An empirical comparison of multi-agent optimization algorithms

In the past decade a large number of distributed algorithms for solving large-scale convex optimization problems have been proposed and analyzed in the literature, especially from the perspective of multi-agent systems. Although it is fairly well understood which algorithms have the most desirable theoretical properties, there has been very little work investigating and evaluating practical implementations of these algorithms, and there is a non-trivial gap between theory and practice. For example, many of the theoretical analyses ignore important practical issues such as asynchronism and communication delays. In this paper we perform an empirical evaluation of non-doubly stochastic multi-agent distributed optimization algorithms for large-scale convex optimization and open source the code. We find that a first order asynchronous subgradient optimization algorithm can actually out-perform state-of-the-art synchronous algorithms in a practical scenario for both small and large multiagent networks running on a high performance cluster.