Reinforcement Learning for Analog Sizing Optimization

Abstract

This paper proposes a novel adaptation of Machine Learning (ML) techniques based on reinforcement learning for the resolution of analog sizing optimization problems. The paper details the proposed solution and highlights its performances using benchmark tests based on classical analog designs such as two-stages, rail-to-rail and folded-cascode amplifiers. This novel adaptation of the ML technique is used for the optimal design of analog circuits. SPICE simulations are used to verify the viability of the proposed algorithm and novel solutions. The GNN-FCNN approach is validated on a complex circuit with more than 50 devices