Over the past few year, machine learning has attracted the attention of antenna engineers. Generally, the process of antenna design requires to find out the EM characteristics of antenna by observing the current distributions through simulations. These EM properties are then used for the parameters optimization. Machine learning (ML) can be combined with simulations to design an antenna. The inclusion of artificial intelligence (AI) can give promising results in the field of antenna designing. In recent years, antenna synthesis or design optimization through evolutionary algorithms (EAs) has been applied widely. At present, differential evolution (DE) and particle swarm optimization (PSO) are top two popular algorithms in the antenna synthesis area.

In smart antenna array, the objective of the gracefully degradation of the beamforming and beamsteering performance, can be achieved by reconfiguring the array when an element is found to be defective. This reconfiguration can be obtained by optimization using Machine learning and Support Vector Machines (SVM). SVMs are a set of supervised learning algorithms used for classification and regression problems. SVMs are a good candidate for the solution of antenna array processing problems such as beamforming and the angle of arrival estimation, because these algorithms provide superior performance in generalization ability and computational complexity. The basic idea is to change the excitation coefficient for each array element (magnitude and phase) to optimize for changes due to the environment surrounding an array antenna. Using Support Vector Machines, the antenna array is trained to change its elements phase or excitation distribution to maintain a certain radiation pattern or to enhance its beam steering and nulling properties and solve the DOA problem as well.

Optimization technique based SVMs algorithm is also used to match the measured far field radiation intensity to the corresponding antenna array structure. The SVM classifier is firstly trained by a set of input feature vectors extracted from measured radiation data of different array structures in various scenarios. Then, it is tested and the SVM parameters are adjusted during the learning process to help approach the optimal classification result. Then the trained SVM classifier is used to locate malfunctioning antenna elements of the array in real-time.

Machine learning based methods are used for calculating the radar cross section (RCS) of the antenna-radome system. In this method, the back-propagation algorithm can be used to train the machine learning model. The machine leaning based RCS calculation becomes a promising area, especially for the industrial sector with large amount of measurement data. Compared to traditional methods, machine learning methods have the potential to handle complex.

Machine learning techniques are also efficient in shaped-beam array designs. Here, SVM is applied for the characterization of the reflection coefficient matrix, which provides an efficient way for deriving the scattering parameters associated with the unit cell dimensions. ML is one of the most promising and salient research areas in artificial intelligence, as it has become a powerful tool in a wide range of antenna designs and related applications.

###
Genetic algorithm (GA)

Genetic algorithm (GA) is one of the global optimization algorithms that is used widely by antenna designers for the optimization of the antenna shape and size to achieve better overall performance of the antenna. GA has been used to enhance the performance of microstrip patch antennas by optimizing the bandwidth, multi-frequency, directivity, gain, size etc. The concept of the GA, first formalized by Holland and extended to functional optimization by De Jong, involves the use of optimization search strategies patterned after the Darwinian notion of natural selection and evolution.

During a GA optimization, the parameters of each individual of the population are encoded as a string of bits (chromosomes). The first group of individuals (generation) is created randomly. The fitness of each individual is determined by the cost function. Mating these individuals forms a new generation. The more fit individuals are selected and given greater chance of reproducing. Crossover and mutation are used to allow global exploration of the cost function. The best individual may be passed unchanged to the next generation. This iterative process creates successive generations until a stop criterion is reached. A block diagram of a genetic algorithm optimizer is shown here.