Soft Actor-Critic Reinforcement Learning-Based Optimization for Analog Circuit Sizing

Abstract

This research proposes a Soft Actor-Critic (SAC) based RL approach to optimize analog circuit sizing. The SAC algorithm efficiently addresses challenges in continuous state and action spaces, providing stable learning and sample efficiency. Comparative experiments were conducted on a 2-stage OTA and a 3-stage TIA, showing that SAC outperforms DDPG and TD3 in terms of success rate, average FoM, and minimum power consumption. The results demonstrate the effectiveness of the proposed SAC-based RL architecture for analog circuit optimization.