Cognitive Reinforcement Learning For Autonomous Driving

Most existing reinforcement learning methods are combined with deep neural networks in autonomous driving. There is a well-known trouble named ‘black-box’ of deep neural networks, which lacks of interpretability, occurring in the decision-making process. That will affect the safety of autonomous vehicle. In this work, we propose a cognitive reinforcement learning framework. This framework illustrates a cognitive model that transfers the state space to cognitive signals. Subsequencely, based on these signals, the model simulates the human decision-making process. As a result, the outcome of decision provides reward to the reinforcement learning. The computational experiments conducted in CARLA demonstrate that our framework performs equally well as the conventional reinforcement learning methods and provides interpretability under the same circumstance in the reinforcement learning.