Graduation Term
Spring 2025
Degree Name
Master of Science (MS)
Department
Department of Mathematics
Committee Chair
Mehdi Karimi
Committee Member
Fusun Akman
Committee Member
Papa Sissokho
Abstract
This thesis explores an application of reinforcement learning (RL) in maintenance optimization. Recent advances in hardware-accelerated computation and deep learning have made RL a powerful tool for solving optimization problems which are too complex for traditional methods. Maintenance optimization involves improving the efficiency and effectiveness of maintenance activities through data-driven approaches, ultimately reducing costs and increasing asset availability. Making informed maintenance decisions is crucial to long-term sustainability.
A desirable maintenance policy maximizes a utility signal while minimizing the cost of maintenance. Techniques in sequential decision making such as dynamic programming (DP) and RL have found success in optimizing these maintenance policies for individual or small groups of components. However, the high dimensionality makes optimizing maintenance of many components computationally intensive.
This thesis designs a novel computationally efficient RL approach to optimize multi-component maintenance policies. We propose a $q$-value approximation method which scales efficiently with a large number of components. This method reduces the size of the action space through centralized execution, and heuristical method of assigning actions through proto-actions we call triage. The method further reduces the search space through a state-weighted Q basis over the action domain. Through several numerical experiments, we show that this method achieves less and more stable loss during training and produces better policies than standard Deep Q-learning.
Access Type
Thesis-Open Access
Recommended Citation
Wittrock, Joseph W., "Centralized Deep Reinforcement Learning for Homogeneous Multi-Component Maintenance Optimization" (2025). Theses and Dissertations. 2095.
https://ir.library.illinoisstate.edu/etd/2095
Included in
Civil Engineering Commons, Construction Engineering and Management Commons, Control Theory Commons, Data Science Commons, Mathematics Commons
