A Self-Organizing Fuzzy Polynomial Neural Network - Multistage Classifier

A fuzzy polynomial neural network multistage classifier (FPNN-MC) is suggested in this paper, suitable for handling complex classification problems with large feature spaces. The multilayered FPNN-MC structure is developed in a self-organizing way, using a structure learning procedure. The network´s neurons are realized through fuzzy rule-based TSK systems, considered as generic fuzzy neuron classifiers (FNC´s). Parent FNC´s are combined to develop new higher-level descendant classifiers at the subsequent layer. Hence, sequential multistage decision is implemented, leading to improved classification results. To exploit the information acquired by FNC´s at each layer and achieve an effective data flow, a fusion scheme is developed associated with a data reduction mechanism. Upon termination of the structure building, parameter learning is carried out using a genetic algorithm platform. A remarkable asset of the approach is that it resolves the feature selection task, providing the most relevant features of a problem. Simulation results on a well known classification problem indicate the efficiency of the proposed model