Development of reinforcement learning strategies for multi agent robotic systems

OData support
Dr. Harmati István
Department of Control Engineering and Information Technology

When applying multiagent robot systems, the way we achieve our goal may vary; in noisy or inconstant environments, the optimum may not even be identified. Achieving the same goal under different circumstances require a flexible approach from the applied strategy. Where repetitional situations occur, the opportunity to try multiple solutions appears, thus enabling machine and reinforcement learning. This may assist in problem solving, resulting in more optimal solutions over time. By finetuning the strategy and the task distribution among the agents, the optimal solution can also be reached in diverse environments.

The interaction with one or more groups of agents, let that be cooperative or competitive, creates a dynamic environment, thus games with multiagent teams where the opponent team or its strategy changes may call for a dynamic system that is capable of learning. With appropriate learning features, a strategy can achieve proper general problem solving properties that may mean a solid advantage over specialized strategies that are optimized for a specific task.

This thesis aims to theoretically and practically study different reinforcement learning methods, with special attention to their performance and usability with multiagent systems, and their application in a specific task, in this case controlling a robot soccer team.


Please sign in to download the files of this thesis.