The Study Of General Regression Neural Network And Genetic Algorithms And Their Application For Chemical Engineering

Accurate models are important to the research and application of chemical engineering process. However, most problems in chemical engineering process's internal mechanisms are too complex to build accurate models directly by the principles. So in many complex practical produce processes, engineers prefer to build models based on the observation data. The most important function of these models is to fit the produce data as accurately as possible. Neural networks build models without the mechanisms, it modeling chemical engineering process by sample data. This dissertation focuses on the network architecture of the general regression neural network (GRNN) and the optimization of the GRNN's network parameter. This thesis includes the following parts mainly:(1) The GRNN automatically extracts the appropriate regression model (linear or nonlinear) from the data, and the training of the GRNN is optimization of the smoothing factors. This work addresses several improvements of GRNN. A smoothing factor was associating with each independent feature. The corresponding vector of smoothing factors (size of the vector is equal to the number of independent features) was optimized by the conjugate gradient method and the Powell method, and the GRNN-CGrad model and the GRNN-Powell model were built. As the prediction importance of each independent feature is separately taken into account, the performance of these two models was improved observably.(2) As the architecture of GRNN is similar to the radial basis function networks (RBFN), this work studied on RBFN too. The partial least square regression (PLSR) methods was applied to determine the weight of the RBFN, the RBF-PLSR model was built. The PLSR picks-up the orthogonal components from the primary independent variable data matrix, and neglects the components with very little variance. PLSR eliminates the multicollinearity between the primary independent variables and ensure the regression process more robust.(3) A eugenic evolution strategy was proposed to improve the efficiency of the conventional simple genetic algorithm (SGA) searching. The Eugenic evolution genetic algorithm (EGA) collects the .population information along the evolution of children generations and constructs a deterministic optimization algorithm, which will be embedded in the evolution process at appropriate stage to speed up the local searching. For the possible deterministic searching methods, the Powell method was found to be feasible in integrating with the genetic algorithm. Besides, proposing an adaptive variation factor to keep the diversity of population, and a novel crossover rule to widen the distribution space of children generations also effectively modified the SGA. Atypical example indicated the good performance of the proposed method. Finally, the EGA was applied successfully to the nonlinear parameter estimation of the GRNN in modeling for Delay cooking process, and the GRNN-EGA model was built.(4) An intelligent agent technology was proposed to implement the GA, and the MAgent-GA model was built. The model can acquire the environment information from the evolution procedure, then manages and guides the GA to proceed and adjusts coefficients dynamically in order to find out the global optimum value quickly and efficiently. The simulation experiments indicate it improves the characteristic of the GA a lot and is better than the SGA. The MAgent-GA was applied successfully to the training of the GRNN in modeling for Delay cooking process, and the GRNN-MAgent-GA model was built.In the end of this dissertation, we make a summary and describe the further works.