EURO 2025 Leeds
Abstract Submission

2789. Joint Issuing and Replenishment for Perishable Products Using Deep Reinforcement Learning

Invited abstract in session TB-34: Advancements of OR-analytics in statistics, machine learning and data science 3, stream Advancements of OR-analytics in statistics, machine learning and data science.

Tuesday, 10:30-12:00
Room: Michael Sadler LG10

Authors (first author is the speaker)

1. Achal Goyal
OM, Indian Institute of Management Udaipur
2. Vinay Reddy
Mahindra University

Abstract

Online grocery stores commonly sell perishable products such as bread, milk, and meat similar to brick-and-mortar stores. There is one key difference in the operations of an online retailer and a brick-and-mortar store: the online retailer gets to decide which unit to issue whereas this decision is customer-driven in the case of a brick-and-mortar store. This gives more flexibility to the online retailer as they can now issue the units in First-In-First-Out (FIFO) orders and reduce wastage. However, this could backfire as customers desire freshness. We consider a periodic review inventory model where every period the manager makes two decisions: 1) How much fresh inventory to order? 2) Which units to use to fulfil the demand? A goodwill penalty is incurred upon selling an old unit which increases as the remaining lifetime of the unit decreases. We apply Deep Reinforcement Learning (DRL) in our problem context. The major difference in our study is that the action space is high dimensional due to the presence of issuing as a decision variable. We implement two state-of-the-art policy gradient RL algorithms: Reinforce and Actor-critic. We observe that the algorithms are unstable while training because of the high dimensionality of action space. We explore transfer learning to stabilise the training process and improve performance. Overall, the DRL policy developed in this paper outperforms the existing heuristics.

Keywords

Status: accepted

Back to the list of papers