Let bandits tell you when to maintain your  machines: A study of the maintenance planning problem

In this thesis algorithms are developed that learn when to plan maintenance. In essence, the main challenge is optimizing the planning problem wherewith cost can be saved by maintaining the part until just before it breaks. The core of the challenge lies in the fact that we do not know when the part will break. The different algorithms that are developed in this paper have the choice between exploring: gaining more knowledge about the failure time by letting the part break and paying for the repair, or exploiting: to plan maintenance early and thereby save cost but not knowing how long the part would have functioned. Notice that letting the part break is expensive, but planning maintenance too early, and thereby too often, is expensive as well. Three algorithms were designed, which all use a different strategy. The first algorithm has an exploration phase and then an exploitation phase. Which means that this algorithm lets the part break for a certain amount of times, and after that...Show moreIn this thesis algorithms are developed that learn when to plan maintenance. In essence, the main challenge is optimizing the planning problem wherewith cost can be saved by maintaining the part until just before it breaks. The core of the challenge lies in the fact that we do not know when the part will break. The different algorithms that are developed in this paper have the choice between exploring: gaining more knowledge about the failure time by letting the part break and paying for the repair, or exploiting: to plan maintenance early and thereby save cost but not knowing how long the part would have functioned. Notice that letting the part break is expensive, but planning maintenance too early, and thereby too often, is expensive as well. Three algorithms were designed, which all use a different strategy. The first algorithm has an exploration phase and then an exploitation phase. Which means that this algorithm lets the part break for a certain amount of times, and after that it plans maintenance on the day that would have given the smallest loss in the explore phase. The second algorithm has an ’explore while exploiting’ strategy. It does so by being optimistic about the expected loss of the possible actions: not choosing the best known option, but choosing what could be the best option. And the third algorithm exploits only: every round it chooses the best known option. These algorithms are then compared by mathematical analysis and simulations of the pseudo regret. From this is concluded what the strong and weak points of every algorithm are. The discussion contains suggestions for further analysis.Show less

Leiden University Student Repository

Documents

In Collections

Let bandits tell you when to maintain your machines: A study of the maintenance planning problem