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

Please visit the MIT System Dynamics Seminars page for more information.

You are invited to attend the System Dynamics Seminar being held this Friday from 12:00-1:30 pm ET in the Jay W. Forrester conference room, E62-450, or via Zoom:: https://mit.zoom.us/j/97116456932 (password: SDFall2022).

Our guest speaker will be James Paine (MIT Sloan) presenting Simpler is (Sometimes) Better: A Comparison of Cost Reducing Agent Architectures in a Simulated Behaviorally-Driven Multi-Echelon Supply Chain Lunch will be provided to those attending in person.

If you would also like to schedule a 30-minute 1:1 meeting with Brent Moritz, please fill out this Doodle poll https://doodle.com/meeting/participate/id/b2vWmxAb by COB Wednesday and I will confirm times with a calendar invite.

Please check here for the latest updates on MIT’s COVID policies: https://now.mit.edu/policies/

Abstract

Supply chains partially consist of, and almost exclusively exist for, people. Behavioral Operations Management has endeavored to identify how the behavioral responses of these people, decision makers in supply chains, differ from the fully rational and to identify policies incorporating on these differences. The complexity of such policies, and the underlying assumptions of rationality, can vary widely. This work utilizes a model of a multi-echelon supply chain, captured by the classic Beer Game inventory management simulation, to compare the features of policies that can reduce costs from bullwhip when placed in such a system while still allowing other entities to behave in a behaviorally. This work contributes to existing supply chain management literature by applying a dueling-DQN structure and Model-Predictive learning structure to this multi-echelon supply chain system in a manner that can be leveraged for other research. However, this is secondary to the main observation of this work that relatively simple ordering policies, including static base-stock rules, in these behaviorally-driven systems can have large cost-reducing effects only marginally behind more complex methods. Additionally, for model-predictive learning agents, even myopic approaches with limited information about the overall system and greedy objectives can be cost reducing globally. This has direct managerial implications by showing how a decision maker embedded in a supply chain with other behavioral actors does not need to be perfectly rational and can be locally focused while achieving global benefits.

James Paine is a fifth-year doctoral candidate at the Sloan School of Management at MIT, studying System Dynamics and its applications to product and service delivery systems. Prior to coming to the System Dynamics group, James gained experience in the nuclear, reverse logistics, and consumer apparel industries, as both an engineer and product lifecycle-focused marketer. Currently, James focuses on behavioral operations management questions, including human-algorithm interactions, supply chain research and analytics, and dynamic modeling of product and service delivery systems. More information about James and his research can be found at https://jpaine.info/.