Dynamic restaurant ranking as demand control in on-demand meal delivery systems

Abstract

On-demand meal delivery (OMD) platforms influence customer choices by how restaurants are ranked in their app interface. Dynamically adjusting these rankings offers a powerful tool for real-time demand shaping. At the same time, ensuring efficient deliveries and fair courier workload distribution requires the platform to make coordinated decisions across both demand and supply sides, including restaurant ranking and courier routing. Since demand-shaping affects order distribution—which in turn impacts courier availability—jointly optimizing these decisions can enhance overall system performance. We propose a multi-stage optimization framework that determines restaurant rankings and courier routing decisions throughout the day while accounting for endogenous customer demand, congestion effects, the order bundling strategy, and the workload fairness of couriers. Equity is incorporated into dispatch through workload-based dominance constraints. We derive the structural properties that characterize when ranking adjustments improve a congestion-aware objective and use these properties to guide a learning-to-optimize algorithm. The proposed algorithm uses offline-trained machine learning models to predict whether a candidate solution will improve the objective and by how much. Dozens of candidate solutions are evaluated in parallel, without relying on commercial solvers. This enables real-time decision-making at scale. Additionally, equity-aware order-courier matching and couriers’ routing decisions are constructed using a fast insertion-based algorithm with dominance rules that balance efficiency and courier workload fairness. Computational experiments on a city-scale meal delivery network covering the entire city of Delft, the Netherlands, show that dynamic ranking considerably improves system performance relative to fixed-ranking and non-learning benchmarks, particularly under constrained courier capacity. Our method produces higher-quality solutions within seconds, and achieves at least a 7.7% improvement in the platform’s profit over the benchmark heuristic algorithm.