Applications of Support Vector Machines in Electromagnetic Problems.

摘要

The emphasis of this dissertation is to demonstrate how Support Vector Classifiers (SVCs) and Support Vector Regressors (SVRs) can be applied to optimize the solution of Electromagnetic problems. SVCs perform optimization by classifying novel inputs into the best matched category subject to maximum separation margin and minimum empirical risk. SVRs extract the underlying mapping relationship from noise contaminated training data set, and try to find the optimal and the most flat hyperplane that fits the data distribution. First, the application of SVCs in Automatic Target Recognition (ATR) is presented as a paradigm of integrating machine learning techniques into ATR optimizations. Next, the SVR-based linear/nonlinear array beamforming and Direction of Arrival (DoA) estimation are presented in comparison with various conventional optimization approaches such as Least Square (LS) methods, Minimum Variance Distortionless Response (MVDM) methods, Multiple Signal Classification (MUSIC) and etc. The SVM-based approaches are proved to have significant advantages over these existing algorithms. Another innovative application of SVMs proposed is in the area of antenna design optimization. An example of applying SVMs for Composite Right and Left Handed (CRLH) Metamaterial Ultra Wide Band (UWB) antenna design optimization is illustrated for the first time. Based on the applications and experiment data analysis we can see that, the SVM-based approaches demonstrate appealing characteristics including their reliability against Gaussian and Non-Gaussian noise interference, the remarkable generalization ability as well as the sparse structure of solution. All mathematical derivations, simulations and experiment work are included for each application.