Adaptive Transient Stepping Policy on Reinforcement Learning

Transient analysis (TA) is the foundation for nonlinear electronic circuit simulation, which determines the time-domain response of the circuit over a specified time interval. However, it is usually computationally intensive and quite time-consuming without careful parameter tuning and proper stepping policy. In this paper, reinforcement learning (RL) is introduced to design a effective transient stepping policy (TSP). The online RL-based TSP works with bidirectional agents and state-switch samples, enabling our model to adaptively and intelligently adjust the bidirectional step sizes in TA. The proposed RL-based TSP has been implemented in an open source SPICE-like simulator, and has achieved remarkable simulation efficiency speedup while maintaining reasonable simulation accuracy.