Active pattern recognition using genetic programming

The need for faster and robust methods for pattern recognition and data mining is ever increasing. Classical machine learning algorithms have always been used in a variety of domains like optical character recognition (OCR), speech recognition and information extraction.; Different levels of informative detail can be present in different regions of a pattern image. Classifiers which selectively use features corresponding to discriminating regions in making decisions for particular classes are called active classifiers. Design of active classifiers requires the pattern recognition technique to blend feature discovery within the classifier training phase. This dual task of feature discovery and classifier training can be combined to make the learning algorithm adaptive.; This dissertation titled ‘Active Pattern Recognition using Genetic Programming’ highlights the need for applications to be adaptive. Traditional machine learning algorithms for classification can be made dynamic in terms of feature selection, computational resource utilization and scalability. This dissertation describes how to make one such algorithm (Genetic Programming) active, scalable and recurrent. The proposed extensions are used to develop classifiers for handwritten digit recognition. Genetic programming (GP) is a biologically motivated machine learning technique like genetic algorithms (GA). The essential idea is to represent states (classification models in our case) as chromosomes (encoded as expression trees) and to evolve a population of new offspring trees by selectively pairing parent trees.; We first illustrate how GP based active classifiers are developed for handwritten digit recognition. A two-stage classification method motivated by pair-wise confusion between digits is then explored. Inspired by the performance for off-line hand written digit classification, a strategy to classify on-line handwritten digits based on off-line features and GP is developed. We then present a recurrent-GP framework which extends the proposed active pattern recognition paradigm for applications where the length of the feature vector is dynamic. One of the key deterrents in using evolutionary computation techniques for complex real-world applications in pattern recognition and data mining is their non-scalable nature in terms of computational requirements. We have designed a new ‘Efficient-GP’ technique to address these issues. The dissertation concludes by discussing the role of this paradigm in computational machine learning theory.