Seminar on Computational Learning and Adaptation

 
Learning to Drive

Jeffrey Forbes
Computer Science Division
UC Berkeley
jforbes@cs.berkeley.edu

Autonomous vehicle control presents artificial intelligence with significant challenges.  In driving, there is rarely a known optimal trajectory, rather the goal is to maximize general performance according to a given set of factors. Reinforcement learning is one method whereby the agent successively improves control policies through experience and feedback (i.e.,  reward) from the system.  Reinforcement learning techniques have shown some promise in solving complex control problems. However, these methods sometimes fall short in environments with continuous state and action spaces, such as driving. This talk presents a method for learning and maintaining a value function approximation using instance-based learning.  I propose a hierarchical model-based solution integrating a number of new methods from reinforcement learning to address vehicle control. The algorithm is effective on canonical control domains as well as more complex driving tasks.

Return to the seminar schedule