Echo State Networks for Accurate and Efficient Modeling of Dynamic Circuits

Neural network-based architectures have long been used as black-box regressors for developing macromodels of circuits and systems. From the more basic feed-forward artificial neural networks to the more recent complex and powerful recurrent structures, increasingly powerful structures have been proposed over the years. Unfortunately, the added modeling power has often been paired with equally increased training requirements and computational costs. In this paper, we introduce a novel macromodeling method, based on the well-known concept of reservoir computing, which can efficiently produce accurate models. When compared to commonly used recurrent neural network architectures, the proposed method is characterized by much lower training requirements, both in terms of training data and training time necessary for model development. Furthermore, the proposed architecture, based on Echo State Networks is not encumbered by the problem of vanishing or exploding gradients which is the cause of much difficulty in training recurrent architectures. The resulting model is shown to be adequately accurate, exhibiting small relative error on the examples shown, requires only modest training time when compared to commonly used recurrent alternatives, and is simple enough that its evaluation can be performed very rapidly, therefore proving to be a very interesting and general alternative for macromodeling dynamic circuits and systems.