Evolutionary Studies Of Genetic Neural Networks

This paper has presented a new method to combine GA and ANN. This method applied with the two-step-strategy can be used to train the weights and choose the driving function of every neuron simultaneously. A common single-input-single-output pH process of a CSTR reactor is simulated by a three-layer network, and in the process of evolution, GA is used to find the optimal assemble of Sigmoid function and RBF function at the first stage, and then adjust the weights and the parameters of each transfer function at the second stage. Simulation result has shown that this new method is simple and effective, which can enhance modeling precision of ANN greatly.The characteristics of most actual objects are asymmetrical. Aiming at that, this paper studied the improvement of the subject function of fuzzy neural network, suggested to use asymmetrical Gauss functions instead of common Gauss functions to improve the performance of fuzzy neural network, and introduced GA to train the connection weights and the parameters c, crl, 2. Simulation results showed that the fuzzy neural network with asymmetrical Gauss function had higher precision.The research showed that there are close relations between the nonlinear characteristics of simulation objects and the transfer functions of neural network, as well as the subject function of fuzzy neural network. Their selection, assemblage and improvement have much influence on the performance of network, and they are a very important aspect in the study of the theories of neural network.