Despite recent advances in algorithms, analog circuit sizing optimization remains a challenging task that demands numerous circuit simulations, hence extensive CPU times. Many simulation-based methods have been proposed in recent decades to address this, using black-box optimization tools such as Genetic Algorithms (GA) and Bayesian Optimization (BO). However, GAs require vast number of simulations offering no guaranteed convergence due to inherent stochasticity. The Gaussian Processes models used in BO methods also suffer from cubic computational scaling, which makes them less practical for problems with more than 20 variables. Recent approaches in analog circuit sizing automation employ Reinforcement Learning (RL) or RL-inspired algorithms and have been shown to be more sample-efficient and less time-consuming. In this seminar, I will show the importance of using an ensemble of probabilistic performance models in an Actor-Critic RL framework for a fast but extensive exploration of the design space.
22/3/2024 11:00 - 12:00
ESAT B91.200
