Simpler is (Sometimes) Better: A Comparison of Cost Reducing Agent Architectures in a Simulated Behaviorally Driven Multi-Echelon Supply Chain

Abstract

Bullwhip remains a persistent and costly challenge in multi-echelon supply chains, especially those driven by behavioral decision-makers with low degrees of vertical integration. While significant research exists, it's often fragmented, with different streams focusing on different aspects of the problem and proposing distinct policy recommendations. This work performs side-by-side comparisons of these diverse approaches to determine which policy features (such as policy complexity, dynamic learning, incentive structure, and information availability) are most effective at reducing supply chain costs. The study employs a simulated, discrete-time, and behaviorally-realistic environment to experimentally compare different policies. Complexity varies from simple base-stock replenishment, to model predictive control, to Deep-Q networks. The environment is built using real runs of an inventory management game with students and professionals from 2021 and 2022. The key findings are: Simple polices are often effective, especially when dynamic learning is not feasible. When feasible, incorporating behavioral assumptions leads to further cost reductions. Finally, under certain conditions, locally focused and limited-information polices can still lead to global cost reductions. This work offers practical guidance for managers, and provides a bridge between existing streams of bullwhip management literature. Managers need not always employ highly complex, data-intensive approaches. When resources for dynamic learning are limited, simple base-stock policies remain robust. When dynamic learning is possible, this research emphasizes the importance of considering behavioral elements. Finally, the paper shows that under plausible conditions, being locally focused can still achieve global benefits, reducing managerial pressure to have complete information or a holistic view of the entire supply chain.

Keywords: Behavioral Operations Management, Bullwhip, Dynamic Decision Making, Supply Chain & Inventory Management, Model Predictive Control, Deep Q-Network, Supply Chain Research

JEL Classification: M11,C44,C63,C45