Open shop scheduling in a manufacturing company using machine learning

Abstract

Scheduling jobs in a manufacturing company that delivers custom products is challenging. Aarbakke is a company that manufactures advanced assemblies for the oil and gas industry. Its existing resource planning tool frequently produces unrealistic job schedules, leading to substantial delays. In this thesis, we describe a reinforcement learning agent that optimizes scheduling by utilizing historical data. Our aim is to minimize the time spent processing jobs past their deadlines - the tardiness. The problem can be modelled as an open shop scheduling problem. Existing research in this area has only looked at other performance measures, such as makespan. Due to the NP-hardness of the problem, we used ant colony optimization and genetic algorithms to produce heuristic scheduling algorithms that can make efficient decisions under uncertainty. Based on cross-validation of 18 algorithms, a candidate model was selected and tested in a hypothesis test. It reduced the mean tardiness by 2.6 % compared to the scheduling algorithm currently in use when testing on historical processing times and by 14.6 % when testing on sets of forecasted processing times. Implementing it in production can potentially lead to savings in manufacturing cost. The approach can be applied to similar problems in other custom job shops.