Learning Optimal Control Strategies from Interactions with a PADAS

Abstract

International audienceThis paper addresses the problem of finding an optimal warning and intervention strategy (WIS) for a partially autonomous driver assistance system (PADAS). An optimal WIS here is defined as the minimizing the probability of collision with a leading vehicle while keeping the number of warnings and interventions as low as possible so as to not distract the driver. A novel approach to this problem is proposed in this paper. The optimal WIS will be considered as solving a sequential decision making problem. The adopted point of view comes from machine learning where the answer to optimal sequential decision making is the Reinforcement Learning (RL) paradigm