Voltage Regulation of DC-DC Buck Converters Feeding CPLs via Automatic Curriclum Learning

Abstract

Curriculum Reinforcement Learning (CRL) conducts agent learning through predefined training courses, improving the learning speed and stability of the agent. However, predefined courses rely too much on the quality of prior experience and ignore feedback to learners. To solve the stability and load uncertainty problems of DC-DC buck converters with constant power load. First, sampling goal on the boundary where an agent can reach a set of targets, then the method will provide a stronger learning signal compared to the target of random sampling. Therefore, we import the Goal Proposal Module to consider more boundary goals and automatically generate effective courses. In the state of relatively large conversion of the constant power load. The simulation results of control strategy based on automatic curriculum learning in reference to PI. The results show that the automatic curriclum learning has higher dynamic performance and learning speed and can track.