Enhanced Reinforcement Learning Methods for Optimization of Power Delivery Networks on PCB

In this paper, reinforcement learning (RL) is used to find design solutions for power delivery networks (PDN) on printed circuit boards (PCB). Two methods, i.e. deep deterministic policy gradient (DDPG) with parameterized action space and advantage actor critic (A2C) are implemented and compared. The design parameters, which are considered, are the placement of decoupling capacitors (Decaps) and the size of the power plane. By appropriate definition and implementation of the RL algorithms PDNs are obtained, which are acceptable within the design constraints. This holds even when dealing with the optimization of objectives that are in trade-off at some levels.