Project Description

We consider a distributed networked optimization problem in which every node of the network has access to one component of the objective function. A new cyclically averaged incremental gradient method that requires partial information sharing between adjacent nodes is proposed. The new method features a simple cyclic information sharing protocol and converges in linear rate. The cyclically averaged incremental gradient method is based on a virtual stack that contains the local gradients of the objective function components. Each sensor holds one component of this stack, corresponding to its local component, and the sensors update their content in turns. The sum of the elements of the stack, which is the only measure shared by the network, is traveling across the network as the algorithm converges to the optimal solution. Figure 1 shows a comparison between the new method and the available incremental gradient method with constant and diminishing step sizes.