The research on Artificial Neural Network (ANN) is one of the mainstreams of the intelligent computing development.The information processing mechanism of the neu-ral network has matured, thus it has been successfully applied in various fields,and has become a major intelligent information technology. The rapid development of wireless communication system requires the more advanced intelligent processing technology be-ing matched with it, and recently ANN has been widely used in communication system to deal with a variety of optimization problems. However, the problems faced in com-munication system are closely resolved in complex-valued domain, such as channel e-qualization, channel estimation and so on, which require us to further study the complex-valued neural network (CVNN). Applications of Real-valued neural network (RVNN) and CVNN in communication fields are studied in this paper, and the specific contents are as follows.1. MIMO OFDM is one of the key technologies of the next generation of wireless communication, and how to make better use of the resource allocation for the OFDM system has become the focus of the communication system research. An adaptive re-source allocation scheme based on Hopfield Neural Network (HNN) is proposed in this paper satisfying the usersâ€™requirement, the transmission rate and its performance. First-ly, the resource allocation is converted into a combinational optimization problem, and then achieve the ultimate goal through constructing HNN model, energy functions and dynamic equations. On the basis of research for the resource allocation, the interference alignment (IA) clustering for the interference network based on HNN is studied contin-ually, in which the interference network is looked as a graph, each user as a vertex, the interference between users as the connection between the vertices. Making use of the set partitioning, the IA clustering problem can be transformed into an optimization prob-lem maximizing the sum of the intra-clustering weights. Thus, according to the HNN optimization theory, the clustering problem can be solved by mapping HNN model, con-structing HNN energy functions and dynamic equations. The clustering problem further states that the HNN can directly solve the Non-linear Programme (NLP) problem or non-convex optimization problem. Comparing with the traditional exhaustive method, HNN optimization method has faster convergence, better performance and greater stability.2. Due to the time-variant of the wireless channel, there are always time-delay errors between the obtained CSI of the transmitter and the actual channel, a channel tracking and prediction algorithm based on CVNN is proposed to compensate the time delay. Two-stage design is used in the proposed algorithm, which are the channel tracking and channel prediction. The channel parameters can be obtained in the first stage, and then the parameters are passed to the second stage to achieve the final channel prediction. The feasibility of the algorithm is demonstrated through the application in the communication system. Comparing with the Kalman Filter, the proposed algorithm has less prediction error, and for the high-speed parallel performance of the CVNN, the running speed is increased, the computational complexity is reduced and real-time processing of time-varying channel can be achieved.3. When the outputs of the network are close to the extreme values (0 or 1), the outputs tend to fall into local optima or optimization failing. To avoid the above, an error-modified CVNN is proposed by introducing a logarithmic function. When the outputs are closed to the extreme values, for eliminating the formulation of the difference be-tween 1 and the output, the back propagation (BP) of the error can be directly spread to avoid the above disadvantages. Combining the error-modified CVNN and KF method, a new estimation method is proposed. Firstly, the network tracking model is constructed, then using KF method the complex-valued connection weights are estimated, thus we can get the optimal estimated values of the connection weights, finally the estimated weight-s is transferred to the CVNN model to carry out the prediction of the complex-valued function. The simulation results show that the proposed algorithm is effective to find the optimal estimates. Most of all the proposed algorithm can improve the stability of the network. Combining the advantages of the wavelet transform a new complex-valued wavelet neural network (CVWNN) is introduced, which can process large data in parallel but also has a strong learning ability, fault tolerance and nonlinear approximation ability. The reliability of the algorithm is further verified by the solutions of the complex-valued function approximation and XOR problem.4. Recurrent Neural Network (RNN) which has feedback features can obtain dy-namic characteristics of the system accordingly. Being different from traditional feed-forward neural network (FNN), RNN can do well with the overall logic order of the high-dimensional information just introducing directional cycle, without constructing it between the levels. The RNN proposed in this paper is constructed on the basis of the KKT conditions by introducing integrators to avoid the inadequate of traditional gradient descent algorithm, thus the new RNN network is easy to implement. Based on semi-definite relaxation (SDR) technology in the MIMO detection, the non-convex quadratic constrained quadratic programming problem can be solved. |