Research On Model Management In Decision Support Systems: Some Issues

Model management is a most important research area in decision support systems (DSSs), for the reason that model management system (MMS) is one of the core components in DSSs, and is also the key factor which affects the development of the DSSs.Researches on model management can be roughly classified into two categories: modeling-in-the-small and modeling-in-the-large. Modeling-in-the-small addresses model conceptualization and formulation for a particular model, and the key issue is the problem of model representation. In contrast, modeling-in-the-large considers the issues of administering a model base, model linkage representation (viz., the logic structure representation of a model base), model composition, model selection, model solution, and MMS design, etc. Therefore, the problem of logic structure representation of a model base, model composition, and MMS design are key issues for the practical application of MMS.This study carried out an in-depth study on several issues related to the model management in DSS. These issues include the representation and modeling of the logic structure of a model base, model composition, general-purpose MMS architecture design. In addition, this study also discussed the approaches of decision attribute selection for decision tree construction. The following four aspects are carefully discussed in this paper:1. Logic Structure Representation for Model BasesThe pivotal problem in modeling-in-the-large is how to represent the logic structure of a model base. An effective model base structure representation is the foundation of model management. It is helpful for the implementation of most operations, such as model composition, query, maintenance, and execution, as well as the MMS architecture design. Several approaches have been proposed. Although each of them has certain advantages and limitations, the graph based approaches are considered the most intuitionistic and effective.The concept of model net proposed in this article is a kind of directed graph, which can be used to represent and analyze the logic structure of a model base. The main efforts in this part are: (1) a concept of generalized data type match is proposed to realize the generalized linkages among models in a model base; (2) the pivotal factors which must be considered in model base construction are analyzed, together with the definitions of their structures; (3) the formal definition of the concept of model net is given; (4) an algorithm of model net construction is worked out.Model net has several prominent characteristics as follows. First of all, the models in a model base are linked together using date type conversions. This type of linkage is more flexible compared with other approaches in which the linkages of the models are assumed to be well formed as long as the variables that are connected have the same name. In the next place, the introduction of the concepts of source data and decision objectives in model net keep the model bases independent from its environment. That is, the structure and implementation of a model base will not be affected by the changes of its environment. Again, model net can provide uniform logic views of different model bases by covering up the implementation detail of them. In addition, model net has an opening structure. That is, the structure of the nodes in a model net can be easily extended to hold additional information. Even, the typeMatch function can also be freely redefined according the specific decision applications without changing the meaning of model net. Finally, a composite model from the model net can be treated as a general basic model, so as that hierarchical modeling is natural in model net.2. Model CompositionInstead of using individual models in practical applications, model composition mechanism is most frequently applied in decision problem solution. Therefore, one of the most important functions of MMS is to provide the decision makers a method to facilitate the construction of composite models.One of the vital demands in model composition is the efficiency of the composition process. It comes to light that finding out the composite model(s) for a given decision problem by searching the entire model base is computational complex. However, not all the models in a model based will participate in the process of model composition for a given decision problem. So, the model base should first be simplified before model composition starts.Composition automation is another important issue in model composition, for the reason that the users always lack the knowledge about the models they have. However, full automation can not be realized because this process requires the knowledge and judgments which can not be solved by computers today. Hence, how to improve the automation level turns into an important issue in model composition researches.To settle the problems mentioned above, the means of model composition based on model net is carefully studied in this article. This part is discussed accordingly in the following: (1) algorithms of model net simplification according to the different categories of decision problems are worked out; (2) a data structure named duplicate path tree is introduced to facilitate finding out the set of reduplicate paths and composite model core for a given decision problem. And the corresponding algorithm is also discussed; (3) an algorithm of model composition using reduplicated path tree and composite model core is introduced.Several advantages of the proposed model composition approach can be concluded as follows. Except for the prominent advantage of effective reduction of search space and computational complex, model composition based on model net also has some other useful characteristics. Firstly, it is capable of finding out all the potential composite models for a given decision problem. Secondly, the automation level of model composition process can be improved by the flexible linkage representation, which is realized using data type conversion in model net, and the introduction of the concepts of source data and decision objectives. In addition, model nodes in a composite model can be executed parallelly to improve the problem solution efficiency. finally, some useful information can be achieved using this algorithm without model solution. 3. MMS Architecture DesignA general-purpose architecture is the hinge for the practical application of MMS, as well as the key factor affecting the development of the DSSs. The lack of efficient representation methods for model base logic structure, which is the foundation of model management, is one of the most important factors which slow down the development steps of MMS.Model net is capable of representing and analyzing the logic structure of model bases efficiently. A new general-purpose MMS architecture based on model net is designed in this article. The discussion in this part can be concluded in four steps: (1) the advantages of model net are first summarized to show how this model is suitable for MMS architecture modeling; (2) a layered MMS architecture is proposed, and the functions of each layer are discussed; (3) the functions of each subsystem and procedure contained in the proposed architecture is studied; (4) decision making process based on this architecture is described roughly.The main characteristics of this architecture are summarized as follows. Firstly, this architecture is designed in layered means. Since the specification of a layer says nothing about its implementation, the implementation details of one layer are hidden from others. As a result, with the well-defined interfaces of each layer, changes can be made in one layer without affecting the others. In addition, the model net provides a uniform representation for the model bases, so that standard operations can be defined and implemented without having to consider the details of a specific model base. Therefore, a general-purpose model management system can be realized based on this architecture. Furthermore, it will come true to define a SQL-like non-procedural language, with which the users need only to specify what information they want without having to tell the system how to find them.4. Representation of Data Type Match Knowledge and Decision Tree Construction Approach Based on Disperse DegreeDatatype match reasoning is the key for the application of model net. The most important problems in it are the efficiency and automation level. Decision tree approach, which has several prominent advantages, is suitable for the representation of data type match knowledge. In the construction process, the precision of a decision tree can be greatly influenced by the mechanism of partition attributes selection.For the improvement of the efficiency and automation level of data type match reasoning, the approach of representing data type match knoeledge using decision tree is dicussed. Then a concept of disperse degree for partition attributes selection in constructing decision trees is proposed, and the correlative algorithm, denoted by DSD, is also discussed in this article. The following four aspects are discussed in this part: (1) the approach of representing data type match knoeledge using decision tree is dicussed; (2) the limitations of entropy-based and rough set based approaches are analyzed; (3) a concept of disperse degree of condition attribute in information system is proposed; (4) an algorithm for decision tree construction based on disperse degree is worked out; (5) experiments are performed to show the efficiency of this algorithm. In conclusion, DSD can overcome the limitations of rough set based approaches. And the results of the experiments on the UCI datasets show that the precision of the decision trees constructed using DSD are approximate to that of the ones constructed using entropy-based approach. However, the time complexity of DSD approach is lower than that of entropy-based approach.