Physically Constrained PCB Placement Using Deep Reinforcement Learning

Abstract

This thesis provides an in depth exploration of Reinforcement Learning (RL) based PCB component placement with emphasis on physically verified placements. Unlike prior methods that rely on heuristic proxies for placement quality, this work focuses entirely on routing based metrics that result in functioning placements without the need for fine tuning. Additionally, this exploration considers true use cases of PCB auto-placement where a human-in-the-loop pre-places a set list of components and the auto-placer places the remaining. This is achieved by first restricting the placement domain to only ring placements; a domain where routing calculations become accessible. Within the ring placement domain, an RL agent is trained to place components on a simulated PCB canvas such that there are no component overlaps or wire crossings upon manufacture. Through the use of an unbounded reward system, the agent is trained progressively with PCB complexity gradually increasing as training steps are run. The resulting placements are robust to varying numbers of components as well as component shape and size. Finally, this thesis concludes with a discussion about further work and challenges facing the future of PCB auto-placement.