A review of decentralized optimization focused on information flows of decomposition algorithms

Abstract

Decentralized decision-making can be represented as a connected decision network of agents collaboratively optimizing their local objective functions over common coupling constraints. In this setting, solving large-scale mathematical programming centrally is undesirable or impossible because the data storage and decision authority are already decentralized, the communication bandwidth for information exchange is limited, and privacy concerns with information may exist. We introduce a taxonomy of mathematical programming-based decentralized optimization problems and decentralized algorithms based on the degree of information sharing, information exchange and existence of a central coordinator. We synthesize the literature and identify the shortcomings of a decentralized algorithm, the trends, and the potential research directions based on the proposed taxonomy.