A Deep Reinforcement Learning Approach for Proactive Traffic Incident Management

The growing demand for people and products transit and the enormous intensity of population in metropolitan areas have raised the need for suitable infrastructure and system development to manage the continually rising and varied traffic conditions effectively. Ramp metering is one of the traffic management techniques that employ signals placed at freeway on-ramps to actively control traffic accessing the freeway. Ramp metering has proven to be successful at reducing recurrent congestion. However, more than half of all congestion is caused by non-recurring events such as accidents, construction zones, adverse weather conditions, special events, and others. Intelligent systems and their theoretical foundations in artificial intelligence have advanced tremendously in recent years. As a result, their application in everyday living has lately grown in popularity. This paper aims to demonstrate the ability of a machine learning control method known as reinforcement learning to react to changing real-time traffic conditions and adjust while managing highway entrance access during incident time. The guidance was developed using a microscopic simulation model that considered non-recurring traffic events. It also considers event characteristics such as incident location, clearance time, and the number of lanes obstructed. The recommendations were created using sensitivity analysis and statistical experiments. The findings are encouraging, demonstrating that the machine learning control method can provide optimum coordinated management of freeway entry ramps.