Diplom- und Master-Arbeiten (eigene und betreute):
D. Bechtold:
"Self-Learning Embedded System";
Betreuer/in(nen): A. Jantsch, A. Wendt;
E384,
2019;
Abschlussprüfung: 10.04.2019.
Kurzfassung englisch:
In recent years, interest in self-learning methods has increased significantly. A driving factor was the growing computing power, which first enabled machines to carry out such computationally intensive methods. Nowadays, machine learning is used in many areas,
such as prediction, pattern recognition, and anomaly detection.
In this thesis, a self learning embedded system (SLES) is supposed to learn solving tasks completely independently and with as little prior knowledge about itself, the task and the environment, as possible. The learning process is guided by a reward signal, which
punishes or rewards the performed actions. Our main focus is on the task of surviving as long as possible. For this purpose charging stations must be located. Subsequently, they must be approached properly to allow a successful charging of the battery. In order to enable the independent learning of tasks, various methods from the
field of Reinforcement Learning (RL), in particular from Q-learning, are used. In addition, several replay memories and exploration methods are implemented and modified. Further, completely new approaches and ideas are realized with the aim of achieving better results. Evaluations help to find the most appropriate methods for our problem. Finally, they are tested in a simulation environment to ensure that they can be applied to the final hardware without significant changes. With the help of evaluations and simulations, this work shows the entire process, from the selection of the methods to the determination of which process parameters to use for the final SLES. In the future, it is planned to improve the selected methods, to reduce the memory requirements and to handle other tasks besides survival. In addition, the simulation model should be improved to be even closer to the real model.
Schlagworte:
Reinforcement Learning, Machine Learning, Deep Neural Networks
Elektronische Version der Publikation:
https://publik.tuwien.ac.at/files/publik_279637.pdf