Abstract

We propose and motivate the use of vicinal-risk minimization (VRM) for training genetic programming classifiers. We demonstrate that VRM has a number of attractive properties and demonstrate that it has a better correlation with
generalization error compared to empirical risk minimization (ERM) so is more likely to lead to better generalization performance, in general. From the results of statistical tests over a range of real and synthetic datasets, we further demonstrate that VRM yields consistently superior generalization errors compared to conventional ERM.