A pruning algorithm for training neural network ensembles

This paper presents a pruning algorithm i.e., dynamic ensemble pruning algorithm (DEPA) by utilizing the knowledge of overfilling and importance of hidden node. The generalization performance of a machine learner depends on how much it avoids the overfilling. The main idea of this algorithm is to reduce the complexity of ensemble networks according to overfilling on progress toward training. DEPA emphasizes on avoiding "overfilling" by dynamically deleting individual neural networks and their hidden nodes starting from a large number of individual neural networks. DEPA has been tested on several standard benchmark problems in machine learning and neural networks, including breast cancer, diabetes and heart disease problems. The experimental results show that DEPA can produce neural network ensembles with good generalization ability.