Interpretable reinforcement learning with structured policies for prescriptive supply chain analytics.

Area of Science:

• Operations Research
• Artificial Intelligence
• Supply Chain Management

Background:

• Reinforcement Learning (RL) shows promise for complex supply chain decisions like inventory control.
• Deep RL methods are often "black boxes," hindering trust and adoption in industrial settings.
• Interpretability remains a significant challenge for applying RL in high-stakes environments.

Purpose of the Study:

• To bridge the interpretability-performance gap in RL for supply chain management.
• To introduce a novel framework, Structured Policy Reinforcement Learning (SPRL), for transparent RL.
• To enable trustworthy autonomous decision-making in real-world supply chains.

Main Methods:

• Developed SPRL, a framework embedding transparency into the RL agent's learning process.
• Utilized a decoupled architecture to distill Q-learning agent value estimates into a Decision Tree (DT).
• Incorporated a hard Inventory Position Cap constraint to stabilize training and enforce lean policies.

Main Results:

• The SPRL-DT policy achieved a mean total cost reduction of 45% compared to the constrained DQN baseline.
• Maintained an excellent service level exceeding 91%.
• Resulted in a low-complexity, interpretable structure with [Formula: see text] nodes.

Conclusions:

• SPRL offers a transparent and verifiable solution for RL in supply chain management.
• Demonstrates the potential for trustworthy autonomous systems through interpretable RL.
• Paves the way for practical deployment of RL in dynamic inventory control and other supply chain operations.