ANN-Powered Reinforcement Learning-Based Analog Circuit Optimization

One of the challenging problems that Electronic Design Automation (EDA) dealing with is automatic circuit sizing, which is basically optimizing circuits to meet various performance metrics, such as gain, bandwidth, and reliability. Recently, it has been shown that Reinforcement learning (RL) has the potential to revolutionize the circuit sizing and improve the efficiency of the existing algorithms. To this end, several RL-based approaches have been proposed in the literature; however, they suffer from expensive simulations performed in every steps of optimization that degrades the efficiency in terms of execution time. To overcome this problem, we propose a novel RL-based optimization approach, where we empower the RL with artificial neural network (ANN) that substitutes the simulation environment, thus avoiding expensive SPICE simulations. According to the synthesis results, the proposed tool can achieve more than $6 \times$ runtime speed-up compared to the simulation-based approach without sacrificing the accuracy.