Deep Reinforcement Learning For Control

Autonomous cars must be capable to operate in various conditions and learn from unforeseen scenarios. Driving a car with a human driver may be a challenging undertaking. As a result, autonomous driving seeks to reduce hazards in comparison to human drivers. Furthermore, autonomous driving is difficult in terms of the outcomes and safety judgments that must be taken. In this thesis work, a method using deep reinforcement learning to train a controller with proper driving behavior has been proposed. In essence, the method is to use a reward-based learning environment to watch how the agent makes decisions. Potential actions must be taken based on prior experiences using a trial and error process. However, determining the essential behavioral outputs for autonomous driving vehicle systems or selecting the optimal output features to learn from them is not easy. Deep Neural Networks were chosen as function estimators because of their capacity to solve the complexity of high-dimensional system issues. As a consequence, the agent is expected to have trained behaviors and navigation without crashing. The complete project is carried out in the CARLA simulator to determine how to operate in discrete action space using Deep Reinforcement Learning (DRL) algorithms. Gathering and evaluating a large amount of data is time and effortintensive. Learning a model in a virtual environment might potentially fail to generalize to the actual world. As a result, the simulation environment makes it possible to collect massive training datasets. Improving learning driving policies can be adopted fast in the actual world. To generate the visual simulation in the simulator, the Python programming language is employed. The improved algorithm will help encourage the real-world implementation of DRL in many autonomous driving applications