Enhancing the prediction accuracy and the reliability of a genetics-based inductive learning environment
Christoph F. Eich, Yeong‐Joon Kim · 1996
Inductive learning is one of the most intensively studied learning paradigms in machine learning research. It is a process of acquiring knowledge by drawing inductive inferences from facts provided by a teacher or the environment. Learning of class definitions from examples is one of the primary tasks in inductive learning. In this task, for a given training set of classified examples drawn from a certain domain, the goal of inductive learning algorithms is to produce a classification algorithm that will correctly classify new examples drawn from the same domain. Over the last three years, we have worked on developing a genetics-based inductive learning environment, called DELVAUX, that learns Bayesian classification rules from given examples using genetic algorithms. The dissertation centers on developing methodologies, concepts, and techniques to deal with the various problems DELVAUX faces in the rule learning process and to enhance the prediction accuracy and the reliability of DELVAUX. We explore the usage of constructive induction in the context of the DELVAUX learning environment and we also develop a meta-learning approach, a hierarchical two-phase learning approach, and an N-version programming approach to enhance the prediction accuracy and the reliability of DELVAUX, Several genetics-based learning environments for the N-version programming approach. The learning performance of these various approaches is evaluated empirically and DELVAUX is compared with two other inductive learning environments: C4.5 and a neural network approach.