Intelligent Delivery System Architecture and Optimization at Meituan

Online and offline business constraints in many industries have created abundant opportunities for operations‑research techniques. Meituan’s intelligent delivery system uses such techniques as a core technology across various business scenarios.

The system is organized into three layers: a foundational structural‑optimization layer (network and capacity planning), a market‑adjustment layer (pricing and marketing), and a real‑time matching layer (dispatch). Supporting subsystems for data collection, perception, forecasting, machine‑learning, IoT, and LBS provide accurate parameters for optimization.

Smart Area Planning – The goal is to reduce the average travel distance per order by optimizing merchant aggregation, order clustering, and the deviation between order and merchant centroids. Constraints include limits on order volume per area, non‑overlapping regions, and full coverage of all merchants. The solution combines multi‑objective optimization, road‑network‑based polygon generation, and simulation to evaluate the impact. A pilot showed a 5% reduction in average rider travel distance, saving over 100 m per order.

Intelligent Rider Scheduling – To handle 24‑hour service while respecting labor regulations, riders are grouped and assigned fixed shift windows. Decision variables are discretized into 30‑minute slots, and the objective is to maximize the proportion of time slots where capacity meets order demand. A constraint‑driven heuristic with local search yields millisecond‑level solutions and reduces scheduling time by about two hours per site.

Rider Route Planning – The problem is a large‑scale, constrained vehicle‑routing task where a rider must serve many orders. It is modeled as a pipeline‑scheduling problem, converting each order into a job with pickup and delivery operations. A deterministic, knowledge‑driven heuristic replaces random iterative methods, achieving a 70% speed‑up while maintaining solution quality.

Order‑Level Intelligent Dispatch – Dispatch is formulated as a combinatorial optimization and Markov decision process. Challenges include ultra‑high performance (solving >10⁴ orders against >10⁴ riders in seconds), dynamic environments, and stochastic factors such as merchant preparation time. Solutions involve robust optimization, stochastic programming, and learning‑based pruning of the search space.

The Meituan delivery team continues to explore operations research, machine learning, reinforcement learning, and data‑mining to improve these systems.