Investigating the use of an ensemble of evolutionary algorithms for letter identification in tremulous medieval handwriting

Ensemble classifiers are known for performing good generalization from simpler and less accurate classifiers. Ensembles have the ability to use the variety in classification patterns of the smaller classifiers in order to make better predictions. However, to create an ensemble it is necessary to determine how the component classifiers should be combined to generate the final predictions. One way to do this is to search different combinations of classifiers with evolutionary algorithms, which are largely employed when the objective is to find a structure that serves for some purpose. In this work, an investigation is carried about the use of ensembles obtained via evolutionary algorithm for identifying individual letters in tremulous medieval writing and to differentiate between scribes. The aim of this research is to use this process as the first step towards classifying the tremor type with more accuracy. The ensembles are obtained through evolutionary search of trees that aggregate the output of base classifiers, which are neural networks trained prior to the ensemble search. The misclassification patterns of the base classifiers are analysed in order to determine how much better an ensemble of those classifiers can be than its components. The best ensembles have their misclassification patterns compared to those of their component classifiers. The results obtained suggest interesting methods for letter (up to 96% accuracy) and user classification (up to 88% accuracy) in an offline scenario.