Offline meta black-box optimization framework for intelligent traffic light management system

Abstract

This paper proposed a novel framework for effective intelligent traffic light management system. This research focuses on searching for an optimal phase combination and phase time allocation scheme for diverse traffic patterns adaptively. This framework first collect an offline meta dataset consists of different phase combination or phase time allocation scheme and corresponding traffic congestion measure across diverse traffic patterns. Then, it trains an Attentive Neural Process (ANP) to predict the congestion measure when deploying a certain traffic light scheme on various traffic patterns. Finally, it uses Bayesian optimization with the trained ANP as a surrogate model, to find an optimal scheme for unseen traffic pattern with a few number of online simulations. Extensive simulation-based experiments show that our framework surpasses prior methods. Furthermore, the suggested framework is deployed into real-world traffic light management system and makes a real improvement of traffic flow compared to original method.