Surrogate-based Real-time Curbside Management for Ride-hailing and Delivery Operations

The present work investigates surrogate model-based optimization for real-time curbside traffic management operations. An optimization problem is formulated to minimize the congestion on roadway segments caused by vehicles stopping on the segment (e.g., ride-hailing or delivery operations) and implemented in a model predictive control framework. A hybrid simulation approach where main traffic flows interact with individually modeled stopping vehicles is adopted. Due to its non-linearity, the optimization problem is coupled with a meta-heuristic. However, because simulations are time expensive and hence unsuitable for real-time control, a trained surrogate model that takes the decision variables as inputs and approximates the objective function is employed to replace the simulation within the meta-heuristic algorithm. Several modeling techniques (i.e., linear regression, polynomial regression, neural network, radial basis network, regression tree ensemble, and Gaussian process regression) are compared based on their accuracy in reproducing solutions to the problem and computational tractability for real-time control under different configurations of simulation parameters. It is found that Gaussian process regression is the most suited for use as a surrogate model for the given problem. Finally, a realistic application with multiple ride-hailing vehicle operations is presented. The proposed approach for controlling the stop positions of vehicles is able to achieve an improvement of 20.65% over the uncontrolled case. The example shows the potential of the proposed approach in reducing the negative impacts of stopping vehicles and favorable computational properties.