AI-DRIVEN DYNAMIC SHELF-SPACE OPTIMIZATION IN RETAIL: A DEEP REINFORCEMENT LEARNING APPROACH USING GROCERY CHAIN DATA

Abstract

We propose a novel approach using Deep Reinforcement Learning (DRL) with data from a major grocery store to adjust shelf spaces on-the-fly. Inspired by prior work on retail analytics and machine learning (Ahumada & Villalobos, 2009; Ferreira et al., 2015), we model shelf allocation as a multi-stage decision problem under uncertainty. Retail shelf-space management: Retail shelf-space assignment is a fundamental problem in operations and supply chain management, which can have substantial influence on sales, customer satisfaction and profit (Chen et al., 2015). Conventional optimization algorithms are insufficiently adapted to the volatile nature of consumer tastes and conflicting trade-offs prevalent in GM, with erratic demand and limited shelf-space. In this work, by making the trade-off among targeted sales volume, stockout rate and inventory turnover, a DQN as well as PPO algorithm are used. We observed that DRL-based mechanisms can potentially improve (up to 18% in case of revenue-optimization and up to 12% in case of stockout-mitigation) over heuristic/lp based policies. These findings highlight the potential of DRL in on-line, adaptive retail decision making (Mnih et al., 2015). The research contributes to the literature by demonstrating the potential use of DRL for making shelf space management a high velocity act, with practical significance for retail planners keen on gaining competitive advantage in fast moving markets.

Keywords: Retail shelf-space, Deep reinforcement learning, Grocery dataset, Inventory optimization, Dynamic allocation, Machine learning, Operations management