Study on The Property of Training Samples and Learning Space with Genetic Algorithms

Historically, the empirical risk of a pattern classifier was asked to be made zero, therefor the default property of training samples were limited to a separable ones. Nowadays on the contrary, the major idea of learning classification no longer ask the empirical risk of classifier must be made zero. In this situation, inseparable feature sets may not be detrimental to the performance of classifier. However, so far no experimental studies and analytical results show whether an inseparable feature set is available or not. This paper firstly analyzes the interaction between learning algorithms and feature selection, and gives a proof by both the analytical analysis and experimental studies.