Auto-Driving Policies in Highway based on Distributional Deep Reinforcement Learning

Autonomous vehicles need to make intelligent decisions in complex environments such as highways or even less organized environments within cities. Driver Assistance Systems are intelligent systems in autonomous vehicles that assist drivers in such environments. In this article, we aim to employ an intelligent supervisor by means of reinforcement learning algorithms to adopt appropriate driving policies in relation to vehicle transmission on highways. To make the real situation similar to our problem and take into account the probabilistic situation of the vehicle movements, we utilize the Implicit Quantile Network, which is suitable for stochastic environments. Implicit Quantile Network is adopted so that we can estimate the full return distribution rather than a single expected return value. The quantile function with differentiable functions can also be calculated by the Implicit Quantile Network. Different training and validation experiments have been performed in a highway driving simulator developed in the unity environment to demonstrate the performance of the proposed approach. The results verify that the proposed method can select the best driving policies for driving scenarios compared to the state-of-the-art studies and can deliver better performance in terms of both speed and lane changing metrics.