K-coreclassic Mistake-bounded- Identiable Et Al

C-CLASSIC is a variant of CLASSIC in which concept learning is theoretically tractable (following the PAC-learning settings). C-CLASSIC is an extension of C-CLASSIC with the connec-tives (default) and (exception) allowing to incorpore default knowledge within concept denitions. Previous work was concerned with both deductive aspects (semantics, subsumption) and in-ductive aspects (PAC-learnability) of C-CLASSIC. Our purpose here is to discuss practical aspects of concept learning from instances within C-CLASSIC. We present an algorithm learning a disjunc-tive denition of a target concept from positive and negative instances. Since the bottom-up search of the concept space relies on Least Common Subsumers computations , the way how instances are represented in C-CLASSIC is crucial. In this paper we propose the use of a domain theory divided into default rules and incoher-ence rules in order to extend the descriptions of instances with excepted properties such as. The presence of in the description of an instance means the instance should have the A propery but has not. One of the central problems studied in Machine Learning is the task of inducing a denition of a concept from a set of positive and negative instances of this concept. The choice of an appropriate representation formalism is very important for learning since it should be expressive, useful, applicable and ecient. For these reasons , Description Logics have been receiving increased attention in the Machine learning community (e.g. 1994a], many results about learnability of DLs are given. In [Cohen and Hirsh, 1994a], the authors describe which is and consequently PAC-Learnable (i.e. learning is ecient in the Valiant's sense of PAC-learnablity [Valiant, 1984]). It has been shown in [Cohen and Hirsh, 1994b] that CLASSIC is not PAC-learnable but that C-CLASSIC is. In [Ventos, 1996], we presented C-CLASSIC which extends C-CLASSIC with two non classical con-nectives and introduced in the toy DL [Coupey and Fouqueré, 1997]). C-CLASSIC makes it possible to express default and excepted properties in the denition of concepts while keeping concept classication monotonic and polynomial. In [Ventos , 1997], we have proven that C-CLASSIC is PAC-learnable. The goal of this paper is twofold. On one hand, we describe how it is possible to learn in C-CLASSIC. On the other hand, we highlight the advantages of the connectives and by comparing learning in C-CLASSIC and learning in C-CLASSIC. Learning of concept denitions, from positive and negative instances, in C-CLASSIC requires four steps. In the rst two steps, the descriptions …