HITL-GB framework cuts cold-start learning in dynamic pricing by 80%

Dynamic pricing in short-term rental markets faces a chicken-and-egg problem: online learning algorithms need data to make good pricing decisions, but bad initial decisions carry significant financial risk. Traditional contextual bandits suffer a cold-start period of weeks or months (around 150 episodes) before they converge. Researchers led by Oleg Miroshnichenko propose the Human-in-the-Loop Gated Bandit (HITL-GB) framework, where a bandit algorithm generates price recommendations but a human agent must approve each recommendation before it is applied. The key insight is that historical pricing data—collected under prior deterministic policies—is structurally equivalent to on-policy warm-up data for initializing the bandit's posterior. This allows the system to bypass the cold-start entirely, using a regularized ridge-regression warm-up procedure from historical episodes.

Validated on real anonymized STR production data (April 2022 – April 2026, 1,461 nightly pricing episodes), the warm-up compressed effective cold-start from ~150 episodes to just ~30 episodes when initializing agents from the Hierarchical Factored Thompson Sampling (HF-TS) family. The authors argue this structural equivalence result is domain-agnostic: any high-stakes domain where human approval is legally or operationally required—including clinical drug dosing, credit origination, content moderation, and radiological diagnosis—satisfies the same conditions. Mandatory human oversight thus becomes a statistical asset rather than a deployment constraint, enabling faster, safer AI deployment in regulated industries.