Items are categorised to select the most appropriate forecasting and Inventory Management procedure for the relevant cluster (Syntetos et al 2005).
Regarding control of independent demand inventories it is often not possible to pay equal attention to all items on stock. Focus on the most important parts is therefore necessary.
A typical classification approach used here is the ABC Analysis (Chakravarty 1981, Partovi and Anandarajan 2002, Chen et al 2008). Parts are clustered into
three categories - A, B, and C - in terms of their annual inventory value (Flores and Whybark 1986). According to the Pareto principle, 20 percent of the items on stock account for 80 percent of the overall annual inventory value. The application of this principle is widely observed in practical inventory systems (Chen et al 2009). By managing these items intensively it is possible to control most of the inventory value (Schroeder 1993, Heizer and Render 2010). As it is easy to implement and use, ABC Analysis is the very popular and widely used in industry (Guvenir and Erel 1998, Ramanathan 2006). Sometimes additional categories are applied (e.g. Nagarur et al 1994).
Flores and Whybark (1986) criticise the classical ABC approach as it only employs the cost value of the items and defines that as the most important criterion. They demonstrate that different units of an organisation regard different classification criteria as most important. Engineering may view
obsolescence of parts to be significant and would therefore concentrate on the items with the highest obsolescence risk. Purchasing by contrast might select procurement lead times for classification. Maintenance, in turn, may favour substitutability, reparability, or criticality as criteria for item categorisation. The authors introduce a concept of a joint criteria matrix which allows items to be classified by two criteria such as criticality and cost value of parts. This technique is not suitable when more than two criteria should be applied (Guvenir and Erel 1998, Partovi and Anandarajan 2002, Ramanathan 2006, Chen et al 2009).
Applying more than two criteria in a multi-criteria ABC classification requires the use of complex computational tools (Ramanathan 2006, Chen et al 2008).
In practice the selection of criteria for multiple ABC classification is often an ad hoc process based on the decision maker’s experience, knowledge and judgment. This can be effective but can also lead to an oversight of relevant criteria (Partovi and Anandarajan 2002). Several approaches to support the decision process for selecting the relevant classification criteria are available in literature.
Cohen and Ernst (1988) and Ernst and Cohen (1990) apply cluster analysis for parts classification. In contrast to the approach developed by Flores and Whybark (1986), here the full range of relevant criteria can be used which
results in a large variety of criterion combinations. The clusters have to be re
evaluated when new items are added to the existing inventory. For Partovi and Anandarajan (2002) this approach has limitations for practical application.
Cluster analysis requires solid data and the need for re-evaluating the categories can lead to different classification of items which can result in a change of control procedure. The model is also known as multiple discrete analysis (Partovi and Anandarajan 2002).
Many authors adopt the Analytical Hierarchy Process for ABC classification (Partovi and Anandarajan 2002, Ramanathan 2006, Chen et al 2008). Gajpal et al (1994) utilise the Analytical Hierarchy Process for clustering spare parts based on criticality. With the Analytical Hierarchy Process both qualitative and quantitative criteria are taken into account for item categorisation. The
significant amount of subjectivity involved in the comparison and weighting of criteria is recognised to be a serious limitation of this method (Partovi and Anandarajan 2002).
To overcome the drawbacks of the Analytical Hierarchy Process, artificial intelligence methods are used (Partovi and Anandarajan 2002, Ramanathan 2006, Chen et al 2009). Artificial Neural Networks can simulate decision maker’s understanding of relationships for both quantitative and qualitative attributes. Inventory classification problems are essentially non-linear. Artificial Neural Networks are capable of approximating these non-linear functions.
Artificial Neural Networks are capable of learning. Back propagation algorithm and genetic algorithm are two commonly used learning methods (Partovi and Anandarajan 2002, Klein 2008).
Guvenir and Erel (1998) apply Artificial Neural Networks with a genetic algorithm as learning method for multi-criteria inventory classification. Genetic algorithms are inspired by natural population genetics. They follow the natural selection process as observed in biological evolution (Partovi and Anandarajan 2002). Within a population of knowledge structures, the fittest structure in terms of the optimum setting of parameters is searched by means of a genetic
algorithm. In their study Guvenir and Erel (1998) compare the classification of a human decision maker with the results obtained by both computed the Analytic
Hierarchy Process and genetic algorithms. Their findings show that the
classification made by the genetic algorithm computation is more similar to the one made by the human decision maker than the one achieved by using the Analytic Hierarchy Process.
Partovi and Anandarajan (2002) utilise both back propagation algorithm and genetic algorithm in their Artificial Neural Networks. With back propagation the input-output relations presented by data sets are loaded in the Artificial Neural Networks model. The model thus learns about the past to make predictions of future events. In their study the authors compare the results obtained by Artificial Neural Networks with the ones derived from the multi- discriminate analysis developed by Ernst and Cohen (1990). Their findings show that Artificial Neural Networks outperform the technique of Ernst and Cohen (1990).
It also becomes evident that the genetic algorithm is the better learning method compared to the back propagation algorithm.
Ramanathan (2006) criticises the Analytic Hierarchy Process in terms of the subjectivity involved in the process, and Artificial Neural Networks in terms of their heuristic nature which do not lead to optimal solutions in all situations. The author proposes a simple weighted linear optimisation model to support the multiple-criteria decision making. In this model a weighted additive function is applied to accumulate the performance of an item in terms of different criteria to the optimal inventory score. By solving the maximisation objective function, the model computes the optimal inventory score for each item. In the study
Ramanathan (2006) compares his model to the traditional ABC classification and the Analytic Hierarchy Process approach. Classification results are different for each model. The author concludes that the proposed model is easy to understand and to use by decision makers. He fails to prove the limitations of the Analytic Hierarchy Process and Artificial Neural Networks approaches he has claimed. Therefore his critique is not regarded to be substantial.
Chen et al (2009) propose a model that allows more than two criteria to be handled simultaneously for inventory classification and thus overcomes the limitations of the model developed by Flores and Whybark (1986). The authors develop a model that uses weighted Euclidean distances to represent decision maker’s preferences over alternatives. Application of the model derives
classification results similar to those obtained by Flores and Whybark (1986).
The categorisation of items according to the Pareto principle is also evident.
Chen et al (2009) thus conclude that their model is robust and can produce solid classification results.