Learning Distributed Strategies for Traac Control

In this paper, we cast the problem of managing traac ow in terms of a distributed collection of independent agents that adapt to their environment. We describe an evolutionary algorithm that learns strategies for lane selection, using local information, on a simulated highway that contains hundreds of agents. Experimental studies suggest that the learned controllers lead to better traac ow than ones constructed manually, and that the learned controllers are robust with respect to to blocked lanes and changes in the number of lanes on the highway.