@phdthesis{,
    author = {Zhang, Tao},
    title = {Material Flow Control on Shop Floors in Job Shops: Release, Routing, and Sequencing},
    editor = {},
    booktitle = {},
    series = {},
    journal = {},
    address = {},
    publisher = {},
    edition = {},
    year = {2018},
    isbn = {},
    volume = {},
    number =  {},
    pages = {},
    url = {},
    doi = {},
    keywords = {Material Flow Control, Sequential Decision-making, Simulation, Machine Learning, Markov Decision Process },
    abstract = {Material flows in manufacturing are the movements of materials through a defined route in a plant for producing a final product. Material flow control (MFC) oversees the movements of all materials. Our study focuses on the release, routing, and sequencing processes. The release process decides if a job will be released while the workload changes. The routing process specifies which machine a job will go to when the job is ready. The sequencing process determines which job will be processed first when machines become idle. These processes are sequential decision-making processes in which decisions are made in a sequence and aim to optimize a long-term objective. Due to the randomness and complexity of the material flows, these decisions are usually made by some decision rules. However, these rules often lack an overall view of the system, which in turn lead to an unstable performance. 
In our study, each decision-making process is an alternative selection procedure in which a priority value is computed for each possible alternative, and the alternative with the highest priority is chosen. The priority values are subject to the constraint from the sequential decision-making, i.e., all decisions made based on the priorities result in the good long-term performance. Apparently, the priority value is a function of the concerned alternative and the current and future states of the material flows. Because simulation is always the first choice to make predictions and analyses of the complex system while machine learning dedicates itself to finding knowledge from raw data, our basic idea is drawn to calculate the priority values by the machine learning according to the current and future information generated from the simulation. Three simulation-based methods are proposed, including a simulation try-then-decide method (STTD), an intelligent method based on the simulation try-then-decide method (INT1), and another intelligent method based on Markov decision process (INT2). Lots of techniques from the machine learning are utilized, such as clustering, neural networks, reinforcement learning, and so on. Because the methods highly depend on the simulation, an agent-based simulator for the material flows is developed first, which is also used to evaluate the methods at last. The three methods are employed to a sample manufacturing line and compare with each other as well as some decision rules. The results show that the STTD and INT1 methods always outperform the rules. The STTD method performs best but consumes much time. Contrarily, the INT1 method takes less time while the performance is just a little bit worse than the STTD. Thus, the STTD is more suitable for offline applications, and the INT1 can be used in the real-time control. Unfortunately, the INT2 method performs unsteadily. It will be further studied in the future.},
    note = {},
    
    school = {Universität der Bundeswehr München},
    
}