Local Bayesian Optimization For Analog Circuit Sizing
Konstantinos Touloupas, Nikos Chouridis, Paul P. Sotiriadis
Abstract
This paper proposes a Bayesian Optimization (BO) algorithm to handle large-scale analog circuit sizing. The proposed approach uses a number of separate Gaussian Process (GP) models approximating the objective and constraint functions locally in the search space. Unlike mainstream BO approaches, it is able to traverse high dimensional problems with ease and provide multiple query points for parallel evaluation. To extend the method to large sample budgets, GP regression and sampling are enhanced by using kernel approximations and GPU acceleration. Experimental results demonstrate that the proposed method finds better solutions within given budgets of total evaluations compared to state-of-the-art approaches.