Background
ABSTRACT:
Dynamic control of traffic light systems has the potential to foster great gains in the efficiency of a congested motor vehicle system. Static controls such as timed lights do not consider the state of the system and thus are not able to react to it. Most commonly, lights only consider whether or not there is a car waiting at a light and do not take into account how many of them there are or the dynamics of each car within the system. Reinforcement Learning has the capability to bridge these issues. In our research, a four-road four-intersection simulation was developed with the purpose of using it to implement and improve a reinforcement learning control system and compare it with a static model. Our algorithm realized significant reductions in traveling time for both congested and sparse systems, showing that it is almost universally superior in its efficacy to the standard static implementation.
Dynamic control of traffic light systems has the potential to foster great gains in the efficiency of a congested motor vehicle system. Static controls such as timed lights do not consider the state of the system and thus are not able to react to it. Most commonly, lights only consider whether or not there is a car waiting at a light and do not take into account how many of them there are or the dynamics of each car within the system. Reinforcement Learning has the capability to bridge these issues. In our research, a four-road four-intersection simulation was developed with the purpose of using it to implement and improve a reinforcement learning control system and compare it with a static model. Our algorithm realized significant reductions in traveling time for both congested and sparse systems, showing that it is almost universally superior in its efficacy to the standard static implementation.