Management of repair debt with reinforcement learning
Puro, Niko (2026)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-202604106081
https://urn.fi/URN:NBN:fi:amk-202604106081
Tiivistelmä
The topic of this development-oriented thesis has been commissioned by Marjetas academy Oy, with focus on finding a solution to manage budgeting of repair debt in favor of improving productivity and decision making accuracy with use of reinforcement learning. It centers on the practical aspects of creating a reinforcement learning agent and custom environment with core features like reward shaping, data management and controlling dimensionality of the observation space and action space.
The development process required extensive research on reinforcement learning in forms of scientific articles and publicly published projects on creation of custom environments and their capabilities. This work acknowledges that decisions were based on knowledge available during the time of development and research which may require adaptation as new information arises. Libraries like Stable_Baselines3 and Farama Gymnasium's environment interface made simultaneous experimentations on multiple different models possible during the testing and validation phase without further adjustments to the environment, while relying heavily on models like DDPG, A2C, PPO and SAC in their present form.
Development of this work achieved a testing and validation phase which included a separately developed interface to visualize results of the models. This work does not cover development beyond this point.
The development process required extensive research on reinforcement learning in forms of scientific articles and publicly published projects on creation of custom environments and their capabilities. This work acknowledges that decisions were based on knowledge available during the time of development and research which may require adaptation as new information arises. Libraries like Stable_Baselines3 and Farama Gymnasium's environment interface made simultaneous experimentations on multiple different models possible during the testing and validation phase without further adjustments to the environment, while relying heavily on models like DDPG, A2C, PPO and SAC in their present form.
Development of this work achieved a testing and validation phase which included a separately developed interface to visualize results of the models. This work does not cover development beyond this point.