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 (Guangzhou) · DiDi Chuxing

* Equal contribution. † Corresponding author.

Paper Code
DispatchEvolve evolution path A mature engine is diagnosed, evolved locally, and integrated into a Pareto-feasible engine. Mature engine Diagnose Evolve locally Integrate globally Globally feasible objective space
Read the paper story

01

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.

We formulate the task as constrained multi-objective 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.

On frozen test logs from four cities, DispatchEvolve is the only evaluated method that improves all six objectives over the incumbent in every city. Only 25 of 147 complete-engine compositions improve all six objectives while satisfying the broadcast-rate constraint, confirming that local gains are not reliably additive.

Why

Whole-engine evolution stalls near a mature system

Direct program evolution assumes that nearby edits provide direction, the target is already known, and local improvements compose. Industrial dispatch violates all three assumptions.

01

Feedback is diluted

Narrow gains disappear in full-data replay, while broad edits often violate coupled objective guardrails.

02

Opportunities are hidden

A weak scenario may be exogenous; actionable gaps require trace evidence and policy attribution.

03

Local gains interact

Policies can reinforce, cancel, duplicate, or conflict once assembled into a complete engine.

DispatchEvolve composes improvements evolved across multiple local regions, moving beyond the short-edit neighborhood around the expert-tuned engine.

02

From trace-grounded opportunities to globally feasible engines

DispatchEvolve separates the problem into two connected search spaces: focused local evolution where the signal is visible, then full-engine integration where coupled constraints are measurable.

Evidence

  • 01 Decision traces
  • 02 Replay outcomes
  • 03 Editable policies
  • 04 Historical A/B records
Stage 1

Trace-Grounded Local Policy Evolution

  1. DiscoverPartition scenarios and expose objective gaps.
  2. AttributeJoin outcomes with policy-level decision traces.
  3. CritiqueReject opportunities that are unlikely to be actionable.
  4. EvolveEdit only related policies under local guardrails.

Output · locally validated policy candidates

Stage 2

Pareto-Guided Global Engine Integration

  1. RelateModel hard conflicts and soft interactions.
  2. ComposeUse archive context to propose complete engines.
  3. ReplayMeasure joint effects on full evolution data.
  4. ArchiveRetain feasible, non-dominated engines across rounds.

Output · globally feasible Pareto archive

A/B-ready
engine

The online-uplift assessment ranks only engines already proven feasible by offline replay. It never changes the objective vector or the Pareto relation.

03

Evidence across the evolution-to-deployment path

The only evaluated method to improve all six objectives in every offline city.

4 / 4 cities with six-objective offline feasibility
15 evaluations to improve all six objectives on average
25 / 147 complete-engine compositions globally feasible

Search trajectory

Offline evaluation efficiency

DispatchEvolve reaches six improved objectives on average after 15 engine evaluations; the strongest baseline reaches five after 30.

City A ablation · Avg. Δ

Every mechanism contributes

Full DispatchEvolve+0.616
w/o Combination LLM+0.542
w/o Opportunity Critic+0.498
w/o Trace Evidence+0.457
w/o Pareto References+0.398
w/o Candidate Relations+0.000

Without explicit candidate relations, independently evolved policies cannot be reliably integrated into a feasible engine.

Opportunity actionability

A gap is not yet an opportunity

Across 560 local searches, a larger scene performance gap does not predict whether evolution will succeed. The DPO-tuned Qwen3-8B critic reaches 63.2% decision accuracy, compared with 42.7% for the base model.

982
unique opportunities discovered
771
retained by the critic
59.8%
local-search success rate

17.0%

Why local gains are not additive

Pareto-archive context and global composition feasibility across full-engine evaluations.

25 / 147

complete-engine compositions satisfy every condition.

Local candidates can overlap in scenes, policies, and metrics. DispatchEvolve therefore never sums their isolated gains: each composition is installed into a runnable engine and evaluated against all six objectives plus the broadcast-rate constraint.

Richer Pareto context accompanies more feasible proposals, but it never replaces complete-engine replay.

04

From one trace to a globally balanced engine

A trace-localized ETA filter edit improves its target scene but carries an ETA trade-off. Global integration combines it with 15 complementary candidates and absorbs that local cost.

Policy-trace case study

Conditional relaxation of one ETA filter

Local candidate · paired scene replay

  • AR+0.756%
  • CR+0.570%
  • DCAA+1.090%
  • ETA−0.255%

Integrated engine · full replay

  • CR+0.325%
  • AR+0.751%
  • GMV+0.138%
  • ETA+0.285%
  • PCAA+0.148%
  • DCAA+0.526%

All values are direction-aligned: positive is favorable after accounting for whether higher or lower is preferred.

05

Deployment evidence

Four production A/B tests connect offline evolution to live marketplaces.

Selected relative changes from the incumbent engine. Stars denote nominal significance as reported by the platform.
City CR ↑ GMV ↑ AR ↑ DCAA ↓ ETA ↓
City A+0.72%*+1.32%***+0.49%−3.09%**+0.82%***
City B+0.58%+0.18%+0.51%+1.04%−0.71%
City E+1.07%**+0.94%**+0.97%***+1.98%−0.03%
City F+0.46%+0.48%+0.37%−0.82%+0.32%

Six nominally significant improvements and one nominally significant ETA degradation. City-level monitoring remains essential after deployment.

$103–115offline LLM cost per city
≤ 11h 37moffline evolution time
+0.68%–4.98%mean serving-latency change

No LLM inference, model loading, or additional service is required in the online dispatch path.

06

Explore the paper

PaperRead the full KDD 2027 ADS submission ↗ CodeImplementation and reproducibility artifacts ↗