Fast Simulation Method with Reinforcement Learning for Automated Optimization of Electronic Systems

Abstract

Design automation of electronic systems is challenging due to the growing design space, high performance tradeoffs, and rapid technological advances. To solve this problem, this paper presents an automated optimization framework that combines Fast Simulation with deep reinforcement learning for automatic circuit design. Fast Simulation can quickly and accurately evaluate circuit performance by neural networks. Deep reinforcement learning is used to find optimal parameters in the design space. Compared with existing reinforcement learning methods, the proposed method can automatically generate labels for the optimization results of the reinforcement learning agent by the simulator to retrain the neural network. To this end, the proposed optimization method performs better designs and reduces the required number of simulations.