Uncertainty-Aware Reinforcement Learning for Safe Control of Autonomous Vehicles in Signalized Intersections

This paper proposes a reinforcement learning approach for the control of autonomous vehicles at signalized intersections. The proposed method is a modified version of the Q-learning approach that takes into account the risky scenarios that might arise in the control of an autonomous vehicle due to the inherent uncertainties in the system. The proposed algorithm enables robust and risk-aware decision-making in uncertain and sensitive environments. The proposed algorithm is evaluated in a simulated autonomous vehicle scenario, where it outperforms the standard Q-learning in terms of safety.