Trustable Reinforcement Learning for Asset Integrity Management

SUMMARY & CONCLUSIONSSignificant recent accomplishments and immense future potentials of Reinforcement Learning (RL) are behind the surge in interest in this class of decision-making algorithms within the Artificial Intelligence domain. However, RL application to real-world problems such as technical asset integrity management has remained very limited. One of the main obstacles to the real-world deployment of RL can be broadly characterized as the lack of trust. Assets operators are reluctant to deploy policies derived from RL-based software because they are not currently designed to earn the user confidence. This research is an effort to address the "trustability" concerns in RL for safety and asset integrity management type of problems. We propose a methodology that sets up multiple risk barriers in the RL output policy and provides an estimation of the RL policy's performance prior to deployment. Our proposed method is compatible with any RL algorithm. The methodology was tested on a scaled-down integrity management case-study in a simulated pipeline environment and successfully derived a safer and superior policy for the pipe's corrosion-related maintenance management.