Feature Selection for Classification Based on Sequential Data

We consider the problem of selecting features from a sequence of transient waveforms, with the goal of improved classification performance. For the example studied here, the waveforms are representative of multi-aspect acoustic scattering from an underwater elastic target. The feature selection is performed via a traditional genetic algorithm (GA), with the principal focus on definition of an appropriate cost function for sequential data. We consider cost functions based on informationtheoretic measures, while separately also considering a sequential classifier as an integral component of the cost function. For the latter we consider a hidden Markov model (HMM) classifier, with this also utilized subsequently to assess the performance of the GA-selected features.