Reinforcement Learning for MDP-based autonomous driving support system

The mobility sector has been revolutionized in the latest years. One of the most interesting aspects is the autonomous driving challenge that many big players in the automotive field are trying to solve. A good autonomous driving system must accomplish the driving task in a safe and effective way, to correctly replace the human behavior and possibly also to improve it. In this thesis, we propose different approaches to implement a high level autonomous driving support system, capable of understanding the state of the ego vehicle and suggesting, or directly applying, the most desirable behavior in different scenarios. The final model was developed using Reinforcement Learning to train an agent over a Markov Decision Process framework. This method provides a good trade-off between computational cost and versatility. The model has been simulated, tested and validated in a SIMULINK environment, while different key performance indicators have been used to highlight the suitability to handle different driving scenarios.