Understanding Didi's Ride‑Hailing Dispatch Algorithms: Challenges and Strategies

Didi's dispatch algorithm, often perceived as mysterious, has evolved from simple taxi hailing to a sophisticated matching system that handles tens of millions of daily requests. This article explains how the problem is analyzed and modeled, the algorithmic challenges involved, and the typical dispatch strategies used.

Why a Better Dispatch Algorithm Is Needed

Early ride‑hailing required passengers to wait for a taxi that might never appear. The introduction of Didi’s platform increased passenger certainty, but the original greedy, on‑demand matching ("first‑come‑first‑served") limited the response rate. After launching the fast‑car service in 2015, a shift to a dispatch ("assignment") model raised the response rate by over 20 percentage points, reaching 90 %+ during peak periods. Similar dispatch mechanisms are now core to other on‑demand services such as Uber, Lyft, food delivery, and logistics.

Dispatch Problem Overview

At any moment there are N passenger orders and M available drivers. The platform knows each driver’s real‑time location and the navigation distance to each order. The goal is to assign drivers to orders while minimizing overall pickup time and respecting business rules.

Simple Cases

1 driver & 1 order : The assignment seems trivial, but rules (e.g., driver availability, vehicle type) may prevent a direct match.

1 order & 2 drivers : If distances are equal, the system prefers the driver with a higher service score. If one driver is closer, the closest driver is chosen, unless timing considerations (e.g., a slightly farther driver becomes available earlier) suggest a different choice.

Business Rules (examples)

Rule A: Fast‑car drivers cannot take premium‑car orders.

Rule B: Ensure drivers do not enter restricted zones after accepting an order.

Rule C: Filter out drivers whose real‑time destination is far from the order.

Rule D: Do not send real‑time orders to drivers who have only accepted pre‑booked rides.

Rule E: An order is sent to a driver only once.

…

From Greedy to Global Matching

Purely greedy dispatch (assigning each order immediately to the nearest driver) fails to achieve global optimality because orders and drivers arrive at different times. Batch Matching (also called delayed or centralized matching) collects orders and driver states over a short window (a few seconds) and solves a bipartite‑graph matching problem to minimize total pickup distance. This approach can significantly reduce average waiting time compared to pure proximity‑based assignment.

Supply‑Demand Prediction Based Dispatch

Even batch matching cannot fully address the latency constraint: users expect a response within a few seconds. By forecasting future order and driver distributions, the system can proactively hold drivers in high‑demand zones or prioritize certain matches, further improving efficiency.

Chain Dispatch

Chain dispatch assigns an order to a driver who is about to finish a current ride, provided the driver’s destination is close to the new order’s pickup location. This reduces idle time and shortens response latency, especially during peak periods.

Advanced Techniques

The core challenges—uncertain future supply/demand, spatio‑temporal dynamics, and user behavior—are tackled with deep learning models for demand forecasting and reinforcement learning for long‑term reward optimization. The system also integrates multi‑service resources (e.g., fast‑car, premium‑car) and balances short‑term versus long‑term objectives across passengers, drivers, and the platform.

Summary

Didi’s dispatch system processes over 30 million daily ride requests, handling more than 60 000 requests per minute during peaks. Recent algorithmic improvements have increased the number of satisfied passengers by over one million per day. Ongoing research focuses on fairness, supply‑demand prediction, reinforcement learning, and large‑scale optimization to further enhance the ride‑hailing experience.