Simulación de estrategias de optimización de producción usando reinforcement learning

Abstract

In this study, two algorithms trained using Reinforcement Learning are evaluated in a simulation environment for a process of changing saw blades among three different saw models. The simulations took into account real data from the process, which includes processing times for each process depending on the type of saw. A total of three different simulations of the process were conducted: the first simulation was the original simulation where the process was simulated as described in the document from which the data were obtained; for the second simulation, an algorithm trained to apply a strategy for selecting saw types to process and increase the number of units processed within a work session was implemented; and finally, for the third simulation, an algorithm trained to modify the priority of workstations to reduce operator downtime was applied. To evaluate the performance of the trained algorithms, the main comparison metrics were the number of units processed during a normal work session (8 hours) and the downtime of the workers. For this purpose, 100 different iterations were performed for each simulation, from which data were collected for interpretation and comparison. The productivity results showed that the saw type selection algorithm improved upon the procedure simulation, whereas the process prioritization algorithm did not achieve any improvements. Unfortunately, no improvements in operator downtime were observed.