Control graphs are used in multi-robot systems to maintain information about which robot senses another robot, and at what position. Control graphs allow robots to localize relative to others, and maintain stable formations. Previous work makes two critical assumptions. First, it assumes edge weights of control graphs are deterministic scalars, while in reality they represent complex stochastic factors. Second, it assumes that a single robot is pre-determined to serve as the global anchor for the robots’ relative estimates. However, optimal selection of this robot is an open problem. In this work, we address these two issues. We show that existing work may be recast as graph-theoretic algorithms inducing control graphs for more general representation of the sensing capabilities of robots. We then formulate the problem of optimal selection of an anchor, and present a centralized algorithm for solving it. We evaluate use of these algorithm on physical and simulated robots and show they very significantly improve on existing work.