Combining Deep Reinforcement Learning with AI for Optimal Selling Strategies on Hog Farms

By Olivia Hall

Assistant Professor of Supply Chain Management Ye Liu and colleagues have developed a new, AI-powered approach to hog farm management. Their support tool, introduced in the Journal of Operations Management, builds on deep reinforcement learning and allows farmers to make better selling decisions in a complex environment.

Decision making in hog farming is rife with uncertainty. Every week, farmers must choose whether and where to sell their hogs, but they face volatility on multiple fronts, from unpredictable growth to varying operating costs and changing numbers of available hogs at optimal weight. They may fulfill long- term contracts with meatpackers, sell on the spot markets or hold animals for sale the following week. Prices for all options fluctuate from week to week.

To tackle this problem, the researchers—Liu; Panos Kouvelis, Emerson Distinguished Professor of Supply Chain, Operations and Technology a Washington University in St. Louis; and Danko Turcic, associate professor
of operations and supply chain management at the University of California, Riverside—used inventory and pricing data from a large U.S. farm, along with publicly available pricing information for various agricultural commodities, to
build a new tool.

Fed with market prices and quantities of at-weight and underweight hogs, the machine delivers a near-optimal selling strategy. But users may not trust the tool without insights into its inner workings. So Liu and her colleagues developed
heuristics using classification trees to understand how the machine makes decisions. They call this method managerial learning. Currently the study farm uses an always-fulfill policy to prioritize its long-term contracts, even if this means selling underweight hogs. The new, more flexible heuristic, on the other hand, will sometimes dictate holding some hogs to hedge the risks of production and sometimes shift trading of hogs to markets with higher prices.

“What the farmers are doing now, based on their experience, is too simple, leading to losses of 26% of possible profits,” Liu says. “With the heuristic, the loss is only 8% compared to machine output. So, if they want to combine it with
their experience, if they feel more trust this way, that is also very good.”

Applications of this approach extend beyond hogs to other types of agricultural operations. “This framework can be used more broadly for decisions involving input and output uncertainty and extracts useful insights from the AI model to
improve decision making,” Liu says.