Comparison of Supervised and Reinforcement Learning Based Current Control in Power Electronic Circuits

The well-known classical control strategies are often limited to the exact design, based on the system and load model and do not consider application-based performance indicators. AI control algorithms are the proposed alternative and have the potential to replace extensive adaptive control strategies. AI algorithms use machine learning to independently adapt to changes in the electronics environment and in the load characteristics. In this paper, the two most promising neural network-based AI controller designs are compared based on applicability. The basic methods, principals, and differences between supervised learning and reinforcement learning are compared by a simulation and an experimental setup using a full-bridge power electronics circuit. The neural network models are trained within the same simulation environment and are then migrated on a hardware. Both AI controllers show similar results in the simulations and in the experiment measurements.