Danshuei Jen, Taipei, Taiwan Tel: 02-26212121 ext 5512 Fax: 02-26212121 ext 5512 email: chiu@email.au.edu.tw Abstract In this paper, the problem of organisation and re-use of hierarchical knowledge in the pilot production stage is described and a proposed architecture to aid this problem called Hierarchy-Oriented Case-Based Reasoning (HOCBR) is discussed. HOCBR consists a hierarchical case-based structure, multi-stage retrieval and hybrid adaptation. The hierarchical case-based structure provides a knowledge representation format for retaining hierarchical knowledge. Multi-stage retrieval is used to search through hierarchical case bases for the most similar cases to the input problem. Hybrid adaptation transforms the information from these cases into a suggested solution. A new system framework based on this architecture is put forward. A description of the implementation in the pilot production stage is also provided. An evaluation of the HOCBR architecture is provided along with possible future research. Keywords: HOCBR (Hierarchy-oriented case-based reasoning), Multi-stage retrieval, Hybrid adaptation, FMEA data, Pilot production. 階層導向案例基礎推理架構於試作生產階段之 應用 邱賜福 真理大學工業管理學系 台北縣淡水鎮真理街 32 號 摘 要 本文描述了在試作階段中階層式知識的組織與再使用問題,並提出了 一個協助解決此問題的階層導向案例基礎推理(HOCBR)架構。HOCBR 架 構包含了:階層式案例庫結構、多階段尋取及混合式轉化等三個方法。階 層式案例庫結構,提供了一個儲存階層式知織的知識表示格式;多階段尋 取,是用來在階層式案例庫中搜尋出一個與輸入問題最相似的案例;混合 75 Tzu-Fu Chiu An application of hierarchy-oriented case-based reasoning in the pilot production stage 式轉化,是將最相似案例的內涵轉換為適用於輸入問題的建議解答。基於 HOCBR 架構作者建立了一個雛型系統,而此雛型系統即應用於試作階段 ,最後對此一 HOCBR 架構進行評估,並提出未來可能研究方向。 關鍵詞:階層導向案例基礎推理,多階段尋取,混合式轉化,FMEA 資,試 作生產 1. Introduction This paper attempts to organize and re-use the hierarchical knowledge in the pilot production stage. Hierarchical knowledge is usually complex and great in volume. Simple representations of knowledge, such as predicate rules, are not rich enough to efficiently encapsulate all of the factors within this type of information. Complex knowledge is also difficult to formulate into general models. Case-Based Reasoning (CBR), a paradigm which solves a new problem by remembering a previous similar situation and by reusing information and knowledge of that situation [1], is a particularly useful approach for use in the retrieval and adaptation of hierarchically complex knowledge. The reason for this is that CBR offers a more suitable structure for compound knowledge and provides a partial matching mechanism for searching out similar existing cases [2, 10]. Pilot production is a stage of the production cycle that prepares available material, existing machines and suitable methods for examining the feasibility of the planned process. It also finds any possible failures and is usually a complicated and time-consuming procedure. The essential knowledge items contained in pilot production are manufacturing data and FMEA 1 data. The relationship between the data items is hierarchical, as a product may include several parts, each part may contain a number of processes and every process may hold a set of processing d ata and a set of FMEA data. Due to the characteristics of the problem domain and the features of the CBR approach, the Hierarchy-Oriented Case-Based Reasoning (HOCBR) architecture, which comprises a hierarchical case-based structure, multi-stage retrieval and hybrid adaptation, is proposed. Utilizing multi-stage retrieval, HOCBR efficiently seeks out the most similar cases from the complex knowledge organised in a hierarchical case-based structure. Using hybrid adaptation, HOCBR also effectively transforms the similar data into more suitable suggested information for the input problem. In the next section the HOCBR architecture is described. This is followed by a description of the system framework for the HOCBR architecture that facilitates the building of a prototype system. The prototype system is subsequently discussed. The discussion continues with an evaluation of the HOCBR architecture. Finally, conclusions are presented and future work outlined. 2. Proposed Hierarchy-Oriented CBR Architecture The basic concept behind the HOCBR architecture is derived from the Material 1 Failure Modes and Effects Analysis (FMEA) involves the investigation and assessment of all causes and effects of all possible failure modes on a system, in the earliest development phases [18]. 76 中華管理學報 第二卷 第一期 第75-87頁 民國九十年 Requirements Planning (MRP) procedure used in industry, which obtains the independent demand from the customer order and then calculates the dependent demand level-by-level, according to the bill of material [12]. The bill of material, a structured parts list, shows the hierarchical relationship between the finished product and its various components [5]. Case-based reasoning is suitable for problem areas where the knowledge (or data) is complex and not easily organized into patterns or models [9, 17]. Using the concept of MRP and the capabilities of CBR, case-based reasoning methods are modified and developed to deal with the hierarchical data structure. The three main components of HOCBR are presented below. 2.1 Hierarchical Case-Based Structure Hierarchical data can be decomposed into different, but, related types. A car part product can be broken down into product data (a sub-assembly is called a product, since it is a finished item from a factory), part data and process data. A product may contain several parts, with each part fabricated by several processes. Each process for a part consists of a group of processes and the previously mentioned FMEA data. Product Product_id Product_description Constituent_part Part01 Part02 . . Part_n Figure 1 Part Part_id Part_description Constituent_process Proc01 Proc02 . . Proc_n Process Process_id Process_description FMEA data Failure_data Failure_mode Cause_of_failure Effect Corrective_action Hierarchical case-based structure with three-level case bases Thus, it was decided that hierarchical knowledge could be stored in separate (but related) case bases using the object-attribute-value triplets format [6] and could also be linked together via a linking function (see Figure 1). A linking function is a processing procedure in which a seeking operation is employed to connect a field of a case in one case base onto a case in another case base, then from that case base to another one and so on. This hierarchical structure also facilitates the retrieval and adaptation operations. Other benefits of the hierarchical case-based structure are that redundancy of storage is avoided entirely and the linking function links the different level case bases together dynamically (starting at any level) and flexibly (linked and unlinked easily). 2.2 Multi-Stage Retrieval As stated, data in HOCBR can be organised into a hierarchical case-based structure. Multi-stage retrieval is proposed to search through multi-level case bases 77 Tzu-Fu Chiu An application of hierarchy-oriented case-based reasoning in the pilot production stage to get a set of related similar cases. The conceptual diagram of multi-stage retrieval is depicted in Figure 2. Input problem First-stage retrieval (Product case base, A) tmp-case (Retrieval) case_a2 case_a1 case_a3 ..... case_a4 case_am (Similar case) (Linking) Second-stage retrieval (Part case base, B) case_b1 case_b2 case_b7 case_b6 (Edited) case_b3 case_b8 case_b4 case_b5 (Similar case) case_b9 . . . . . case_bn tmp-case (Retrieval) (Linking) (Linking) Third-stage retrieval (Process case base, C) case_c1 case_c2 (Similar case) case_c6 case_c7 (Retrieval) tmp-case case_c4 case_c3 case_c8 (Edited) case_c9 (Edited) case_c5 ..... (Similar case) case_cp tmp-case (Retrieval) Hybrid adaptation Figure 2 A conceptual diagram of multi-stage retrieval Multi-stage retrieval searches through the hierarchical case bases, from the highest level case base down to the lowest level case base. Nearest Neighbour Retrieval (NNR) [9, 13] is employed to search through the three case bases for the most similar case to the temporary case at each stage. This is shown in Figure 2, using three-stage retrieval to demonstrate the conceptual logic. Initially first stage retrieval accepts the input problem from the user and keeps it as a temporary case at in memory. The case as is found in the top level case base A, such that Ma (s, t) = max {Ma (x, t)} for x A, where Ma is the similarity score, associated with the case base A, given by n n k 1 k 1 Ma (x, t) = ( wk fkxt) / ( wk) Here fkxt is the value of an appropriate similarity function for feature k of cases ax, at such that 0 fkxt 1, k = 1, ..., n. The vector of weights w' = (w1, …, wn) reflects the importance assigned to the features and is called a retrieval view. In Figure 1, s = 3. Using the constituent parts of as and the linking function, first -stage 78 中華管理學報 第二卷 第一期 第75-87頁 民國九十年 retrieval links the first-level case base A (e.g. in Figure 2, cases a3) to second-level case base B (e.g. cases b3, b6, and b7). As there is no editing required for case b3 and case b5 in this example, second-stage retrieval links to the third-stage directly, using the constituent processes of case b3 and case b6 and the linking function. The user may wish to edit a case (to change portions of its fields, or to add a necessary new case) in order to fit the input problem more closely, as for case b7. Second-stage retrieval then stores this edited case as a temporary case in memory and utilizes this temporary case to search out a similar case (e.g. case b9) in the part case base. Using the constituent processes of the similar case and the linking function, second-stage retrieval links the second-level case base B (e.g. case b9) to third-level case base C (e.g. case c3, c4, c5, and c9). In the third-stage retrieval shown in the example, two cases (i.e. case c8 and c9) have been edited by the user. Case c8 and case c9 are then stored as temporary cas es in memory in order to find the most similar cases to them (using NNR). Multi -stage retrieval (including hybrid adaptation) is performed iteratively on the process and FMEA data until the search of all of the constituent elements is completed. The advantages of multi-stage retrieval are: the interactive editing function at each level for modifying portions of the intermediate data and the dynamic searching procedure through hierarchical case bases for the most similar cases. Thus the entire case bases are searched thoroughly and efficiently. 2.3 Hybrid Adaptation Within multi-stage retrieval, when retrieval proceeds down to the third-stage in this system, the retrieved case data will be passed onto the next operation, hybrid adaptation, in order to generate the suggested solutions for the input problem. The different hybrid adaptation methods which are suitable to the application area of pilot production and include formula adaptation, rule-based adaptation and shift-view adaptation, are discussed below. Formula adaptation: this method uses parameter adjustment [14] to substitute the processing time and cutting speed for a similar case let us say S (retrieved at 3rd-stage retrieval). The ‘formula field’ is used off-line to calculate the processing time and cutting speed for each case in the process case base, according to the relevant data kept in the case fields. The calculation formulae for processing time t (minutes) and cutting speed V (mm/min) are listed below [8]: t = L / (fN); V = DN where L is the length of the workpiece (mm), f is the feeding speed (mm/revolution), N is the rotation speed (rpm), and D is the bore diameter (mm). During system execution, at the editing stage of multi-stage retrieval, if the process specifications (i.e. the length of a workpiece and/or the rotation speed of a machine) have been changed and differ from the similar case S, the calculation formula is employed again to re-calculate a new processing time and cutting speed to reflect these parameter changes. This will provide adjusted values for the similar case S. Rule-based adaptation: this method uses domain-specific adaptation [9] to search out the most suitable machine and tool associated with a specific part and 79 Tzu-Fu Chiu An application of hierarchy-oriented case-based reasoning in the pilot production stage process for the similar case S. The machine selection data and tool changing data collected from the factory is used to generate a knowledge base. During the editing procedure, if the main specifications of a process are changed, the reasoner passes this data on onto the rule-based environment. A set of rules for searching out a suitable machine and tool for the “Body” of an oil pump in the plane milling process are shown as an example below. IF Part_name = “Body (OP)” AND Proc_name = “Plane milling” AND Hole_distance < 150 THEN Machine_selected = “MC02; Machine centre” IF Proc_name = “Plane milling” AND Machine = “MC02; Machine centre” THEN Tool_selected = “T0120110; Milling tool” Subsequently, the inference engine is activated to seek out a solution which exactly matches the input condition. The information obtained is sent back to complete the adaptation procedure. If no exact match exists no value is returned. Shift-view adaptation: this method applies case-based substitution [7] to replace FMEA data in the similar case S with a more useful set of FMEA data (or a combination of two sets of FMEA data), which match the edited case let us say E more closely. A retrieval view w may be partitioned into subviews w1, w2, wn. Each subview corresponds to the features in a particular section of the data. For example a process case contains data of two kinds: process data and FMEA data. Thus, as well as the general view wg, where w'g = (w'1, w'2), to emphasise the different data types the process view is defined, wp, where w'p = (w'1, 0') and the FMEA view, wf, where w'f = (0', w'2). Here w1 and w2 are the subviews corresponding to the process and FMEA data respectively and 0 is the vector of zeros. The relationships between different kinds of retrieval views is shown below in Figure 3. Retrieved case Sp [retrieval] Process view, Mp (Sp, E) (Similarity_score_Sp) [comparison] Edited case E General view Similar case S FMEA view, Mf (Sf, E) (Similarity_score_Sf) [retrieval] Figure 3 Process view, Mp (Sp, Sf) (Process_score) FMEA view, Mf (Sp, Sf) (FMEA_score) [comparison] Retrieved case Sf Relationships between different kinds of retrieval views 80 中華管理學報 第二卷 第一期 第75-87頁 民國九十年 Case-based substitution is used to replace FMEA data in the similar case S (retrieved using wg). A case which is found to be similar using wp may consist of a process closely related to that of E and may also contain useful FMEA data for E. Similarly the case retrieved using wf will contain a set of FMEA data closely related to that in E. Therefore two cases Sp, Sf are retrieved using wp and wf respectively. These two cases are then compared using the NNR method (also through the process view and FMEA view). The aim of the comparison is to measure the degree of similarity of fields between the two retrieved cases (i.e. Sp and Sf), based on the different views. The different comparison scores between the two views will facilitate the following selection procedure for finding a useful set of FMEA data for the similar case S. Essentially, these two retrieved cases are similar to the edited case E in some aspects, depending on the view used. If the two retrieved cases are also similar to each other, their two sets of FMEA data may be useful for predicting possible future FMEA situations in the edited case E (i.e. the edited process). The two retrieved cases are retained or discarded as follows. Let Mv (c1, c2) be the similarity (or comparison) score of cases c1, c2 using view wv. Firstly, the threshold value is that Mp (Sp, E) 0.5 and Mf (Sf, E) 0.5. Then we retain case Sp if and only if max {Mp (Sp, Sf), Mf (Sp, Sf)} 0.5 or Mp (Sp, E) 0.75, and similarly we retain case Sf if and only if max {Mp (Sp, Sf), Mf (Sp, Sf)} 0.5 or Mf (Sf, E) 0.75. These decision criteria can also be explained as in Table 1. Table 1 Selection criteria table for the retrieved cases Criterion Condition 1 Condition 2 Condition 3 Condition 4 Conclusion No. Mp (Sp, Sf) Mf (Sp, Sf) Mp (Sp, E) Mf (Sf, E) Recommended retrieved case 1 >= 50% -- -- -- Retrieved case Sp & case Sf 2 < 50% >= 50% -- -- Retrieved case Sp & case Sf < 50% >= 75% >= 75% Retrieved case Sp & case Sf < 75% Retrieved case Sp >= 75% Retrieved case Sf < 75% No case 3 4 5 < 75% 6 3. System Framework for the HOCBR Architecture Based on the HOCBR architecture, a system framework has been developed. The system framework, incorporating the background environment, development tools and main operations, is shown in Figure 4. According to the HOCBR system framework, the final prototype system will be able to perform the following activities. Firstly, the system accepts the manufacturing and FMEA records as initial data to form the case bases. Process data is then entered as an input (or current) problem. The multi-stage retrieval (the first shaded operation in Figure 4) is then activated to do the retrieving operation upon the hierarchical case bases, by searching for the most similar cases to the input case. Subsequently, the hybrid adaptation (the second shaded operation in Figure 4) is triggered to do the 81 Tzu-Fu Chiu An application of hierarchy-oriented case-based reasoning in the pilot production stage transformation process associated with the case bases and knowledge base, by transforming the most similar cases into successive temporary solutions. The temporary solutions are then integrated by the data integration unit into a final suggested solution for the user. The main functions of the system are: single-stage retrieval, multi-stage retrieval, formula adaptation, rule-based adaptation, shift-view adaptation and data integration [3]. Application area: pilot production in the car part industry Problem: retrieval and adaptation of hierarchical knowledge Input problem Data entry Modified data Problem modification Users No Multi-stage retrieval Hybrid adaptation Down to the lowest level? Yes Case base Knowledge base (Knowledge Archive) Successive solutions Programming language: ReMind API & Visual Basic No More alternatives needed? Solution display Yes Suggested solution Data integration Using: manufacturing & FMEA data Based on: ReMind development system Figure 4 HOCBR system framework for the problem domain Key: : an operation : main operations of the system : logical flow : data flow related to the knowledge archive : judgement : knowledge archive (a set of data structures) : various types of data used in the operations; : various types of data structure in the archive. 4. Implementation in Industry A prototype system for this research has been developed from the HOCBR system framework and successfully installed in an automobile component factory in Taiwan. Real data from factory has been entered into the system. The system is currently being used by engineers who work in the interactive manufacturing environment of pilot production within the factory. The system provides the four main functions for browsing, multi-stage retrieval, hybrid adaptation (a screen of the shift-view adaptation as an example shown in Figure 5) and integration. Feedback about the system has been positive. Summarised questionnaire evaluation from the engineers states that they found the system useful, improved efficiency and 82 中華管理學報 第二卷 第一期 第75-87頁 民國九十年 would continue to be of use after the initial testing period [4]. Rule-based Shift-view Shift-view S(p) S(f) Figure 5 A screen of the shift-view adaptation result 5. Evaluation of the HOCBR Architecture In order to perform an evaluation of the HOCBR architecture in detail, this section is divided into three sub-sections. Firstly, the hierarchical case-based structure is compared to the individual structure (normally used in the flat-record case base), in order to highlight its advantages with respect to data storage. Multi-stage retrieval is then compared to other hierarchical systems, in order to show its capability in linking and editing dynamically. Finally, shift-view adaptation is compared to the CLAVIER method [7], in order to demonstrate its functionality in data transformation using different retrieval views. 5.1 Comparison of the Number of Data Units in the Hierarchical and Individual Case-Based Structures The individual case-based structure has been employed in many existing CBR systems (e.g. CHEF, CASEY and HYPO) for data storage [19]. Thus it is used as a basis for comparison with the hierarchical case-based structure. Each cluster of product, part, or process data is called a data unit. The number N1 of data units in the individual case-based structure is calculated as in formula (1); and the number N2 of data units in the hierarchical case-based structure is calculated as in formula (2). N1 = 3 mnp, (1) N2 = (m + mn + mnp) – c, (2) 83 Tzu-Fu Chiu An application of hierarchy-oriented case-based reasoning in the pilot production stage where m is the number of data units in the product level, n is the average number of constituent data units in the part level, p is the average number of constituent data units in the process level and c is the number of common units in the part and process levels. For c = 0, n > 1 and p 1, it can be said that p > (1 + n) / 2n; => 3np > 1 + n + np; => 3mnp > m + mn + mnp; that is N1 > N 2 therefore, the number (N2) of data units in the hierarchical case-based structure is less than the number (N1) of data units in the individual case-based structure. The ratio (i.e. the relative efficiency) for the number of data units between the individual case-based and hierarchical case-based structures is expressed below. R = (3mnp) / (m + mn + mnp); = (3np) / (1 + n + np); = (3) / ((1 / np) + (1 / p) + 1); The ratio will approach 3 as p approaches infinity; meanwhile the ratio will approach (3p) / (p + 1) as n approaches infinity. In the prototype system, there are 22 data units (i.e. 8 oil pumps, 8 water pumps and 6 front axle assemblies) in the product base, 54 data units in the part base and 171 data units in the process base. The total number of data units in the hierarchical case-based structure is: N1 = 22 (product base) + 54 (part base) + 171 (process base) = 247 However, to keep all of this data in the individual case-based structure, the number of data units will increase as follows (according to the BOM of products): ((8*1*5), (8*1*2), (8*1*2), (8*1*2)) = 88 (oil pump) ((8*1*4), (8*1*1)) = 40 (water pump) ((6*1*5), (6*1*4), (6*1*3)) = 72 (front axle) N2 = 3 (88 + 40 + 72) = 600 (data units) Consequently, the number of data units for keeping the manufacturing and FMEA data in the hierarchical case-base structure is much less than in the individual case-base structure. The redundancy of data storage is avoided successfully. Meanwhile, this will also improve the searching efficiency of the retrieval operation (while the case base size gets larger and larger). 5.2 Comparison of Multi-Stage Retrieval with Other Hierarchical CBR Systems Of some other systems organised in a hierarchical way, the more notable ones are the Deja Vu system [15, 16] and the CBRefurb system [11]. In Deja Vu, complex problems are stored as hierarchical collections of cases and individual cases describe part of a more complex solution at some given level of abstraction through the network-based structure. It then uses ‘adaptation-guided retrieval’ which ensures that adaptable cases are always retrieved. In CBRefurb, a whole building task is broken 84 中華管理學報 第二卷 第一期 第75-87頁 民國九十年 down into smaller tasks (building items) by organising the refurbishment cases as a hierarchical structure composed of cases and subcases. It uses ‘multiple case retrieval’ to collate information from several old cases for estimating the cost of refurbishment. The adaptation-guided retrieval of Deja Vu uses adaptation knowledge during retrieval to determine the adaptation requirements of cases and the multiple case retrieval of CBRefurb relies on the indices and the organisation of the memory to direct the search to potentially useful cases. By contrast, multi-stage retrieval used in the HOCBR architecture searches through different level case bases to find the required pieces of a product (i.e. parts and processes) and is accompanied by linking and editing functions. The advantages of multi-stage retrieval are that the constituent parts of a product and constituent processes of a part guide the linking function to connect case bases efficiently, stage-by-stage; the editing function allows the user to add a new intermediate case or to change portions of an existing one in order to retrieve a more closely matching similar case. 5.3 Comparison of Shift-view CLAVIER Method Adaptation with the Shift-view adaptation, using different retrieval views (i.e. weighted vectors) within nearest neighbour matching to find the most similar cases within a case base, produces a more suitable solution for an input problem. The CLAVIER method [7] uses two types of knowledge to divide cases into two compatibility groups with respect to unmatched parts: global case compatibility and local spatial compatibility. Global case compatibility is used to find compatible substitutions for a given part in cases that are globally similar to the current situation. Local spatial compatibility is utilised to select a specific part as a substitute according to its relative position in the case piece. By contrast, three different views (i.e. general view, process view and FMEA view) are used in shift-view adaptation in order to focus on various attribute groups (i.e. all attributes of a case, attributes regarding process features and attributes regarding FMEA features). As a set of failure modes and corrective actions were generated for a specific process, its similarity can be alternatively measured in three ways: using all attributes, attributes regarding process features, or attributes regarding FMEA features. Consequently, based on the various views, all of the potential failure modes and corrective actions in the case base will be screened out successfully. 6. Conclusions and Future Work This research has identified the problem as the organisation and re-use of hierarchical knowledge in the pilot production stage, with particular reference to the car part industry. The HOCBR architecture has been proposed to tailor the general CBR methodology to the selected problem domain. This architecture organises several case bases into a hierarchical case-based library, searches out the most similar case from the multi-level case bases and adapts this case (using different adaptation methods) into a suitable suggested solution for a current problem. Furthermore, based on the proposed HOCBR architecture, the system framework has been developed and evaluated in an industrial environment. The full potential of this approach has not yet been realised in practice. For example, only three levels of retrieval 85 Tzu-Fu Chiu An application of hierarchy-oriented case-based reasoning in the pilot production stage have been demonstrated and the number of cases was relatively small. The current prototype requires that the case bases still need to be created off-line and some processes within the CBR cycle are not yet fully implemented. Potential research areas suggested by these limitations are listed below, along with other areas for future investigation: to strengthen the functions of the HOCBR architecture for the problem domain, especially the revision and retainment operations; to develop further application systems based on the HOCBR architecture in other industries which have similar features to those of the problem domain; and to collect more real data from the collaborating factory to increase the processing capacity of the application system. The implementation and evaluation of the HOCBR architecture have shown that the system functions successfully. The author believes the proposed methods and system architecture to be more efficient, in certain respects, than existing approaches and to offer considerable practical advantages. References 1. Aamodt, A. and E. Plaza, “Case-Based Reasoning: Foundational Issues, Methodological Variations and System Approaches,” AICOM, 7(1), 39-59(1994). 2. Bradley, J. and U.G. Gupta, “A Classification Framework for Case-Based Reasoning Systems,” The New Review of Applied Expert Systems, 1, 35-48(1995). 3. Chiu, T.F., Hierarchy-Oriented Case-Based Reasoning with an Application to Pilot Production in the Car Component Industry, Ph.D. Dissertation, School of Computing and Information Systems, University of Sunderland, UK(1998a). 4. Chiu, T.F., “An Evaluation of the HOCBR System in the Car Component Industry,” Tamsui Oxford Journal of Management Sciences, Tamsui Oxford University College, 13-14, 15-43(1998b). 5. Evans, J.R., Anderson, D.R., Sweeney, D.J. and T.A. Williams, Applied Production and Operations Management, West Publishing Company(1990). 6. Giarratano, J. and G. Riley, Expert Systems: Principles and Programming, PWS Publishing Company(1994). 7. Hennessy, D. and D. Hinkle, “Applying Case-Based Reasoning to Autoclave Loading,” IEEE Expert, 7(5), 21-26(1992). 8. Kalpakjian S., Manufacturing Engineering and Technology, Addison-Wesley Publishing Co.(1992). 9. Kolodner, J., Case-Based Reasoning, Morgan Kaufmann Publishers, Inc.(1993). 10. Kumar, H.S. and C.S. Krishnamoorthy, “A Framework for Case-Based Reasoning in Engineering Design,” Artificial Intelligence for Engineering Design, Analysis and Manufacturing, 9, 161-182(1995). 11. Marir, F. and I. Watson, “Representing and Indexing Building Refurbishment Cases for Multiple Retrieval of Adaptable Pieces of Cases,” Case-Based Reasoning Research and Development, M. Veloso and A. Aamodt (eds.), 55-66(1995). 12. Martinich, J.S., Production and Operations Management: An Applied Modern 86 中華管理學報 第二卷 第一期 第75-87頁 民國九十年 Approach, John Wiley & Sons, Inc.(1997). 13. Richter, M.M., “Introduction,” Case-Based Reasoning Technology: From Foundation to Application, M. Lenz, B. Bartsch-Sporl, H. Burkhard & S. Wess (eds.), 1-15(1998). 14. Riesbeck, C.K. and R.C. Schank, Inside Case-Based Reasoning, Lawrence Erlbaum Associates, Inc.(1989). 15. Smyth, B. and P. Cunningham, “A Hierarchical Case-Based Reasoning System for Software Design,” Proceedings of the 10th European Conference on Artificial Intelligence, Vienna, Austria, 587-589(1992). 16. Smyth, B. and M. Keane, “Experiments on Adaptation-Guided Retrieval in Case-Based Design,” Case-Based Reasoning Research and Development, First International Conference, ICCBR-95, M. Velosoand and A. Aamodt (eds.), 313-324(1995). 17. Watson, I. and F. Marir, “Case-Based Reasoning: A Review,” The Knowledge Engineering Review, 9(4), 327-354(1994). 18. Wirth, R., Berthold, B., Kramer, A. and G. Peter, “Knowledge-Based Support of System Analysis for the Analysis of Failure Modes and Effects,” Engineering Applications of Artificial Intelligence, 9(3), 219-229(1996). 19. Zito-Wolf, R.J., Case-Based Representations for Procedural Knowledge, Ph.D. Dissertation, Brandeis University, USA(1993). 87