Offline Dynamic Inventory and Pricing Strategy: Addressing Censored and Dependent Demand
Korel Gundem, Zhengling Qi·April 14, 2025
Summary
The paper tackles offline sequential feature-based pricing and inventory control for censored, dependent demand. It introduces two data-driven algorithms to solve Bellman equations, addressing challenges like missing profit data and non-Markovian properties. The first algorithm adapts based on recent censoring, while the second prevents more than n consecutive censorings. The paper also discusses policy classes for managing censoring in sequential decision-making, offering efficient algorithms for dynamic pricing and demand learning. It outlines equations for sequential decision-making in censored states, emphasizing expected action values and value functions under various policies. The text summarizes proofs for key lemmas, focusing on expected future rewards and conditional survival functions estimation using the Kaplan-Meier method.
Introduction
Background
Overview of offline sequential feature-based pricing and inventory control
Challenges in dealing with censored, dependent demand
Objective
Aim of the paper: addressing challenges in pricing and inventory control with missing profit data and non-Markovian properties
Method
Data Collection
Techniques for gathering data on censored, dependent demand
Data Preprocessing
Methods for handling missing profit data and non-Markovian properties in the dataset
Algorithms for Solving Bellman Equations
Algorithm 1: Recent Censoring Adaptation
Description of the algorithm's approach to learning from recent censoring events
Algorithm 2: Prevention of Consecutive Censorings
Explanation of the algorithm designed to avoid more than n consecutive censorings
Policy Classes for Managing Censoring
Dynamic Pricing and Demand Learning
Overview of policy classes for managing censoring in sequential decision-making
Efficient algorithms for dynamic pricing and demand learning
Sequential Decision-Making in Censored States
Equations for Sequential Decision-Making
Formulation of equations for making decisions in censored states
Focus on expected action values and value functions under various policies
Proofs and Lemmas
Expected Future Rewards
Summarized proofs for expected future rewards in censored states
Conditional Survival Functions Estimation
Use of the Kaplan-Meier method for estimating conditional survival functions
Discussion on the estimation of survival probabilities in censored data
Basic info
papers
statistics theory
machine learning
artificial intelligence
applications
Advanced features
Insights
What are the limitations of the proposed algorithms in handling non-Markovian properties and missing profit data?
What are the main challenges addressed by the paper in the context of offline sequential feature-based pricing and inventory control?
How do the two data-driven algorithms introduced in the paper solve the Bellman equations for censored, dependent demand?
What innovative approaches does the paper propose for managing censoring in sequential decision-making?