An Argumentation Framework for Concept Learning

Concept learning is an important problem in AI that consists of, given a set of training examples and counter-examples on a particular concept, identifying a model that is coherent with the training examples, i.e that classifies them correctly. The obtained model is intended to be reused for the purpose of classifying new examples. Version space is one of the formal frameworks developed for that purpose. It takes as input a consistent set of training examples on the concept to learn, a set of possible models, called hypothesis ordered by generality, and returns the hypothesis that are coherent with the training examples. The returned set of hypothesis is called version space and is described by its lower and upper bounds. This paper provides an argumentation-based framework that captures the results of the version space approach. The basic idea is to construct arguments in favor of/against each hypothesis and training example, to evaluate those arguments and to determine among the conflicting arguments the acceptable ones. We will show that the acceptable arguments characterize the version space as well as its lower and upper bounds. Moreover, we will show that an argumentationbased approach for learning offers an additional advantage by allowing the handling of common problems in classical concept learning. Indeed, it is possible to reason directly with sets of hypothesis rather than one, and to deal with inconsistent sets of training examples. Lastly, the framework translates the problem of classifying examples into a decision one.