KDD 2027 ADS Submission

DispatchEvolve: Autonomous Policy Evolution for Industrial Ride-Hailing Dispatch Engines with Multi-Objective Constraints

Zirui Yuan^*, Tengfei Lyu^*, Kai Wan, Xu Liu, Zhui Sun, Zihao Lu, Li Ma, Hao Liu^†

HKUST(GZ) Didichuxing Co. Ltd.

^* Equal contribution. Work done during internship at Didichuxing Co. Ltd. ^† Corresponding author.

Method Overview

DispatchEvolve's two-stage framework: trace-grounded local policy evolution followed by Pareto-guided global engine integration — **Figure 2.** Stage I discovers and locally evolves trace-grounded opportunities. Stage II assembles and validates complete engines, maintains a cross-round Pareto archive, and selects the next incumbent or final A/B-ready engine.

Abstract

Ride-hailing dispatch engines must adapt continually to changing demand, supply, and driver behavior while protecting several coupled business objectives. Production improvement, however, still relies heavily on experts to identify a scenario, propose a policy change, validate it by offline replay, and advance it to an online A/B test. Automating this loop requires more than code generation. Around a mature engine, whole-engine edits provide little informative feedback, high-value scenarios and their responsible policies are not specified in advance, and locally beneficial edits may conflict when combined. We formulate the task as constrained multiobjective dispatch engine evolution and propose DispatchEvolve. Trace-Grounded Local Policy Evolution constructs scenario-specific opportunities from decision traces and replay outcomes, filters them with an opportunity critic, and evolves only the relevant policies under local guardrails. Pareto-Guided Global Engine Integration models relations among the resulting candidates, evaluates their combinations as complete engines, and retains globally feasible, non-dominated improvements in a cross-round archive. An assessment trained on historical A/B records is used only to rank archived engines for the next online test.

Across held-out replay logs from four cities, DispatchEvolve is the only evaluated method to improve all six objectives over the incumbent in every city. Across cities, it improves all six objectives on average after 15 engine evaluations, whereas the best-performing baseline improves only five after 30. Four production A/B tests show positive estimated completion-rate and GMV changes in every city, while also demonstrating that offline feasibility does not guarantee online non-regression and must be followed by city-level monitoring.

Contributions

1
We formulate industrial dispatch improvement as constrained, six-objective program evolution under hard operational guardrails.
2
We connect trace-grounded local search with relation-aware, Pareto-guided integration of complete dispatch engines.
3
We validate the complete workflow on held-out replay logs from four cities and in four production A/B tests.

Offline Results

DispatchEvolve improves all six objectives over the incumbent in every evaluated city.

4 / 4cities improve all six objectives

15evaluations to reach all six on average

0.350%macro FAI, versus 0.088% for the strongest baseline

Table 1 from the paper showing held-out offline performance across four Brazilian cities — **Table 1.** Direction-aligned relative changes from the incumbent engine. FAI is zero unless all six objectives strictly improve and the broadcast-rate constraint is satisfied.

City A component ablation measured by FAI — **Figure 3.** Full DispatchEvolve reaches 0.616% FAI; removing candidate relations reduces FAI to zero.

Component evidence

Each part of the two-stage workflow carries measurable value.

Replacing the fine-tuned Critic with Gemini 3 Flash reduces FAI to 0.364%. Removing trace evidence or the Opportunity Critic lowers it to 0.457% and 0.498%, while withholding Pareto references reduces it to 0.398%.

Opportunity outcomes and Critic accuracy from Figure 4 of the paper — **Figure 4.** The DPO-tuned Opportunity Critic reaches 63.2% accuracy.

Engine feasibility by Pareto archive size from Figure 5 of the paper — **Figure 5.** Feasibility rises from 8.3% to 42.9% as the archive grows from 1–4 to 12–14 entries.

982 discovered 771 retained 560 searched 335 successful 59.8% success

Production A/B Tests

Four city-level tests evaluate engines selected after offline feasibility screening.

Table 2 from the paper showing production A/B-test changes across four cities — **Table 2.** Estimated completion-rate and GMV changes are positive in every city; bold entries denote nominal significance.

6 nominally significant improvements

+1.32% largest GMV gain

−3.09% largest DCAA reduction

City A also shows a nominally significant 0.82% ETA regression, reinforcing that offline feasibility is a disciplined promotion criterion—not a guarantee of online non-regression.

Deployment practicality

One complete matched-budget run costs USD 103.32–115.37 per city and finishes within 11 hours 37 minutes. Mean serving latency changes by 0.68%–4.98%, with a 3.09% macro average.