Bayesian direct-binary-search algorithm for the efficient design of mosaic-based power splitters
Takeshi Fujisawa, Kunimasa Saitoh
Abstract
An efficient design method of mosaic-based power splitters based on Bayesian optimization technique is proposed. First, learning characteristics of Gaussian process (GP), which is one of the Bayesian inference techniques, is investigated to show its high regression performance. The transmission characteristics of 1×2 mosaic-based power splitters can be learned with an error of only 0.5%, which is comparable or better than simple ANN. Next, it is demonstrated that an efficient design of 1×2 mosaic-based power splitter with various splitting ratios is possible by using Bayesian optimization based on GP for selecting the next pixel. In the conventional direct-binary-search (DBS) design of mosaic-structure, the next pixel is chosen randomly. On the other hand, in the proposed method it is chosen based on the statistical information obtained by Bayesian inference. By accumulating the information of the transmission characteristics of the device obtained by electromagnetic (EM) simulation as training data, 70% reduction of the number of EM simulation compared with conventional DBS design is demonstrated. Furthermore, by using Bayesian optimization technique, it is shown that the device structure with better characteristics is obtained, compared with those obtained by conventional DBS design for the same number of EM simulation. There results indicate that the proposed method is useful for the design of mosaic-based devices.