Distributed Conditional Gradient Algorithm for Two-Network Saddle-point Problem

We consider a two-network saddle-point problem with constraints, whose projections are expensive. We propose a projection-free algorithm, which is referred to as Distributed Frank-Wolfe Saddle-Point algorithm (DFWSP), which combines the gradient tracking technique and Frank-Wolfe technique. We prove that the algorithm achieves O(1/k
²
) convergence rate for strongly-convex-strongly-concave saddle-point problems. We empirically shows that the proposed algorithm has better numerical performance than the distributed projected saddle-point algorithm.