Application of Deep Reinforcement Learning for Measuring the Efficiency of Autonomous Vehicles under a Mixed-Traffic Condition in Non-Signalized Intersections

Abstract

The objective of this dissertation is to develop deep reinforcement learning for multiple autonomous vehicles under mixed traffic conditions in non-signalized junctions. To achieve the objective, firstly, the responsibility-sensitive safety-based partially observable Markov decision process model can be proposed to enhance the efficiency of one autonomous vehicle at a non-signalized intersection according to traffic safety guarantee, delay time, and smooth driving. Secondly, a deep reinforcement learning method evaluated the efficiency of multiple autonomous vehicles at a non-signalized intersection. This proposed method integrated reinforcement learning with multilayer perceptron algorithms. Additionally, a set of proximal policy optimization hyperparameters improved the performance of deep reinforcement learning training. Lastly, an advanced deep reinforcement learning was conducted to analyze the efficiency of multiple autonomous vehicles in the urban network with nine non-signalized intersections and propose a set of proximal policy optimization hyperparameters for enhancing deep reinforcement learning training's performance. The experimental results showed that an advanced deep reinforcement learning was getting a promising approach for optimizing self-driving behaviors under a mixed traffic condition regarding average reward, average speed, delay time, fuel consumptions, and emissions. A significant improvement in traffic perturbations was demonstrated as a higher the market penetration rate of autonomous vehicles.