Application Of RBF Network Based On Immune Secondary Response Pricinple To Dynamic Systems

The RBF Neural Network is a feed-forward neural network model with the capacity of local approximation, distributed processing and flexible training, so it can be used in complicated non-linear systems. However, present RBF training algorithms are applied mainly in static training, which are unsuitable for applications in dynamic tracing problems; furthermore, these algorithms have the disadvantage of slow convergence. The main work in this paper summarized as follows:Firstly, a novel RBF training algorithm based on immune secondary response is proposed. Two innovations are: 1. Inspired by immune principles, especially the immune memory mechanism and secondary response mechanism, immune library is constructed and immune operators are designed to make full use of related information during training. Immune operators improve the efficiency of RBF training, while immune library accelerates the training speed through rapid response to similar problems. 2. Two-step training based on different usage of different neuron width is proposed, which takes different intensive level of training to the stable part and changing part of the dynamic model in order to lower computational complexity and enhance convergence.Secondly, a CDMA multiuser detector based on immune secondary response RBF network is designed. The detector achieves better performance at a lower cost than other present multiuser detectors based on RBF training algorithms by balancing the training efficiency and computational complexity. Simulation demonstrates its better practicability and real-time performance in the dynamic CDMA multiuser detect problem.Finally, an OFDM channel equalizer based on immune secondary response RBF network is designed. Spectrum overlap of sub-channels is allowed in OFDM system, which has no interference to other channels. This method has the characteristics of high bandwidth efficiency and robust performance in multipath dispersive environment. The equalizer achieves better performance at a lower cost than other equalizer based on RBF training algorithms by balancing the training efficiency and computational complexity. Simulation demonstrates its better practicability and real-time performance in the dynamic OFDM channel equalization problem.