Machine Learning Seminar

Machine Learning Seminar

Abstract

We consider learning-based approaches to decision-making for behavioral planning in autonomous driving, that is, decisions at the scale of maneuvers (change lane, yield/take way, etc.). Given the critical nature of the task, we show the limitations of a naive application of the traditional Reinforcement Learning setting, and highlight the need for the evaluation and control of the level of risk, otherwise known as Safe Reinforcement Learning. We present two different approaches, namely Budgeted and Robust Reinforcement Learning, and introduce several novel algorithms in both the model-free and model-based families. Throughout this work, we put a strong emphasis on the practical efficiency and implementability of these methods, and in particular of planning algorithms which do not always reflect the insights of sample complexity analyses.

Date
Location
Lille, France