Research On Key Technologies Of Wireless Communication Physical Layer Based On Deep Learning

Deep learning and machine learning are key technologies in future intelligent wireless communications.A typical wireless communication system is model-driven,and its corre-sponding system design is derived from domain knowledge.When the mathematical model is inaccurate or difficult to be established,and the optimization problem is inherently nonconvex or non-deterministic polynomial hard(NP hard),the model-driven method usually sacrifices the optimality for the tractability.On the other hand,the data-driven method is applicable for end-to-end system design,and it learns to solve an optimization problem through data train-ing.Nowadays,deep learning has achieved some breakthroughs in the wireless communication field,and also has great potential for further investigation.In addition,the rapid development of distributed storage and large-scale parallel computation hardwares,guarantees the execution of deep learning algorithm to be fast and efficient.We consider several key problems of wireless communication physical layer which include:·We consider an ultra narrow band communication system using m-ary phase position shift keying modulation and infinite impulse response filter.At the receiver,the joint op-timization problem of matched filtering,equalization and demodulation is difficult to be solved by a model-driven method.Therefore,we propose an end-to-end receiver based on neural network model,i.e.,decision feedback receiver(DFR).DFR iteratively detect the transmitted symbols,and utilizes the soft information from the last dection as feedback.In addition,DFR has low computation complexity and model complexity,and parallel detects the symbols frame-by-frame.·We consider the mismatch problem between the training set and the test set.We establish a mathematical model and quantitatively analyze the caused test error,and then the upper and lower bounds of the test error are given.The analysis indicates that the test error is caused by two issues:data divergence and data missing.We extend mismatch problem to the case where a single model is trained on an integral training set.With the assumption of sufficient model complexity,we prove that the data divergence no longer exists in deep unfolding with parameter sharing;when the model complexity is insufficient,we model the inference process of iterative algorithm as a Markov decision process,and the condition that the data divergence does not exist is then given.Based on the above theoretical analysis,we study the mismatch problem in DFR and propose an improved DFR.·We consider the automatic modulation classification(AMC).Generally,the AMC algo-rithm based on the maximum likelihood,i.e.,ML-AMC,has high computation complex-ity and thus is difficult for practical application.Meanwhile,the feature-based AMC algorithm needs artificial feature engineering to extract the features from the observed sequence.In this paper,we propose an AMC algorithm based on a convolutional neural network,namely CNN-AMC.With utilization of parallel computation,the CNN-AMC works much faster than the ML-AMC.In comparison to the feature-based AMC,the CNN-AMC has closer approximation to the ML-AMC,learns to extract the features through data training,end-to-end realize the classification task.In practical training the model cannot converge from the beginning,and we propose a two-step training method;to improve the training efficiency on a range of related but different AMC problems,we propose a training method based on the transfer learning.In addition,we propose a unit CNN-AMC to flexibly process the observation sequences with varying dimensions.·We consider a wireless cellular network with single-input single-output interference broad-cast channel.Under the maximum emitting power constraint,the maximization of the network sum rate with respect to the downlink power,is investigated.This optimiza-tion problem is nonconvex and NP-hard.The centralized model-driven algorithms,such as fractional programming and weighted minimum mean square error,have high com-putation complexity.Meanwhile,the existing reinforcement learning algorithm is not designed for the static optimization problems.We both develop the centralized and dis-tributed reinforcement learning algorithms for the static optimization problems.Further-more,we propose a framework of the multi-agent reinforcement learning algorithm using coordinations,which includes centralized training and distributed execution.The algo-rithm design includes cooperations across the cells,offline and online training,and a dynamic environment tracking mechanism.The concrete deep reinforcement learning algorithms include:the policy gradient-based REINFORCE,value-based deep Q learn-ing,and actor-critic-based deep deterministic policy gradient.