Experimental platform for hierarchical reinforcement learning

OData support
Supervisor:
Dr. Dobrowiecki Tadeusz Pawel
Department of Measurement and Information Systems

The subject of my thesis is the construction of a simulation framework, which can host the testing of hierarchical multi-agent reinforcement learning (HMARL) on tasks defined by the user. I present the theoretical background of HMARL, then I implement a cooperative version based on task decomposition. The tasks of the agents are arranged into a hierarchy according to the logic employed by the programmer, and in the case of certain subtasks, the agents may even communicate with each other. The communication cost can be adjusted, which allows the testing of its effect on agent behavior. The system assumes a two-dimensional „grid world” as the foundation for the agent environment, for which I implement a visualization interface that can be used to observe agent behavior, as well as querying charts and other performance statistics. I present the system in action by testing scenarios, including the task of controlling cleaner robots, and the classic taxi problem. The results of these experiments can be used to gauge the efficiency of the system.

Downloads

Please sign in to download the files of this thesis.