FUZZY NEURAL NETWORK FOR SOLVING MULTI-CLASS CLASSIFICATION PROBLEM

• Main Author: Rofi, Syahrul
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39171
• Institution: Institut Teknologi Bandung

A construction of machine learning programs that can solve multi-class classification problems is studied in this final project. Fuzzy neural network, i.e. an artificial neural network whose feed-forward process is the stages of fuzzy logic control system that is constructed based on available input and output data, will be used as machine learning models. The stages are fuzzification, fuzzy inference, and defuzzification. The stage of fuzzy inference requires fuzzy rules and the results of fuzzification stage. Fuzzy rules consist of several interrelated fuzzy implications. Two phases are required to get the fuzzy neural network model: structure identification and parameter identification. The structure identification phase produces the initial structure of fuzzy rules which includes number of implications for fuzzy rules and the initial values of each parameter involved in the fuzzy neural network model. The parameter identification phase refines the values of the parameters involved in the model. To get the most optimal fuzzy neural network model, a cross-validation process is needed repeatedly. The cross validation process produces an average accuracy for a model. The model with a cross validation process that produces the highest average accuracy is the most optimal model. For this final project, three types of data are simulated, i.e.: cartesian coordinate data, iris plant data, and car evaluation data. Three schemes of multi-class classification are applied: one against all, one against one, and one against higher order. The fuzzy neural network model obtained with a one against all classification scheme has an accuracy of more than 95%. The fuzzy neural network model obtained with a one against one classification scheme has an accuracy of at least 94%. The fuzzy neural network model obtained with a one against higher order scheme has an accuracy of more than 84%.