Differential evolution and meta-learning for dynamic ensemble of neural network classifiers

Diversity is deemed to be a key issue in classifier combination. For this reason, not every classifier is an expert for every query pattern. Thus, many researchers have focused on dynamic ensemble selection. Most works, however, use only one criterion to perform the dynamic selection. Hence, multiple criteria can provide a decision more effective than the one produced by any of the criteria. Another important issue is accuracy of the classifiers, which strongly depends on the adequate choice of its parameters, including, for example, learning algorithm, structure and input feature vector. Therefore, we present a hybrid intelligent system to generate automatically a pool of classifiers, and choose dynamically an ensemble to predict each query pattern. The method evolves simultaneously the classifier parameters and trains, via a learning algorithm, the candidate solutions. Meta-features are extracted and used to build meta-classifiers to predict whether a base classifier is competent enough to classify the query pattern. Experimental results show that the proposed method improves classification accuracy when compared against current state-of-the-art techniques.