Optimizing Hysteresis-Based Reinforcement Learning to Improve Collision Avoidance for Self-Driving Vehicles

This paper considers an optimization for fully automated driving systems. We analyze and optimize the performance of two different types of hysteresis-based feedback mechanisms based on noisy sensor inputs. We show how hysteresis can be used in conjunction with Reinforcement Learning (RL) algorithms to make them more resilient to noise. Our results show that by using hysteresis-based RL, we reduce the likelihood of collisions by over 50% in comparison to non-hysteresis-based RL.