Abstract: | Decentralized optimization aims to solve optimization problems in a multi-agent network. They do not require a central data storage or even a central control station and thus fits many real network situations, such as limited energy or bandwidth and short wireless communication range. This talk introduces first-order decentralized algorithms for solving convex unconstrained optimization problems defined on a network. In addition, we address a severe limitation of existing methods by proposing a decentralized algorithm that achieves the same convergence rate as their centralized counterparts up to constant factors determined by the network topology. This is joint work with Qing Ling, We Shi, Gang Wu, and Kun Yuan at USTC. |