Learning dispatch structure and driver sequencing preferences from historical operations.
CUSTOMER DECISION / APPLIED RESEARCH · REAL-DATA VALIDATED ·
HISTORICAL OPERATIONAL DATA
Decision question: What structure in historical dispatch decisions is stable enough to learn, reuse, and combine with explicit optimization?
Human dispatchers do not optimize only geometric distance. Their decisions can reflect:
- delivery zones;
- vehicle capacity;
- habitual driver areas;
- road / operational knowledge;
- sequence preferences;
- customer and time constraints.
This repository studies whether those patterns can be learned from historical dispatch data.
It should not be positioned as "AI beats VRP." The more accurate framing is:
Learn useful behavioral structure from operations, then decide which parts should become constraints, priors, or objectives in a deployable dispatch engine.
Historical orders + dispatch assignments
↓
Stage 1 — learn grouping structure
↓
capacity-feasible order groups
↓
Stage 2 — learn sequence preferences
↓
driver / zone ordering prior
↓
future hybrid dispatch decision
The current implementation combines H3 spatial structure, graph / community methods, capacity rules, and sequence prediction.
The repository reports evaluation on 5 months / 161 independent operating days using leave-one-day-out validation.
Published grouping metrics include:
- mean daily ARI: 0.5161;
- best reported daily ARI: 0.7299.
These metrics indicate partial recovery of historical grouping structure. They do not justify saying the model "restores 80%+ of dispatcher logic" unless a separately defined and reproducible metric supports that statement.
Similarly, statements such as "decision consistency 85--90%," "SOTA," or "perfectly solves cross-zone empty running" should be removed or tied to an explicit metric, baseline, and evaluation protocol.
- historical operations contain learnable spatial / behavioral structure;
- capacity constraints can be enforced while learning grouping patterns;
- leave-one-day-out validation is preferable to random leakage-prone splits for day-level dispatch behavior;
- H3 / hierarchical zone representations are useful candidate abstractions.
- global route optimality;
- business-outcome superiority over a strong VRP baseline;
- causal proof that copying historical behavior is desirable;
- universal transfer to new depots / cities / fleets.
Orders + weight + pieces + history
↓
H3 spatial representation
↓
capacity-aware grouping model
↓
historical grouping similarity
↓
driver / zone sequence model
↓
route-order prior
↓
hybrid dispatch layer
These two repositories should cross-link explicitly:
logistics-dispatch-clustering
learns from historical human operations
↓
preferences / grouping priors
open-dispatch
solves explicit routing / dispatch constraints
↓
deterministic executable plan
The long-term architecture is likely learning + optimization, not either one alone.
Detailed research material lives alongside the code in docs/. This
README keeps the two-stage architecture, evidence, limits and quick
start visible at the top.
- similarity to historical dispatch does not equal optimality;
- human decisions may encode both expertise and historical inefficiency;
- behavior may drift over time;
- new drivers / zones create cold-start problems;
- operational objectives should be evaluated directly: distance, time, SLA, load balance, stability, overtime, and cost.
topprism:
purpose:
- customer-decision
- applied-research
capability: dispatch-behavior-learning
platform_layer: decision-engine
maturity: real-data-validated
evidence:
type: historical-operational-data
scope: "5 months, 161 operating days"
related:
- open-dispatch