Multi Agent Reinforcement Learning for smart mobility and traffic scenarios

Autonomous Driving is one of the most fascinating and stimulating field in modern engineering. While some partial autonomous cars already exist in the industrial market, they are far from being completely independent in any situation. One of the things that these vehicles lack the most is the ability of handling traffic scenarios and situations in which interaction with other road users is required. The purpose of this work is that of investigating learning techniques that could be exploited in order to face the challenge described above. The focus will be put on the Multi-Agent Reinforcement Learning (MARL) paradigm, which seems particularly appropriate to address this kind of problem, given its ability to learn and solve complex tasks without any prior knowledge requirement. The MARL paradigm has its roots both in the classical single agent Reinforcement Learning setup, but also in the Game Theory field. For this reason, after an introduction to the problem and some literature review of related works, the project will begin with an introduction to those two topics. After that the MARL paradigm will be analyzed, focusing both on the theoretical aspects and on the algorithmic point of view. The thesis will then proceed with an experimental section, where some of the state-of-the-art MARL algorithms will be adapted to the Autonomous Driving setup and tested, making use of a simulator, called SMARTS, specifically developed for this purpose. To conclude this work the results obtained in simulation will be analyzed and discussed, and also some ideas for future development will be presented.