Constrained reinforcement learning for advanced control in powder bed fusion*

Reinforcement Learning (RL) continues to attract considerable attention in academia and industry. Its data-driven nature, combined with its varied and flexible formulation, makes it applicable in a variety of complex control tasks, in which control theory techniques can be challenging to implement. The Powder Bed Fusion (PBF) process, comprises an example of such a complex control task. However, there are still critical challenges in RL that need to be addressed in order to fully enable its use in PBF implementations. For instance, while constraint satisfaction comprises a necessity in PBF process control, there are still gaps in demonstration and analysis of RL algorithmic behaviour and control performance under constraint satisfaction. Existing constraint techniques in the literature, such as radial squashing, can provide zero constraint violation guarantees in process control. However, a constraint framework that also accounts for satisfactory control performance must be established. In this work, an attempt to address the above challenges is presented, providing a thorough analysis on constrained RL for PBF process control, and assessing the impact of the radial squashing technique. The results of this analysis show that tuning the intensity of radial squashing can be vital for maintaining satisfactory control performance under constraints.