Keynote Session 4 — Alexandre Bayen

Abstract

This talk inves­ti­gates Lagrangian (mobile) con­trol of traf­fic flow at local scale (vehic­u­lar lev­el). The ques­tion of how self-dri­v­ing vehi­cles will change traf­fic flow pat­terns is inves­ti­gat­ed. We describe approach­es based on deep rein­force­ment learn­ing pre­sent­ed in the con­text of enabling mixed-auton­o­my mobil­i­ty. The talk explores the grad­ual and com­plex inte­gra­tion of auto­mat­ed vehi­cles into the exist­ing traf­fic sys­tem. We present the poten­tial impact of a small frac­tion of auto­mat­ed vehi­cles on low-lev­­el traf­fic flow dynam­ics, using nov­el tech­niques in mod­­el-free deep rein­force­ment learn­ing, in which the auto­mat­ed vehi­cles act as mobile (Lagrangian) con­trollers to traf­fic flow. Illus­tra­tive exam­ples will be pre­sent­ed in the con­text of a new open-source com­pu­ta­tion­al plat­form called FLOW, which inte­grates state of the art microsim­u­la­tion tools with deep-RL libraries on AWS EC2. Inter­est­ing behav­ior of mixed auton­o­my traf­fic will be revealed in the con­text of emer­gent behav­ior of traf­fic: https://flow-project.github.io/