High-Order Deep Recurrent Neural Network With Hybrid Layers for Modeling Dynamic Behavior of Nonlinear High-Frequency Circuits

Abstract

In this article, a new technique for macromodeling of high-frequency circuits and components called high-order deep recurrent neural network (HODRNN) is proposed. This technique explores an alternative approach to learn RNN for time dependencies in a more efficient way resulting in more accurate model. HODRNN uses more memory units to track previous hidden states, all of which are returned to the hidden layers as feedback through various weight paths. Moreover, a new improved structure called Hybrid-HODRNN is proposed for further increasing the modeling accuracy of HODRNN. The proposed Hybrid-HODRNN uses hybrid layers with both single and high orders for taking advantage of HODRNN and also reducing the overfitting problem, which finally leads to a more accurate model. In addition, the proposed method requires less training signals compared to the conventional shallow and deep RNNs in order to create a model with similar accuracy. Also, the obtained models from the proposed method are considerably faster than the transistor-level models while having similar accuracy. By modeling three high-frequency circuits in this article, we conclude that the HODRNN and its hybrid structure offer the ability to create a better macromodel of high-frequency nonlinear circuits than the conventional RNNs, which verifies the superiority of the new macromodeling techniques.