Pattern recognition using genetic programming for classification of diabetes and modulation data

The field of science whose goal is to assign each input object to one of the given set of categories is called pattern recognition. A standard pattern recognition system can be divided into two main components, feature extraction and pattern classification. During the process of feature extraction, the information relevant to the problem is extracted from raw data, prepared as features and passed to a classifier for assignment of a label. Generally, the extracted feature vector has fairly large number of dimensions, from the order of hundreds to thousands, increasing the computational complexity significantly. Feature generation is introduced to handle this problem which filters out the unwanted features. The functionality of feature generation has become very important in modern pattern recognition systems as it not only reduces the dimensions of the data but also increases the classification accuracy. A genetic programming (GP) based framework has been utilised in this thesis for feature generation. GP is a process based on the biological evolution of features in which combination of original features are evolved. The stronger features propagate in this evolution while weaker features are discarded. The process of evolution is optimised in a way to improve the discriminatory power of features in every new generation. The final features generated have more discriminatory power than the original features, making the job of classifier easier. One of the main problems in GP is a tendency towards suboptimal-convergence. In this thesis, the response of features for each input instance which gives insight into strengths and weaknesses of features is used to avoid suboptimal-convergence. The strengths and weaknesses are utilised to find the right partners during crossover operation which not only helps to avoid suboptimal-convergence but also makes the evolution more effective. In order to thoroughly examine the capabilities of GP for feature generation and to cover different scenarios, different combinations of GP are designed. Each combination of GP differs in the way, the capability of the features to solve the problem (the fitness function) is evaluated. In this research Fisher criterion, Support Vector Machine and Artificial Neural Network have been used to evaluate the fitness function for binary classification problems while K-nearest neighbour classifier has been used for fitness evaluation of multi-class classification problems. Two Real world classification problems (diabetes detection and modulation classification) are used to evaluate the performance of GP for feature generation. These two problems belong to two different categories; diabetes detection is a binary classification problem while modulation classification is a multi-class classification problem. The application of GP for both the problems helps to evaluate the performance of GP for both categories. A series of experiments are conducted to evaluate and compare the results obtained using GP. The results demonstrate the superiority of GP generated features compared to features generated by conventional methods.