Study On The Application Of Machine Learning In The Wireless Communications

The broadcast nature of radio signal results that interference is ubiquitous in wire-less environment, and leads to many technical problems in wireless communications. In this thesis, we investigate two technical issues related to the interference in wireless communication systems:Channel Equalization and Multiuser Detection. The former is adopted to address the inevitable intersymbol interference (ISI) in wireless envi-ronment, which is caused by multipath effects, channel bandwidth limitation and the uncertainty of channel characteristics. The latter is caused by multiple access interfer-ence (MAI) in code division multiple access (CDMA) systems. Channel equalization techniques can improve the performance of the communication system suffering deep fading, by solving the ISI problem caused by multipath delay. And in the CDMA communication system, multiuser detection is one of the most efficient methods to overcome MAI.Machine learning (ML) is an artificial intelligence science that involves many disciplines, such as psychology, biology, neurophysiology, mathematics, automation and computer science. ML is the core topic of artificial intelligence. The main issue of ML is how to get the potential rules from the observed data and then how to use these rules to predict the unknown data. In this thesis, we adopt machine learning technique to estimate the characteristics of ISI and MAI, and then using these characteristics to detect the transmission signals. The major contributions of this thesis are following:1. Analyzed the Gaussian process (GP) for regression and the sparse Gaussian process. Gaussian process is a regression method based on Bayesian framework, which can be seen as the promotion of the minimum mean square error (MMSE) criterion to solve nonlinear problems. In the wireless communications, many problems are solved by using MMSE criterion. If we use GP instead of the MMSE criterion, it can improve the system performance. Furthermore, GP can give an output value to calculate the corresponding input, it can also give the probability distribution of output values, which is more more accurate for prediction. Sparse Gaussian process is to use some sparse approximation method to approximate the Gaussian process (full GP). It can reduce the computational complexity significantly with a approach accuracy, which is necessary for the real-time communication system.2. Proposed the based on sparse Gaussian process multiuser detection. From the analysis of multi-user detection model, we can conclude the problem to the function regression. Because of this we use GP to solve the problem of multiuser detection. We don’t need to know the information of each user’s spreading code, only need to know some of the known transmitted sequence and its corresponding received sequence. We can regress a function by these sequences, and detective the transmission signal of unknown received signal use the function. Gaussian process can regress a nonlinear function, so its performance is superior to the MMSE.3. Proposed the based on sparse Gaussian process channel equalization. Similar to the multiuser detection model, channel equalization model can also be seen as a func-tion regression problem. The equalization based on GP can equal the various nonlinear effects in wireless communication system. Compared to the MMSE equalization, the performance of the GP equalization improves greatly.4. Analyzed each sparse Gaussian process performance. By simulation experi-ment, we analyze the performance of each of the equalizer and give the reasons. From the figures, we can see that sparse spectrum Gaussian process (SSGP) can approximate to the full GP, but the subset of regressors (SoR) is poor. This is because the former is the approximation of the kernel function, the latter is the approximate covariance matrix. In contrast, matrix approximation has a greater impact to the results.In summary, we applied the (sparse) Gaussian process to address the multiuser detection and channel equalization problem. Simulation and analysis results show our proposed methods work better than other conventional detector or equalizer.