Deep Autonomous Agents Comparison for Self-driving Cars

Autonomous driving is one of the most challenging problems of the last decades. The development in recent years is mainly due to the continuous expansion of Artificial Intelligence. Nowadays, most self-driving systems use Deep Learning techniques. In recent years, however, thanks to the successful learning demonstrations of Atari games and AlphaGo by Google DeepMind, new frameworks based on Deep Reinforcement Learning are being developed. The objective is to combine the advantages of image processing and feature extraction of convolutional networks, and the learning process through the interaction of one or multiple agents with their environment. This work aims to deepen and explore these new methodologies applied to autonomous driving cars. In particular, we developed a framework for controlling a car in a simulated environment. The agent learns to drive within a neighborhood with constant speed, variable light conditions, and avoiding collisions with external objects. The proposed techniques are based on Double Deep Q-learning and Dueling Double Deep Q-learning. We implemented two variants of the algorithms: one trained from random weights and one exploiting the concepts of Transfer Learning. After a simulation study, the Dueling Double Deep Q-learning with Transfer Learning has showed promising performance.