Decision support systems

Decision support systems Abstract Nowadays, Decision Support Systems has a significant role in almost all areas of life. These systems go further and use new technologies like data mining and knowledge and data discovery (KDD) to improve and facilitate human decision making. First of all we provide some definitions about decision making, models and processes. Afterwards, we discuss about knowledge and data discovery and also, Intelligent decision support systems. At last, as an empirical survey, we compare two different cultures in using decision making support systems. One of them uses decision support system in clinical environment to improve the decision making and reduce crucial errors significantly; while the other uses the traditional system and relies on the human memory and experience rather than using decision support systems. Keywords: Decision Making, Decision Support, Knowledge and Data Discovery (KDD), Intelligent Decision Support Systems Introduction Information system has a significant role in supporting decision making, and in some special environments like business, health and education, gets the mandatory part. Moreover, such systems go further and use data mining and knowledge and data discovery (KDD) techniques to improve their abilities in supporting decision making. One of the environments that need information systems support for making crucial decisions and have direct effect on human life is clinical and health environment. We are going to look through the effect of decision support system in it. Decision Making Decisions and models Decision making is undeniably an essential and vital part of the human life. A decision problem may consist of numerous smaller decisions inter-related together, and the results of multiple decisions can be consolidated together; or one decision can influence another subsequent one. This influence can be fed as the input to a subsequent decision, or as a decisional choice for the users in determining which decision to make subsequently. This bigger decision, and its smaller decisions embedded within, must be represented in a simple manner for decision makers to read, understand, and communicate with. Each decision can be represented in the form of a model, to represent, describe and depict the decision problem and its interaction under consideration, whether it is simply an abstraction schema, insights to the decisions rather than mere numbers actual model instance, or executable computer program module. Each decision model can be a permanent modeling scenario which can be retrieved and included as part of a bigger scenario. Alternatively, it can be a temporary modeling scenario that is aggregated or pipelined within a bigger scenario. Such model integration treatments are subject to the discretion of users at the time of making such decisions. Even though each of these decisions may have a direct or indirect bearing on other subsequent decisions and can easily influence the overall decision and conclusion, many decision making processes and systems treat these decisions as independent and unrelated. This obscures the users from seeing and discovering the true effects and influen ce of the decision problem and its interaction under consideration, whether they are interrelated and/or interdependent. The element of interdependence may not be discovered until the full picture can be seen and assessed. Even though many decisions do occur in a sequential fashion, there are also many decisions that occur in parallel, evolve over time and converge to a concluding decision, or eventually combine or are interwoven into a final decision. Therefore, the decision making process should neither be fixed nor predetermined beforehand so that the execution order can be created as required. Hence, modeling is an important process in understanding, capturing, representing, and solving these decision models especially in terms of their interrelatedness across multiple models and their instances over a period of time. Furthermore such models should ideally be able to capture functional, behavioral, organizational, and informational perspectives. Decision systems are intended to assist users in making a decision. There are several types of users involved in using decision systems and these users progress as they develop more confidence: from inexperienced/na?ve decision makers, to average decision makers/ analysts, to experienced decision makers/modelers. Each type of user has different needs and should not be restricted by the constraints of any decision system that dictates the steps and techniques behind analyzing and solving a decision problem. Some users may need more decisional and/or system usage guidance while others may prefer to have minimal guidance. Some may wish the decision system will take care of the entire decision making process including prescribing the order in which each set of data is requested as well as the order in which each decision model is executed; while others may wish to intervene to a greater extent in designing the entire decision making process and the execution order to suit, or to a lesser extent in specifying a particular solution method. There are a variety of reasons as to why a human intervention is warranted and needed from the perspective of an experienced user. However, it is interesting to note that the type of guidance may have an adverse effect on decision model selection and ultimately the decision outcome. It is unreasonable and impractical to expect decision makers to operate a different decision making system for each decision and to comprehend the full effects of the consolidation and integration from these decisions. A decision making process is not necessarily about concentrating on the decision itself, but should emphasize the ways in which decisions are made. Therefore, users should be able to choose an optimizing approach and solution as well as a satisfying approach and solution, and not be limited to only one approach and solution that is traditionally incorporated in decision systems. Due to the frequency and complexity of interrelated decisions, some users may recall an existing scenario as input to another scenario, or recall several existing scenarios for comparative purposes. Decision systems need to be built in a flexible way so that decision models and components can be easily assembled and/or integrated together to create new scenarios and specific scenarios can be built and tailored to meet the needs of particular user groups. With all these issues in mind, the framework and architecture of an ideal decision system should have independent components that enable components to be easily assembled and integrated together to form a decision scenario. They should be flexible enough to serves various types of users and accommodate various types of decision making processes. They should also be sufficiently versatile to handle decision problems regardless of paradigms and/or domains under consideration. Good decision making frameworks must therefore be in place f or system framework and architecture to exhibit modeling flexibility, component independence, and versatility in domain and/or paradigm. To overcome the issues and fulfill the requirements discussed above, we first propose a converging decision analysis process, an optimizing?satisfying decision model, and a cyclical modeling lifecycle. Normative decision making processes Decisions can evolve and converge into a concluding decision over time. This can occur within re-evaluating a decision problem, or evaluating across multiple decision problems that are similar. This iterative decision making process is known as the convergence process. As decisions evolve and refine over time, decision makers are able to concentrate on essential factors and eliminate nonessential ones in order to narrow down the scope of the decision problem. Such attention-focused method provides a cut down version of the problem. A decision is subsequently made from these remaining factors of the reduced problem. Such decision-focused method provides an actionable result from the given problem. Since there can be many decisions within a decision problem, several iterations of attention-focused and decision-focused methods are applied while intermediate decisions within the decision problem are made and converged. Such revision and refinement occur irrespective of paradigms and doma ins. This notion of applying the attention focused and decision-focused methods within a convergence decision making process are depicted in Figure 1. Figure 1. Converging decision analysis, as in an 1D-CSP scenario One-Dimensional Cutting Stock Problem (1D-CSP) was used for illustrative purposes in order to design and implement the proposed framework and architecture. 1D-CSP is about cutting strips of raw material into desired sizes according to customer order widths. We often do not have unlimited supplies of raw materials and would therefore need to formulate and decide on which cutting patterns are used. 1D-CSP is a resource management problem with a traditional goal of minimizing wastage. Besides wastage, there may be other objectives that must be considered. For example, minimize machine setups through the changing of cutting knives, minimize machine setups through reducing the number of cutting patterns used, or minimize the number of disruption in the sequence of cutting patterns used. Even though 1D-CSP is considered to be a simple problem in pure mathematical terms, it becomes a reasonably complex decision problem once one considers all the real world constraints and objectives, and th e interrelated decisions involved within its decision making process. The 1D-CSP can be used as a decision problem to illustrate the converging decision analysis process, as depicted in Figure 1. The first decision is a pattern generation heuristic that generates combinations of cutting patterns. This decision concentrates only on generating those cutting patterns that are relevant to the decision problem under consideration (an attention-focused method). The second decision is determining which cutting patterns among the generated ones should be retained or discarded (a decision-focused method). This can be based on specific rules such as an allowable number of cutting knives per cutting pattern. It can also be based on the decision makers personal experience on whether certain cutting patterns should be discarded. The third decision is the creation of linear programming constraints that identifies the feasible area of the problem under consideration (an attention-focused method), while the fourth decision is finding an optimal point within the feasib le area (a decision-focused method). Neither of the focused methods has to produce an optimal or a satisfying solution necessarily. It is entirely up to the decision maker to decide on what sort of solution is desired at the time. Each decision and solution can be encompassed within a decision model that consists of both the optimizing model and satisfying model, as depicted in Figure 2. In a decision problem that consists of multiple interrelated decisions, the result from one model may be fed into another model continuously until an ultimate result is reached, and the result from a model can take on a different solution option. Each decision model may return to itself for refinement, or return to the previous model for additional processing, or feed to the next model for further processing. This return may be due to an infeasible solution, or a better understanding of the model which eventually leads to a change in the parameters of the model. The 1D-CSP can be used to illustrate the optimizing?satisfying decision model, as depicted in Figure 2. The first decision model pattern generation heuristic is a satisfying model that produces only those cutting patterns that are relevant and desirable to the decision problem under consideration. The second decision model is also a satisfying model in selecting or deselecting among the cutting patterns already produced. The third and fourth decision models are optimizing models that optimize using the linear programmings simplex method. Figure 2. Optimizing?satisfying decision model Decision modeling lifecycle The approach of Simon to the decision making process in terms of intelligence, design, and choice is very decision-oriented. However, as Glob has suggested it is about the way in which we model the decision. Therefore, we propose to integrate Simons proposal with MS/ORs modeling proposals that attempt to support every phase and aspects of decisions and modeling lifecycle. Such a design approach is crucial to support the modeling and decision environments and ensure that non-predetermined decision making processes and interrelated decisions characteristics can be modeled. This proposed modeling process is cyclical and iterative, and enables continuous adjustment and refinement especially in storing and retrieving decision problems as decision scenarios, as summarized in Figure 3. Despite the fact that the modeling lifecycle progresses step-by-step in a cycle, it can return to any earlier steps and not just the previous one, and can skip some steps in the later iteration if it has already gone through that particular step earlier on. It is however more difficult to represent these possible movements visually in the modelling lifecycle and is therefore not illustrated in Figure 3. The lifecycle is valuable not only from the point of view of modeling the decision itself but especially for highlighting the role of the system components of the decision, whether it is a data, model, solver, or scenario. Once a problem is understood it can be represented in the form of a model which is then instantiated with data and integrated with solvers so that it can be executed. Such a model is especially beneficial if it is storable and retrievable for later use and comparison. Once a model is represented, a solution can be derived through analyzing and investigating as well as comparing with various model instances. The derived solution is then reviewed and validated. If it is considered unsatisfactory such information can be used to modify and reformulate the decision model. Figure 3. Cyclical modeling lifecycle Even though the decision system will progress through the entire modeling lifecycle in producing the end result, it is important to note however that not all users will execute all the steps of the modeling lifecycle. Depending on the competencies of the decision makers and their permissions, they may interact with certain steps in the modeling lifecycle. For example, the inexperienced decision maker may interact with only step 2; the average decision maker may interact with steps 2, 3 and 4; whereas the experienced decision maker may interact with all 6 steps in the modeling lifecycle, as shown and contrasted in Figure 4. This decision modeling lifecycle provides a sound basis for the decision support and modeling framework and architecture. Figure 4. Interaction between 3 types of user groups and the modeling lifecycle Intelligent Decision Support Systems While IDSS (Intelligent Decision Support Systems) have been receiving increasing attention from the DSS research community by incorporating knowledge- based techniques to provide intelligent and active behavior, the state-of-the-art IDSS architecture provides little support for incorporating novel technologies that serve useful DSS information, such as the results from the knowledge and data discovery (KDD) community. Data Mining and Knowledge Discovery In recent years, the terms knowledge discovery and data mining (commonly referred to as KDD) have been used synonymously. They both refer to the area of research that draws upon data mining methods from pattern recognition (Tuzhilin, 1993), machine learning (Han et al., 1992) and database (Agrawal et al., 1993, 1994) techniques in the context of vast organizational databases. Conceptually, KDD refers to a multiple step process that can be highly interactive and iterative in the following (Fayyad Uthurusamy, 1995): the selection, cleaning, transformation and projection of data; mining the data to extract patterns and appropriate models; evaluating and interpreting the extracted patterns to decide what constitutes ?knowledge?; consolidating the knowledge, resolving conflicts with previously extracted knowledge; making the knowledge available for use by the interested elements within the system. A number of KDD systems are similar to IADSS data miner agents in spirit and in technique. Such work in designing and implementing practical KDD systems is crucial to our research in the sense that their results provide solid KDD pragmatic technologies ready to be integrated into our IADSS architecture. However, the current state of using KDD techniques for decision support remains in its infancy, as preliminary applications that use exclusively KDD techniques. It is our point of view that such isolated applications have limited scope and capabilities, while future KDD techniques will play an integral role in complex business systems that incorporate a wide range of technologies including intelligent agents, multimedia and hypermedia, distributed systems and computer networks such as the internet, and many others. From a DSS perspective, a simple DSS architecture that consists of a single decision maker with single information source knowledge discovery functionality lacks the ability to deal with complex situations in which multiple decision makers or multiple informatio n sources are involved. Most existing DSSs with data mining and knowledge discovery capability fall into this category. Intelligent Agents The concept of intelligent agents is rapidly becoming an important area of research (Bhargava Branley, 1995; Etzioni Weld, 1994; Khoong, 1995). Informally, intelligent agents can be seen as software agents with intelligent behavior, that is, they are a combination of software agents and intelligent systems. Formally, the term agent is used to denote a software-based computer system that enjoys the following properties (Wooldridge Jennings, 1995): Autonomy: Agents operate without the direct intervention of humans. Co-operatively: Agents co-operate with other agents towards the achievement of certain objectives. Reactivity: Agents perceive their environment and respond in a timely fashion to changes that occur. Pro-activity: Agents do not simply act in response to their environment; they are able to exhibit goal-directed behavior by taking the initiative. Mobility: Agents are able to travel through computer networks. An agent on one computer may create another agent on another computer for execution. Agents may also transport from computer to computer during execution and may carry accumulated knowledge and data with them. Furthermore, there has been a rapid growth in attention paid to developing and deploying intelligent agent-based systems to tackle real world problems by taking advantage of the intelligent, autonomous and active nature of this technology (Wang Wang, 1996). Intelligent Decision Support Systems Intelligent decision support systems (Chi Turban, 1995; Holtzman, 1989), incorporating knowledge-based methodology, are designed to aid the decision-making process through a set of recommendations reflecting domain expertise. Clearly, the knowledge-based methodology provides useful features for the application of domain knowledge in decision making. However, the knowledge stored in the knowledge bases is highly domain-oriented and relatively small changes in the problem domain require extensive intervention by the expert. Powerful information communication channels, such as the internet (information superhighway), are continuously changing the decision making process. When decision makers make decisions they not only rely on brittle domain knowledge but also on other relevant information from all over the world. As a result, the challenge of discovering and incorporating new knowledge with existing ones requires us to introduce new techniques (such as intelligent agents and knowledg e discovery) into DSSs. Research into IDSS includes the work by Rao et al. (1994), who presented an intelligent decision support system architecture, IDSS, that stresses active involvement of computer systems in decision making, on the other hand, the work by Sycara at CMU LEI (Laboratory for Enterprise Integration) proposed the PERSUADER (Sycara, 1993), which incorporates machine learning for intelligent support of conflict resolution and the work on NEST which incorporates distributed artificial intelligence (DAI) with group decision support systems by Fox and Shaw (Shaw Fox, 1993). The proposed IDSS architecture is similar in substance to our proposed IADSS, which incorporates distributed artificial intelligence and incorporates the principles of co-operative distributed problem solving in the decision-making process. However, as we have pointed out above, it is necessary for the incorporation of data mining technology which extracts important information from vast amounts of or ganizational data sources in order to provide additional information that may be crucial for the decision-making process. IADSS architectural configuration As we have pointed out in our introduction, there exist numerous obstacles that remain to be overcome in today?s DSSs to fully achieve the vision of IADSS. The integration of intelligent agents with DSSs will be able to address most, if not all, of the articulated issues. However, even within the application of an intelligent agent-based architecture, there exist two different forms (or configurations) of the decision-making process that the particular architecture will be able support: Single decision maker-multiple miners and multiple decision makers-multiple miners. Single Decision Maker-Multiple Miner DSS Processes We have argued in the previous section that a possible configuration of IADSS architecture, namely the single decision maker-single miner form, has severe limitations when it comes to extendibility and the ability to be integrated into an overall organizational decision support framework. However, in many real life cases, the single decision maker situation is still of importance. In today?s organization, there may exist a myriad of organizational information sources on which useful data relationships and patterns may be discovered to support the singular decision maker?s decision process. As a result, the IADSS configuration of a single decision maker with multiple data miners warrants attention and analysis. Under IADSS, the architecture of such a single decision maker, multiple knowledge miners assisted DSS is shown in Figure 5. Figure 5. Multi-Agent-based DSS Figure 6. A Multi-Agent-Based GDSS There are three classes of intelligent agents (we call them decision support agents or DS agents) contained within this architecture: Knowledge miners that discover hidden data relations in information sources, user assistants that act as the intelligent interface agents between the decision maker and the IADSS and a knowledge manager with repository support that provides system co-ordination and facilitates knowledge communication. Further details about the functionality and internal structure about each type of agent is elaborated in the next section. Multiple Decision Maker-Multiple Miner-Assisted GDSS Process The single decision maker configuration discussed above can be easily extended into a group decision support system (GDSS) architecture (as seen in Figure 6. by the introduction of additional user assistants for each additional decision maker). Compared to the single decision maker configuration in Figure 5, each user assistant agent is further augmented to provide support for group-based communication between different decision makers. It is important to observe that with the introduction of each additional DS agent; only an extra knowledge communication channel between the new DS agent and the knowledge manager is needed. This enables a manageable linear increase in the number of knowledge communication links corresponding to the increase in the number of agents in the IADSS system, rather than the quadratic increase in the number of direct communication links in a direct agent-to-agent fashion. Furthermore, our proposed IADSS is an open architecture with potential for the integration of future technologies by the incorporation of additional classes of intelligent agents. IADSS architecture at a glance Intelligent Decision Support Agents As described above, there are three types of intelligent agents in an IADSS system: Knowledge miners, user assistants and knowledge managers. This section will provide a more detailed description of such agents and their internal architectures. Knowledge Miners. The role of knowledge miners in IADSS is to actively discover patterns or models about a particular topic which provides support in the decision-making process. There are four components in a knowledge miner. The IADSS interface component manages the communication between the miner and the knowledge manager. When a knowledge miner receives messages that are represented in a common representation, the IADSS interface translates these messages into the local format based on the common vocabulary. On the other hand, when the knowledge miner sends messages out, the IADSS interface translates them into common format first, then sends them to the knowledge manager. In order to carry out the mining task, the necessary control knowledge as well as domain knowledge is kept in the knowledge base component, while the data interface component serves as a gateway to the external information sources. The knowledge discovery is usually done by discovering special patterns of the d ata, i.e. by clustering together data that share certain common properties. For instance, a knowledge miner may find that within this week, a number of stocks are going up. There are two different types of knowledge mining agents, event-driven knowledge miners and tusk-driven knowledge miners. The event-driven knowledge miners are agents that are invisible to the decision makers, and their results may contribute towards the decision-making process. Based on the specification of the IADSS, such event-driven knowledge miners start when the IADSS starts up. When a particular event comes, an agent will start its knowledge mining. Events may be temporal events, e.g. every day at 1 a.m., every hour, etc. Or, events may be constraint-triggered events, e.g. every 10,000 customers, when a certain type of customer reaches lo%, etc. Usually, such event-driven knowledge miners work periodically. They follow a sleep-work-sleep-work cycle and will be destroyed when the entire IADSS system termina tes. On the other hand, task-driven knowledge miners are created for particular data mining tasks based on requests originated by the decision makers. After a knowledge miner completes its task, it sends the mining results to the knowledge manager and is then terminated automatically. From the view point of decision support, knowledge miners play the role of information extractors which discover hidden relationships, dependencies and patterns within the database, whether the information is discovered by an event-driven knowledge miner or a task-driven knowledge miner, which may be utilized as evidence by decision makers within the GDM process. User Assistants. Interaction between a particular decision maker and the IADSS is accomplished through a user assistant agent. The architecture of a user assistant contains four components. The multimedia user interface component manages the interactions with the decision maker such as accepting requests for a task-driven knowledge miner, while the IADSS interface manages the knowledge communication with the knowledge manager. The necessary knowledge such as the common vocabulary, decision history and others are kept in a local knowledge base component. All three components are controlled by an operational component that provides the facility of differencing, multimedia presentation and collaboration. With regard to the role the user assistant plays in the decision process, it enables the decision maker to view the current state of the decision process and to convey his or her own opinions and arguments to the rest of the decision making group. It also enables the decision maker to i ssue requests for task-driven knowledge miners to attempt to discover some particular type of organizational knowledge from business data. The user assistant will then relay the request to the knowledge manager and interpret the mining result if it is deemed appropriate. Knowledge Manager: The knowledge manager provides management and co-ordination control functions over all the agents in the IADSS architecture. The internal component-wide architecture of the knowledge contains four Components: The decision maker interface, the operational facilities, the miner interface and the agent knowledge base that provides support for localized reasoning. From the functional standpoint, the knowledge manager provides the following functionality in the IADSS architecture: Makes decisions concerning the creation and termination of knowledge miners as provided by the miner interface component of the knowledge manager. Mediates requests from user assistants through the decision maker interface, analyzes these requests through the localized knowledge and inference engine and then initiates an appropriate group of task-driven knowledge miners to collaboratively perform the requested task through the miner interface. Mediates the discovered knowledge from knowledge miners (whether it is an event-driven or a task-driven miner), stores the knowledge into the repository for possible future usage and forwards the relevant knowledge to interested decision maker users through the decision-maker interface. Manages and co-ordinates the knowledge transactions with each individual decision support agent such as common vocabulary, available decision topics, existing mining results and strategic knowledge, as provided by the operational facilities component. Manages the synchronization between the collection of decision support agents such as the progress of the task-driven knowledge miners and the notification of the decision makers when crucial knowledge is discovered. Mediates all other types of communication among decision support agents including the communication among user assistants and supports the retrieval of appropriate evidence from the repository by user assistants. In terms of the decision support process, the knowledge manager plays the role of manager and mediator between two decision makers, between the decision maker and the corresponding task-driven miners and between all decision support agents and the repository to address the issue of knowledge sharing. Current prescription process at the hospital The prescription process is shown in Figure 7. This description is based on interviews (questions 1?3 in the interview guideline, Appendix A) and observations by the first author. Figure 7. Current prescription process in the Ekbatan and Boras Hospital (UML activity diagram) The process starts as the physician in charge takes the patients history, performs physical examinations, and reviews available medical documents, including progress notes, laboratory findings, and imaging. These data sources guide the physician(s) to a set of differential diagnoses or a definitive diagnosis, which help the prescriber(s) to select appropriate treatment for the patient. The prescriber will then register medical records