Development and performance assessment of a novel hybrid fuzzy neural network structure

In this study, as a novel network structure, Hybrid Radial Based Function Neural Network (HybRbfNN) is developed and its performance is investigated. The network has a total of 4 layers of network structure, with 3 hidden layers and 1 output layer. The network has been developed specifically for modelling systems with uneven surfaces and its modelling performance has been compared with the most significant competitors, adaptive-network based fuzzy inference systems (ANFIS) and conic section function neural network (CSFNN). The comparison is made on modelling of a benchmark system having an uneven surface. The parameters of the networks are trained using improved particle swarm optimization (iPSO). The modeling efficiencies of HybRbfNN, ANFIS and CSFNN are investigated using their training measurement graphs, surfaces they modelled, and error surfaces. The obtained results show that the HybRbfNN network developed in this study, especially on uneven surfaces and sharp transitions, shows a better learning outcome than ANFIS and CSFNN networks and a superior modelling performance over the result.