Nonrestrictive Concept-Acquisition by Representational Redescription
Wei Hui,Yuan Liang
2006-01-01
Computer Science
Abstract:A rich concept system, having fine-granularity and abundant semantic relationships, is absolutely necessary for any artificial intelligence system, especially when this system is confronted with a domain-open task, or a task beyond the range of classical expert system, such as let an AI system comprehend sentences in our daily life. These tasks exhibit their characteristics more from cognitive view instead of from engineering view, because problem-solving in these situations involves much more common sense and knowledge from diverse domains, or in other words this kind of domain-open problem-solving requires a comparatively dense concept system to broaden the base of semantic. However, most of existing expert systems and other knowledge-based systems can not meet such requirements because they are designed for those domain-restricted applications. In Cognitive Psychology and Developmental Psychology, the study of concept acquisition offered many cases and evidences in the processes of cognitive skills' maturation in behavior level. And this revealed that the development is one of the most important attributes of man's concept system. And a main lack of the classic expert systems is their knowledge bases are not constructed by development. Inspired by the concept acquisition of Developmental and Cognitive Psychology, we can improve the knowledge base design along the same way. Whereas what those concept development theories can provide are only coarse procedures or very abstract frames from the point of view of algorithm, because some crucial issues like the representation, evolution, storage, and learning process of concept etc. are not described. Maybe the realization details of constructing a concept system in a computer are not the aim of Psychology, but they are the core problems in artificial intelligence. In an operable and practicable level, we propose a concept acquisition and development method, which was inspired by Karmiloff-Smith's Representation Redescription (RR) supposition. They divide the cognition development into three phases and four levels of representation: Implicit (I), Explicit 1 (E1), Explicit 2 (E2) and Explicit 3 (E3). But in primitive RR how it realizes the representation and redescription was not mentioned, that is to say the RR only has basic and abstract sprit and no implementation consideration. Thus we should develop a systematic method to avoid its ambiguity, formalize its architecture, represent its learning results, arrange its learning processes, and match the concept's inner and static structure with its outer and dynamic applicability, and we propose using Object-Oriented (OO) theory to explain RR's basic sprit and use this new RR-OO model in the growth of concept a new mode and make it adept at task-diversity. In detail we firstly use RR to analyze the concept developing from low level to high level, and find that its representation should become more and more general and flexible. Thus we need a well-formed representation to solidify this gradual change of learning a concept both in its inner structure and its outer appearances. The OO theory has accurate definitions in class and data abstraction, encapsulation, visibility of object's attributes and behaviors, polymorphism, overloading, modularity and inheritance. This makes it to be a perfect tool to meet the requirements of Implicit and Explicit representation of RR. We apply an OO-similar frame to formalize and represent the developmental process of a concept, i.e.a concept system in one's mind can be simulated by an object system, because an object system is convenient to describe a real world, and its objects can be seen as element blocks of a knowledge system. In this paper an elementary cognition skill, counting, is taken as an example to demonstrate the development of a concept, and its evolution is materialized by the forms of objects in their attributes and behaviors. This new method is much more computable than classical RR, and more suited for us to build a base-wider semantic system for those domain-open applications of AI.