Investigation of Inspection and Maintenance Optimization with Deep Reinforcement Learning in Absence of Belief States

Abstract

Maintenance of deteriorating infrastructure is a major cost factor for owners and society. Therefore, efficient inspection and maintenance (I&M) strategies are of paramount importance. Deep reinforcement learning (DRL) has been proposed for maintenance optimization of deteriorating systems. For good performance, DRL relies on information rich state representations, but information about the state may only be available through costly inspections. One option to alleviate this is by use of belief states, however this might not always be possible due to incomplete model knowledge or computational constraints. Hence, there is potential for DRL approaches using only the information which is already available. In this work, we investigate several observation representations and compare them by training multiple DRL agents. Our experiments show that the choice of informative observation representations has a strong effect on the performance of the resulting optimized maintenance strategy.