Multi-Agent Reinforcement Learning in Designing the Low-Dropout Regulator Circuits

Low-dropout regulator (LDO) circuit, whose function is to provide a power supply robust to variations in process, voltage, and temperature (PVT), is an essential part in any system-on-chip. Therefore, the design of this circuit must satisfy various intricate specifications, rendering the design process generally perceived as tedious and lengthy. As a result, previous research has been conducted to explore the use of machine learning, particularly reinforcement learning, in speeding up and automating the LDO design process, especially the sizing phase. However, the results of these works are limited in terms of the number of design variables and specifications handled by the automation engine. This study presents the application of single-agent proximal policy optimization (PPO) and multi-agent proximal policy optimization (MAPPO), including both parameter-separated and parameter-sharing methods, to address the LDO sizing automation problem. The experimental result shows that the PPO- and MAPPO-based implementation in LDO sizing automation outperforms that of the classical particle swarm optimization algorithm. We demonstrate that the parameter-separated MAPPO features the most effective learning process compared with other PPO-based benchmarks, resulting in a design result that is competitive to that of a well-known commercial tool.