Feature extraction using coevolutionary genetic programming

We propose a new feature extraction method using evolutionary computations. The extracted features are defined as polynomial expressions, which are composed of the original input pattern. These polynomial expressions are searched by coevolutionary genetic programming. We introduce a new fitness function based on competition between individuals. Experiments are performed for some databases of UCI repository using the proposed method and kappa-Nearest Neighbor rule. Experimental results show that the proposed method can preserve the diversity of populations and improve recognition accuracy on most databases.