The RBF Neural Network Optimization And Application Based On Hybrid Hierarchy Genetic Algorithm

Radial basis function (RBF) neural network is a kind of three-layer feedforward neural network with single hidden layer, there is great difference between it's structure and learning algorithms with BP neural network's. RBF network is a structural simulation of local regularization and mutual overcast in human brain. With local approximation characterize, it can approximate arbitrary continuous function with arbitrary precision. Genetic algorithm is a kind of stochastic global parallel search algorithm, which simulates natural genetics mechanism and biologic evolution theoretic. It has strong robustness and global optimization performance. The RBF network configuration is formulated as a minimization problem with respect to the number of hidden layer nodes, the center locations and the connection weights. Although the objective functions are continuous and differentiable with respect to the center locations and the connection weights, they are discontinuous and non-differentiable with respect to the number of hidden layer nodes. This presents difficulties for most conventional optimization methods, which need derivative information of the objective function. Since genetic algorithms operate directly on the objective functions, this provides an alternative method for constructing RBF networks.In the present study, hybrid hierarchy genetic algorithms is introduced to configure the structure and parameters of RBF network, and the results are compared with which produced by orthogonal least squares learning (OLS) algorithm. Based on the research on the hidden layer structure and the parameter character of RBFnetwork, a new coding method of hidden layer location and width in RBF network is presented, that is, these two parameters are coded individually, thus, they can be optimized by genetic evolution in individual search space. The convergence rate and training-test precision are all improved. In the end of paper, the improved hybrid hierarchy genetic algorithm is applied to identification and forecast of two typical nonlinear systems; the simulation results show its validity.