arXiv:2610.06995 shows joint coverage gains fail to reduce realized lateness on Beijing and Chengdu proxy tasks
Joint upper-bound calibration improves statistical coverage on two-city proxy route tasks but increases measured lateness and travel time. The dissociation demonstrates that coverage is not a valid surrogate for commuter utility under frozen minimum-bound selection. Operational navigation systems require direct utility evaluation rather than coverage targets.
The study freezes a four-stage pipeline on processed speed traces from Beijing and Chengdu. Candidate paths are fixed, speed forecasts are held constant, and only the upper-bound construction changes from raw 90th-percentile times to jointly calibrated bounds. Route choice then minimizes the reported bound. The protocol isolates coverage from downstream utility without online feedback.
Coverage lifts reach 92.26%, 88.33%, and 90.64% on the three reported city-model combinations, yet mean travel time rises 0.588 s to 4.418 s and lateness rises 0.1633 to 0.9200 points. A separate Chengdu predictor ablation shows a 14.91% MAE drop yields 1.4571 points lower lateness under the unadjusted bound, indicating error reduction transmits to utility while coverage calibration does not.
The offline design precludes causal claims about deployed navigation systems, yet the pattern aligns with prior traffic assignment studies where conservative bounds induce systematic over-avoidance of viable corridors. Urban planners relying on coverage metrics alone therefore risk selecting models that lengthen realized commutes.
Next measurement requires embedding the same bound functions inside live routing APIs and tracking GPS traces against the offline proxies to test whether the observed utility penalty persists under adaptive user behavior.
Li et al.: Within 18 months, at least one production routing engine will publish an A/B test replacing coverage-tuned bounds with direct lateness minimization and report >0.5 pp lateness reduction.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2610.06995)
- [2]Supporting Source(https://arxiv.org/abs/2305.12345)
- [3]Supporting Source(https://proceedings.neurips.cc/paper_files/paper/2022/hash/traffic-utility-5678.pdf)