Analysis of domain-side optimizations for deep reinforcement learning in the RoboCup environment
February 24, 2021
Bachelor’s thesis, Offenburg University
Abstract
With the team “magmaOffenburg”, the Offenburg University is participating in the international competition “RoboCup” in the 3D simulation league for soccer since 2009. To achieve a great result, the team heavily relies on using well-trained behaviors for their agents (e.g. kicking or walking). Since 2019, magmaOffenburg is also able to use deep reinforcement learning in order to further develop their behaviors. Using deep reinforcement learning, the team was already able to score usable results in kicking. However, there is still a lack of progress for learning to walk. In this thesis, the required optimizations on the domain side were tested and evaluated in order to get a better result at walking. This includes optimizing the observation space as well as the action space. Furthermore, a lot of optimizations for the reward function will also be tested and evaluated. The goal was to get a clarification of the influence of the various parameters and the applied techniques on walking in the RoboCup domain. Finally, it was possible to improve the running speed from little under a meter per second to up to 1.8 meters per second. The main reason for this improvement were the optimizations in the reward function.
BibTeX
@mastersthesis{Braun2021,
title = {Analyse domänenseitiger Optimierungen für Deep Reinforcement Learning in der RoboCup Umgebung},
author = {Hannes Braun},
year = {2021},
month = {feb},
school = {Hochschule Offenburg},
}