The Research Of Multiuser Detection Algorithm Based On Radial Basis Function Neural Network

Code Division Multiple Access communications system is a interference-limited system, multi-access interference and near-far effect exist because spreading codes between users are not completely orthogonal along with different users random accessing. Multiuser detection is the key anti-interference technology, it can make CDMA communications system have good anti-interference ability, solve the near-far effect problem, lower the requirement for power control and notably enhance the capacity of the system. Multiuser detection is a combinatorial optimization question in nature. However, Radial Basis Function neural network itself has a strong ability of function approximation, the simple network structure, quick and easy method of training, which is an effective method for solving such problems. It combines the two together, hoping to obtain better detection performance. It has been a hotspot of recent research. The major works of this paper are summarized as follows:(1) This paper summarizes the development direction and the current research present situation of multiuser detection technology and classifies multiuser detection technology. Besides these, it compares several kinds of typical multiuser detection technology through the simulation experiments and lays the foundation for the performance comparison of the latter algorithm.(2) This paper introduces the basic principle of RBF network, analyzes and compares the advantages and disadvantages of several commonly used learning algorithms and discusses the principle and the system structure of the multiuser detection technology based on RBF neural network.(3) This paper analyzes effects of learning rate and hidden layer's node number of the RBF multiuser detection technology trained by gradient descent method on arithmetic performance. In view of these above problems, it separately introduces one kind of gradient descent method which has variable learning rate and nearest neighbor clustering algorithm, constitutes a hybrid learning algorithm to train RBF neural network, and applies it to the multiuser detection. The simulation results indicate that the computation speeds up and performance of the RBF neural network multiuser detection trained by hybrid learning algorithm is better than the RBF neural network multiuser detection trained by the traditional algorithm and OLS algorithm.(4 ) This paper proposes one kind of hybrid hierarchy genetic algorithm to train the RBF neural network's structure and parameters simultaneously, introduces the improved chromosome code scheme, uses the least squares method based on singular value decomposition to compute weights of network's output layer, enhances the efficiency of genetic search and simplifies the structure of the network. Finally it uses variable learning rate gradient descent method to optimize optimum network which is trained by genetic algorithm, then applied in the multiuser detection. The simulation indicates that structure of the network trained by the new hybrid learning algorithm is simpler than the network trained by the other algorithms and achieves good performance.