Reproducible Sensitivity and Ablation Results

This document reports only experiments backed by machine-readable artifacts. Static metrics use the public reference objective; rollout fare is simulator-specific.

Planning horizon

Horizon

NDCG@3

Hit@3

Coverage

Mean daily fare

1

0.9582

0.9783

13.69%

$569.78

2

0.9565

0.9714

14.07%

$570.61

3

0.9549

0.9759

14.45%

$573.47

5

0.9525

0.9741

14.45%

$575.97

Adaptive

0.9559

0.9735

14.07%

$573.31

The horizon experiment shows that static reference NDCG and simulated fare are not monotonic with each other.

Executable parameter grid

The current runner evaluates half-saturation \(\{120,240,360\}\), gamma \(\{0.25,0.5,0.75\}\), and candidate pools \(\{50,100\}\) against the public static reference objective.

The best static configuration is:

Half-saturation

Gamma

K

NDCG@3

Hit@3

240

0.25

50

0.9577

0.9762

At half-saturation 240 and K=50:

Gamma

NDCG@3

Hit@3

0.25

0.9577

0.9762

0.50

0.9565

0.9714

0.75

0.9557

0.9732

K=50 and K=100 produce the same rounded NDCG/Hit for these settings, so the smaller pool is preferred on computational grounds. The production-facing default remains gamma=0.5 because the static public reference is a diagnostic rather than the sole deployment objective.

Stress tests

Scenario

NDCG@3

Hit@3

Top-3 overlap

Unperturbed

0.9565

0.9714

100%

Manhattan demand +50%

0.9451

0.9667

88.03%

Random 10% missing cells

0.9042

0.7872

81.37%

Drop OD probabilities below 0.001

0.9565

0.9714

99.84%

Remove bottom-demand 10% zones

0.9565

0.9714

100%

Random missing zone-time cells are the most damaging tested perturbation. The rare-zone result also shows that bottom-demand zones have effectively no policy influence.

Exposure concentration

Strategy

Coverage

Gini

Effective zones

Airport exposure

Hot Zone

7.98%

0.974

9.22

27.74%

Single-Step

14.07%

0.970

9.21

50.62%

Two-Step

14.07%

0.982

5.51

70.33%

Nominal coverage increases, but effective exposure becomes more concentrated. A production-oriented objective should add supply, queue capacity, and concentration penalties.

Unsupported ablations removed

Earlier versions listed precise contributions for cleaning, transition probabilities, duration modeling, K=150/263, and “regret versus optimal” without executable artifacts. Those tables have been removed. Future ablations must save configs, per-run outcomes, and paired uncertainty.