The main objective of this project is to develop Multi-Agent Reinforcement Learning (MARL) techniques for large state spaces. For this purpose we will make use of single-agent techniques, which are known to work well in such environments. We can divide this objective in the following sub problems:
- Analysis of large state spaces. The first objective of this research is to obtain a clearer picture of the problems that arise when dealing with MAS and large state spaces. Therefore we will first look at several realistic problems and analyse how is dealt with large state spaces up to now. The mapping of these issues will form the starting point for our next objective.
- Extend single-agent techniques to MAS. The second and main objective of this research project consists in extending single-agent algorithms, which are known to perform well in large state spaces, to situations where multiple agents are present. Learning in a MAS is even more challenging because the environment the agents experience is non-stationary. Our techniques should also be able to cope with the additional complexity that comes to surface if agents have conflicting goals. We will gradually increase the difficulty. I.e. we will begin with a simple environment with a relatively small state space, before moving on to more complex situations.
- Analysis and evaluation. The third objective of this research is to analyse the performance and quality of the developed multi-agent techniques. We will look at convergence speed as well as the quality of the solution. We will analyse the developed techniques with different settings in order to show the strengths and weaknesses of the various techniques.