Garrett van Ryzin, VP/Distinguished Scientist, Amazon
Session Chair: Nick Arnosti, University of Minnesota
Title: Consensus planning protocol (CPP): A market-based approach to coordination at Amazon
Abstract:
Managing Amazon’s fulfillment network is extraordinarily complex due to its vast scale and scope. To handle this complexity, Amazon has divided management responsibility into multiple organizations (e.g., buying, placement, fulfillment, labor planning, transportation planning), each of which owns its own portfolio of planning and execution systems. Optimizing overall supply chain performance requires a high degree of coordination across these various organizations and systems. At the same time, we believe local ownership of business activities reduces complexity and enables teams to maximize innovation and flexibility (e.g., quickly experimenting with new technologies).
Consensus planning protocol (CPP) solves this dilemma using a distributed, market-based approach to coordination. CPP uses a central coordinator that interfaces with multiple planning systems (called agents) to iteratively “negotiate” over a common set of public variables through the use of agent-specific prices and proximal penalties. Prices in CPP have the interpretation of subsidies (either positive or negative) that an agent receives (or pays) for its public variable choices. Each agent retains responsibility for its own local optimization and can use private information and private variable choices that are hidden from the coordinator. Indeed, agents in CPP are essentially “black boxes” that only report their preferred choices for the public variables when queried by the coordinator. The coordinator, in turn, uses the responses of agents to adjust subsidies to induce agents to reach consensus on the public variables – a process akin to a Walrasian auction. Because the process corresponds to consensus ADMM, it is provably convergent if agents are convex.
In this talk, we describe CPP and how is being used at Amazon to build an automated planning system (APS) for supply chain capacity planning. APS coordinates complex capacity decisions that were previously made separately and uses scenario-based optimization to jointly hedge capacity decisions against supply and demand risk. The modular, agent-based design enables us to build APS quickly, and easily scale and expand the system in the future. And because CPP is a generic framework for coordination, it has many potential applications beyond APS.
Federico Echenique, UC Berkeley
Session Chair: Laura Doval, Columbia University
Title: TBA
Abstract: TBA