Thesis (M.Sc.) - Cairo University -Faculty of Engineering - Department of Electronics and Communications

Autonomous Driving is one of the difficult problems faced the automotive applications. This is due to many corner cases which can be formulated in the unexpected behavior of the autonomous vehicles during the interaction with the other vehicles. The presented work used the Reinforcement Learning field, a strong Artificial Intelligence paradigm that teaches machines through the environment interaction and learning from their mistakes, in order to reach to having an Autonomous Driving vehicle. This work compared between two main categories: Discrete Action Algorithms like: Q-Learning, Double Q-Learning Algorithms, and Continuous Action Algorithms like: Deep Deterministic Policy Gradient (DDPG) Algorithm. It was proven as expected that Continuous Action Algorithms have better performance, so we applied some enhancements over the DDPG algorithm like solving the Long learning time which faced all machine learning problems. These enhancements were depending on disabling some restricted conditions and compensating them with the Reward term. The proposed work depends on Simulator called TORCS to reach to our aim

Issued also as CD