Intersection management in fully autonomous driving era : Attention based multi agent reinforcement learning

Abstract

For the last decade, AV technology has arisen as a countermeasure to urbanization-related traffic congestion and traffic safety issues. Along with different improvements to the transportation system, the age of fully autonomous driving is anticipated to be implemented rapidly by 2035, alongside technical advancements. Unsignalized intersections by communication between cars is an effective means of enhancing traffic flow efficiency. However, it is difficult to determine the intention of each vehicle’s driving route, so the possibility of an accident is higher. Therefore the likelihood of an accident is greater than on conventional roads, and blind spots can compromise driving efficiency and safety. Consequently, in this thesis, we use a rule-based collision prediction algorithm and an attention mechanism-based multi-agent reinforcement learning model to identify the location and speed between each vehicle in real time and present a safe and efficient unsignalized intersection management methodology by means of a cooperative multi-agent reinforcement learning design. The two lane unsignalized intersection and circular intersection environments were constructed using the OpenAI Gym-based simulator, and model parameters were adjusted according on the circumstance. By learning rule-based collision prediction algorithms and attention-based multi-agent reinforcement learning models, we create a safe and effective fully autonomous driving-based unsignalized intersection management model by minimizing average speed and accident rate in a variety of traffic situations.