Comprehensive Automated Driving Maneuvers under a Non-Signalized Intersection Adopting Deep Reinforcement Learning

Abstract

Automated driving systems have become a potential approach to mitigating collisions, emissions, and human errors in mixed-traffic environments. This study proposes the use of a deep reinforcement learning method to verify the effects of comprehensive automated vehicle movements at a non-signalized intersection according to training policy and measures of effectiveness. This method integrates multilayer perceptron and partially observable Markov decision process algorithms to generate a proper decision-making algorithm for automated vehicles. This study also evaluates the efficiency of proximal policy optimization hyperparameters for the performance of the training process. Firstly, we set initial parameters and create simulation scenarios. Secondly, the SUMO simulator executes and exports observations. Thirdly, the Flow tool transfers these observations into the states of reinforcement learning agents. Next, the multilayer perceptron algorithm trains the input data and updates policies to generate the proper actions. Finally, this training checks the termination and iteration process. These proposed experiments not only increase the speeds of vehicles but also decrease the emissions at a higher market penetration rate and a lower traffic volume. We demonstrate that the fully autonomous condition increased the average speed 1.49 times compared to the entirely human-driven experiment.