Distributed Online Optimization via Kernel Reproduced Gradient Descent

This paper focuses on a stochastic optimization problem in distributed online learning. To accelerate the op-timization in a distributed framework, a reproducing gradient method is proposed based on kernel learning. Considering the practicality of distributed settings, this method utilizes random Fourier features for gradient learning, which enables privacy protection as model parameters and gradients do not need to be transmitted over the network. Further, a distributed online reproduced gradient descent algorithm is proposed, and the convergence rate of the algorithm is provided. Numerical results show that our method outperforms the existing distributed stochastic gradient descent based on gradient tracking.