Utilization of a Reinforcement Learning Algorithm to Optimize a Multi-Level Gate Driver Circuit

Abstract:
This work presents a reinforcement learning approach, the Q-learning algorithm, to optimize a multilevel gate driver for GaN-HEMTs in terms of minimizing switching energy and current/voltage overshoot, controlling the values of di/dt and dv/dt. Parasitic inductances, resistances, pulse widths, and voltages are considered to achieve this goal. The Q-learning algorithm is written in Python and tested in a LTspice double-pulse simulation. The results show that by applying the Q-learning algorithm to the multi-level gate driver, the reward parameters can be reduced by up to 19% for IS and 30% for Eon when choosing an individually optimized pulse width for a three-level turn-on driver combining RG1 , RG2 and RG3 when compared to a conventional gate driver.