An Empirically Grounding Analytics (EGA) Approach to Hog Farm Finishing Stage Management: Deep Reinforcement Learning as Decision Support and Managerial Learning Tool

45 Pages Posted: 28 Nov 2023 Last revised: 15 Apr 2024

See all articles by Panos Kouvelis

Panos Kouvelis

Washington University in St. Louis

Ye Liu

Washington University in St. Louis - John M. Olin Business School

Danko Turcic

University of California, Riverside (UCR) - A. Gary Anderson Graduate School of Management

Date Written: October 30, 2023

Abstract

The final growth period of farm hogs, known as the ``finishing stage," occurs just before the hogs are sent off for trade and processing in the pork supply chain. This stage presents a complex, dynamic inventory problem that deals with unpredictable prices and amounts of resources and products. Each week, the farmer faces crucial decisions: selecting hogs for immediate sale on the open market, determining which to keep for another week, and deciding which to send to the meatpacker. The farmer is committed to delivering a certain number of hogs to the meatpacker each week at a price set by a specific market index. Any shortfall in delivery incurs a penalty, calculated according to the shortfall's size and indexed to the market.

This paper introduces a novel approach to inventory management by leveraging Deep Reinforcement Learning (DRL), specifically the Actor-Critic method, to optimize inventory policy. Our approach falls into the Empirically Grounding Analytics (EGA) domain, as we empirically validate model assumptions and assess model results within a real context with concise lessons for successful practice. Unlike traditional methods such as dynamic programming, which struggle with the curse of dimensionality and complex policy structures, DRL offers a more scalable solution and a powerful decision-support tool with a robust, practical application. We address two key challenges: First, we develop a method of integrating inventory constraints into neural networks by adding hidden layers. Second, we focus on policy interpretability, a growing concern in machine learning, to identify crucial system parameters for optimal control. Our trained DRL agent significantly outperformed existing heuristic policies in tests, achieving an efficiency gain of almost 25% and trailing an exact dynamic solution by less than 1%, underscoring its potential as a robust optimization tool in pig farming inventory management.

Suggested Citation

Kouvelis, Panos and Liu, Ye and Turcic, Danko, An Empirically Grounding Analytics (EGA) Approach to Hog Farm Finishing Stage Management: Deep Reinforcement Learning as Decision Support and Managerial Learning Tool (October 30, 2023). Available at SSRN: https://ssrn.com/abstract=4617964 or http://dx.doi.org/10.2139/ssrn.4617964

Panos Kouvelis

Washington University in St. Louis ( email )

One Brookings Drive
Campus Box 1156
St. Louis, MO 63130-4899
United States

HOME PAGE: http://www.panoskouvelis.info

Ye Liu

Washington University in St. Louis - John M. Olin Business School ( email )

One Brookings Drive
Campus Box 1208
Saint Louis, MO 63130-4899
United States

Danko Turcic (Contact Author)

University of California, Riverside (UCR) - A. Gary Anderson Graduate School of Management ( email )

Riverside, CA 92521
United States

Do you have a job opening that you would like to promote on SSRN?

Paper statistics

Downloads
60
Abstract Views
219
Rank
661,406
PlumX Metrics