An Empirically Grounded Analytical Approach to Hog Farm Finishing Stage Management: Deep Reinforcement Learning as Decision Support and Managerial Learning Tool
Author(s):
Panos Kouvelis, Washington University in St. Louis
Ye Liu, Syracuse University
Danko Turcic, University of California, Riverside
Journal (Year):
Journal of Operations Management (forthcoming)
Summary:
The paper describes a new approach to hog farm management using Deep Reinforcement Learning that outperforms current practices and provides valuable insights for farmers through a process called "managerial learning."
Research Questions:
1. Can Deep Reinforcement Learning (DRL) be used to make better hog selling decisions than current practices?
2. How can insights from complex DRL models be made interpretable and actionable for hog farmers?
These questions lead the authors to develop a novel approach combining DRL with "managerial learning" to extract useful insights and improve decision-making in hog farm management.
What We Know:
Optimizing hog farming is complex due to the many factors involved, like fluctuating prices and unpredictable hog growth. Traditional methods, which rely on simplified rules, often result in lower profits. This paper presents a new AI-powered solution: a Deep Reinforcement Learning (DRL) model trained on real hog farm data. This DRL model learns to make optimal decisions, potentially leading to increased profitability. While focused on hog farming, this approach could also be applied to optimize other types of agricultural operations.
Novel Findings:
The authors develop a novel approach combining DRL with "managerial learning," a process that extracts useful managerial insights from the AI model, to extract useful insights and improve decision-making in hog farm management.
Implications for Practice:
The model can be used as a decision support tool on livestock farms.
Implications for Policy:
Using DRL as a managerial learning tool leads to the creation of simple policies that are superior to existing methods. This results in insights that managers can easily interpret, improving planning performance within our context without the need for the DRL model. The superior quality of these empirical insights helps to expand the understanding of planning practices.
Implications for Society:
This study has broader implications for operations management, where many problems can be framed as dynamic optimization exercises. We provide a framework for how such problems can be solved using AI and how the results can be translated into a policy that humans can understand.
Implications for Research:
Future research can apply our method to other dynamic problems to obtain implementable, effective, and interpretable policies.
Full Citation:
Kouvelis, Panos and Liu, Ye and Turcic, Danko, An Empirically Grounded Analytical (EGA) Approach to Hog Farm Finishing Stage Management: Deep Reinforcement Learning as Decision Support and Managerial Learning Tool. (Forthcoming) Journal of Operations Management
Abstract:
In hog farming, optimizing hog sales is a complex challenge due to uncertain factors such as hog availability, market prices, and operating costs. This study uses a Markov Decision Process (MDP) to model these decisions, revealing the importance of the final weeks in profit management. The MDP's intractability due to the curse of dimensionality leads us to employ Deep Reinforcement Learning (DRL) for optimization. Using real-world and synthetic data, our DRL model outperforms existing practices. However, it lacks interpretability, hindering trust and legal compliance in the food industry. To address this, we introduce “managerial learning”, extracting actionable insights from DRL outputs using classification trees that would have been difficult to obtain otherwise. We leverage these insights to devise a smart heuristic that significantly beats the heuristic currently used in practice.
This study has broader implications for operations management, where DRL can solve complex dynamic optimization problems that are often intractable due to dimensionality. By applying methods such as classification trees and DRL, one can scrutinize solutions for actionable managerial insights that can enhance existing practices with straightforward planning guidelines.
Web URL for the Article:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4617964