A Constrained Genetic Algorithm for Efficient Dimensionality Reduction for Pattern Classification

In automated pattern recognition systems, the two main challenges are feature selection and extraction. The fea- tures selected directly affects the number of measurements required; and extracting low-dimensional features from the selected ones reduces the computational complexity of the classifier. In traditional approaches, human expertise is obligatory for feature selection and statistical techniques are employed for feature projection. In this paper, a con- strained genetic algorithm for performing these two tasks simultaneously, in conjunction with the k-nearest neighbor classifier is proposed. This algorithm requires minimal hu- man intervention as it realizes good tradeoff solutions be- tween classification accuracy, feature measurement require- ments, and computational complexity.