Machine learning for multi-jurisdictional optimal traffic corridor control

Urban traffic corridors are often controlled by more than one agency. Typically in North America, a state of provincial transportation department controls freeways while another agency at the municipal or city level controls the nearby arterials. While the different segments of the corridor fall under different jurisdictions, traffic and users know no boundaries and expect seamless service. Common lack of coordination amongst those authorities due to lack of means for information exchange and/or possible bureaucratic 'institutional grid-lock' could hinder the full potential of technically-possible integrated control. Such institutional gridlock and related lack of timely coordination amongst the different agencies involved can have a direct impact on traffic gridlock. One potential solution to this problem is through integrated automatic control under intelligent transportation systems (ITS). Advancements in ITS and communication technology have the potential to considerably reduce delay and congestion through an array of network-wide traffic control and management strategies that can seamlessly cross-jurisdictional boundaries. Perhaps two of the most promising such control tools for freeway corridors are traffic-responsive ramp metering and/or dynamic traffic diversion possibly using variable message signs (VMS). Technically, the use of these control methods separately might limit their potential usefulness. Therefore, integrated corridor control using ramp metering and VMS diversion simultaneously might be synergetic and beneficial. Motivated by the above problem and potential solution approach, the aim of the research presented in this paper is to develop a self-learning adaptive integrated freeway-arterial corridor control for both recurring and non-recurring congestion. The paper introduces the use of reinforcement learning, an Artificial Intelligence method for machine learning, to provide optimal control using ramp metering and VMS routing in an integrated agent for a freeway-arterial corridor. Reinforcement learning is an approach whereby the control agent directly learns optimal strategies via feedback reward signals from its environment. A simple but powerful reinforcement learning method known as Q-learning is used. Results from an elaborate simulation study on a key corridor in Toronto are very encouraging and discussed in the paper.