Artificial intelligence for resource allocation tasks

Strauss, Niklas (2024): Artificial intelligence for resource allocation tasks. Dissertation, LMU München: Fakultät für Mathematik, Informatik und Statistik

Abstract

In recent years, innovations in artificial intelligence have led to advances in many different areas, ranging from natural language processing to computer vision. One area that is of special interest are resource allocation tasks. The field of resource allocation includes a diverse range of tasks. Oftentimes, resource allocation tasks are complex sequential decision making problems like portfolio optimization, the traveling officer problem, or redeploying ambulances to base stations. Deep reinforcement learning offers a way to solve these tasks. However, in order to achieve state-of-the-art performance, existing neural network architectures are not sufficient and, in this thesis, we propose several novel architectures. While spatial resource allocation tasks typically have discrete action spaces, some allocation tasks, like portfolio optimization, have continuous action spaces. The contributions in this thesis handle three types of allocation tasks: discrete resource allocation, resource collection, and continuous resource allocation with allocation constraints. First, we focus on discrete resource allocation and resource collection. More specifically, we look into different spatial resource allocation tasks, presenting a spatial-aware reinforcement learning-based approach for the traveling officer problem, a prominent resource collection task, achieving state-of-the-art performance. We also propose an approach for multi-agent stochastic resource collection featuring a novel neural network architecture. After that, we focus on dynamic ambulance redeployment. We develop a high-performance event-based simulator, conduct comparisons and benchmarks of existing approaches using real-world data. After that, we are the first to develop an approach for dynamically redeploying electric ambulances. Our method deploys the ambulance to the base station which minimizes the energy deficit. Afterward, we continue with continuous resource allocation tasks featuring simplex action spaces accompanied with allocation constraints. One of the most important continuous action resource allocation tasks is portfolio optimization, where a portfolio manager allocates its wealth across various assets in each time step over an investment horizon. In practice, these tasks often come with allocation constraints. We develop approaches to efficiently and effectively incorporate one and two allocation constraints using a decomposition of the simplex, allowing us to learn a policy using standard deep reinforcement learning approaches. Our methods never violate the constraints, even during training. Furthermore, we propose an approach capable of handling an arbitrary number of constraints by iteratively sampling actions in each dimension autoregressively, while utilizing linear programming to compute the action bounds. Our approach remains trainable using existing reinforcement learning algorithms.

• Dokumententyp: Dissertationen (Dissertation, LMU München)
• Themengebiete: 000 Allgemeines, Informatik, Informationswissenschaft ; 000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik
• Fakultäten: Fakultät für Mathematik, Informatik und Statistik
• Sprache der Hochschulschrift: Englisch
• Datum der mündlichen Prüfung: 20. Dezember 2024
• 1. Bericht­erstatter:in: Schubert, Matthias