A methodology for scheduling within-day roadway work zones using deep neural networks and active learning

City infrastructure agencies routinely implement road projects that address various elements of urban infrastructure. The majority of these projects are short-term in nature (e.g., utility repair), as they are completed in a few hours within 8:00 a.m. to 5:00 p.m. of a workday. The implementation of these projects during working hours, in spite of the inconvenience imposed on road users, helps the agency avoid extra labor costs associated with nonregular working hours. Careful scheduling of these projects can prevent unduly increased travel delays (road users' interest) while keeping project costs low (the agency's interest). This study presents a bi-level framework for scheduling short-term urban road projects that analyzes the implicit tradeoffs between the two stakeholders' interests. The upper-level model establishes the optimal schedule considering the project characteristics, such as cost and duration. The lower-level model captures the dynamic user equilibrium conditions that yield the road users' path and departure time choices. The bi-level model is a mixed-integer program with nonlinear constraints. Recognizing the relatively low efficiency of traditional solution methods, this paper proposes a deep-neural-network-ensemble-assisted active learning (DN2EA2L) algorithm and adopts a fixed-point algorithm for solving the bi-level model. The numerical experiment uses the Sioux Falls network to demonstrate the efficiency of the DN2EA2L, compared to conventional metaheuristic methods. It is shown that travel time increases due to the project implementation during the peak hours can outweigh the agency's saving in wage costs. Further, it is shown that a significant reduction in the road users' value of time enables the agency to schedule more projects during regular working hours.