Candidate oversampling prefers two to tango
David Wallin, Conor Ryan, Raja Muhammad Atif Azad · 2011
Recent work has enhanced the Evolutionary Bayesian Classifier-based Optimization Algorithm (EBCOA) by oversampling the next generation and identifying promising solutions without actually evaluating their fitness values. In order to model the existing generation, that work considered two classes of solutions, that is, high performing solutions (H-Group) and poorly performing solutions (L-Group). In this study, we test the utility of using two classes instead of using a single class, as is the norm in standard Estimation of Distribution Algorithms (EDAs). Our results show that a dual class model is preferable when oversampling is used.