Bayesian Optimization Combining Parallel Local Sampling for Microwave Design Applications

For electromagnetic (EM) optimization of microwave components, with a bad computational starting point, local optimization can easily fall into local optima that cannot satisfy design specifications. In this case, global optimization methods can be utilized. However, global optimization methods are usually time-consuming because of its relatively low convergence rate. In this paper, an efficient EM optimization method combining Bayesian optimization (BO) and parallel local sampling method is presented to solve this problem. In each optimization iteration, the presented method utilizes both the predicted optimal solution and several local samples in an adaptive range around it to construct the surrogate model which is used to guide the optimization process. In this way, the presented method keeps a good balance of exploration and exploitation during optimization. Compared to existing BO which only utilizes the predicted optimal solution, the presented BO effectively reduces the number of optimization iterations, thus increasing the convergence ability of optimization. An example of a microwave filter optimization is presented to illustrate the presented method.