| Summary: | In the thesis presented here, variations of two very prominent machine learning techniques, the Neural Network (NN) and Support Vector Machine (SVM) are used in an attempt to solve two classification problems. Classification involves the assignment of an unknown object into a pre-determined group which consists of a set of preclassified objects with similar features to that unknown object. The main theme of the research conducted in this thesis involves investigation into existing and proposed classifier architectures to improve the classification performance for certain research problems. The aim of the research conducted is to develop new classifiers that are robust and able to show a high level of classification accuracy to the problems that are being considered. The problems being considered in this thesis are material surface classification and epilepsy seizure phase classification. The material surface classification problem involves the classification of a material based on its surface features which are obtained from a tactile-sensing robotic arm. Feature extraction is carried out on this input and the classifier is then used to classify based on the extracted feature inputs. Epileptic seizure is a common neurological disorder which causes the sudden discharge of cortical neurons in the brain. This results in the onset of seizures lasting from a few seconds to around a minute. The input consists of data obtained from the electroencephalograph (EEG) of patients who suffer from epilepsy. The input is then subjected to feature extraction and the extracted feature inputs are applied to the classifier. Four traditional classifiers, namely SVM, NN, k-nearest neighbour (kNN) and naive Bayes classifier are utilised for comparison purposes to evaluate the performance of the proposed classifiers during the research conducted. To evaluate the robustness property of the classifier, the original data is contaminated with Gaussian white noise at various levels. The results of the research carried out are presented in three parts: 1)The performance of six commonly used neural-network-based classifiers are investigated in solving the material surface classification problem. The significant contribution from the research conducted in this section is in the application of the neural network architectures to a novel problem (i.e material classification). The neural network architectures are also altered and re-structured in order to t the problem space. Experimental results show that the parallel-structured, tree-structured and naive Bayes classifier outperform the others based on the average classification accuracy of the classifier when under the original data. The tree-structured classifier demonstrates the best robustness property under the noisy data. 2) In continuation of the research conducted in the previous section, a novel neural network having variable weights is proposed to deal with the material classification problem. The aim of doing this is to compare its performance to the best out of the 6 neural network architectures applied in dealing with the material classification problem. The epilepsy seizure phase classification problem is also introduced with the proposed variable weight neural network being implemented to deal with this problem. It is shown that the variable weight neural network (VWNN) classifier outperforms the traditional methods in terms of classification accuracy and robustness property when the input data is contaminated with noise. 3) A novel Interval Type-2 Fuzzy Support Vector Machine (IT2FSVM) classifier has been proposed to deal with the epilepsy seizure phase classification problem. The performance of the classifier is measured based on its classification accuracy for each of the epilepsy phases. Three traditional classifiers (SVM, kNN and naive Bayes) are used for comparison purposes. The results obtained from simulations show that the novel IT2FSVM is able to show improved performance in terms of the average classification accuracy when compared to the other classifiers under the original dataset and also shows a high level of robustness when compared to other classifiers under a noisy dataset.
|
|---|
