Road traffic demand estimation and traffic signal control

A transport network is a complex non-stationary open environment with multiple heterogeneous stochastic agents such as intersection controllers and road users. Changes in such a network may occur due to random demand fluctuations, supply degradation, or actions of agents. Thus, it makes the task of adaptive Traffic Light Control (TLC) extremely challenging. This paper is our attempt to analyze the complexity of TLC on a single isolated intersection by considering it from the perspective of supervised learning. We explicitly split the problem into two parts: demand estimation and traffic light control itself. For each task we train offline a model (Neural Network) on synthesized data. Our experiments show that even Shallow Neural Network (SNN) models can estimate and control traffic well on a single intersection. The only bottleneck of this method resides in the process of data generation and annotation - the number of simulations needed for data generation polynomially grows with the increase of number of intersections. This might prevent the current approach from being applied to transport networks with multiple intersections.