Mixed-Variable Bayesian Optimization for Analog Circuit Sizing using Variational Autoencoders
Konstantinos Touloupas, Paul Peter Sotiriadis · 2022
Bayesian Optimization (BO) has recently gained popularity within the context of automatic sizing of analog and Radio-Frequency (RF) Integrated Circuits (ICs). However, its reliance on Gaussian Process models, which operate only on continuous-valued spaces, reduces its applicability in real-world scenarios, where multiple discrete-valued variables often exist. In this paper, we propose an approach to mitigate this issue by using a Deep Learning scheme to transform devices parametrizations to continuous ones, where classic BO can be applied. Specifically, a composite architecture that consists of a Convolutional Variational Autoencoder (VAE) and a dense Neural Network is built to define a continuous representation of integrated inductors in a TSMC 90nm process. By optimizing using these representations, we overcome the limitation of discrete-valued variables. Experimental results on a Low Noise Amplifier highlight the efficiency of the proposed approach.