Machine Learning in Logistics
In the course of the module Machine Learning in Logistics (Master, 3rd semester), Jupyter notebooks are used in addition to traditional media such as presentation slides and worksheets. They are used for visualization and evaluation of data. At the same time, the file format Jupyter Notebook offers the possibility to document the individual work steps cleanly. Such a procedure is very important in order to make the data from one's own company or institution understandable for all those involved.
In the field of logistics, various special features have to be taken into account which are not (always) covered in the introductory literature on machine learning: There are elementary differences when working with temporal data, spatio-temporal data and image data. Utilization forecasts at a logistics node are a typical example of temporal data. Traffic data, especially GPS tracks of goods and vehicles, on the other hand, are referred to as spatio-temporal data and are treated accordingly. This also includes ship movements of a (in future possibly autonomous) fleet of container ships. Automatic image recognition in autonomous vehicles and vessels already plays a major role today. Due to this variety of data, Jupyter notebooks are suitable, because the programming language Python, which works in the background, has a large selection of proven libraries which this data can be visualized and evaluated with. Therefore, a large part of the exercises are done with Jupyter notebooks.
At the end of the course, a part of the exam is executed digitally. Therefore the exam takes place in a room prepared for e-exams. Each student is provided with a specially prepared laptop. During the entire time of the exam, students are free to choose when they want to work on the tasks on the laptop and when the tasks on paper. At the end of the exam, the digital task sheets are automatically collected and used for evaluation in addition to the paper exam.
Competences are tested on the laptop using interactive programming. With the help of the IDE Jupyter Notebook, programming tasks are set. The code that the students write can be tried out in the IDE interactively. Error messages of the interpreter point out syntax errors. Visualizations of data sets are displayed interactively. This makes the exam situation much more similar to the everyday work of data scientists than when programming with pen and paper. The didactic possibilities are explained in more detail with the digitaler Freischwimmer of the ZLL.
For the technical execution, the application JupyterHub has been installed on a virtual machine provided by the IT department of the TUHH. On each of the laptops, the exam browser is set up to connect to the virtual machine. This minimises the effort required to set up the laptops and the virtual machine can be set up in advance and thoroughly tested by the examiner. Details on the implementation have been explained in the INSIGHTS-Blog of the Institute for Technical Education & University Didactics. The sample implementation listed in the article can be obtained by interested parties from the git repository on the TUHH GitLab instance.
Besides the technical feasibility, other aspects have to be considered, such as data protection or archiving of examination results. This requires careful preparation and follow-up of the exam. These aspects of the first run were presented as a poster at the e-Prüfungs-Symposium in Siegen in 2019. For further questions or for a general exchange of views, please contact Marvin Kastner.