Hybrid Ensemble Polynomial Neural Network Classifier: Analysis and Design

In this paper, we propose a hybrid ensemble polynomial neural network (HEPNN) with the aid of polynomial neural network (PNN) and hybrid ensemble polynomials neurons (HEPNs). Two types of HEPNs including ensemble radial-based-function polynomial neuron (ERPN) and ensemble polynomial neuron (EPN) are proposed. ERPN and EPN are generalized polynomial neurons based on ensemble architecture. To address the problem of multiple covariance in traditional PNN neural networks, correlation coefficients and performance are utilized to select neurons replacing the original selection of nodes by performance only. The main strategies of HEPNN design are as follows: First, the first layer of the network consists of ERPN that are utilized to reflect the structure encountered between the data, while the second and higher layers consists of EPN, which reflect higher polynomial-order relationships between input and output data. Second, particle swarm optimization (PSO) is adopted to optimize the architecture of HEPNN. A comparative study shows the proposed HEPNN has better performance than other state-of-art models reported in literature.