A Machine Learning Method for Dynamic Traffic Control and Guidance on Freeway Networks

A distributed approach to reinforcement learning in tasks of ramp metering and dynamic route guidance is presented. The problem domain, a freeway integration control application, is formulated as a distributed reinforcement learning problem. The DRL approach was implemented via a multi-agent control architecture where the decision agent was assigned to each of the on-ramp or VMS. The return of each agent is simultaneously updating a single shared policy. The control strategypsilas efficiency is demonstrated through its application to the simple freeway network. Analyses of simulation results using this approach show significant improvement over traditional local control, especially for the case of large traffic demand. Using the DRL approach, the TTS of the Network has been reduced by 20% under the heavy demands.