Reinforcement Learning pro ovládání robotů

Abstract

The goal of this paper is to find a suitable way to design and test Reinforcement Learning methods for the robotics domain. This method will then be suitably described so that, based on it, the reader can independently design his/her own Reinforcement Learning tasks for robotics, in a comprehensive way, from the task design and model to the actual implementation and testing. As a result, the work should also find practical applications in academia, for example for teaching. Included is an extensive theoretical background that introduces the reader to the topic of Reinforcement Learning in a broader context, both from the perspective of artificial intelligence and from the perspective of neuroscience and psychology. Here, currently used and historical methods of Reinforcement Learning are also outlined and explained. A survey of existing solutions is presented, followed by a selection and detailed description of the most suitable one. Thus, it presents the reader with an extract of information on how to work with the environment and how to conduct their own experiments. To this end, the thesis includes one demonstrative experiment. A prerequisite for the solution is the disposition of powerful computing resources, especially in terms of graphical performance.