Research On Machine Learning-Based Channel-Adaptive MIMO Communication System

The nature of communication research is transmitting information precisely from one end to another end in high speed.Given that performance of communication system in communication process essentially depends on the quality of channels,it is necessary to design the communication systems based on channel models.Yet huge differences between real channels and reference channel models do exist in real-world communication scenarios,especially in mobile communication scenarios.On the other hand,Doppler effect under mobile wireless communication scenarios could allow channel to be timevarying,could also bring difficulties to channel estimation and equalization.Since MIMO(Multiple-Input Multiple-Output)can considerably improve the transmission reliability by diversity gain,this thesis proposed a DNN(Deep Neural Network)-based MIMO communication,aiming to slove channel-adaptive problem.Machine learning is a field of computer science which enables computer to improve its performance by exploiting calculation methods and experiences without explicitly programmed.Not only has it evolved from symbolic machine learning to statistical machine learning,but also it has been applied in various fields.As a branch of machine learning,deep learning has been a heated topic among researcher in recent decades,allowing revolutionary breakthrough to occur.Deep learning is defined as facilitating computers from constrcuting simple concepts to learn complicated conceptes by structured concepts.Although more and more research has been focusing on using reinforcement learning to make behavioral decision in communication domain,applying machine learning in physical layer has just risen.By combining with the advantages of machine learning and features of MIMO system,the main work of this thesis includes:(1)Since the modules of source encoding,channel edcoding,channel estimation and equaliztion are considered separately in the designing process of tradition communication systems,we proposed an AE(Autoencoder)-based MIMO communication system.In training process,cost function is defined as the mean square error between the output of decoder and the input of encoder,invloving algorithms of gradient descent and Adam to optimize weight.By using a five-layer neural network,the proposed AE-based system can achieve better performance with traditional Alamouti system.Hopefully,its performance advantage compared with traditional system is more obvious under the condition of high signal-to-noise ratio.(2)To solve problems like difficult channel state information feedback,low performance of channel estimation as well as equalization on ill-conditioned matrix,an DNNbased MIMO system has been proposed to achieve adaptive channel equalization by using equalization layer in receiver.Simulation results have shown that DNN-based MIMO system can can effectively recover channel-caused faded signals and achieve better performance comapred with traditional system.