Christoph Globig

On case-based learnability of languages

Case-based reasoning is deemed an important technology to alleviate the bottleneck of knowledge acquisition in Artificial Intelligence (AI). In case-based reasoning, knowledge is represented in the form of particular cases with an appropriate similarity measure rather than any form of rules. The case-based reasoning paradigm adopts the view that an Al system is dynamically changing during its life-cycle which immediately leads to learning considerations.Within the present paper, we investigate the problem of case-based learning of indexable classes of formal languages. Prior to learning considerations, we study the problem of case-based representability and show that every indexable class is case-based representable with respect to a fixed similarity measure. Next, we investigate several models of case-based learning and systematically analyze their strengths as well as their limitations. Finally, the general approach to case-based learnability of indexable classes of formal languages is prototypically applied to so-called containmet decision lists, since they seem particularly tailored to case-based knowledge processing.

Case-based and symbolic classification

Contrary to symbolic learning approaches, that represent a learned concept explicitly, case-based approaches describe concepts implicitly by a pair (CB,sim), i.e. by a measure of similarity sim and a set CB of cases. This poses the question if there are any differences concerning the learning power of the two approaches. In this article we will study the relationship between the case base, the measure of similarity, and the target concept of the learning process. To do so, we transform a simple symbolic learning algorithm (the version space algorithm) into an equivalent case-based variant. The achieved results strengthen the hypothesis of the equivalence of the learning power of symbolic and casebased methods and show the interdependency between the measure used by a case-based algorithm and the target concept.

Symbolic Learning and Nearest-Neighbor Classification

The Nearest-Neighbor Classification has a long tradition in the area of pattern recognition while knowledge-based systems apply mainly symbolic learning algorithms. There is a strong relationship between Nearest-Neighbor Classification and learning. The increasing number of cases and the adaptation of the similarity measure are used to improve the classification ability. Nowadays, Nearest-Neighbor Classification is applied in knowledge-based systems by a technique called case-based reasoning. In this paper we present first results from a comparison of case-based and symbolic learning systems.

On Case-Based Representability and Learnability of Languages

Within the present paper we investigate case-based representability as well as case-based learnability of indexed families of uniformly recursive languages. Since we are mainly interested in case-based learning with respect to an arbitrary fixed similarity measure, case-based learnability of an indexed family requires its represent ability, first.

Learning in Case-Based Classification Algorithms

While symbolic learning approaches encode the knowledge provided by the presentation of the cases explicitly into a symbolic representation of the concept, e.g. formulas, rules, or decision trees, case-based approaches describe learned concepts implicitly by a pair (CB, d), i.e. by a set CB of cases and a distance measure d. Given the same information, symbolic as well as the case-based approach compute a classification when a new case is presented. This poses the question if there are any differences concerning the learning power of the two approaches. In this work we will study the relationship between the case base, the measure of distance, and the target concept of the learning process. To do so, we transform a simple symbolic learning algorithm (the version space algorithm) into an equivalent case-based variant. The achieved results strengthen the conjecture of the equivalence of the learning power of symbolic and case-based methods and show the interdependency between the measure used by a case-based algorithm and the target concept.