H. Zhang and M. Ishikawa (Japan)
ensemble learning, real-coded genetic algorithm, pruning, classification, generalization ability
We have proposed an ensemble method using hybrid real coded genetic algorithm with pruning (HRGA/P) for su perior generalization ability in classification. In order to further improve its performance, this paper proposes to use a novel hybrid real-coded genetic algorithm with prun ing (HRGA/Pr) instead of HRGA/P for estimating classi fiers. A crucial idea here is to replace the evaluation of the entire classifier by the original Rumelhart’s regularizer to that of each unit as an additive criterion term for reduc ing the complexity. It is intended for improving the gen eralization ability of the classifier with efficiently explor ing the simple structure of the classifier by execution of the additional criterion. Accordingly, the resulting classi fiers are expected to be structurally simple and have supe rior generalization ability in classification. Applications of the proposed method to an iris classification problem well demonstrate its effectiveness. Our experimental results in dicate that it has superior generalization ability for test data (classification rate: 98.3%) than the conventional algo rithms such as backpropagation (classification rate: 94.1%) and structural learning with forgetting (classification rate: 95.0%).
Important Links:
Go Back
