Policy Gradient Methods for Automated Driving

Automated driving requires designing a system capable of maintaining safety while simultaneously maintaining passenger comfort. Models of highway driving have high dimensionality and stochasticity traditionally specifying the histories for a large, varying number of agents in a continuous state and action space. Traditional value-based reinforcement learning methods require exponential time in the size of the state space and thus often require coarse state and action space discretization to remain tractable. Policy gradient methods instead maintain constant complexity proportional to the complexity of their parameterization and are thus candiates for use in such applications if suitable parameterizations can be found. This project explores several policy gradient methods against the problem of optimal autonomous highway driving framed as a Markov decision process.