Research On The Framework Of Multiple Abstraction Model

The modeling process of the physical world is an important step for model-based reasoning. The question is how to find an appropriate framework to represent abstraction, in which the abstraction model representation of physical world W, fit for some reasoning mode, can be constructed automatically.Abstraction is a pervasive behavior of people's perception, conceptualization and reasoning. There is such a consensus as follows: the ability capturing essence from outward form is a pivotal problem, however, when it comes to build an intelligent system, finding an appropriate representation usually becomes very difficult. Many literatures have made a lot of research to abstraction of AI field, such as problem solving, problem reconfiguration. Abstraction is defined by most researchers not in a structural way but as an induction of specific problem changes for reducing the computational complexity. Most representation changes called abstraction reduce information, like syntactic mapping and semantic mapping of abstraction, almost all of which are trying to capture the generic concept of abstraction, but they don't provide the constructive way of the problem. In 1998, after Korf, Saitta and Zucker proposed a model to represent changes, which includes both syntactic reconfiguration (the same forms, but the different information contents) and abstraction (the different forms, but the same information contents). The model generates abstraction as well as provides a general framework to formalize the physical world W and applies to conceptualization representation and learning. The model is called KRA (Knowledge Reformulation and Abstraction) model designing to accomplish the process of problem conceptualization and automatic operation of abstraction operators.In the field of qualitative modeling of physical systems, many researchers have made a lot of researches to accomplish particular reasoning tasks. The milestone in this field is the component-based way of ENVISION, the process-based way of qualitative theory and the constraint-based way of QSIM. All above ideas are using structural and behavioral knowledge to support various tasks, such as prediction, diagnosis, causal explanation and design. Later, some researchers started to focus on the qualitative reasoning theory based on functional and teleological knowledge. At the same time, for improving the validity and efficiency, the problem about mutual collaboration of multiple models of the same systems has been paid interest in. Whereas there still exit some problems: Firstly, there is no explicit, organic framework to represent the distinct models of the same system and to support the strict, valid and consistent representation of them. Secondly, the problem of the collaboration of multiple models of the same systems is not discussed. Thirdly, there is no discussion about how to output the partial results received in some models to other ones. In 1989-1995, Chittaro and colleagues proposed a new way to represent physical systems: Multimodelling. This method takes the reasoning tasks of physical systems as a kind of collaboration behavior and uses the contributions the multiple models making to reasoning. Every model of the multiple model framework surrounds a particular kind of knowledge and representation, as well as using some specific theory and aiming at a specific artificial system.In this paper, the research is mostly based on two fields: the general theory of abstraction in AI and the qualitative modeling of physical systems. Since it takes the people's cognizing process to the whole world including artificial systems as main body and focuses on the abstraction framework theory of physical world, it shows greatly the generalization. The generic formal abstraction framework of the physical world W is combined with the method of multiple knowledge in the modeling process of W to explore a generalized modeling framework representing W, in which the hierarchical abstraction model of various kinds of physical worlds can be automatically generated using multiple knowledge and at the same time, the corresponding reasoning system is built. In this paper, the abstraction process of the physical world is extended, which means the abstraction representation framework has potential to unite the worlds of different fields, to automatically percept the elements constructing the physical world, to build an abstraction mechanism and as a result to generate abstraction models automatically. Furthermore, multi-knowledge is introduced to build the mapping relation between models guided by various knowledge and to enrich the reasoning mechanism of abstraction models. The mainly research contents are as following:1. The abstraction operators of all levels in the KRA model framework are functionally defined and a dependent operating procedure of the abstraction operators is proposed. First of all, the abstraction operators are applied to each level of the basic framework of the system. Then, compare the behavior mode of the new super component, generated by applying the abstraction operators to T level, to the behavior modes of all the components of the most basic system. If the behavior mode of the super component is consistent with some kind of component, modify the results of P level and S level, keep its information like its name, and update its information such as type with the corresponding information of it. If there are no basic components whose behavior is consistent with the super component, the procedure stops. This operation always generates different component types not only to reduce space but also to increase the probability of operator reusing and improve the efficiency of abstraction operation. The abstraction hierarchical process of model-based diagnosis is described and two ways to construct operator sets, Static and Dynamic, are also proposed. With the static idea, all possible abstraction operators are pre-built according to the given basic framework Rg. It is just a relative way that predefines some operators right before the abstraction operation, however if some new types of components is generated during the abstraction operation, then new operators needs to be added to the operator sets dynamically. In addition, with the situation that there exist more than one component are mutually related, the operator must also be dynamically constructed. In the dynamic way, it only constructs the aggregation operators of connecting (serial or parallel) component sets. During the abstraction operation, the operator set will be complemented. Most abstraction operators based-on the basic components are pregenerated with the static way, so there is less chance to define new operators dynamically during the hierarchical abstraction process and the reusing possibility of abstraction operators is greatly improved. Only partial abstraction operators need to be constructed with the dynamic way and the time and space for predefining accordingly decreased, nevertheless new operators may be constructed more frequently to reduce the efficiency of the hierarchical process. The dynamic and static way to construct operator sets respectively adapt well to the systems with simple and relatively complex connecting structure. In this paper, a procedure is proposed to automatically generate the hierarchical representation of the system to be diagnosed by using the abstraction operator sets, and at the same time, its validity and complexity are also analyzed.2. The KRA model framework is extended on the Perception Level from two directions. On one side, the perception is divided into primary perception and abstraction perception. The process of primary perception is a common phase that is the first operation to the physical world W and the primary perception P is built by percepting the objects of W according to the particular task. Primary perception should cut less down the approximation conditions in order to describe W more general and meet more task requirements. Abstraction perception will be different from the agents'different reasoning requirements. With abstraction perception, the objects of P are mapped to the objects of the abstraction objects database Oa predefined by some agent, so the new perception P* is generated and the abstraction models adapting to various reasoning are constructed. On the other side, primary perception is extended to multi-perception which generates multi-domain abstraction model basing on the domains the objects of W belonging to. The domain relations between multi-abstraction models Wi and Wj are formally defined from two points of view: the intersection of their objects sets is empty (there is no such an entity involved in two different domains); there are at least one entity involved in two different domains. Knowledge of various domains can be applied to achieve multi-abstraction models and the reasoning ability is enhanced.3. In this paper, the concept of ontology is introduced in the General KRA model and the Abstraction Object Database is extended to Ontology Class which makes the knowledge sharing and reuse possible in the Perception level, the Language level and the Theory level of the KRA model. The three kinds of ontology abstraction operator (set) are also defined: fundamental operator, entity operator set and connection operator set, working on W and ontology class. Furthermore, the mapping between two different ontology classes is given to realize model abstraction and model reverse-abstraction. The abstraction degree of ontology class is proposed to represent the abstraction degree of ontology class and it is as well pointed out that in such a framework, the models of W with different abstraction degrees can be automatically constructed by using abstraction mapping. The hierarchical process of the ontology class and the modeling process of W based on the ontology class are formally described.4. The concept of process is extended and the concept of Flow Fragment is proposed. Moreover, the ontology class is redefined as Object-based Ontology Class and the definition of Flow Fragment is extended to propose the system-centered concept of Flow-based Ontology Class. The hierarchical representation of the flow-based ontology class is discussed in two directions: the abstraction hierarchy of flow fragment and the aggregation of indistinguishable flow fragments. In the hierarchical structure, the abstraction mapping relation between flow-based ontology classes is constructed to realize the hierarchical modeling process based on system-centered ontology. At the same time, the hierarchical relation of flow-based ontology classes also provides a mechanism of system-centered knowledge sharing and reuse. The modeling process based on the flow ontology class is formally described by using a good example. Three types of flow fragment groups are proposed. Moreover, the concepts of Phenomenon and Goal are described and the phenomenon-based indistinguishable flow fragments (or flow fragment groups) are defined by which some flow fragments, achieved in the process of flow-based perception, can be aggregated. The mapping relation is constructed between the set of flow fragments and of phenomenons, as well as between the set of phenomenons and of goals. The hierarchical process of the set of flow fragments, based on multi-knowledge, is formally described. This can provide the sharable and reusable knowledge database and the transformation mechanism for the hierarchical abstraction process of the physical world.