سیستم پشتیبانی تصمیم گیری هوشمند برای سرمایه گذاری های فن آوری تولید
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|10724||2006||12 صفحه PDF||سفارش دهید||محاسبه نشده|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : International Journal of Production Economics, Volume 104, Issue 1, November 2006, Pages 179–190
Making strategic decision on new manufacturing technology investments is difficult. New technologies are usually costly, affected by numerous factors, and the potential benefits are often hard to justify prior to implementation. Traditionally, decisions are made based upon intuition and past experience, sometimes with the support of multicriteria decision support tools. However, these approaches do not retain and reuse knowledge, thus managers are not able to make effective use of their knowledge and experience of previously completed projects to help with the prioritisation of future projects. In this paper, a hybrid intelligent system integrating case-based reasoning (CBR) and the fuzzy ARTMAP (FAM) neural network model is proposed to support managers in making timely and optimal manufacturing technology investment decisions. The system comprises a case library that holds the details of past technology investment projects. Each project proposal is characterised by a set of features determined by human experts. The FAM network is then employed to match the features of a new proposal with those from historical cases. Similar cases are retrieved and adapted, and information on these cases can be utilised as an input to prioritisation of new projects. A case study is conducted to illustrate the applicability and effectiveness of the approach, with the results presented and analysed. Implications of the proposed approach are discussed, and suggestions for further work are outlined.
Manufacturing companies are constantly striving to improve their competitive capability, and to lower production costs by investing in new or proven manufacturing technologies. In general, a typical manufacturing technology (MT) delivery process can be seen as a three-stage process, namely: MT Concept Identification; Proof of Concept; and MT Roll-Out (see Fig. 1). How can managers decide whether to invest in a new MT project when benefits are hard to justify prior to its initial proof of concept stage, and when financial consideration still plays a dominant role?There are many evaluation approaches available for MT investment. Most of them assess an MT project based on its alignment to the identified corporate objectives, and fulfilment of a set of investment assessment criteria. In general, the evaluation of the criteria is a very subjective and unstructured process. It relies heavily on managers’ experience, knowledge, as well as intuition. However, managers can hardly consider all the relevant criteria due to bounded rationality and a limited capacity for information processing (Deng, 1994). Thus, the evaluation approach is often not effectively carried out, as managers do not make effective use of their knowledge and experience of previously delivered technologies and projects as an input to the prioritisation of future projects. The impact of this is that managers are not confident that resources are being optimised and applied to a mixed portfolio of projects to maximise benefits. Thus, how could managers retain and reuse their knowledge of previously (successful and unsuccessful) delivered technologies and projects to support the future decision making in MT investments? Knowledge-based intelligent systems, including case-based reasoning (CBR), have played an increasingly important role in today's industrial applications (Lee et al., 2004; Mezgar et al., 2000; Yen et al., 1995). The reasons for this growth are manifold: intelligent systems based on the artificial intelligence (AI) methodology are becoming increasingly popular and mature in solving real-life problems; knowledge-based systems have favourable characteristics compared to conventional approaches or pure, symbolic AI systems; and the tools for developing AI systems are becoming easier to use nowadays (Mezgar et al., 2000). A review of literature shows that intelligent systems such as CBR and neural networks could give many benefits to domain users, such as providing consistency in decision making, improving learning capability, supporting knowledge reuse and retention, and yielding instantaneous decision support. Two main objectives are focused in this research. First, an intelligent system based on a hybrid CBR and fuzzy ARTMAP (FAM) (Carpenter et al., 1992)—a supervised adaptive resonance theory (ART) (Carpenter and Grossberg, 1987) neural network model—is proposed for undertaking the incremental learning problem in CBR. Second, the proposed intelligent system is applied to decision support in MT project evaluations—a complex decision-making task in business investment which is often affected by changing environments. A case study in the pharmaceutical industry, where injection of new technologies into manufacturing processes is frequent, and critical to success, is conducted to demonstrate the effectiveness of the proposed intelligent system in supporting the MT project evaluation process. The idea of combining CBR and neural networks is not new. From the literature review on similar work with this research, i.e., integration of CBR and clustering-based and/or prototype-based neural network models (Malek and Amy, 1997; Reategui et al., 1997; Gutpa and Montazemi, 1997; Kim et al., 2002; Park et al., 2004; Yang et al., 2004), one can see that the main objective of this line of research is to instil a more efficient learning procedure into CBR. Malek and Amy (1997) proposed an integration of CBR and a prototype-based neural network to make the indexing system, hence the retrieval process, of CBR more efficient and, at the same time, to maintain a continuous learning process. Good performance was reported as compared to two other inductive indexing approaches. Use of a neural network to overcome the difficulty of indexing and retrieval in CBR was also proposed by Reategui et al. (1997). The effectiveness of their hybrid CBR and neural network system was demonstrated with a medical diagnosis problem. Gutpa and Montazemi (1997) used a connectionist approach for similarity assessment in a CBR system, and their work was applied to fault diagnosis in electric motors. Similarly, the self-organising map was integrated with CBR and applied to fault diagnosis (Kim et al., 2002). Park et al. (2004) developed a neural network model to guide a CBR system to learn the case-specific local weighting patterns for reasoning, while Yang et al. (2004) used an ART–Kohonen network to make hypotheses and to search for a similar previous case that supports one of the hypotheses in a CBR system. When compared with other approaches, the work by Yang et al. (2004) is the closest to ours. They designed an unsupervised ART network coupled with the learning strategy of the Kohonen network in which the Euclidean distance was used to search for a winner. In our case, however, the supervised FAM model is integrated with a CBR system, but the original learning algorithm of FAM is retained. The main reason is to inject the CBR system with the ability to overcome the stability–plasticity dilemma, i.e., how to incorporate new knowledge without forgetting/corrupting previously learned knowledge. The ART family of neural networks, including FAM, has been specifically designed to tackle this dilemma, which poses the fundamental question in designing intelligent learning systems (Carpenter and Grossberg, 1987 and Carpenter and Grossberg, 1988), i.e. “how a learning system can remain adaptive in response to significant events and, at the same time, can remain stable against irrelevant events”. The FAM network possesses an added advantage whereby it combines the ART property (useful for overcoming the stability–plasticity dilemma) and fuzzy set theory (useful for handling imprecise, uncertain information in linguistic variables comprehensible by humans) into a common platform. By combining CBR and FAM, both incremental learning and uncertainty adaptation capabilities can be realised in an integrated framework. These features allow managers to change the behaviours of the system in response to significant events in business environments, while maintaining the capability of the system to gain and adapt to new knowledge online. Our proposed system is therefore able to capture the important features required for a complex decision-making environment for MT investment analysis. It is also useful for capturing a company's strategic information, providing facilities to quantify qualitative features, and analysing them alongside the quantitative features in an evaluation framework. This paper is structured into three parts. Firstly, the fundamentals of CBR and FAM are introduced. Then, the integration of CBR and FAM is explained. The detailed development of the system for MT project evaluation in a pharmaceutical firm is described and the results obtained are explained. Finally, the feasibility and practicality of this approach is analysed and discussed, and plans for further work are outlined.
نتیجه گیری انگلیسی
In this paper, the development of an intelligent decision support system combining CBR and FAM for evaluation of MT projects is described. In the proposed system, CBR is utilised to support managers in retaining and reusing their previous knowledge on both successfully and unsuccessfully delivered projects to evaluate future projects. To address the stability–plasticity dilemma, FAM is incorporated into the developed system to support incremental learning in dynamic environment. Given a new project proposal, one (or more) similar case(s) will be retrieved from the case library. Information of the retrieved project(s) will be reused as a reference in evaluating the new project. Through the system, similar cases can be retrieved to enable managers to make effective use of their knowledge and experience. A series of experiments were conducted to assess the performance of proposed system. Two operating strategies of FAM, namely averaging and voting, were examined. The vigilance parameter was altered systematically to optimise the network performance. Good accuracy rates were achieved, with most of the results ranging from 80% to 90%. The best accuracy rate was 92.00% from voting FAM with View the MathML source. In addition, bootstrapping was employed to estimate the 95% confidence intervals of the average results. The small confidence bound associated with the best accuracy rate implied that the network performance was stable statistically. Evaluation by the domain experts on the usefulness of the proposed system was also conducted. Positive comments were obtained, and commitments to implement the proposed system in the real environment were given. Although encouraging results have been achieved, there are a number of aspects that need further investigations. First, more data samples have to be collected to further ascertain the effectiveness of the proposed system in MT project evaluation. It would be interesting if statistical estimates including risk factors could be incorporated into the network predictions. On the other hand, the outputs from FAM could be augmented with rules to explain the rationale used to reach a prediction. In this aspect, the FAM rule extraction technique exemplified in Carpenter and Tan (1995) can be examined. Further work will also focus on testing of the developed system in the real environment in collaboration with Company X. All these undertakings will lay a more comprehensive platform for managers to use the hybrid CBR—FAM-based system as a useful decision support tool in MT investment analysis.