Reproduce “In-Context Learning for Data-Driven Censored Inventory Control” on OpenResearch

About this paper

Sohom Mukherjee, Anh-Duy Pham, Richard Pibernik, Yunbei Xu

The paper introduces an In-Context Generative Posterior Sampling (ICGPS) framework, implemented via ChronosFlow, to address censored inventory control problems with continuous actions and decision-dependent feedback. It achieves sublinear Bayesian regret guarantees for this challenging setting and demonstrates enhanced robustness to demand distribution shifts and prior mismatch compared to traditional methods.

Key insights

• The theoretical framework establishes a decomposition of Bayesian regret for ICGPS, linking it to ideal Thompson Sampling performance and a deployment penalty derived from the learned generative model's quality.
• The deployment penalty is shown to be controllable by the observable offline predictive mismatch under conditions of 'max-order coverage' and 'self-contraction along AR trajectory', providing guidance on effective model learning.
• In-context generative posterior sampling provides a data-driven approach that is robust to misspecification of demand distributions and prior mismatch, a common pitfall for traditional Bayesian methods.

Summary by alphaXiv.