Feedback Control of Traffic Signal Network of Less Traffic Sensors by Help of Machine Learning

As a way of resolving vehicle congestion, there is a feedback control approach which models a traffic network as a discrete dynamical system and derives feedback gain for controlling green light times of each junction. Since the input is the sensory observed traffic flow of each link, and since the state equation models both the topology and the parameters of the network, it is effective for adaptive control of a wide area traffic in real-time. One of the essential factors in a state equation is the vehicles’ turning ratio at each junction. However, in a normal traffic sensor layout, it is impossible to directly measure this value in real-time, and values from traffic census are used. This paper is to propose a method that predicts this value in real-time through machine learning and gives more appropriate feedback control. Out idea is to find the turning ratio through probabilistic search by Reinforcement Learning referring to the degree of improvement of the entire traffic flow. At this moment we have finished formulation of the scheme and the verification for the performance by a traffic simulator is on the way. Keywords-component; Reinforcement Learning, Traffic light Control, Feedback Control, Discrete Dynamical Syste, Split