Research On Blind Equalization Algorithm Based On Feed-forward Neural Network Optimized By Genetic Algorithm

In modern communication systems, transmission channel is greatly distorted because of the exist of factors such as channel-fading, multi-path propagation and side-channel interference, and thus inter-symbol interference (ISI) emerges, which severely reduces the performance of transmission channel. The adoption of equalization technique greatly suppresses inter-symbol interference and improves the performance of the system. The blind equalization is a kind of self-adaptive equalization method, which can realize channel convergence relies solely on the prior-information of received output sequence. It can effectively compensate the non-ideal characters of transmission channel and reduce error rate, and subsequently improves communication performance. Recently it has become a new hot pursuit in relevant field.In modern transmission environment, since there isn't any completely linear channel, the algorithms aiming at solving linear channel processing become invalid, so blind equalization technology that can overcome the problem draws more and more attention. As we all know, multi-layer neural network can form a static network, which can map static inputs into static outputs, so nowadays more and more attention is paid to blind equalization algorithm based on artificial neural network.Conventional blind equalization algorithms based on Feed-forward neural networks obtain some achievements to some degree, for example, the algorithms can be used in any kind of channels, no matter linear or non-linear ones, they also overcome influences caused by uncertainty of channel orders, and besides that, they own notable fault tolerance performance against additive noise. However, because of inherent deficiency of neural network, they are only good at local searching, when confronting with problems with more than one pole; the algorithms incline to trap in local optimum. Genetic algorithm is a random searching method with global and parallel characteristics, and it owns powerful robust and global convergent properties, after summarizing and analyzing disadvantages and limitations of traditional algorithms, the paper adopts Genetic Algorithm in neural network blind equalization by optimizing the structures and weights of these networks, the computer simulations show that the improved algorithm obtains good convergence performance and has super equalization effects.The main works of this paper can be summarized as follows:(1) The paper analyzes disadvantages and limitations of existing neural network blind equalizations, and it presents new thoughts and methods in optimizing of neural network structures and its weights. And then the paper determines bonding points between GA and neural network blind equalization.(2) In order to overcome shortcomings of conventional Genetic Algorithm, the paper proposes a new GA which can maintains population diversity, and then applies the algorithm in optimizing weights and structures of Feed-forward neural network respectively, simulation results prove that comparing to traditional neural network blind equalization algorithm, the proposed algorithms obtain great improvements in convergence speed and residual error.(3) Based on analyzing limitations of previous structure optimization algorithm of neural network, in order to overcome its shortcomings such as complex operation, long searching time and large memory, the paper adopts elitism-based compact genetic algorithm in optimizing structure of neural network, simulation results show this neural network blind equalization algorithm achieves faster convergence speed and smaller residual error in linear channels and nonlinear ones.