End-to-End Reinforcement Learning for Self-driving Car

Most of the current self-driving cars make use of multiple algorithms to drive. Furthermore, most of the approaches use supervised learning to train a model to drive the car autonomously. This approach leads to human bias being incorporated into the model. We implement the Deep Q-Learning algorithm to control a simulated car, end-to-end, autonomously. The algorithm is based on reinforcement learning which teaches machines what to do through interactions with the environment. The application of reinforcement learning for driving is of high relevance as it is highly dependent on interactions with the environment. Our model incorporates a CNN as the deep Q network. The system was tested on an open-source car-racing simulator called TORCS. The Deep Q-Learning approach allows the system to be more efficient and robust than a system that has been trained solely through supervised training. Our simulation results show that the system is able to drive autonomously and maneuver complex curves.