Results

Retention across shift severity

Retention across the severity quartiles for standard against robust RL, online against offline, and the five mechanism families.
Figure 1. Retention is the score under shift divided by the method's nominal score. Left: Standard RL (6 algorithms) against Robust RL (14 algorithms); the labels give the gap at Q1 and Q4. Middle: online against offline; solid lines are robust methods and dashed lines standard methods. Right: the five mechanism families.

Retention decays at a similar rate overall, serving as a complementary metric against the absolute score. Figure 1 shows standard and robust RL, online and offline, and all five families share similar parallel retention curves. Online robust methods retain more than online standard methods at every level, so their mechanisms preserve nominal performance under shift. Offline robust methods retain about as much as offline standard methods (within 0.05 at every level) despite a positive absolute gap (see the gains), so offline robustness is realized through higher nominal scores. Online, the data based family (environment-centric) retains less than the learner-centric family at every level, consistent with data based mechanisms that may augment more qualified data to improve nominal performance. Offline, the data-centric and learner-centric families retain about the same.