Comparison study of pattern‐synthesis techniques using neural networks

Artificial Neural Network Analysis of Sierpinski Gasket Fractal Antenna: A Low Cost Alternative to Experimentation

“…In recent years, a number of studies are done on performance comparison of ANNs for a range of applications [19,21,26,[33][34][35][36][37][38][39]. The performance of the proposed models is compared on the basis of two different measures that is, (1) mean absolute error (MAE), which gives the quality of training the model.…”

Analysis, synthesis, and diagnostics of antenna arrays through complex-valued neural networks

“…(6). Figure 2 shows a power pattern synthesized using the Orchard-Elliott method [8,9] which consists of a "flat-topped beam" pattern radiated by the array with sidelobes Ϫ25-dB below the maximum. The circles on the power pattern are indicating the angular positions taken for "measuring" (in a numerical simulation) the amplitude and phase of the field pattern.…”

“…A direct comparison of this CVANN architecture with those given in earlier works for similar purposes (radial-basis-function architecture in ANNs prepared to estimate the direction of signal arrival, or beamforming for interference cancellation, for example, applied to array antennas [2,3], or the synthesis procedures compared with several ANN architectures given in [8], in which the same linear array radiating the same pattern as that considered here is presented) is not possible. Nevertheless, CVANNs have some advantages with respect to conventional ANNs, as the number of layers needed in each case: in real valued ANNs (RVANNs) it is often required the use of one or more hidden layers, mainly when the problem to solve is nonlinear [1] such as those related to radiation of antenna arrays [2][3][4]8], whereas CVANNs require a reduced number of layers (for example, no hidden layers are used in the examples presented here).…”

“…Nevertheless, CVANNs have some advantages with respect to conventional ANNs, as the number of layers needed in each case: in real valued ANNs (RVANNs) it is often required the use of one or more hidden layers, mainly when the problem to solve is nonlinear [1] such as those related to radiation of antenna arrays [2][3][4]8], whereas CVANNs require a reduced number of layers (for example, no hidden layers are used in the examples presented here). Some features related to nonlinearity are avoided when using CVANNs [5,6].…”

“…A number of additional features were further simplified in the example presented in this work, such as the computational time of the training process, which was reduced to a matrix (pseudo) inversion. The simplification of the transfer functions was also advantageous: the ones used in RVANNs are related to exponential, hyperbolic tangential or sigmoid functions for hidden layers (that is, those layers inserted between the first and last layers of the neural network, often called input and output layers, respectively) [1][2][3][4]8], whose selections always imply additional computational effort; the CVANNs presented in this work use the identity as the only transfer function. Finally, the number of neurons in the CVANN presented here is equal to the number of outputs; therefore, no trial-and-error processes of selection of their architectures are needed, as is usually the case with RVANNs [1].…”