Over the last two decades, enterprise architecture has grown into an established approach for holistic management of information systems in organizations [1,2].
A number of enterprise architecture initiatives have been proposed, such as The Open Group Architecture Framework (TOGAF) [3], the Zachman Framework [4], and military architectural frameworks such as DoDAF [5] and NAF [6]. The core concept of the enterprise architecture approach is the employment of models.
Diagrammatic descriptions of IT systems and their environment are heavily used. However, enterprise architecture models are not limited to descriptive use only, but can also be employed to predict the behavior and eects of decisions. Rather than modifying enterprise information systems using trial and error, models allow predictions about the behavior of future architectures.
One prominent challenge to rational decision making is uncertainty. Therefore, a good enterprise architecture model should be able to capture uncertainties about assessment theory, system conguration or data quality, thus providing better decision support and risk management.
What constitutes a good enterprise architecture model is dependent on its purpose, i.e. the type of analysis it is intended to support [7]. For instance in the case of analyzing data quality, the property of whether the data objects are accurate with respect to the real world they describe is of interest. This property however, is irrelevant for a number of other analyses, such as performance evaluation.
Several enterprise architecture software tools are available on the market, including Metis [8], System Architect [9] and Aris [10]. These tools generally focus on the modeling of an architecture whereas the analysis functionality is generally limited to performing an inventory or to sum costs over the modeled architecture. None of the mentioned tools has signicant capabilities for system quality analysis based on an elaborated theory. Furthermore, these tools do not support the consideration of uncertainty as described above.
In this paper an enterprise architecture software tool is presented. This tool does not only provide functionality to model enterprise architectures, but also supports the analysis of them. In order to support enterprise architecture analysis as it has been outlined in [7] the tool consists of two main components. In the rst component the theory relevant to analyze a certain system quality, such as data quality or modiability, is modeled. One can consider this as the denition of a language tailored to describe a certain aspect, e.g. data quality. The second component supports the application of the theory to evaluate a specic enterprise architecture. This is done by modeling the as-is or to-be architecture of the enterprise. Based on the created models it is possible to determine how the architecture fullls the requirements as they have been dened in the theory. The two-component architecture encourages the reuse of the developed theory as it is possible to use the same language to describe several architecture instances. The presented tool makes use of the Probabilistic Relation Models (PRM) formalism as it has been presented in [11] and can thereby manage the uncertainty aspects discussed above.
Enterprise architecture models have several purposes. Kurpjuweit and Winter [12] identify three distinct modeling purposes with regard to information systems, viz. (i) documentation and communication, (ii) analysis and explanation and (iii) design. The present article focuses on the analysis and explanation since this is necessary to make rational decisions about information systems [7].
An analysis-centric process of enterprise architecture is illustrated in Fig. 1. In the rst step, assessment scoping, the problem is described in terms of one or a set of potential future scenarios of the enterprise and in terms of the assessment criteria with its theory (the PRM in the gure) to be used for scenario evaluation. In the second step, the scenarios are detailed by a process of evidence collection, resulting in a model (instantiated PRM, in the gure) for each scenario. In the nal step, analysis, quantitative values of the models' quality attributes are calculated, and the results are then visualized in the form of e.g. enterprise architecture diagrams.
More concretely, assume that a decision maker in an electric utility is contemplating changes related to the maintainance of the power grid. The introduction of an maintenance management system would improve the quality of the maintenance process and allow more cost ecient grid maintenance. The question for the decision maker is whether this change is feasible or not.
As mentioned, in the assessment scoping step, the decision maker identies the available decision alternatives, i.e. the enterprise information system scenarios. In this step, the decision maker also needs to determine how the scenario should be evaluated, i.e. the goal of the assessment. One such goal could be to assess the data quality of an information system. Often several quality attributes are desirable goals.
During the next step, to identify the best alternative, the scenarios need to be detailed to facilitate analysis of them. Information about the involved systems and their organizational context is required for a good understanding of their data quality. For instance, it is reasonable to believe that a more precise data object attribute would increase the probability that the data quality of an information system is high. The impact of a certain data object attribute is thus one factor that can aect the data quality and should therefore be recorded in the scenario model. The decision maker needs to understand what information to gather, and also ensure that this information is indeed collected and modeled.
When the decision alternatives are detailed, they can be analyzed with respect to the desirable system quality. The pros and cons of the scenarios are then traded against each other in order to determine which alternative ought to be preferred.
Calculation Function
Visualization Function As it is succinctly put in [11], PRMs are to Bayesian networks as relational logic is to propositional logic.
A PRM model may be instantiated as a relational skeleton, σ r , containing just objects, object relationships, and attributes. Furthermore, a qualitative dependency structure S denes the details of the attribute relationships, i.e. the sets of probabilistic parents inuencing each attribute. Finally, the PRM is completed by the set of parameters θ S specifying the full conditional probabilistic dependencies between attributes in the form of numbers in Conditional Probability Matrices (CPM). The following expression thus denes the conditional probability of an instance I, given σ r , S, and θ S :
Compared to the standard chain rule for Bayesian networks, this equation is dierent in three ways: (i) the random variables are the attributes of a set of objects, (ii) the parents of a random variable depend on the model context of the object, and (iii) the parameters are shared between the attributes of objects in the same class. In other words, the variables in the dependency structure are the properties of the objects in the instantiated information model, and their causal relations are expressed by the CPM [11].
A PRM thus constitutes a formal machinery for calculating the probabilities of various architecture instantiations. This allows us to infer the probability that a certain attribute assumes a specic value, given some (possibly incomplete) evidence of the rest of the architecture instantiation. In addition to expressing and inferring uncertainty about attribute values as specied above, PRMs also provide support for specifying uncertainty about the structure of the instantiations.
PRMs further allow specializing classes through inheritance relationships.
Classes can be related to each other using the subclass relation ≺, and each class X is associated with a nite set of subclasses C
both Z and Y are subclasses of X. If Z ≺ Y then Z is a subclass of Y , and vice versa Y is a superclass of Z. A subclass Z always contains all dependencies and attributes of its superclass Y . PRMs further allow the dependencies and conditional probability distributions of inherited attributes to be specialized in subclasses.
The presented tool is implemented in Java based on a Model-View-Controller architecture. The data model for PRMs and instantiated PRMs, is specied in XSD and stored in XML les. The user interface is build upon the NetBeans Visual Library [14] with usage of the JApplication framework. The tool is separated into two units, one supporting the modeling of the PRM, whereas the other one makes the tool user able to instantiate and analyze this dened structure. These parts have to be used sequentially, reecting the method that has been described in section 2, starting with the modeling of classes and their attributes as well as the relationships and dependencies between them. Thereby the focus of the This section will illustrate the application of the tool through a case study performed at an electric grid operator in Sweden [16]. The case study focused on assessing data quality in a maintenance management system. In this paper a reduced version of the study is presented, and in the correspondingly reduced theory data quality is dened as the quality of the content and the quality of the values for the data objects used by the system. Quality of content is dened in terms of relevance, the degree to which the information objects have a purpose for the users, and precision, i.e. that the information objects are detailed enough for the application using them. Turning to quality of values, this is dened in terms of accuracy and completeness of the data objects, or more precisely the attributes contained within a data object. Accuracy is measured as the degree to which the values found in a system correspond to the actual values they represent, whereas completeness is measured as the amount of values stored in the system compared to the domain they represent.
Based on the theory outlined above the PRM for data quality analysis can be constructed. The PRM has four classes, rstly the Information System for which the assessment is performed. This system uses a set of Data Objects that constitute the abstract information model employed by the system. Each object contains one or more Data Object Attributes and in the operational system these attributes are instantiated to Data Object Values. Turning to the attributes of the PRM, each of the concepts dened in the theory, e.g. relevance and completeness, corresponds to one attribute. This can be seen in Fig. 2, for instance the attribute Data quality is associated to the class Information System and depends on the attributes quality of values and quality of content of the class Data Object. The conditional probabilistic dependencies between attributes were also dened, for details, see [16].
The data in this case study was collected primarily through interviews but for the acurracy of the data objects direct observations were made and compared with values in the system, i.e comparing the data object values with the actual items they represent to assess the accuracy of an attribute. The instantiated PRM is shown in Fig. 3 where the maintenance system, two data objects, three attributes and a set of values of these attributes are modeled. Based on (i) the model and (ii) the values of the descriptive attributes in the PRM (as found during the data collection), the assessment can be performed. The data quality of the maintenance system, measured in percent of complete fullment of data quality requirements, was found to be 62 percent. 5 Discussion and Future works This paper presents a tool which supports enterprise architecture analysis with the use of the PRM formalism. While providing a powerful mechanism for the use of discrete variables in an analysis, the PRM formalism in its initial form has a few weaknesses that deserve further studies.
Several system qualities are typically analyzed through the usage of continuous variables e.g. in [17] continuous variables are used for performance analysis.
In order to perform those evaluations with support of the presented tool, it is necessary to discretize all continuous variables. At the moment we are investigating how the PRM formalism can be extended so that it can be used with combinations of continuous and discrete variables, so called hybrid networks [18],
as well as a corresponding tool implementation.
Another weakness of the PRM formalism is that it does not provide any means to query the models for structural information such as given an information system, how many elements does the set of related data objects contain?
The Object Constraint Language (OCL) [19] is a formal language developed to describe constraints on UML models. OCL provides a means to specify such constraints and perform queries on the models in a formal language. OCL in its original form is side eect free, but currently an imperative version of OCL is being added to the tool. Thereby the analysis functionality can be extended to consider the structure of the PRM instantiation more comprehensively.
Besides the two mentioned shortcomings of the used formalism there are some improvements with respect to usability. Regarding the user interface of the tool, we are planning to make the models more intuitive and the information provided easier to understand. Enterprise architecture models are more graspable if they only depict the interesting parts of the model (in a goal-sense). Therefore, the tool should be extended to support views and viewpoints, e.g. as presented in [17]. Additionally we plan the support of iconic visualization of typical enterprise architecture elements, such as applications or data objects, to present the models in an easily understandable way.
Finally we are planning to integrate the support of predened model components. As models based on the same metamodel are likely to have common parts, the modeling process can be sped up if common building blocks are oered by the metamodel provider and used by the person that creates a certain model.
In this paper a tool and method for analysis of enterprise architecture scenarios was presented. To fulll this purpose the tool consist of two separate parts, one for dening analysis theory and one for enterprise architecture modeling, and makes use of the PRM formalism for specifying theory. Applying this formalism allows for the consideration of uncertainty, an aspect that so far is uncommon in the eld of enterprise architecture analysis. The paper describes the PRM formalism as well as the underlying architecture of the tool briey.
In the paper an example of data quality assessment was outlined, but the tool supports the analysis of various quality attributes, such as maintainability, information security, and interoperability. The tool supports information system decision making as it allows the comparison of several scenarios with regard to a system quality. Thereby the as-is as well as several to-be architecture of an enterprise can be compared quantitatively in order to nd the one that best satises decision maker requirements.
Availability Reliability Customer Support : Function Availability Fault Management : Function Availiability Joseph : System administrator Responsiveness Experience Maintenance system : System Availability Reliability provide d by provides administrate d by administrates provide d by provides Availability 11