Abstract:
Traffic problems are an essential part of our everyday lives, from daily commuter traffic to the development of urban infrastructures and global freight transportation. A key challenge is to optimize the allocation of resources to minimize costs or maximize efficiency. Optimal Transportation (OT) theory provides a mathematical framework to formally describe such problems and their solution, taking into account complex objectives and constraints. Over the years, various methods have been developed to solve OT problems. However, traditional methods reach their limits when solution spaces are continuous and the consideration of traffic density is important. The Dynamic Monge-Kantorovich (DMK) algorithm developed in recent years offers a solution to these challenges. It models traffic problems using dynamic equations and enables the solution of transportation problems in continuous two-dimensional solution spaces. It also allows traffic constraints due to traffic density to be taken into account, such as prioritizing busy routes over many smaller, similar routes. With these properties, DMK offers a practical method for solving real-world problems.
Nevertheless, the DMK algorithm reaches its limits in certain situations. Although many practical scenarios are defined in continuous space and therefore require a continuous solution, a discrete solution, for example in the form of a graph, is often required for implementation in the real world. In addition, the algorithm often requires a longer execution time, which can affect its practical applicability in time-critical scenarios. Furthermore, the DMK model has not yet been sufficiently tested in its practical application, especially in the field of machine learning, where OT problems play a major role.
In this thesis, we extend the aforementioned limitations of the DMK and show the practical application of the model in classical machine learning contexts. We first present different methods to transform the continuous solutions of the original two-dimensional space in which the DMK model is defined into practically realizable discrete solutions. In the process, we transform DMK solutions, which are available as density distributions, into graphs and hypergraphs that represent discrete solutions. Furthermore, we show that the DMK algorithm can be stopped before its termination without strongly affecting the solution quality. In time-critical practical applications, the DMK algorithm can thus find effective solutions in a short time by determining an optimal stopping time.
The second part of this thesis focuses on the practical applicability of the DMK model. Based on the theoretical findings described above, we develop algorithms for the practical use of a variant of the DMK model, the Graph-DMK algorithm, in various machine learning tasks. Specifically, we show how DMK can be used for the clustering of networks, the extraction of networks from images, and the classification of images. Our results show the potential of the DMK algorithm to successfully solve a variety of machine learning tasks.
Thus, this thesis extends the previous limits of the DMK model and demonstrates its practical application.