Drawback-Parameter-Free Decentralized Nonconvex Stochastic Optimization
Authors: Jiaxiang Li, Xuxing Chen, Shiqian Ma, Mingyi Hong
Summary: Present decentralized algorithms often require information of downside parameters for updating native iterates. For instance, the hyperparameters (resembling studying charge) often require the information of Lipschitz fixed of the worldwide gradient or topological info of the communication networks, that are often not accessible in follow. On this paper, we suggest D-NASA, the primary algorithm for decentralized nonconvex stochastic optimization that requires no prior information of any downside parameters. We present that D-NASA has the optimum charge of convergence for nonconvex goals underneath very delicate circumstances and enjoys the linear-speedup impact, i.e. the computation turns into quicker because the variety of nodes within the system will increase. Intensive numerical experiments are carried out to help our findings.