Optimizing FELM ensembles using GA-BIC
Wei Shiung Liew, Chu Kiong Loo, Takenori Obo · 2017
Performance of neural network learners is often dependent on initializing conditions. There is often a trade-off between processing speed and the generalization ability of the classifier. Using extreme learning machines, a two-stage evolutionary algorithm is proposed for feature selection and setting hidden layer size, as well as to select classifiers for ensembles. Bayesian Information Criterion was used for evaluating network generalization and complexity. The objective of the algorithm is to be able to initialize relatively simple classifiers with good generalization and subsequently construct a good ensemble of classifiers. Benchmark experiments showed improved ensembles in four of the eight tested datasets.