This paper explains the attempted development of a deep reinforcement learning-based self-driving car system for a simulated, 3D environment. As a relatively new deep learning paradigm with a lot of potential, the interest in developing this system is to draw conclusions about the place for deep reinforcement learning in production-ready self-driving car systems. The deep reinforcement learning algorithm called double deep Q-learning, which uses a double deep Q-network with convolutional and simple recurrent layers, is used to steer the self-driving car. As such, the requisite foundational material, that of reinforcement learning and deep learning, are explored so as to make the developed system understandable. The results of the attempted development are surprisingly negative, and accordingly, the conclusions reflect this.


Fox, Nathan


Computer Science; Mathematics


Artificial Intelligence and Robotics | Geometry and Topology | Navigation, Guidance, Control, and Dynamics | Other Applied Mathematics | Probability | Statistical Theory | Theory and Algorithms


self-driving car, deep reinforcement learning, reinforcement learning, deep learning, neural networks, q-learning, kjv

Publication Date


Degree Granted

Bachelor of Arts

Document Type

Senior Independent Study Thesis



© Copyright 2019 Eric Michael Gabriel