Complexity reduction for analog circuit performance models using random forests

Most existing optimization methods for analog circuits focus on the accuracy of their performance modeling techniques, which results in complicated and non-intuitive models such as high order polynomials or high rank matrices. These models may give designers good estimates on circuit performance, but not sufficient insight into the circuit behavior, as “black-box” optimization approaches often lead to pathological cases. In addition, having high-order polynomial terms will also cause problems related to over-fitting. In this paper a new approach of systematically selecting critical polynomial terms for any given equation is proposed. Utilizing random forests, this approach examines the importance of each polynomial term in a model, and removes terms according to the current and target fit values of the model. The results of applying this approach to a folded differential-in-differential-out amplifier and a two-stage differential-in-differential-out amplifier with common-mode feedback show that the sizes of performance models can be reduced without sacrificing modeling accuracy. Such a complexity reduction will provide a better insight into the trends and behavioral characteristics of the circuit under optimization.