2019/04/17 by Filipe Rodrigues, Rodrigues, Filipe, Carlos Lima Azevedo +1
Engineering · Social Sciences · #FOS: Computer and information sciences #FOS: Electrical engineering #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Systems and Control (eess.SY) #Traffic Prediction and Management Techniques #Traffic control and management #Transportation Planning and Optimization #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.1904.08353
openalex publication_date 2019/04/17 · openalex created_date 2022/07/29 · openalex updated_date 2026/07/28
Reinforcement learning (RL) constitutes a promising solution for alleviating\nthe problem of traffic congestion. In particular, deep RL algorithms have been\nshown to produce adaptive traffic signal controllers that outperform\nconventional systems. However, in order to be reliable in highly dynamic urban\nareas, such controllers need to be robust with the respect to a series of\nexogenous sources of uncertainty. In this paper, we develop an open-source\ncallback-based framework for promoting the flexible evaluation of different\ndeep RL configurations under a traffic simulation environment. With this\nframework, we investigate how deep RL-based adaptive traffic controllers\nperform under different scenarios, namely under demand surges caused by special\nevents, capacity reductions from incidents and sensor failures. We extract\nseveral key insights for the development of robust deep RL algorithms for\ntraffic control and propose concrete designs to mitigate the impact of the\nconsidered exogenous uncertainties.\n