An Asynchronous Gossip Algorithm with Restart Strategy in Distributed Minimax Optimization

Abstract

An asynchronous gossip algorithm with a restart strategy is proposed to an approximated distributed minimax optimization in this paper. The restart strategy controls a step length of a subgradient method and resets a local clock under a certain criteria in the algorithm. The convergence property is analyzed for the algorithm with the restart strategy. Numerical examples illustrate that the proposed algorithm works well in various instances and high approximation ratios.