Novel HME Model Based On Discriminant Analysis And Multilayer Quadratic Perceptron

Modeling is an important method to analyze systems and solve problems. For single modular NNs, it is too hard to achieve satisfying performance for a complication system that exhibit strong nonlinear behavior. So a novel HME (NHME) model based on committee machine structure with a good performance was presented. The model is a modular network with a fast training speed and a high precision.Child modules influence on NHME performance directly. A NHME is composed of Multilayer Quadratic Perceptron (MLQP) experts adopted first and improved modular gating networks with a few parameters and a plain structure. A NHME has a better performance and the difficulty of training is reduced because of MLQP experts adopted and new gating networks improved. The typical pH-mixture tank system with badly nonlinear character had been applied as the simulated models for NHME network. The results of simulative experiments indicated that NHME can model for static and dynamic systems processes well.On realizing and using NHME function, cross sorts arithmetic of dividing input-space and layer training were introduced. Model error reduced because of minus probability and its mathematical modeling presented. The adaptive study rate was used in BP arithmetic and the rule was proposed. The paper also elementarily studied Weights initialization methods and found the seemly criterion. The results of experiments on NHME with pH-mixture tank validated the efficiency of new methods.The NHME model based on new methods exhibit many merits. It has a faster training speed, high modeling precision, good generalization ability, a simple network structure and the satisfying performance. Experiments proved that NHME is a good model that adapts to solve increasingly difficult problems with complicated circumstance information, fuzzy background knowledge and ambiguous reasoning rules. A NHME model is an available and efficient modeling method for the realistic scabrous systems processes.