Decision Support System for Modelling of Systems and Control Systems Design J. Sebestyénová Institute of Informatics, Slovak Academy of Sciences, Dúbravská 9, Bratislava, Slovakia
[email protected]
Abstract Agent-based decision support system for modelling and control of continuous, discrete event and distributed systems is described. Ontology-based database of the system contains methods and tools for modelling and control synthesis as well as complete models of some systems specified by attributes. The system provides three possibilities of reasoning: search according to specifications and requirements, search of accomplished similar cases, and case-based reasoning of models and control methods used in the cases that are similar to specified task.
1. Introduction Decision support systems are used by people who are skilled in their jobs and who need to be supported rather than replaced by a computer system. The broadest definition states that decision support system is an interactive computer-based system or subsystem intended to help decision makers use communications technologies, data, documents, knowledge and/or models to identify and solve problems, complete decision process tasks, and make decisions. According to [7], five specific decision support system types include: - Communications-driven DSS (group DSS, interactive video, virtual organizations, e-meeting systems, workflow systems) - Data-driven DSS (data warehouses, online analytical processing, data structures) - Document-driven DSS (full text search applied) - Knowledge-driven DSS (production rules, inference methods, data mining, case-based reasoning, data visualization, genetic algorithms, neural networks, semantic networks, frame concepts, databases, …) - Model-driven DSS (accounting and financial models, decision trees and multi attribute utility models, influence diagrams, forecasting models,
network and optimization models, simulation models, modelling languages). Another classification of decision support systems in a knowledge-oriented way: - Symbiotic systems (static systems, knowledge explicitly predefined) - Expert systems (static systems, reasoning using explicit and implicit knowledge in form of rules) - Adaptive systems (dynamic systems, inductive inference to generate new knowledge) - Holistic systems (dynamic systems capable of holistic problem processing). The concept of decision automation is deceptively simple and intriguingly complex. From a narrow perspective, a decision is a choice among defined alternative courses of action. From a broader perspective, a decision involves the complete process of gathering and evaluating information about a situation, identifying a need for a decision, identifying or in other ways defining relevant alternative courses of action, choosing the “best”, the “most appropriate” or the “optimum” action, and then applying the solution and choice in the situation. Automation refers to using technologies including computer processing to make decisions and implement programmed decision processes. Typically decision automation is considered most appropriate for well-structured, clearly defined, routine or programmed decision situations. A decision support system can be approached from two major disciplinary perspectives, those of information systems science and artificial intelligence. We present in this paper an extended ontology for a decision support system in control theory domain. The ontology explicates relevant constructs and presents a vocabulary for a decision support system. It emphasizes the need to cover environmental and contextual variables as an integral part of decision support system development and evaluation methodologies. These results help the system developers to take the system's context into account
Proceedings of the 2005 International Conference on Computational Intelligence for Modelling, Control and Automation, and International Conference on Intelligent Agents, Web Technologies and Internet Commerce (CIMCA-IAWTIC’05) 0-7695-2504-0/05 $20.00 © 2005
IEEE
through the set of defined variables that are linked to the application domain. With these extensions the focus in decision support systems development shifts from task ontology towards domain ontology. Agent-based decision support system for modelling and control of continuous as well as discrete event systems developed in MARABU project [3] is used to illustrate the ontology-based database creation and reasoning. The database of the support system contains methods and tools for modelling of the systems, control synthesis, and simulation. Further, the database contains complete models of some systems specified by attributes.
2. Domain ontology for control theory Specifications of domain conceptualizations, called ontologies, are essential for the development and use of intelligent systems [4] as well as for the interoperation of heterogeneous systems. They provide a vocabulary for a domain and can be used as building block components of knowledge bases, object schema for object-oriented systems, conceptual schema for databases, structured glossaries for human collaborations, vocabularies for communication between agents, class definitions for conventional software systems, etc. We can consider ontology as a collection of agreements upon a vocabulary of common terms and meanings in a given domain. Concept is an entity representing some "thing" in the real world. Main directions in usage of ontologies are: - Semantic web (computer - understandable semantics) - Multi-agent systems communication (meaning of sent and received messages) - Database systems (in the process of distributed databases reasoning and in the process of ontologybased database creation). Ontologies are of three types: terminological, information and knowledge modelling ontologies. The difference between them is defined by the amount and type of structure in ontology. Terminological ontologies define a lexicon, information ontologies define record structures of databases, and knowledge modelling ontologies specify knowledge with a richer internal structure than information ontologies. In practical terms, developing an ontology includes: (1) defining classes in the ontology, (2) arranging the classes in a taxonomic (subclass–superclass) hierarchy, (3) defining slots and describing allowed values for these slots, (4) filling in the values for slots for instances.
Domain ontologies capture the knowledge valid for a particular type of domain. Generic ontologies are valid across several domains. Domain of control theory is wide and use of different control laws is appropriate on different levels of plant control. Various approaches to systems description, model forms, identification methods, and control methods have to be involved [5, 9]. For example, conventional control methods range from manual control to feedback control, cascade control, ratio control, feed forward control, combined FB, FFW and cascade control. Classical methods of advanced control contain adaptive control, gain scheduling methods, time delay compensation control methods. Model predictive control refers to a class of computer control algorithms that utilize an explicit process model to predict the future response of a plant. Advanced control methods: adaptive and self tuning control methods, optimal control methods, non-linear control methods, robust, hybrid and discrete event control methods, fuzzy control methods, artificial neural and genetic control methods, fuzzy-neuro and fuzzy-genetic predictive control methods. Multi-dimensional control methods, e.g. dynamic matrix control, have to be applied in large-scale systems where interactions between subprocesses cannot be compensated by independent control loops. If we do not know mathematical description of the controlled process, robust and adaptive controls solve the problem. In addition to methods used for continuous variable dynamic systems, modelling and control methods for discrete event dynamic systems should be involved in control theory domain ontology. Function of discrete event systems is in coordination level hierarchically placed over process control. In development of modern technologies, special interest is paid to matter motion control over complex definition areas as distributed parameter systems [6]. Distributed parameter systems can be found in various technical and non-technical branches in the form of lumped-input/distributed-output systems. For a domain of control theory, following terms have to be defined in the ontology: system, model, control, system description method, control method, and tool. Each term is characterized by attributes. Concept hierarchy plus attributes gives ontology. Relations used in the described ontology are: part of, attribute of, value of, is a (subclass of), instance of. There are two kinds of conceptual knowledge: concept and set. A concept is defined by the essence of the objects it subsumes and not by their state. Such a definition allows us to focus on the essence of the concepts and not on their state. An essence is invariant,
Proceedings of the 2005 International Conference on Computational Intelligence for Modelling, Control and Automation, and International Conference on Intelligent Agents, Web Technologies and Internet Commerce (CIMCA-IAWTIC’05) 0-7695-2504-0/05 $20.00 © 2005
IEEE
which is not the case of state. On the other hand, a set makes it possible to put together objects whose state shares some common properties. For instance, if “Human Being” refers to a concept, “Teenagers” refers to a set composed of human beings whose age is in given constraints. Differences are elementary units from which the meaning of terms is built. This means they have no meaning in themselves. A difference belongs to the essence of objects. Unlike an attribute it cannot be removed from the definition of an object without changing its nature; nor can it be valued. For example, for human beings “mortal” is a difference whereas “age” is an attribute. A difference is a unit that builds meanings and divides concepts. Classification of concepts used in a wide domain is not a trivial task [8].
3. Ontology - based database Properties represent the internal data storage of an object. The state of an object at any time is determined by values stored in each of its properties. Properties can be: - Attributes: primitive variables such as numbers, strings, or Boolean values (e.g. age, name, …) - Single-valued references: references to other objects - Multiple-valued references: references to collections of other objects (for example a collection of employees in a company object). The DB schema is the structure of data, whereas the data are the facts. We can map the data structures in the database to ontology: table ↔ class, table column ↔ property, value of table column ↔ literal or resource, foreign key ↔ property pointing to other resource, table row ↔ instance of class. Relational databases are the most common DB today. If we want to create a relational database schema according to a given domain ontology, first we have to create a table (or tables) of properties with all of their possible values. Secondly, we have to create a table for representing a given class of the ontology. An instance of the class (individual) will be represented as a row in the table, and it will be characterized by the properties with given values (valuation is restricted to previously specified possible values). Such approach can provide variable creation of database schema from an application, though it is not usual or obvious for relational databases. In database design, it is important to properly arrange and index the attributes to achieve effective reasoning. The proposed database consists of three parts: DB of methods and tools that are available, case-
Figure 1 Specification of required support in MARABU
base of concrete examples, and knowledge base of control theory domain. They are arranged in 37 tables of relational database. For any model, we need to know which system is the model of, what modelling method is used, which tool was used to create the model, and what modelling requirements were given. Similar specifications are used for control, too. In database, each method is stored with values of its attributes that specify the system and requirements for which the given method can be used. If valuation of the attribute is not important from the point of view of specified method, tool, or case usability, the attribute need not be specified. In such situation, the decision support systems reasoning algorithm regards this attribute matching any requirements. The same is done for tools. Proposal of our database schema is based on systems classification described in [8]. Lists of attributes are organized into a set of questionnaires. The user specifies the characteristics of the system he wants to model and/or control by filling in the questionnaires. The questionnaires are organized hierarchically from general characteristics to more detailed attributes. The user need not specify all of the available attributes in the questionnaires, only those that are important for a given task. Knowledge base of a multi-agent system contains an ontology-based representation of data relevant to structure of a domain as well as supplementary functional data. Persistent data are stored in database. Temporary information is stored in system variables. For example, type of a required support is stored in variables model, control, simulation, or similar cases as given for illustration in Fig. 1.
4. Agent - based decision support system In knowledge engineering, agents offer the flexibility to integrate many different categories of processing within a single system. Agent definitions
Proceedings of the 2005 International Conference on Computational Intelligence for Modelling, Control and Automation, and International Conference on Intelligent Agents, Web Technologies and Internet Commerce (CIMCA-IAWTIC’05) 0-7695-2504-0/05 $20.00 © 2005
IEEE
Figure 2 Model and control methods searched according to user requirements in context of distributed systems
range from descriptions based on a functional analysis of how agents are used in technology to far more ranging expositions based on different interpretations of the role and objectives of artificial intelligence and cognitive science. Artificial intelligence is a very diverse field and agents are used as metaphors for work in many areas [1, 2]. Multi-agent systems are appropriate for domains that are naturally distributed and require automated reasoning. Agents should perform the following capabilities to some degree: - Planning or reacting to achieve goals - Modelling the environment to properly react to situations - Sensing and acting - Inter-agent coordination - Conflict resolution. To design a multi-agent system for a given problem, the designer has to understand how should agent and AI techniques be applied to the domain, what competencies agents need, and which techniques implement those competencies. Thus, multi-agent
system design consists of (1) dividing resources and domain responsibilities among agents, (2) determining which core competencies satisfy which domain responsibilities, and (3) selecting techniques to satisfy each core competency. According to distributed domain-specific responsibilities agent-based systems may be heterogeneous, with each agent responsible for a different set of goals or homogeneous, where agents share the same goals. Agents in the proposed system work according to simple workflow that is specified by user in terms of required support. Decision theory is a means of analyzing which of a series of options should be taken when it is uncertain exactly what the result of taking the option will be. Decision theory provides a powerful tool to analyze scenarios in which an agent must make decisions in an environment. Most AI systems operate on a first-principles basis, using rules or axioms plus logical inference to do their work. Those few reasoning systems that include analogy tend to treat it as a method of last resort, something to use only when other forms of inference have failed. The exceptions are case-based reasoning systems, which started out to provide computational mechanisms similar to those that people seem to use to solve everyday problems. Unfortunately, case-based reasoning systems generally have the opposite problem, tending to use only minimal first-principles reasoning. In most case-based reasoning systems, cases are stored as named collections of facts in a memory. They are designed for a specific range of problems. Each case is a set of features, or attribute-value pairs. The case retrieval algorithm is mostly a simple k-nearest neighbors algorithm. The basic case-based learning algorithm performs poorly when cases contain many irrelevant attributes. Unfortunately, deciding which features are important for a particular learning task is difficult. In proposed decision support system, there are three possibilities of reasoning. First, according to requirements specified in questionnaires, a user can obtain a list of appropriate methods and existing tools, with their detailed characteristics and web-connection to a given tool’s provider. After receiving a list of methods satisfying the requirements (see Fig. 2), the user can choose from them, or require, as an advice, information from the system about which of the methods fulfils the requirements the best or which of them is mostly used for similar tasks. The other possibility is to obtain examples of using some suitable tools for modelling a system similar to the system
Proceedings of the 2005 International Conference on Computational Intelligence for Modelling, Control and Automation, and International Conference on Intelligent Agents, Web Technologies and Internet Commerce (CIMCA-IAWTIC’05) 0-7695-2504-0/05 $20.00 © 2005
IEEE
Figure 3 Model forms and control methods reasoned from similar cases in context of distributed systems
specified by the user. Third reasoning possibility of the support system is case-based reasoning [3]. This branch of reasoning has been applied if no model or control method matches exactly the specifications and requirements. The web-based portals prove to be very suitable for knowledge management. Knowledge portals are flexible and easy to use and may provide almost any kind of content or functionality. To structure the architecture of a knowledge portal, the following threelayer model is being used: user interface and navigation, functions (personalization, active process support, coordination of agents, document management), and knowledge base. Knowledge portals provide a flexible knowledge environment to a potentially large number of users. The mission of a knowledge portal is not only to provide a library-like pool of information, but to actively support the user in his or her decision processes. Fig.1 shows specification of required support, system and model requirements in context of distributed systems. Fig. 2 is a screenshot of the searched models and control methods in MARABU. Fig. 3 illustrates case-based reasoning. Whenever no model or control method matches exactly the user requirements, the model forms and control methods are reasoned from similar cases. If the user requires support in a form of accomplished similar case, similar cases with references to a tool where simulation can be done are provided as shown in Fig. 4.
Figure 4 Similar cases to user specified system in context of distributed systems
Generation of questionnaires as well as database search works differently in context of continuous, discrete event, or distributed systems. For example, model requirements for DES also need to deal with verification. The control requirements in context of continuous systems are visible in Fig. 5; a form for PID tuning method selection is available only if selected controller type is PID. Fig. 6 shows the same in context of smart control.
5. Distributed computing In distributed DB environments, the combination of resources from multiple sources requiring different interfaces is a universal problem. The current solution requires an expert to generate ontology, or mapping, which contains all interconnections between the various fields in the databases. Research goes on automating the ontology creation for distributed database environments. Web services are self-contained, modular applications, accessible via the web, that provide a set of functionalities. Grid computing is a services-oriented architectural approach that uses open standards to enable distributed computing over the Internet or a private network. This
Proceedings of the 2005 International Conference on Computational Intelligence for Modelling, Control and Automation, and International Conference on Intelligent Agents, Web Technologies and Internet Commerce (CIMCA-IAWTIC’05) 0-7695-2504-0/05 $20.00 © 2005
IEEE
specifications and requirements or to search for accomplished similar cases. In addition, they can also use various tools, web resources and remote labs to provide a simulation of a specified task. Acknowledgment
Figure 5 Control requirements for continuous systems
approach can help research organizations and universities aggregate resources, data storage, and devices to create a single, unified system. At its core, grid computing enables devices to be virtually shared, managed, and accessed across an enterprise, consortium or workgroup. Although the physical resources that compose a grid may reside in multiple locations, users have seamless and uninterrupted access to these resources and remote labs [9].
The author is grateful to the Slovak Science and Technology Assistance Agency (grant No. APVT-51011602) for partial support of this work. We also gratefully acknowledge the many useful discussions with control theory domain experts Š. Kozák, B. Hrúz, G. Hulkó from Slovak University of Technology in Bratislava and J. Liguš from University of Technology in Košice.
7. References [1] Davis, D.N., “Synthetic Agents: Synthetic Minds? Frontiers of Cognitive Agents”, In: IEEE Symposium on Systems, Man and Cybernetics, San Diego, 1998. [2] Franklin, S.P., A. G. Graesser, “Is It an Agent, or Just a Program?: A Taxonomy for Autonomous Agents”, In: Intelligent Agents III, J.P. Muller, M.J. Wooldridge and N.R. Jennings (Editors.), 1996, Springer-Verlag, Heidelberg. [3] Frankovič B., I. Budinská, J. Sebestyénová, Dang T. Tung, V. Oravec, “MARABU - Multiagentový podporný systém pre modelovanie, riadenie a simuláciu dynamických systémov”, In: AT&P Journal 4, 2005, ISSN 1335-2237, pp. 57-59, in Slovak. [4] Heijst, G., “The Role of Ontologies in Knowledge Engineering”, PhD. Thesis, University of Amsterdam, Netherlands, 1995. [5] Hrúz B., Š. Kozák, V. Veselý, “Trendy vo výučbe spojitých a diskrétnych systémov riadenia”, In: Cybernetics and Informatics, New Trends in Education of Automation and Information Technology, September 2004, ISSN: 13364774, pp.50-62, in Slovak.
Figure 6 Control requirements and searched methods in context of smart control
6. Conclusion Support system for modelling and control of the continuous, as well as discrete event systems has been described. The database of the system contains methods and tools for modelling and control synthesis as well as a set of complete models of systems specified by their attributes. Users can opt to search for models and control methods according to given
[6] Hulkó G., M. Antoniová, C. Belavý, J. Belanský, J. Szuda, P. Végh, Modeling, Control and Design of Distributed Parameter Systems, STU Bratislava, 1998, ISBN 80-227-1083-0. [7] Power, D. J., Decision Support Systems Hyperbook. Cedar Falls, IA: DSSResources.COM, HTML version, Fall 2000, http://dssresources.com/dssbook/ [8] Sebestyénová J., “Usage of Domain Ontology in e-Learning”, In: Proc. 5th International Conference Virtual University VU’04, December 2004, Bratislava, ISBN: 80227-2171-9, pp. 272-277. [9] Zolotová I., J. Liguš, J. Jadlovský, J. Horváth, M. Duľa, S. Laciňák, “Remote Labs – Industrial Portal”, In: Proc. 5th Int. Conf. Virtual University, 2004, Bratislava, ISBN: 80227-2171-9, pp. 238-240.
Proceedings of the 2005 International Conference on Computational Intelligence for Modelling, Control and Automation, and International Conference on Intelligent Agents, Web Technologies and Internet Commerce (CIMCA-IAWTIC’05) 0-7695-2504-0/05 $20.00 © 2005
IEEE
