Reinforcement learning for autonomous vehicles

Autonomous vehicle control presents a significant challenge for artificial intelligence and control theory. The act of driving is best modeled as a series of sequential decisions made with occasional feedback from the environment. Reinforcement learning is one method whereby the agent successively improves control policies through experience and feedback from the system. Reinforcement learning techniques have shown some promise in solving complex control problems. However, these methods sometimes fall short in environments requiring continual operation and with continuous state and action spaces, such as driving. This dissertation argues that reinforcement learning utilizing stored instances of past observations as value estimates is an effective and practical means of controlling dynamical systems such as autonomous vehicles. I present the results of the learning algorithm evaluated on canonical control domains as well as automobile control tasks.