The Application Of Growing And Pruning RBF(GAP-RBF) In Mobile Robots

In recent years, neural network is widely used in control system, but the performance of control system depends largely on the approximation properties of identification network. And the structure design is the key of radial basis function neural network. The network structure design refers to the selection of the number and center of hidden units, which have great influence on the convergence speed of the network and the identification accuracy as well as the generalization ability. Therefore, aiming at the problem which is the selection of the number and center of hidden units, this paper makes a further study and introduces three kinds of sequential, local, incremental algorithm such as RAN, MRAN, GAP-RBF in detail. Through the comparison with back propagation(Back Propagation, BP) algorithm in the approximation of Hermite polynomial and the prediction of Mackey-Glass time series as well as the application of robot navigation along the wall, it is proved that the three kinds of sequential, local, incremental algorithm mentioned above are far better than the batch learning algorithm. The specific work which needs to complete in this paper is as follows:Firstly, we introduce the significance and background of research on sequential, local, incremental algorithm in detail. Then several kinds of optimization algorithms of the structure of RBF neural network are compared, and the advantages and disadvantages between the batch learning algorithm and sequential, local, incremental algorithm are also compared. And then we describe the research status of several commonly used sequential, local, incremental algorithms briefly.Secondly, we summarize the neuron model of neural network and two kinds of neural network structure, which leads to a detailed introduction to the RBF neural network. Then we describe the RAN, MRAN and GAP-RBF specifically, and present the implementation steps of the three algorithms. Through the comparison with BP neural network in the approximation of Hermite polynomial and the prediction of Mackey-Glass time series, we verify the excellent performance of the three algorithms in nonlinear function approximation.Finally, we briefly introduce the Khepera II robot platform and design a RBF neural network controller. Then the RBF neural network controller is applied to the Khepera II robot platform, and the robot walk along the wall successfully in the simulation environment and real time environment. And then through the analysis of the data, the feasibility and performance of RAN, MRAN and GAP-RBF is verified in robot application.