Incremental Adjustment of Representations for Learning

To remain effective without human interaction, intelligent systems must adapt to their environment. A useful form of adaptation is to incrementally form concepts from examples for the purposes of inference and problem-solving. A number of systems have been constructed for this task, yet their capability is limited by the representational decisions made when the system was designed. This paper presents a concept acquisition system that is able to form conjunctive, disjunctive, and negated concept descriptions from potentially noisy, discrete and real-valued attributes. A central focus of this paper is the interaction of a triad of learning methods and their ability to minimize the impact of initial representational decisions. After describing the methods, an empirical section illustrates their interaction.