Offline vs. Online Learning for Deep Driving from Demonstrations

We are studying alternative learning methods for Driving from Demonstrations.  The focus on offline learning dating back almost 30 years has been largely supplanted by online learning methods such as DAgger, which focus on difficult-to-learn areas of state space and require tedious and error-prone corrective feedback from human supervisors. Our initial results suggest that offline learning may be preferable for highly-expressive policy classes such as Deep Learning.